; Copyright 2017 Google Inc.
;
; Use of this source code is governed by a BSD-style license that can be
; found in the LICENSE file.

; This file is generated semi-automatically with this command:
;   $ src/jumper/build_stages.py

IFDEF RAX
_text SEGMENT

PUBLIC _sk_start_pipeline_hsw
_sk_start_pipeline_hsw LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  86                                  ; push          %rsi
  DB  87                                  ; push          %rdi
  DB  83                                  ; push          %rbx
  DB  72,129,236,160,0,0,0                ; sub           $0xa0,%rsp
  DB  197,120,41,188,36,144,0,0,0         ; vmovaps       %xmm15,0x90(%rsp)
  DB  197,120,41,180,36,128,0,0,0         ; vmovaps       %xmm14,0x80(%rsp)
  DB  197,120,41,108,36,112               ; vmovaps       %xmm13,0x70(%rsp)
  DB  197,120,41,100,36,96                ; vmovaps       %xmm12,0x60(%rsp)
  DB  197,120,41,92,36,80                 ; vmovaps       %xmm11,0x50(%rsp)
  DB  197,120,41,84,36,64                 ; vmovaps       %xmm10,0x40(%rsp)
  DB  197,120,41,76,36,48                 ; vmovaps       %xmm9,0x30(%rsp)
  DB  197,120,41,68,36,32                 ; vmovaps       %xmm8,0x20(%rsp)
  DB  197,248,41,124,36,16                ; vmovaps       %xmm7,0x10(%rsp)
  DB  197,248,41,52,36                    ; vmovaps       %xmm6,(%rsp)
  DB  77,137,205                          ; mov           %r9,%r13
  DB  77,137,198                          ; mov           %r8,%r14
  DB  72,137,203                          ; mov           %rcx,%rbx
  DB  72,137,214                          ; mov           %rdx,%rsi
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,199                          ; mov           %rax,%r15
  DB  73,137,244                          ; mov           %rsi,%r12
  DB  72,141,67,8                         ; lea           0x8(%rbx),%rax
  DB  76,57,232                           ; cmp           %r13,%rax
  DB  118,5                               ; jbe           75 <_sk_start_pipeline_hsw+0x75>
  DB  72,137,223                          ; mov           %rbx,%rdi
  DB  235,65                              ; jmp           b6 <_sk_start_pipeline_hsw+0xb6>
  DB  185,0,0,0,0                         ; mov           $0x0,%ecx
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
  DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
  DB  197,204,87,246                      ; vxorps        %ymm6,%ymm6,%ymm6
  DB  197,196,87,255                      ; vxorps        %ymm7,%ymm7,%ymm7
  DB  72,137,223                          ; mov           %rbx,%rdi
  DB  76,137,230                          ; mov           %r12,%rsi
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  65,255,215                          ; callq         *%r15
  DB  72,141,123,8                        ; lea           0x8(%rbx),%rdi
  DB  72,131,195,16                       ; add           $0x10,%rbx
  DB  76,57,235                           ; cmp           %r13,%rbx
  DB  72,137,251                          ; mov           %rdi,%rbx
  DB  118,191                             ; jbe           75 <_sk_start_pipeline_hsw+0x75>
  DB  76,137,233                          ; mov           %r13,%rcx
  DB  72,41,249                           ; sub           %rdi,%rcx
  DB  116,41                              ; je            e7 <_sk_start_pipeline_hsw+0xe7>
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
  DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
  DB  197,204,87,246                      ; vxorps        %ymm6,%ymm6,%ymm6
  DB  197,196,87,255                      ; vxorps        %ymm7,%ymm7,%ymm7
  DB  76,137,230                          ; mov           %r12,%rsi
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  65,255,215                          ; callq         *%r15
  DB  76,137,232                          ; mov           %r13,%rax
  DB  197,248,40,52,36                    ; vmovaps       (%rsp),%xmm6
  DB  197,248,40,124,36,16                ; vmovaps       0x10(%rsp),%xmm7
  DB  197,120,40,68,36,32                 ; vmovaps       0x20(%rsp),%xmm8
  DB  197,120,40,76,36,48                 ; vmovaps       0x30(%rsp),%xmm9
  DB  197,120,40,84,36,64                 ; vmovaps       0x40(%rsp),%xmm10
  DB  197,120,40,92,36,80                 ; vmovaps       0x50(%rsp),%xmm11
  DB  197,120,40,100,36,96                ; vmovaps       0x60(%rsp),%xmm12
  DB  197,120,40,108,36,112               ; vmovaps       0x70(%rsp),%xmm13
  DB  197,120,40,180,36,128,0,0,0         ; vmovaps       0x80(%rsp),%xmm14
  DB  197,120,40,188,36,144,0,0,0         ; vmovaps       0x90(%rsp),%xmm15
  DB  72,129,196,160,0,0,0                ; add           $0xa0,%rsp
  DB  91                                  ; pop           %rbx
  DB  95                                  ; pop           %rdi
  DB  94                                  ; pop           %rsi
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  197,248,119                         ; vzeroupper
  DB  195                                 ; retq

PUBLIC _sk_just_return_hsw
_sk_just_return_hsw LABEL PROC
  DB  195                                 ; retq

PUBLIC _sk_seed_shader_hsw
_sk_seed_shader_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,249,110,199                     ; vmovd         %edi,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
  DB  197,252,88,2                        ; vaddps        (%rdx),%ymm0,%ymm0
  DB  196,226,125,24,16                   ; vbroadcastss  (%rax),%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,236,88,201                      ; vaddps        %ymm1,%ymm2,%ymm1
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
  DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
  DB  197,204,87,246                      ; vxorps        %ymm6,%ymm6,%ymm6
  DB  197,196,87,255                      ; vxorps        %ymm7,%ymm7,%ymm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_constant_color_hsw
_sk_constant_color_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,226,125,24,0                    ; vbroadcastss  (%rax),%ymm0
  DB  196,226,125,24,72,4                 ; vbroadcastss  0x4(%rax),%ymm1
  DB  196,226,125,24,80,8                 ; vbroadcastss  0x8(%rax),%ymm2
  DB  196,226,125,24,88,12                ; vbroadcastss  0xc(%rax),%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clear_hsw
_sk_clear_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcatop_hsw
_sk_srcatop_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,204                       ; vmulps        %ymm4,%ymm8,%ymm9
  DB  196,194,69,168,193                  ; vfmadd213ps   %ymm9,%ymm7,%ymm0
  DB  197,60,89,205                       ; vmulps        %ymm5,%ymm8,%ymm9
  DB  196,194,69,168,201                  ; vfmadd213ps   %ymm9,%ymm7,%ymm1
  DB  197,60,89,206                       ; vmulps        %ymm6,%ymm8,%ymm9
  DB  196,194,69,168,209                  ; vfmadd213ps   %ymm9,%ymm7,%ymm2
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  196,194,69,168,216                  ; vfmadd213ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstatop_hsw
_sk_dstatop_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  196,226,101,184,196                 ; vfmadd231ps   %ymm4,%ymm3,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  196,226,101,184,205                 ; vfmadd231ps   %ymm5,%ymm3,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  196,226,101,184,214                 ; vfmadd231ps   %ymm6,%ymm3,%ymm2
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,194,69,168,216                  ; vfmadd213ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcin_hsw
_sk_srcin_hsw LABEL PROC
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstin_hsw
_sk_dstin_hsw LABEL PROC
  DB  197,228,89,196                      ; vmulps        %ymm4,%ymm3,%ymm0
  DB  197,228,89,205                      ; vmulps        %ymm5,%ymm3,%ymm1
  DB  197,228,89,214                      ; vmulps        %ymm6,%ymm3,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcout_hsw
_sk_srcout_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstout_hsw
_sk_dstout_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,252,92,219                      ; vsubps        %ymm3,%ymm0,%ymm3
  DB  197,228,89,196                      ; vmulps        %ymm4,%ymm3,%ymm0
  DB  197,228,89,205                      ; vmulps        %ymm5,%ymm3,%ymm1
  DB  197,228,89,214                      ; vmulps        %ymm6,%ymm3,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcover_hsw
_sk_srcover_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  196,194,93,184,192                  ; vfmadd231ps   %ymm8,%ymm4,%ymm0
  DB  196,194,85,184,200                  ; vfmadd231ps   %ymm8,%ymm5,%ymm1
  DB  196,194,77,184,208                  ; vfmadd231ps   %ymm8,%ymm6,%ymm2
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstover_hsw
_sk_dstover_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
  DB  196,226,61,168,196                  ; vfmadd213ps   %ymm4,%ymm8,%ymm0
  DB  196,226,61,168,205                  ; vfmadd213ps   %ymm5,%ymm8,%ymm1
  DB  196,226,61,168,214                  ; vfmadd213ps   %ymm6,%ymm8,%ymm2
  DB  196,226,61,168,223                  ; vfmadd213ps   %ymm7,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_modulate_hsw
_sk_modulate_hsw LABEL PROC
  DB  197,252,89,196                      ; vmulps        %ymm4,%ymm0,%ymm0
  DB  197,244,89,205                      ; vmulps        %ymm5,%ymm1,%ymm1
  DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_multiply_hsw
_sk_multiply_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,212                       ; vmulps        %ymm4,%ymm8,%ymm10
  DB  196,98,53,184,208                   ; vfmadd231ps   %ymm0,%ymm9,%ymm10
  DB  196,194,93,168,194                  ; vfmadd213ps   %ymm10,%ymm4,%ymm0
  DB  197,52,89,209                       ; vmulps        %ymm1,%ymm9,%ymm10
  DB  196,98,61,184,213                   ; vfmadd231ps   %ymm5,%ymm8,%ymm10
  DB  196,194,85,168,202                  ; vfmadd213ps   %ymm10,%ymm5,%ymm1
  DB  197,52,89,210                       ; vmulps        %ymm2,%ymm9,%ymm10
  DB  196,98,61,184,214                   ; vfmadd231ps   %ymm6,%ymm8,%ymm10
  DB  196,194,77,168,210                  ; vfmadd213ps   %ymm10,%ymm6,%ymm2
  DB  197,52,89,203                       ; vmulps        %ymm3,%ymm9,%ymm9
  DB  196,66,69,168,193                   ; vfmadd213ps   %ymm9,%ymm7,%ymm8
  DB  196,194,69,168,216                  ; vfmadd213ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_plus__hsw
_sk_plus__hsw LABEL PROC
  DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
  DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
  DB  197,236,88,214                      ; vaddps        %ymm6,%ymm2,%ymm2
  DB  197,228,88,223                      ; vaddps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_screen_hsw
_sk_screen_hsw LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  196,194,93,172,192                  ; vfnmadd213ps  %ymm8,%ymm4,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  196,194,85,172,200                  ; vfnmadd213ps  %ymm8,%ymm5,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  196,194,77,172,208                  ; vfnmadd213ps  %ymm8,%ymm6,%ymm2
  DB  197,100,88,199                      ; vaddps        %ymm7,%ymm3,%ymm8
  DB  196,194,69,172,216                  ; vfnmadd213ps  %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_xor__hsw
_sk_xor__hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,212                       ; vmulps        %ymm4,%ymm8,%ymm10
  DB  196,194,53,168,194                  ; vfmadd213ps   %ymm10,%ymm9,%ymm0
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  196,226,61,184,205                  ; vfmadd231ps   %ymm5,%ymm8,%ymm1
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  196,226,61,184,214                  ; vfmadd231ps   %ymm6,%ymm8,%ymm2
  DB  197,180,89,219                      ; vmulps        %ymm3,%ymm9,%ymm3
  DB  196,98,69,168,195                   ; vfmadd213ps   %ymm3,%ymm7,%ymm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,195                      ; vmovaps       %ymm8,%ymm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_darken_hsw
_sk_darken_hsw LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,100,89,204                      ; vmulps        %ymm4,%ymm3,%ymm9
  DB  196,193,124,95,193                  ; vmaxps        %ymm9,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,100,89,205                      ; vmulps        %ymm5,%ymm3,%ymm9
  DB  196,193,116,95,201                  ; vmaxps        %ymm9,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
  DB  196,193,108,95,209                  ; vmaxps        %ymm9,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lighten_hsw
_sk_lighten_hsw LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,100,89,204                      ; vmulps        %ymm4,%ymm3,%ymm9
  DB  196,193,124,93,193                  ; vminps        %ymm9,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,100,89,205                      ; vmulps        %ymm5,%ymm3,%ymm9
  DB  196,193,116,93,201                  ; vminps        %ymm9,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
  DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_difference_hsw
_sk_difference_hsw LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,100,89,204                      ; vmulps        %ymm4,%ymm3,%ymm9
  DB  196,193,124,93,193                  ; vminps        %ymm9,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,100,89,205                      ; vmulps        %ymm5,%ymm3,%ymm9
  DB  196,193,116,93,201                  ; vminps        %ymm9,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
  DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_exclusion_hsw
_sk_exclusion_hsw LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,196                      ; vmulps        %ymm4,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,205                      ; vmulps        %ymm5,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colorburn_hsw
_sk_colorburn_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
  DB  197,52,89,216                       ; vmulps        %ymm0,%ymm9,%ymm11
  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,228                       ; vmulps        %ymm4,%ymm8,%ymm12
  DB  197,68,92,236                       ; vsubps        %ymm4,%ymm7,%ymm13
  DB  197,20,89,235                       ; vmulps        %ymm3,%ymm13,%ymm13
  DB  197,20,94,232                       ; vdivps        %ymm0,%ymm13,%ymm13
  DB  196,65,68,93,237                    ; vminps        %ymm13,%ymm7,%ymm13
  DB  196,65,68,92,237                    ; vsubps        %ymm13,%ymm7,%ymm13
  DB  196,66,101,168,235                  ; vfmadd213ps   %ymm11,%ymm3,%ymm13
  DB  196,65,28,88,237                    ; vaddps        %ymm13,%ymm12,%ymm13
  DB  197,28,88,224                       ; vaddps        %ymm0,%ymm12,%ymm12
  DB  196,193,124,194,194,0               ; vcmpeqps      %ymm10,%ymm0,%ymm0
  DB  196,195,21,74,196,0                 ; vblendvps     %ymm0,%ymm12,%ymm13,%ymm0
  DB  197,92,194,231,0                    ; vcmpeqps      %ymm7,%ymm4,%ymm12
  DB  197,36,88,220                       ; vaddps        %ymm4,%ymm11,%ymm11
  DB  196,195,125,74,195,192              ; vblendvps     %ymm12,%ymm11,%ymm0,%ymm0
  DB  197,52,89,217                       ; vmulps        %ymm1,%ymm9,%ymm11
  DB  197,60,89,229                       ; vmulps        %ymm5,%ymm8,%ymm12
  DB  197,68,92,237                       ; vsubps        %ymm5,%ymm7,%ymm13
  DB  197,20,89,235                       ; vmulps        %ymm3,%ymm13,%ymm13
  DB  197,20,94,233                       ; vdivps        %ymm1,%ymm13,%ymm13
  DB  196,65,68,93,237                    ; vminps        %ymm13,%ymm7,%ymm13
  DB  196,65,68,92,237                    ; vsubps        %ymm13,%ymm7,%ymm13
  DB  196,66,101,168,235                  ; vfmadd213ps   %ymm11,%ymm3,%ymm13
  DB  196,65,28,88,237                    ; vaddps        %ymm13,%ymm12,%ymm13
  DB  197,28,88,225                       ; vaddps        %ymm1,%ymm12,%ymm12
  DB  196,193,116,194,202,0               ; vcmpeqps      %ymm10,%ymm1,%ymm1
  DB  196,195,21,74,204,16                ; vblendvps     %ymm1,%ymm12,%ymm13,%ymm1
  DB  197,84,194,231,0                    ; vcmpeqps      %ymm7,%ymm5,%ymm12
  DB  197,36,88,221                       ; vaddps        %ymm5,%ymm11,%ymm11
  DB  196,195,117,74,203,192              ; vblendvps     %ymm12,%ymm11,%ymm1,%ymm1
  DB  197,52,89,202                       ; vmulps        %ymm2,%ymm9,%ymm9
  DB  196,65,108,194,210,0                ; vcmpeqps      %ymm10,%ymm2,%ymm10
  DB  197,60,89,222                       ; vmulps        %ymm6,%ymm8,%ymm11
  DB  197,68,92,230                       ; vsubps        %ymm6,%ymm7,%ymm12
  DB  197,28,89,227                       ; vmulps        %ymm3,%ymm12,%ymm12
  DB  197,28,94,226                       ; vdivps        %ymm2,%ymm12,%ymm12
  DB  197,164,88,210                      ; vaddps        %ymm2,%ymm11,%ymm2
  DB  196,65,68,93,228                    ; vminps        %ymm12,%ymm7,%ymm12
  DB  196,65,68,92,228                    ; vsubps        %ymm12,%ymm7,%ymm12
  DB  196,66,101,168,225                  ; vfmadd213ps   %ymm9,%ymm3,%ymm12
  DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
  DB  196,227,37,74,210,160               ; vblendvps     %ymm10,%ymm2,%ymm11,%ymm2
  DB  197,76,194,215,0                    ; vcmpeqps      %ymm7,%ymm6,%ymm10
  DB  197,52,88,206                       ; vaddps        %ymm6,%ymm9,%ymm9
  DB  196,195,109,74,209,160              ; vblendvps     %ymm10,%ymm9,%ymm2,%ymm2
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colordodge_hsw
_sk_colordodge_hsw LABEL PROC
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  197,52,92,215                       ; vsubps        %ymm7,%ymm9,%ymm10
  DB  197,44,89,216                       ; vmulps        %ymm0,%ymm10,%ymm11
  DB  197,52,92,203                       ; vsubps        %ymm3,%ymm9,%ymm9
  DB  197,100,89,228                      ; vmulps        %ymm4,%ymm3,%ymm12
  DB  197,100,92,232                      ; vsubps        %ymm0,%ymm3,%ymm13
  DB  196,65,28,94,229                    ; vdivps        %ymm13,%ymm12,%ymm12
  DB  197,52,89,236                       ; vmulps        %ymm4,%ymm9,%ymm13
  DB  196,65,68,93,228                    ; vminps        %ymm12,%ymm7,%ymm12
  DB  196,66,101,168,227                  ; vfmadd213ps   %ymm11,%ymm3,%ymm12
  DB  196,65,20,88,228                    ; vaddps        %ymm12,%ymm13,%ymm12
  DB  197,20,88,232                       ; vaddps        %ymm0,%ymm13,%ymm13
  DB  197,252,194,195,0                   ; vcmpeqps      %ymm3,%ymm0,%ymm0
  DB  196,195,29,74,197,0                 ; vblendvps     %ymm0,%ymm13,%ymm12,%ymm0
  DB  196,65,92,194,224,0                 ; vcmpeqps      %ymm8,%ymm4,%ymm12
  DB  197,36,88,220                       ; vaddps        %ymm4,%ymm11,%ymm11
  DB  196,195,125,74,195,192              ; vblendvps     %ymm12,%ymm11,%ymm0,%ymm0
  DB  197,44,89,217                       ; vmulps        %ymm1,%ymm10,%ymm11
  DB  197,100,89,229                      ; vmulps        %ymm5,%ymm3,%ymm12
  DB  197,100,92,233                      ; vsubps        %ymm1,%ymm3,%ymm13
  DB  196,65,28,94,229                    ; vdivps        %ymm13,%ymm12,%ymm12
  DB  197,52,89,237                       ; vmulps        %ymm5,%ymm9,%ymm13
  DB  196,65,68,93,228                    ; vminps        %ymm12,%ymm7,%ymm12
  DB  196,66,101,168,227                  ; vfmadd213ps   %ymm11,%ymm3,%ymm12
  DB  196,65,20,88,228                    ; vaddps        %ymm12,%ymm13,%ymm12
  DB  197,20,88,233                       ; vaddps        %ymm1,%ymm13,%ymm13
  DB  197,244,194,203,0                   ; vcmpeqps      %ymm3,%ymm1,%ymm1
  DB  196,195,29,74,205,16                ; vblendvps     %ymm1,%ymm13,%ymm12,%ymm1
  DB  196,65,84,194,224,0                 ; vcmpeqps      %ymm8,%ymm5,%ymm12
  DB  197,36,88,221                       ; vaddps        %ymm5,%ymm11,%ymm11
  DB  196,195,117,74,203,192              ; vblendvps     %ymm12,%ymm11,%ymm1,%ymm1
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  197,100,89,222                      ; vmulps        %ymm6,%ymm3,%ymm11
  DB  197,100,92,226                      ; vsubps        %ymm2,%ymm3,%ymm12
  DB  196,65,36,94,220                    ; vdivps        %ymm12,%ymm11,%ymm11
  DB  197,52,89,230                       ; vmulps        %ymm6,%ymm9,%ymm12
  DB  196,65,68,93,219                    ; vminps        %ymm11,%ymm7,%ymm11
  DB  196,66,101,168,218                  ; vfmadd213ps   %ymm10,%ymm3,%ymm11
  DB  196,65,28,88,219                    ; vaddps        %ymm11,%ymm12,%ymm11
  DB  197,28,88,226                       ; vaddps        %ymm2,%ymm12,%ymm12
  DB  197,236,194,211,0                   ; vcmpeqps      %ymm3,%ymm2,%ymm2
  DB  196,195,37,74,212,32                ; vblendvps     %ymm2,%ymm12,%ymm11,%ymm2
  DB  196,65,76,194,192,0                 ; vcmpeqps      %ymm8,%ymm6,%ymm8
  DB  197,44,88,214                       ; vaddps        %ymm6,%ymm10,%ymm10
  DB  196,195,109,74,210,128              ; vblendvps     %ymm8,%ymm10,%ymm2,%ymm2
  DB  196,194,69,184,217                  ; vfmadd231ps   %ymm9,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hardlight_hsw
_sk_hardlight_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,220                       ; vmulps        %ymm4,%ymm8,%ymm11
  DB  196,98,45,184,216                   ; vfmadd231ps   %ymm0,%ymm10,%ymm11
  DB  197,124,88,200                      ; vaddps        %ymm0,%ymm0,%ymm9
  DB  197,52,194,227,2                    ; vcmpleps      %ymm3,%ymm9,%ymm12
  DB  197,124,89,204                      ; vmulps        %ymm4,%ymm0,%ymm9
  DB  196,65,52,88,233                    ; vaddps        %ymm9,%ymm9,%ymm13
  DB  197,100,89,207                      ; vmulps        %ymm7,%ymm3,%ymm9
  DB  197,68,92,244                       ; vsubps        %ymm4,%ymm7,%ymm14
  DB  197,228,92,192                      ; vsubps        %ymm0,%ymm3,%ymm0
  DB  196,193,124,89,198                  ; vmulps        %ymm14,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
  DB  196,195,125,74,197,192              ; vblendvps     %ymm12,%ymm13,%ymm0,%ymm0
  DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
  DB  197,44,89,217                       ; vmulps        %ymm1,%ymm10,%ymm11
  DB  196,98,61,184,221                   ; vfmadd231ps   %ymm5,%ymm8,%ymm11
  DB  197,116,88,225                      ; vaddps        %ymm1,%ymm1,%ymm12
  DB  197,28,194,227,2                    ; vcmpleps      %ymm3,%ymm12,%ymm12
  DB  197,116,89,237                      ; vmulps        %ymm5,%ymm1,%ymm13
  DB  196,65,20,88,237                    ; vaddps        %ymm13,%ymm13,%ymm13
  DB  197,68,92,245                       ; vsubps        %ymm5,%ymm7,%ymm14
  DB  197,228,92,201                      ; vsubps        %ymm1,%ymm3,%ymm1
  DB  196,193,116,89,206                  ; vmulps        %ymm14,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,180,92,201                      ; vsubps        %ymm1,%ymm9,%ymm1
  DB  196,195,117,74,205,192              ; vblendvps     %ymm12,%ymm13,%ymm1,%ymm1
  DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  196,98,61,184,214                   ; vfmadd231ps   %ymm6,%ymm8,%ymm10
  DB  197,108,88,218                      ; vaddps        %ymm2,%ymm2,%ymm11
  DB  197,36,194,219,2                    ; vcmpleps      %ymm3,%ymm11,%ymm11
  DB  197,108,89,230                      ; vmulps        %ymm6,%ymm2,%ymm12
  DB  196,65,28,88,228                    ; vaddps        %ymm12,%ymm12,%ymm12
  DB  197,68,92,238                       ; vsubps        %ymm6,%ymm7,%ymm13
  DB  197,228,92,210                      ; vsubps        %ymm2,%ymm3,%ymm2
  DB  196,193,108,89,213                  ; vmulps        %ymm13,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,180,92,210                      ; vsubps        %ymm2,%ymm9,%ymm2
  DB  196,195,109,74,212,176              ; vblendvps     %ymm11,%ymm12,%ymm2,%ymm2
  DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_overlay_hsw
_sk_overlay_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,220                       ; vmulps        %ymm4,%ymm8,%ymm11
  DB  196,98,45,184,216                   ; vfmadd231ps   %ymm0,%ymm10,%ymm11
  DB  197,92,88,204                       ; vaddps        %ymm4,%ymm4,%ymm9
  DB  197,52,194,231,2                    ; vcmpleps      %ymm7,%ymm9,%ymm12
  DB  197,124,89,204                      ; vmulps        %ymm4,%ymm0,%ymm9
  DB  196,65,52,88,233                    ; vaddps        %ymm9,%ymm9,%ymm13
  DB  197,100,89,207                      ; vmulps        %ymm7,%ymm3,%ymm9
  DB  197,68,92,244                       ; vsubps        %ymm4,%ymm7,%ymm14
  DB  197,228,92,192                      ; vsubps        %ymm0,%ymm3,%ymm0
  DB  196,193,124,89,198                  ; vmulps        %ymm14,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
  DB  196,195,125,74,197,192              ; vblendvps     %ymm12,%ymm13,%ymm0,%ymm0
  DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
  DB  197,44,89,217                       ; vmulps        %ymm1,%ymm10,%ymm11
  DB  196,98,61,184,221                   ; vfmadd231ps   %ymm5,%ymm8,%ymm11
  DB  197,84,88,229                       ; vaddps        %ymm5,%ymm5,%ymm12
  DB  197,28,194,231,2                    ; vcmpleps      %ymm7,%ymm12,%ymm12
  DB  197,116,89,237                      ; vmulps        %ymm5,%ymm1,%ymm13
  DB  196,65,20,88,237                    ; vaddps        %ymm13,%ymm13,%ymm13
  DB  197,68,92,245                       ; vsubps        %ymm5,%ymm7,%ymm14
  DB  197,228,92,201                      ; vsubps        %ymm1,%ymm3,%ymm1
  DB  196,193,116,89,206                  ; vmulps        %ymm14,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,180,92,201                      ; vsubps        %ymm1,%ymm9,%ymm1
  DB  196,195,117,74,205,192              ; vblendvps     %ymm12,%ymm13,%ymm1,%ymm1
  DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  196,98,61,184,214                   ; vfmadd231ps   %ymm6,%ymm8,%ymm10
  DB  197,76,88,222                       ; vaddps        %ymm6,%ymm6,%ymm11
  DB  197,36,194,223,2                    ; vcmpleps      %ymm7,%ymm11,%ymm11
  DB  197,108,89,230                      ; vmulps        %ymm6,%ymm2,%ymm12
  DB  196,65,28,88,228                    ; vaddps        %ymm12,%ymm12,%ymm12
  DB  197,68,92,238                       ; vsubps        %ymm6,%ymm7,%ymm13
  DB  197,228,92,210                      ; vsubps        %ymm2,%ymm3,%ymm2
  DB  196,193,108,89,213                  ; vmulps        %ymm13,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,180,92,210                      ; vsubps        %ymm2,%ymm9,%ymm2
  DB  196,195,109,74,212,176              ; vblendvps     %ymm11,%ymm12,%ymm2,%ymm2
  DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_softlight_hsw
_sk_softlight_hsw LABEL PROC
  DB  72,131,236,56                       ; sub           $0x38,%rsp
  DB  197,252,17,20,36                    ; vmovups       %ymm2,(%rsp)
  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
  DB  197,44,194,223,1                    ; vcmpltps      %ymm7,%ymm10,%ymm11
  DB  197,92,94,199                       ; vdivps        %ymm7,%ymm4,%ymm8
  DB  196,67,45,74,224,176                ; vblendvps     %ymm11,%ymm8,%ymm10,%ymm12
  DB  196,65,28,88,196                    ; vaddps        %ymm12,%ymm12,%ymm8
  DB  196,65,60,88,232                    ; vaddps        %ymm8,%ymm8,%ymm13
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,66,21,168,237                   ; vfmadd213ps   %ymm13,%ymm13,%ymm13
  DB  196,65,28,92,240                    ; vsubps        %ymm8,%ymm12,%ymm14
  DB  184,0,0,224,64                      ; mov           $0x40e00000,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  196,65,28,89,249                    ; vmulps        %ymm9,%ymm12,%ymm15
  DB  196,66,21,184,254                   ; vfmadd231ps   %ymm14,%ymm13,%ymm15
  DB  196,65,124,82,236                   ; vrsqrtps      %ymm12,%ymm13
  DB  196,65,124,83,237                   ; vrcpps        %ymm13,%ymm13
  DB  196,65,20,92,236                    ; vsubps        %ymm12,%ymm13,%ymm13
  DB  197,92,88,244                       ; vaddps        %ymm4,%ymm4,%ymm14
  DB  196,65,12,88,246                    ; vaddps        %ymm14,%ymm14,%ymm14
  DB  197,12,194,247,2                    ; vcmpleps      %ymm7,%ymm14,%ymm14
  DB  196,67,21,74,239,224                ; vblendvps     %ymm14,%ymm15,%ymm13,%ymm13
  DB  197,124,88,240                      ; vaddps        %ymm0,%ymm0,%ymm14
  DB  197,12,92,251                       ; vsubps        %ymm3,%ymm14,%ymm15
  DB  196,65,60,92,228                    ; vsubps        %ymm12,%ymm8,%ymm12
  DB  196,98,5,168,227                    ; vfmadd213ps   %ymm3,%ymm15,%ymm12
  DB  197,28,89,228                       ; vmulps        %ymm4,%ymm12,%ymm12
  DB  197,4,89,255                        ; vmulps        %ymm7,%ymm15,%ymm15
  DB  196,65,4,89,237                     ; vmulps        %ymm13,%ymm15,%ymm13
  DB  196,98,101,184,236                  ; vfmadd231ps   %ymm4,%ymm3,%ymm13
  DB  197,12,194,243,2                    ; vcmpleps      %ymm3,%ymm14,%ymm14
  DB  196,195,21,74,212,224               ; vblendvps     %ymm14,%ymm12,%ymm13,%ymm2
  DB  197,84,94,239                       ; vdivps        %ymm7,%ymm5,%ymm13
  DB  196,67,45,74,237,176                ; vblendvps     %ymm11,%ymm13,%ymm10,%ymm13
  DB  196,65,20,88,245                    ; vaddps        %ymm13,%ymm13,%ymm14
  DB  196,65,12,88,246                    ; vaddps        %ymm14,%ymm14,%ymm14
  DB  196,66,13,168,246                   ; vfmadd213ps   %ymm14,%ymm14,%ymm14
  DB  196,65,20,92,248                    ; vsubps        %ymm8,%ymm13,%ymm15
  DB  196,65,4,89,246                     ; vmulps        %ymm14,%ymm15,%ymm14
  DB  196,66,53,184,245                   ; vfmadd231ps   %ymm13,%ymm9,%ymm14
  DB  196,65,124,82,253                   ; vrsqrtps      %ymm13,%ymm15
  DB  196,65,124,83,255                   ; vrcpps        %ymm15,%ymm15
  DB  196,65,4,92,253                     ; vsubps        %ymm13,%ymm15,%ymm15
  DB  197,84,88,229                       ; vaddps        %ymm5,%ymm5,%ymm12
  DB  196,65,28,88,228                    ; vaddps        %ymm12,%ymm12,%ymm12
  DB  197,28,194,231,2                    ; vcmpleps      %ymm7,%ymm12,%ymm12
  DB  196,67,5,74,230,192                 ; vblendvps     %ymm12,%ymm14,%ymm15,%ymm12
  DB  197,116,88,241                      ; vaddps        %ymm1,%ymm1,%ymm14
  DB  196,65,60,92,237                    ; vsubps        %ymm13,%ymm8,%ymm13
  DB  197,12,92,251                       ; vsubps        %ymm3,%ymm14,%ymm15
  DB  196,98,5,168,235                    ; vfmadd213ps   %ymm3,%ymm15,%ymm13
  DB  197,4,89,255                        ; vmulps        %ymm7,%ymm15,%ymm15
  DB  196,65,4,89,228                     ; vmulps        %ymm12,%ymm15,%ymm12
  DB  197,20,89,237                       ; vmulps        %ymm5,%ymm13,%ymm13
  DB  196,98,101,184,229                  ; vfmadd231ps   %ymm5,%ymm3,%ymm12
  DB  197,12,194,243,2                    ; vcmpleps      %ymm3,%ymm14,%ymm14
  DB  196,67,29,74,237,224                ; vblendvps     %ymm14,%ymm13,%ymm12,%ymm13
  DB  197,76,94,231                       ; vdivps        %ymm7,%ymm6,%ymm12
  DB  196,67,45,74,212,176                ; vblendvps     %ymm11,%ymm12,%ymm10,%ymm10
  DB  196,65,44,88,218                    ; vaddps        %ymm10,%ymm10,%ymm11
  DB  196,65,36,88,219                    ; vaddps        %ymm11,%ymm11,%ymm11
  DB  196,66,37,168,219                   ; vfmadd213ps   %ymm11,%ymm11,%ymm11
  DB  196,65,44,92,224                    ; vsubps        %ymm8,%ymm10,%ymm12
  DB  196,65,28,89,219                    ; vmulps        %ymm11,%ymm12,%ymm11
  DB  196,66,45,168,203                   ; vfmadd213ps   %ymm11,%ymm10,%ymm9
  DB  196,65,124,82,218                   ; vrsqrtps      %ymm10,%ymm11
  DB  196,65,124,83,219                   ; vrcpps        %ymm11,%ymm11
  DB  196,65,36,92,218                    ; vsubps        %ymm10,%ymm11,%ymm11
  DB  197,76,88,230                       ; vaddps        %ymm6,%ymm6,%ymm12
  DB  196,65,28,88,228                    ; vaddps        %ymm12,%ymm12,%ymm12
  DB  197,28,194,231,2                    ; vcmpleps      %ymm7,%ymm12,%ymm12
  DB  196,67,37,74,201,192                ; vblendvps     %ymm12,%ymm9,%ymm11,%ymm9
  DB  197,124,16,52,36                    ; vmovups       (%rsp),%ymm14
  DB  196,65,12,88,222                    ; vaddps        %ymm14,%ymm14,%ymm11
  DB  197,36,92,227                       ; vsubps        %ymm3,%ymm11,%ymm12
  DB  196,65,60,92,210                    ; vsubps        %ymm10,%ymm8,%ymm10
  DB  196,98,29,168,211                   ; vfmadd213ps   %ymm3,%ymm12,%ymm10
  DB  197,28,89,231                       ; vmulps        %ymm7,%ymm12,%ymm12
  DB  196,65,28,89,201                    ; vmulps        %ymm9,%ymm12,%ymm9
  DB  197,44,89,214                       ; vmulps        %ymm6,%ymm10,%ymm10
  DB  196,98,101,184,206                  ; vfmadd231ps   %ymm6,%ymm3,%ymm9
  DB  197,36,194,219,2                    ; vcmpleps      %ymm3,%ymm11,%ymm11
  DB  196,67,53,74,202,176                ; vblendvps     %ymm11,%ymm10,%ymm9,%ymm9
  DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,220                       ; vmulps        %ymm4,%ymm8,%ymm11
  DB  196,98,45,184,216                   ; vfmadd231ps   %ymm0,%ymm10,%ymm11
  DB  196,193,108,88,195                  ; vaddps        %ymm11,%ymm2,%ymm0
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  196,226,61,184,205                  ; vfmadd231ps   %ymm5,%ymm8,%ymm1
  DB  196,193,116,88,205                  ; vaddps        %ymm13,%ymm1,%ymm1
  DB  196,193,44,89,214                   ; vmulps        %ymm14,%ymm10,%ymm2
  DB  196,226,61,184,214                  ; vfmadd231ps   %ymm6,%ymm8,%ymm2
  DB  196,193,108,88,209                  ; vaddps        %ymm9,%ymm2,%ymm2
  DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,131,196,56                       ; add           $0x38,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_0_hsw
_sk_clamp_0_hsw LABEL PROC
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
  DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
  DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
  DB  196,193,100,95,216                  ; vmaxps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_1_hsw
_sk_clamp_1_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
  DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
  DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
  DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_a_hsw
_sk_clamp_a_hsw LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
  DB  197,252,93,195                      ; vminps        %ymm3,%ymm0,%ymm0
  DB  197,244,93,203                      ; vminps        %ymm3,%ymm1,%ymm1
  DB  197,236,93,211                      ; vminps        %ymm3,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_set_rgb_hsw
_sk_set_rgb_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,226,125,24,0                    ; vbroadcastss  (%rax),%ymm0
  DB  196,226,125,24,72,4                 ; vbroadcastss  0x4(%rax),%ymm1
  DB  196,226,125,24,80,8                 ; vbroadcastss  0x8(%rax),%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_rb_hsw
_sk_swap_rb_hsw LABEL PROC
  DB  197,124,40,192                      ; vmovaps       %ymm0,%ymm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,194                      ; vmovaps       %ymm2,%ymm0
  DB  197,124,41,194                      ; vmovaps       %ymm8,%ymm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_hsw
_sk_swap_hsw LABEL PROC
  DB  197,124,40,195                      ; vmovaps       %ymm3,%ymm8
  DB  197,124,40,202                      ; vmovaps       %ymm2,%ymm9
  DB  197,124,40,209                      ; vmovaps       %ymm1,%ymm10
  DB  197,124,40,216                      ; vmovaps       %ymm0,%ymm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,196                      ; vmovaps       %ymm4,%ymm0
  DB  197,252,40,205                      ; vmovaps       %ymm5,%ymm1
  DB  197,252,40,214                      ; vmovaps       %ymm6,%ymm2
  DB  197,252,40,223                      ; vmovaps       %ymm7,%ymm3
  DB  197,124,41,220                      ; vmovaps       %ymm11,%ymm4
  DB  197,124,41,213                      ; vmovaps       %ymm10,%ymm5
  DB  197,124,41,206                      ; vmovaps       %ymm9,%ymm6
  DB  197,124,41,199                      ; vmovaps       %ymm8,%ymm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_src_dst_hsw
_sk_move_src_dst_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,224                      ; vmovaps       %ymm0,%ymm4
  DB  197,252,40,233                      ; vmovaps       %ymm1,%ymm5
  DB  197,252,40,242                      ; vmovaps       %ymm2,%ymm6
  DB  197,252,40,251                      ; vmovaps       %ymm3,%ymm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_dst_src_hsw
_sk_move_dst_src_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,196                      ; vmovaps       %ymm4,%ymm0
  DB  197,252,40,205                      ; vmovaps       %ymm5,%ymm1
  DB  197,252,40,214                      ; vmovaps       %ymm6,%ymm2
  DB  197,252,40,223                      ; vmovaps       %ymm7,%ymm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_premul_hsw
_sk_premul_hsw LABEL PROC
  DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
  DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
  DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_unpremul_hsw
_sk_unpremul_hsw LABEL PROC
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,100,194,200,0                ; vcmpeqps      %ymm8,%ymm3,%ymm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  197,44,94,211                       ; vdivps        %ymm3,%ymm10,%ymm10
  DB  196,67,45,74,192,144                ; vblendvps     %ymm9,%ymm8,%ymm10,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_srgb_hsw
_sk_from_srgb_hsw LABEL PROC
  DB  184,145,131,158,61                  ; mov           $0x3d9e8391,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
  DB  184,154,153,153,62                  ; mov           $0x3e99999a,%eax
  DB  197,121,110,216                     ; vmovd         %eax,%xmm11
  DB  196,66,125,88,219                   ; vpbroadcastd  %xmm11,%ymm11
  DB  184,92,143,50,63                    ; mov           $0x3f328f5c,%eax
  DB  197,121,110,224                     ; vmovd         %eax,%xmm12
  DB  196,66,125,88,228                   ; vpbroadcastd  %xmm12,%ymm12
  DB  196,65,125,111,235                  ; vmovdqa       %ymm11,%ymm13
  DB  196,66,125,168,236                  ; vfmadd213ps   %ymm12,%ymm0,%ymm13
  DB  184,10,215,35,59                    ; mov           $0x3b23d70a,%eax
  DB  197,121,110,240                     ; vmovd         %eax,%xmm14
  DB  196,66,125,88,246                   ; vpbroadcastd  %xmm14,%ymm14
  DB  196,66,45,168,238                   ; vfmadd213ps   %ymm14,%ymm10,%ymm13
  DB  184,174,71,97,61                    ; mov           $0x3d6147ae,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  196,193,124,194,194,1               ; vcmpltps      %ymm10,%ymm0,%ymm0
  DB  196,195,21,74,193,0                 ; vblendvps     %ymm0,%ymm9,%ymm13,%ymm0
  DB  197,60,89,201                       ; vmulps        %ymm1,%ymm8,%ymm9
  DB  197,116,89,233                      ; vmulps        %ymm1,%ymm1,%ymm13
  DB  196,65,125,111,251                  ; vmovdqa       %ymm11,%ymm15
  DB  196,66,117,168,252                  ; vfmadd213ps   %ymm12,%ymm1,%ymm15
  DB  196,66,21,168,254                   ; vfmadd213ps   %ymm14,%ymm13,%ymm15
  DB  196,193,116,194,202,1               ; vcmpltps      %ymm10,%ymm1,%ymm1
  DB  196,195,5,74,201,16                 ; vblendvps     %ymm1,%ymm9,%ymm15,%ymm1
  DB  197,60,89,194                       ; vmulps        %ymm2,%ymm8,%ymm8
  DB  197,108,89,202                      ; vmulps        %ymm2,%ymm2,%ymm9
  DB  196,66,109,168,220                  ; vfmadd213ps   %ymm12,%ymm2,%ymm11
  DB  196,66,53,168,222                   ; vfmadd213ps   %ymm14,%ymm9,%ymm11
  DB  196,193,108,194,210,1               ; vcmpltps      %ymm10,%ymm2,%ymm2
  DB  196,195,37,74,208,32                ; vblendvps     %ymm2,%ymm8,%ymm11,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_srgb_hsw
_sk_to_srgb_hsw LABEL PROC
  DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
  DB  196,65,124,83,216                   ; vrcpps        %ymm8,%ymm11
  DB  196,65,124,82,224                   ; vrsqrtps      %ymm8,%ymm12
  DB  184,41,92,71,65                     ; mov           $0x41475c29,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,89,232                       ; vmulps        %ymm0,%ymm8,%ymm13
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  184,194,135,210,62                  ; mov           $0x3ed287c2,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  184,206,111,48,63                   ; mov           $0x3f306fce,%eax
  DB  197,121,110,240                     ; vmovd         %eax,%xmm14
  DB  196,66,125,88,246                   ; vpbroadcastd  %xmm14,%ymm14
  DB  184,168,87,202,61                   ; mov           $0x3dca57a8,%eax
  DB  53,0,0,0,128                        ; xor           $0x80000000,%eax
  DB  197,121,110,248                     ; vmovd         %eax,%xmm15
  DB  196,66,125,88,255                   ; vpbroadcastd  %xmm15,%ymm15
  DB  196,66,13,168,223                   ; vfmadd213ps   %ymm15,%ymm14,%ymm11
  DB  196,66,45,184,220                   ; vfmadd231ps   %ymm12,%ymm10,%ymm11
  DB  196,65,52,93,219                    ; vminps        %ymm11,%ymm9,%ymm11
  DB  184,4,231,140,59                    ; mov           $0x3b8ce704,%eax
  DB  197,121,110,224                     ; vmovd         %eax,%xmm12
  DB  196,66,125,88,228                   ; vpbroadcastd  %xmm12,%ymm12
  DB  196,193,124,194,196,1               ; vcmpltps      %ymm12,%ymm0,%ymm0
  DB  196,195,37,74,197,0                 ; vblendvps     %ymm0,%ymm13,%ymm11,%ymm0
  DB  197,124,82,217                      ; vrsqrtps      %ymm1,%ymm11
  DB  196,65,124,83,235                   ; vrcpps        %ymm11,%ymm13
  DB  196,65,124,82,219                   ; vrsqrtps      %ymm11,%ymm11
  DB  196,66,13,168,239                   ; vfmadd213ps   %ymm15,%ymm14,%ymm13
  DB  196,66,45,184,235                   ; vfmadd231ps   %ymm11,%ymm10,%ymm13
  DB  197,60,89,217                       ; vmulps        %ymm1,%ymm8,%ymm11
  DB  196,65,52,93,237                    ; vminps        %ymm13,%ymm9,%ymm13
  DB  196,193,116,194,204,1               ; vcmpltps      %ymm12,%ymm1,%ymm1
  DB  196,195,21,74,203,16                ; vblendvps     %ymm1,%ymm11,%ymm13,%ymm1
  DB  197,124,82,218                      ; vrsqrtps      %ymm2,%ymm11
  DB  196,65,124,83,235                   ; vrcpps        %ymm11,%ymm13
  DB  196,66,13,168,239                   ; vfmadd213ps   %ymm15,%ymm14,%ymm13
  DB  196,65,124,82,219                   ; vrsqrtps      %ymm11,%ymm11
  DB  196,66,45,184,235                   ; vfmadd231ps   %ymm11,%ymm10,%ymm13
  DB  196,65,52,93,205                    ; vminps        %ymm13,%ymm9,%ymm9
  DB  197,60,89,194                       ; vmulps        %ymm2,%ymm8,%ymm8
  DB  196,193,108,194,212,1               ; vcmpltps      %ymm12,%ymm2,%ymm2
  DB  196,195,53,74,208,32                ; vblendvps     %ymm2,%ymm8,%ymm9,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_2dot2_hsw
_sk_from_2dot2_hsw LABEL PROC
  DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,200                   ; vrsqrtps      %ymm8,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  197,252,89,192                      ; vmulps        %ymm0,%ymm0,%ymm0
  DB  196,65,60,89,208                    ; vmulps        %ymm8,%ymm8,%ymm10
  DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
  DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  196,65,124,82,210                   ; vrsqrtps      %ymm10,%ymm10
  DB  197,244,89,201                      ; vmulps        %ymm1,%ymm1,%ymm1
  DB  196,65,52,89,217                    ; vmulps        %ymm9,%ymm9,%ymm11
  DB  196,65,52,89,203                    ; vmulps        %ymm11,%ymm9,%ymm9
  DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
  DB  197,124,82,202                      ; vrsqrtps      %ymm2,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  196,65,124,82,210                   ; vrsqrtps      %ymm10,%ymm10
  DB  197,236,89,210                      ; vmulps        %ymm2,%ymm2,%ymm2
  DB  196,65,52,89,217                    ; vmulps        %ymm9,%ymm9,%ymm11
  DB  196,65,52,89,203                    ; vmulps        %ymm11,%ymm9,%ymm9
  DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
  DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
  DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_2dot2_hsw
_sk_to_2dot2_hsw LABEL PROC
  DB  197,252,82,192                      ; vrsqrtps      %ymm0,%ymm0
  DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,200                   ; vrsqrtps      %ymm8,%ymm9
  DB  197,252,83,192                      ; vrcpps        %ymm0,%ymm0
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  196,65,124,83,193                   ; vrcpps        %ymm9,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
  DB  197,252,82,201                      ; vrsqrtps      %ymm1,%ymm1
  DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  197,252,83,201                      ; vrcpps        %ymm1,%ymm1
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  196,65,124,83,202                   ; vrcpps        %ymm10,%ymm9
  DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
  DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
  DB  197,252,82,210                      ; vrsqrtps      %ymm2,%ymm2
  DB  197,124,82,202                      ; vrsqrtps      %ymm2,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  197,252,83,210                      ; vrcpps        %ymm2,%ymm2
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  196,65,124,83,202                   ; vrcpps        %ymm10,%ymm9
  DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
  DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_rgb_to_hsl_hsw
_sk_rgb_to_hsl_hsw LABEL PROC
  DB  72,131,236,56                       ; sub           $0x38,%rsp
  DB  197,252,17,60,36                    ; vmovups       %ymm7,(%rsp)
  DB  197,252,40,254                      ; vmovaps       %ymm6,%ymm7
  DB  197,252,40,245                      ; vmovaps       %ymm5,%ymm6
  DB  197,252,40,236                      ; vmovaps       %ymm4,%ymm5
  DB  197,252,40,227                      ; vmovaps       %ymm3,%ymm4
  DB  197,252,40,216                      ; vmovaps       %ymm0,%ymm3
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  65,184,171,170,42,62                ; mov           $0x3e2aaaab,%r8d
  DB  184,0,0,192,64                      ; mov           $0x40c00000,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  65,185,0,0,0,64                     ; mov           $0x40000000,%r9d
  DB  184,0,0,128,64                      ; mov           $0x40800000,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  197,100,95,217                      ; vmaxps        %ymm1,%ymm3,%ymm11
  DB  197,36,95,218                       ; vmaxps        %ymm2,%ymm11,%ymm11
  DB  197,100,93,225                      ; vminps        %ymm1,%ymm3,%ymm12
  DB  197,28,93,226                       ; vminps        %ymm2,%ymm12,%ymm12
  DB  196,65,36,92,236                    ; vsubps        %ymm12,%ymm11,%ymm13
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,65,60,94,197                    ; vdivps        %ymm13,%ymm8,%ymm8
  DB  197,116,194,242,1                   ; vcmpltps      %ymm2,%ymm1,%ymm14
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  196,65,4,87,255                     ; vxorps        %ymm15,%ymm15,%ymm15
  DB  196,67,5,74,201,224                 ; vblendvps     %ymm14,%ymm9,%ymm15,%ymm9
  DB  197,116,92,242                      ; vsubps        %ymm2,%ymm1,%ymm14
  DB  196,66,61,168,241                   ; vfmadd213ps   %ymm9,%ymm8,%ymm14
  DB  197,236,92,195                      ; vsubps        %ymm3,%ymm2,%ymm0
  DB  197,100,92,201                      ; vsubps        %ymm1,%ymm3,%ymm9
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  196,66,61,168,202                   ; vfmadd213ps   %ymm10,%ymm8,%ymm9
  DB  196,193,121,110,209                 ; vmovd         %r9d,%xmm2
  DB  196,98,125,88,210                   ; vpbroadcastd  %xmm2,%ymm10
  DB  196,194,61,168,194                  ; vfmadd213ps   %ymm10,%ymm8,%ymm0
  DB  197,164,194,201,0                   ; vcmpeqps      %ymm1,%ymm11,%ymm1
  DB  196,227,53,74,192,16                ; vblendvps     %ymm1,%ymm0,%ymm9,%ymm0
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,164,194,211,0                   ; vcmpeqps      %ymm3,%ymm11,%ymm2
  DB  196,195,125,74,198,32               ; vblendvps     %ymm2,%ymm14,%ymm0,%ymm0
  DB  196,193,36,88,220                   ; vaddps        %ymm12,%ymm11,%ymm3
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,228,89,209                      ; vmulps        %ymm1,%ymm3,%ymm2
  DB  197,244,194,202,1                   ; vcmpltps      %ymm2,%ymm1,%ymm1
  DB  196,65,44,92,195                    ; vsubps        %ymm11,%ymm10,%ymm8
  DB  196,65,60,92,196                    ; vsubps        %ymm12,%ymm8,%ymm8
  DB  196,195,101,74,200,16               ; vblendvps     %ymm1,%ymm8,%ymm3,%ymm1
  DB  196,193,36,194,220,0                ; vcmpeqps      %ymm12,%ymm11,%ymm3
  DB  197,148,94,201                      ; vdivps        %ymm1,%ymm13,%ymm1
  DB  196,195,125,74,199,48               ; vblendvps     %ymm3,%ymm15,%ymm0,%ymm0
  DB  196,195,117,74,207,48               ; vblendvps     %ymm3,%ymm15,%ymm1,%ymm1
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,228,89,192                      ; vmulps        %ymm0,%ymm3,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,220                      ; vmovaps       %ymm4,%ymm3
  DB  197,252,40,229                      ; vmovaps       %ymm5,%ymm4
  DB  197,252,40,238                      ; vmovaps       %ymm6,%ymm5
  DB  197,252,40,247                      ; vmovaps       %ymm7,%ymm6
  DB  197,252,16,60,36                    ; vmovups       (%rsp),%ymm7
  DB  72,131,196,56                       ; add           $0x38,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hsl_to_rgb_hsw
_sk_hsl_to_rgb_hsw LABEL PROC
  DB  72,129,236,184,0,0,0                ; sub           $0xb8,%rsp
  DB  197,252,17,188,36,128,0,0,0         ; vmovups       %ymm7,0x80(%rsp)
  DB  197,252,17,116,36,96                ; vmovups       %ymm6,0x60(%rsp)
  DB  197,252,17,108,36,64                ; vmovups       %ymm5,0x40(%rsp)
  DB  197,252,17,100,36,32                ; vmovups       %ymm4,0x20(%rsp)
  DB  197,252,17,28,36                    ; vmovups       %ymm3,(%rsp)
  DB  197,252,40,233                      ; vmovaps       %ymm1,%ymm5
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,98,125,88,193                   ; vpbroadcastd  %xmm1,%ymm8
  DB  196,193,108,194,200,1               ; vcmpltps      %ymm8,%ymm2,%ymm1
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,224                     ; vmovd         %eax,%xmm4
  DB  196,98,125,88,212                   ; vpbroadcastd  %xmm4,%ymm10
  DB  197,172,88,229                      ; vaddps        %ymm5,%ymm10,%ymm4
  DB  197,220,89,226                      ; vmulps        %ymm2,%ymm4,%ymm4
  DB  197,84,88,202                       ; vaddps        %ymm2,%ymm5,%ymm9
  DB  196,98,85,188,202                   ; vfnmadd231ps  %ymm2,%ymm5,%ymm9
  DB  196,99,53,74,204,16                 ; vblendvps     %ymm1,%ymm4,%ymm9,%ymm9
  DB  65,184,0,0,0,64                     ; mov           $0x40000000,%r8d
  DB  184,171,170,170,62                  ; mov           $0x3eaaaaab,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,98,125,88,233                   ; vpbroadcastd  %xmm1,%ymm13
  DB  197,148,88,224                      ; vaddps        %ymm0,%ymm13,%ymm4
  DB  184,0,0,0,0                         ; mov           $0x0,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,98,125,88,225                   ; vpbroadcastd  %xmm1,%ymm12
  DB  197,172,194,204,1                   ; vcmpltps      %ymm4,%ymm10,%ymm1
  DB  196,65,92,92,218                    ; vsubps        %ymm10,%ymm4,%ymm11
  DB  196,195,93,74,203,16                ; vblendvps     %ymm1,%ymm11,%ymm4,%ymm1
  DB  196,65,92,194,220,1                 ; vcmpltps      %ymm12,%ymm4,%ymm11
  DB  197,44,88,244                       ; vaddps        %ymm4,%ymm10,%ymm14
  DB  196,195,117,74,206,176              ; vblendvps     %ymm11,%ymm14,%ymm1,%ymm1
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,219                   ; vpbroadcastd  %xmm3,%ymm11
  DB  196,66,109,170,217                  ; vfmsub213ps   %ymm9,%ymm2,%ymm11
  DB  65,184,171,170,42,62                ; mov           $0x3e2aaaab,%r8d
  DB  184,0,0,192,64                      ; mov           $0x40c00000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,65,52,92,243                    ; vsubps        %ymm11,%ymm9,%ymm14
  DB  197,12,89,243                       ; vmulps        %ymm3,%ymm14,%ymm14
  DB  184,171,170,42,63                   ; mov           $0x3f2aaaab,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,98,125,88,251                   ; vpbroadcastd  %xmm3,%ymm15
  DB  197,132,92,217                      ; vsubps        %ymm1,%ymm15,%ymm3
  DB  196,194,13,168,219                  ; vfmadd213ps   %ymm11,%ymm14,%ymm3
  DB  196,193,116,194,255,1               ; vcmpltps      %ymm15,%ymm1,%ymm7
  DB  196,227,37,74,219,112               ; vblendvps     %ymm7,%ymm3,%ymm11,%ymm3
  DB  196,193,116,194,248,1               ; vcmpltps      %ymm8,%ymm1,%ymm7
  DB  196,195,101,74,249,112              ; vblendvps     %ymm7,%ymm9,%ymm3,%ymm7
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,244,194,203,1                   ; vcmpltps      %ymm3,%ymm1,%ymm1
  DB  196,194,13,168,227                  ; vfmadd213ps   %ymm11,%ymm14,%ymm4
  DB  196,227,69,74,228,16                ; vblendvps     %ymm1,%ymm4,%ymm7,%ymm4
  DB  197,172,194,200,1                   ; vcmpltps      %ymm0,%ymm10,%ymm1
  DB  196,193,124,92,250                  ; vsubps        %ymm10,%ymm0,%ymm7
  DB  196,227,125,74,207,16               ; vblendvps     %ymm1,%ymm7,%ymm0,%ymm1
  DB  196,193,124,194,252,1               ; vcmpltps      %ymm12,%ymm0,%ymm7
  DB  197,172,88,240                      ; vaddps        %ymm0,%ymm10,%ymm6
  DB  196,227,117,74,206,112              ; vblendvps     %ymm7,%ymm6,%ymm1,%ymm1
  DB  197,132,92,241                      ; vsubps        %ymm1,%ymm15,%ymm6
  DB  196,194,13,168,243                  ; vfmadd213ps   %ymm11,%ymm14,%ymm6
  DB  196,193,116,194,255,1               ; vcmpltps      %ymm15,%ymm1,%ymm7
  DB  196,227,37,74,246,112               ; vblendvps     %ymm7,%ymm6,%ymm11,%ymm6
  DB  196,193,116,194,248,1               ; vcmpltps      %ymm8,%ymm1,%ymm7
  DB  196,195,77,74,241,112               ; vblendvps     %ymm7,%ymm9,%ymm6,%ymm6
  DB  197,244,194,203,1                   ; vcmpltps      %ymm3,%ymm1,%ymm1
  DB  196,193,124,92,253                  ; vsubps        %ymm13,%ymm0,%ymm7
  DB  196,194,13,168,195                  ; vfmadd213ps   %ymm11,%ymm14,%ymm0
  DB  196,227,77,74,200,16                ; vblendvps     %ymm1,%ymm0,%ymm6,%ymm1
  DB  197,172,194,199,1                   ; vcmpltps      %ymm7,%ymm10,%ymm0
  DB  196,193,68,92,242                   ; vsubps        %ymm10,%ymm7,%ymm6
  DB  196,227,69,74,198,0                 ; vblendvps     %ymm0,%ymm6,%ymm7,%ymm0
  DB  196,193,68,194,244,1                ; vcmpltps      %ymm12,%ymm7,%ymm6
  DB  197,44,88,215                       ; vaddps        %ymm7,%ymm10,%ymm10
  DB  196,195,125,74,194,96               ; vblendvps     %ymm6,%ymm10,%ymm0,%ymm0
  DB  196,194,13,168,251                  ; vfmadd213ps   %ymm11,%ymm14,%ymm7
  DB  197,132,92,240                      ; vsubps        %ymm0,%ymm15,%ymm6
  DB  196,194,13,168,243                  ; vfmadd213ps   %ymm11,%ymm14,%ymm6
  DB  196,65,124,194,215,1                ; vcmpltps      %ymm15,%ymm0,%ymm10
  DB  196,227,37,74,246,160               ; vblendvps     %ymm10,%ymm6,%ymm11,%ymm6
  DB  196,65,124,194,192,1                ; vcmpltps      %ymm8,%ymm0,%ymm8
  DB  196,195,77,74,241,128               ; vblendvps     %ymm8,%ymm9,%ymm6,%ymm6
  DB  197,252,194,195,1                   ; vcmpltps      %ymm3,%ymm0,%ymm0
  DB  196,227,77,74,223,0                 ; vblendvps     %ymm0,%ymm7,%ymm6,%ymm3
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,212,194,232,0                   ; vcmpeqps      %ymm0,%ymm5,%ymm5
  DB  196,227,93,74,194,80                ; vblendvps     %ymm5,%ymm2,%ymm4,%ymm0
  DB  196,227,117,74,202,80               ; vblendvps     %ymm5,%ymm2,%ymm1,%ymm1
  DB  196,227,101,74,210,80               ; vblendvps     %ymm5,%ymm2,%ymm3,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,16,28,36                    ; vmovups       (%rsp),%ymm3
  DB  197,252,16,100,36,32                ; vmovups       0x20(%rsp),%ymm4
  DB  197,252,16,108,36,64                ; vmovups       0x40(%rsp),%ymm5
  DB  197,252,16,116,36,96                ; vmovups       0x60(%rsp),%ymm6
  DB  197,252,16,188,36,128,0,0,0         ; vmovups       0x80(%rsp),%ymm7
  DB  72,129,196,184,0,0,0                ; add           $0xb8,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_1_float_hsw
_sk_scale_1_float_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_u8_hsw
_sk_scale_u8_hsw LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,56                              ; jne           126b <_sk_scale_u8_hsw+0x48>
  DB  197,122,126,0                       ; vmovq         (%rax),%xmm8
  DB  196,66,125,49,192                   ; vpmovzxbd     %xmm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           1273 <_sk_scale_u8_hsw+0x50>
  DB  196,65,249,110,193                  ; vmovq         %r9,%xmm8
  DB  235,167                             ; jmp           1237 <_sk_scale_u8_hsw+0x14>

PUBLIC _sk_lerp_1_float_hsw
_sk_lerp_1_float_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
  DB  196,226,61,168,196                  ; vfmadd213ps   %ymm4,%ymm8,%ymm0
  DB  197,244,92,205                      ; vsubps        %ymm5,%ymm1,%ymm1
  DB  196,226,61,168,205                  ; vfmadd213ps   %ymm5,%ymm8,%ymm1
  DB  197,236,92,214                      ; vsubps        %ymm6,%ymm2,%ymm2
  DB  196,226,61,168,214                  ; vfmadd213ps   %ymm6,%ymm8,%ymm2
  DB  197,228,92,223                      ; vsubps        %ymm7,%ymm3,%ymm3
  DB  196,226,61,168,223                  ; vfmadd213ps   %ymm7,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_u8_hsw
_sk_lerp_u8_hsw LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,76                              ; jne           131b <_sk_lerp_u8_hsw+0x5c>
  DB  197,122,126,0                       ; vmovq         (%rax),%xmm8
  DB  196,66,125,49,192                   ; vpmovzxbd     %xmm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
  DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
  DB  196,226,61,168,196                  ; vfmadd213ps   %ymm4,%ymm8,%ymm0
  DB  197,244,92,205                      ; vsubps        %ymm5,%ymm1,%ymm1
  DB  196,226,61,168,205                  ; vfmadd213ps   %ymm5,%ymm8,%ymm1
  DB  197,236,92,214                      ; vsubps        %ymm6,%ymm2,%ymm2
  DB  196,226,61,168,214                  ; vfmadd213ps   %ymm6,%ymm8,%ymm2
  DB  197,228,92,223                      ; vsubps        %ymm7,%ymm3,%ymm3
  DB  196,226,61,168,223                  ; vfmadd213ps   %ymm7,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           1323 <_sk_lerp_u8_hsw+0x64>
  DB  196,65,249,110,193                  ; vmovq         %r9,%xmm8
  DB  235,147                             ; jmp           12d3 <_sk_lerp_u8_hsw+0x14>

PUBLIC _sk_lerp_565_hsw
_sk_lerp_565_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,179,0,0,0                    ; jne           1401 <_sk_lerp_565_hsw+0xc1>
  DB  196,193,122,111,28,122              ; vmovdqu       (%r10,%rdi,2),%xmm3
  DB  196,98,125,51,195                   ; vpmovzxwd     %xmm3,%ymm8
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
  DB  197,124,91,203                      ; vcvtdq2ps     %ymm3,%ymm9
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,52,89,203                       ; vmulps        %ymm3,%ymm9,%ymm9
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
  DB  197,124,91,211                      ; vcvtdq2ps     %ymm3,%ymm10
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,44,89,211                       ; vmulps        %ymm3,%ymm10,%ymm10
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
  DB  196,226,53,168,196                  ; vfmadd213ps   %ymm4,%ymm9,%ymm0
  DB  197,244,92,205                      ; vsubps        %ymm5,%ymm1,%ymm1
  DB  196,226,45,168,205                  ; vfmadd213ps   %ymm5,%ymm10,%ymm1
  DB  197,236,92,214                      ; vsubps        %ymm6,%ymm2,%ymm2
  DB  196,226,101,168,214                 ; vfmadd213ps   %ymm6,%ymm3,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  197,225,239,219                     ; vpxor         %xmm3,%xmm3,%xmm3
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  15,135,59,255,255,255               ; ja            1354 <_sk_lerp_565_hsw+0x14>
  DB  69,15,182,192                       ; movzbl        %r8b,%r8d
  DB  76,141,13,76,0,0,0                  ; lea           0x4c(%rip),%r9        # 1470 <_sk_lerp_565_hsw+0x130>
  DB  75,99,4,129                         ; movslq        (%r9,%r8,4),%rax
  DB  76,1,200                            ; add           %r9,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,225,239,219                     ; vpxor         %xmm3,%xmm3,%xmm3
  DB  196,193,97,196,92,122,12,6          ; vpinsrw       $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm3
  DB  196,193,97,196,92,122,10,5          ; vpinsrw       $0x5,0xa(%r10,%rdi,2),%xmm3,%xmm3
  DB  196,193,97,196,92,122,8,4           ; vpinsrw       $0x4,0x8(%r10,%rdi,2),%xmm3,%xmm3
  DB  196,193,97,196,92,122,6,3           ; vpinsrw       $0x3,0x6(%r10,%rdi,2),%xmm3,%xmm3
  DB  196,193,97,196,92,122,4,2           ; vpinsrw       $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3
  DB  196,193,97,196,92,122,2,1           ; vpinsrw       $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3
  DB  196,193,97,196,28,122,0             ; vpinsrw       $0x0,(%r10,%rdi,2),%xmm3,%xmm3
  DB  233,231,254,255,255                 ; jmpq          1354 <_sk_lerp_565_hsw+0x14>
  DB  15,31,0                             ; nopl          (%rax)
  DB  241                                 ; icebp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  233,255,255,255,225                 ; jmpq          ffffffffe2001478 <_sk_callback_hsw+0xffffffffe1ffd27e>
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  217,255                             ; fcos
  DB  255                                 ; (bad)
  DB  255,209                             ; callq         *%rcx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,201                             ; dec           %ecx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  189                                 ; .byte         0xbd
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_tables_hsw
_sk_load_tables_hsw LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  76,3,8                              ; add           (%rax),%r9
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,121                             ; jne           151a <_sk_load_tables_hsw+0x8e>
  DB  196,193,126,111,25                  ; vmovdqu       (%r9),%ymm3
  DB  185,255,0,0,0                       ; mov           $0xff,%ecx
  DB  197,249,110,193                     ; vmovd         %ecx,%xmm0
  DB  196,226,125,88,208                  ; vpbroadcastd  %xmm0,%ymm2
  DB  197,237,219,203                     ; vpand         %ymm3,%ymm2,%ymm1
  DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
  DB  72,139,72,8                         ; mov           0x8(%rax),%rcx
  DB  76,139,72,16                        ; mov           0x10(%rax),%r9
  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
  DB  196,226,53,146,4,137                ; vgatherdps    %ymm9,(%rcx,%ymm1,4),%ymm0
  DB  197,245,114,211,8                   ; vpsrld        $0x8,%ymm3,%ymm1
  DB  197,109,219,201                     ; vpand         %ymm1,%ymm2,%ymm9
  DB  196,65,45,118,210                   ; vpcmpeqd      %ymm10,%ymm10,%ymm10
  DB  196,130,45,146,12,137               ; vgatherdps    %ymm10,(%r9,%ymm9,4),%ymm1
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  197,181,114,211,16                  ; vpsrld        $0x10,%ymm3,%ymm9
  DB  196,65,109,219,201                  ; vpand         %ymm9,%ymm2,%ymm9
  DB  196,162,61,146,20,136               ; vgatherdps    %ymm8,(%rax,%ymm9,4),%ymm2
  DB  197,229,114,211,24                  ; vpsrld        $0x18,%ymm3,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  185,8,0,0,0                         ; mov           $0x8,%ecx
  DB  68,41,193                           ; sub           %r8d,%ecx
  DB  192,225,3                           ; shl           $0x3,%cl
  DB  73,199,194,255,255,255,255          ; mov           $0xffffffffffffffff,%r10
  DB  73,211,234                          ; shr           %cl,%r10
  DB  196,193,249,110,194                 ; vmovq         %r10,%xmm0
  DB  196,226,125,33,192                  ; vpmovsxbd     %xmm0,%ymm0
  DB  196,194,125,140,25                  ; vpmaskmovd    (%r9),%ymm0,%ymm3
  DB  233,99,255,255,255                  ; jmpq          14a6 <_sk_load_tables_hsw+0x1a>

PUBLIC _sk_load_tables_u16_be_hsw
_sk_load_tables_u16_be_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,216,0,0,0                    ; jne           1631 <_sk_load_tables_u16_be_hsw+0xee>
  DB  196,1,121,16,4,72                   ; vmovupd       (%r8,%r9,2),%xmm8
  DB  196,129,121,16,84,72,16             ; vmovupd       0x10(%r8,%r9,2),%xmm2
  DB  196,129,121,16,92,72,32             ; vmovupd       0x20(%r8,%r9,2),%xmm3
  DB  196,1,122,111,76,72,48              ; vmovdqu       0x30(%r8,%r9,2),%xmm9
  DB  197,185,97,194                      ; vpunpcklwd    %xmm2,%xmm8,%xmm0
  DB  197,185,105,210                     ; vpunpckhwd    %xmm2,%xmm8,%xmm2
  DB  196,193,97,97,201                   ; vpunpcklwd    %xmm9,%xmm3,%xmm1
  DB  196,193,97,105,217                  ; vpunpckhwd    %xmm9,%xmm3,%xmm3
  DB  197,121,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm8
  DB  197,121,105,202                     ; vpunpckhwd    %xmm2,%xmm0,%xmm9
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,113,105,235                     ; vpunpckhwd    %xmm3,%xmm1,%xmm13
  DB  197,185,108,194                     ; vpunpcklqdq   %xmm2,%xmm8,%xmm0
  DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
  DB  65,184,255,0,0,0                    ; mov           $0xff,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,209                   ; vpbroadcastd  %xmm1,%ymm10
  DB  196,193,125,219,202                 ; vpand         %ymm10,%ymm0,%ymm1
  DB  196,65,37,118,219                   ; vpcmpeqd      %ymm11,%ymm11,%ymm11
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  76,139,72,16                        ; mov           0x10(%rax),%r9
  DB  196,65,29,118,228                   ; vpcmpeqd      %ymm12,%ymm12,%ymm12
  DB  196,194,29,146,4,136                ; vgatherdps    %ymm12,(%r8,%ymm1,4),%ymm0
  DB  197,185,109,202                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm1
  DB  196,226,125,51,201                  ; vpmovzxwd     %xmm1,%ymm1
  DB  196,193,117,219,210                 ; vpand         %ymm10,%ymm1,%ymm2
  DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
  DB  196,194,61,146,12,145               ; vgatherdps    %ymm8,(%r9,%ymm2,4),%ymm1
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  196,193,49,108,213                  ; vpunpcklqdq   %xmm13,%xmm9,%xmm2
  DB  196,226,125,51,210                  ; vpmovzxwd     %xmm2,%ymm2
  DB  196,65,109,219,194                  ; vpand         %ymm10,%ymm2,%ymm8
  DB  196,162,37,146,20,128               ; vgatherdps    %ymm11,(%rax,%ymm8,4),%ymm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,98,125,88,195                   ; vpbroadcastd  %xmm3,%ymm8
  DB  196,193,49,109,221                  ; vpunpckhqdq   %xmm13,%xmm9,%xmm3
  DB  197,177,113,243,8                   ; vpsllw        $0x8,%xmm3,%xmm9
  DB  197,225,113,211,8                   ; vpsrlw        $0x8,%xmm3,%xmm3
  DB  197,177,235,219                     ; vpor          %xmm3,%xmm9,%xmm3
  DB  196,226,125,51,219                  ; vpmovzxwd     %xmm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,1,123,16,4,72                   ; vmovsd        (%r8,%r9,2),%xmm8
  DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,85                              ; je            1697 <_sk_load_tables_u16_be_hsw+0x154>
  DB  196,1,57,22,68,72,8                 ; vmovhpd       0x8(%r8,%r9,2),%xmm8,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,72                              ; jb            1697 <_sk_load_tables_u16_be_hsw+0x154>
  DB  196,129,123,16,84,72,16             ; vmovsd        0x10(%r8,%r9,2),%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  116,72                              ; je            16a4 <_sk_load_tables_u16_be_hsw+0x161>
  DB  196,129,105,22,84,72,24             ; vmovhpd       0x18(%r8,%r9,2),%xmm2,%xmm2
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,59                              ; jb            16a4 <_sk_load_tables_u16_be_hsw+0x161>
  DB  196,129,123,16,92,72,32             ; vmovsd        0x20(%r8,%r9,2),%xmm3
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  15,132,250,254,255,255              ; je            1574 <_sk_load_tables_u16_be_hsw+0x31>
  DB  196,129,97,22,92,72,40              ; vmovhpd       0x28(%r8,%r9,2),%xmm3,%xmm3
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  15,130,233,254,255,255              ; jb            1574 <_sk_load_tables_u16_be_hsw+0x31>
  DB  196,1,122,126,76,72,48              ; vmovq         0x30(%r8,%r9,2),%xmm9
  DB  233,221,254,255,255                 ; jmpq          1574 <_sk_load_tables_u16_be_hsw+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  197,233,87,210                      ; vxorpd        %xmm2,%xmm2,%xmm2
  DB  233,208,254,255,255                 ; jmpq          1574 <_sk_load_tables_u16_be_hsw+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  233,199,254,255,255                 ; jmpq          1574 <_sk_load_tables_u16_be_hsw+0x31>

PUBLIC _sk_load_tables_rgb_u16_be_hsw
_sk_load_tables_rgb_u16_be_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,127                       ; lea           (%rdi,%rdi,2),%r9
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,207,0,0,0                    ; jne           178e <_sk_load_tables_rgb_u16_be_hsw+0xe1>
  DB  196,129,122,111,4,72                ; vmovdqu       (%r8,%r9,2),%xmm0
  DB  196,129,122,111,84,72,12            ; vmovdqu       0xc(%r8,%r9,2),%xmm2
  DB  196,129,122,111,76,72,24            ; vmovdqu       0x18(%r8,%r9,2),%xmm1
  DB  196,129,122,111,92,72,32            ; vmovdqu       0x20(%r8,%r9,2),%xmm3
  DB  197,225,115,219,4                   ; vpsrldq       $0x4,%xmm3,%xmm3
  DB  197,185,115,216,6                   ; vpsrldq       $0x6,%xmm0,%xmm8
  DB  197,177,115,218,6                   ; vpsrldq       $0x6,%xmm2,%xmm9
  DB  197,161,115,217,6                   ; vpsrldq       $0x6,%xmm1,%xmm11
  DB  197,169,115,219,6                   ; vpsrldq       $0x6,%xmm3,%xmm10
  DB  197,249,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm0
  DB  196,193,57,97,209                   ; vpunpcklwd    %xmm9,%xmm8,%xmm2
  DB  197,241,97,203                      ; vpunpcklwd    %xmm3,%xmm1,%xmm1
  DB  196,193,33,97,218                   ; vpunpcklwd    %xmm10,%xmm11,%xmm3
  DB  197,121,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm8
  DB  197,121,105,202                     ; vpunpckhwd    %xmm2,%xmm0,%xmm9
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,241,105,219                     ; vpunpckhwd    %xmm3,%xmm1,%xmm3
  DB  197,185,108,194                     ; vpunpcklqdq   %xmm2,%xmm8,%xmm0
  DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
  DB  65,184,255,0,0,0                    ; mov           $0xff,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,209                   ; vpbroadcastd  %xmm1,%ymm10
  DB  196,193,125,219,202                 ; vpand         %ymm10,%ymm0,%ymm1
  DB  196,65,37,118,219                   ; vpcmpeqd      %ymm11,%ymm11,%ymm11
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  76,139,72,16                        ; mov           0x10(%rax),%r9
  DB  196,65,29,118,228                   ; vpcmpeqd      %ymm12,%ymm12,%ymm12
  DB  196,194,29,146,4,136                ; vgatherdps    %ymm12,(%r8,%ymm1,4),%ymm0
  DB  197,185,109,202                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm1
  DB  196,226,125,51,201                  ; vpmovzxwd     %xmm1,%ymm1
  DB  196,193,117,219,210                 ; vpand         %ymm10,%ymm1,%ymm2
  DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
  DB  196,194,61,146,12,145               ; vgatherdps    %ymm8,(%r9,%ymm2,4),%ymm1
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  197,177,108,211                     ; vpunpcklqdq   %xmm3,%xmm9,%xmm2
  DB  196,226,125,51,210                  ; vpmovzxwd     %xmm2,%ymm2
  DB  196,193,109,219,218                 ; vpand         %ymm10,%ymm2,%ymm3
  DB  196,226,37,146,20,152               ; vgatherdps    %ymm11,(%rax,%ymm3,4),%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,129,121,110,4,72                ; vmovd         (%r8,%r9,2),%xmm0
  DB  196,129,121,196,68,72,4,2           ; vpinsrw       $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  117,5                               ; jne           17a7 <_sk_load_tables_rgb_u16_be_hsw+0xfa>
  DB  233,76,255,255,255                  ; jmpq          16f3 <_sk_load_tables_rgb_u16_be_hsw+0x46>
  DB  196,129,121,110,76,72,6             ; vmovd         0x6(%r8,%r9,2),%xmm1
  DB  196,1,113,196,68,72,10,2            ; vpinsrw       $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,26                              ; jb            17d6 <_sk_load_tables_rgb_u16_be_hsw+0x129>
  DB  196,129,121,110,76,72,12            ; vmovd         0xc(%r8,%r9,2),%xmm1
  DB  196,129,113,196,84,72,16,2          ; vpinsrw       $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  117,10                              ; jne           17db <_sk_load_tables_rgb_u16_be_hsw+0x12e>
  DB  233,29,255,255,255                  ; jmpq          16f3 <_sk_load_tables_rgb_u16_be_hsw+0x46>
  DB  233,24,255,255,255                  ; jmpq          16f3 <_sk_load_tables_rgb_u16_be_hsw+0x46>
  DB  196,129,121,110,76,72,18            ; vmovd         0x12(%r8,%r9,2),%xmm1
  DB  196,1,113,196,76,72,22,2            ; vpinsrw       $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,26                              ; jb            180a <_sk_load_tables_rgb_u16_be_hsw+0x15d>
  DB  196,129,121,110,76,72,24            ; vmovd         0x18(%r8,%r9,2),%xmm1
  DB  196,129,113,196,76,72,28,2          ; vpinsrw       $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  117,10                              ; jne           180f <_sk_load_tables_rgb_u16_be_hsw+0x162>
  DB  233,233,254,255,255                 ; jmpq          16f3 <_sk_load_tables_rgb_u16_be_hsw+0x46>
  DB  233,228,254,255,255                 ; jmpq          16f3 <_sk_load_tables_rgb_u16_be_hsw+0x46>
  DB  196,129,121,110,92,72,30            ; vmovd         0x1e(%r8,%r9,2),%xmm3
  DB  196,1,97,196,92,72,34,2             ; vpinsrw       $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,20                              ; jb            1838 <_sk_load_tables_rgb_u16_be_hsw+0x18b>
  DB  196,129,121,110,92,72,36            ; vmovd         0x24(%r8,%r9,2),%xmm3
  DB  196,129,97,196,92,72,40,2           ; vpinsrw       $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
  DB  233,187,254,255,255                 ; jmpq          16f3 <_sk_load_tables_rgb_u16_be_hsw+0x46>
  DB  233,182,254,255,255                 ; jmpq          16f3 <_sk_load_tables_rgb_u16_be_hsw+0x46>

PUBLIC _sk_byte_tables_hsw
_sk_byte_tables_hsw LABEL PROC
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,127,67                   ; mov           $0x437f0000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  68,137,197                          ; mov           %r8d,%ebp
  DB  77,137,194                          ; mov           %r8,%r10
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,192                 ; vmovq         %xmm0,%r8
  DB  69,137,195                          ; mov           %r8d,%r11d
  DB  77,137,199                          ; mov           %r8,%r15
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  69,137,198                          ; mov           %r8d,%r14d
  DB  77,137,196                          ; mov           %r8,%r12
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,221                          ; mov           %ebx,%r13d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  196,131,121,32,4,25,0               ; vpinsrb       $0x0,(%r9,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,57,1               ; vpinsrb       $0x1,(%r9,%r15,1),%xmm0,%xmm0
  DB  65,15,182,44,41                     ; movzbl        (%r9,%rbp,1),%ebp
  DB  196,227,121,32,197,2                ; vpinsrb       $0x2,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,17                     ; movzbl        (%r9,%r10,1),%ebp
  DB  196,227,121,32,197,3                ; vpinsrb       $0x3,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,41                     ; movzbl        (%r9,%r13,1),%ebp
  DB  196,227,121,32,197,4                ; vpinsrb       $0x4,%ebp,%xmm0,%xmm0
  DB  65,15,182,44,25                     ; movzbl        (%r9,%rbx,1),%ebp
  DB  196,227,121,32,197,5                ; vpinsrb       $0x5,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,49                     ; movzbl        (%r9,%r14,1),%ebp
  DB  196,227,121,32,197,6                ; vpinsrb       $0x6,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,33                     ; movzbl        (%r9,%r12,1),%ebp
  DB  196,227,121,32,197,7                ; vpinsrb       $0x7,%ebp,%xmm0,%xmm0
  DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
  DB  197,124,91,208                      ; vcvtdq2ps     %ymm0,%ymm10
  DB  189,129,128,128,59                  ; mov           $0x3b808081,%ebp
  DB  197,249,110,197                     ; vmovd         %ebp,%xmm0
  DB  196,98,125,88,200                   ; vpbroadcastd  %xmm0,%ymm9
  DB  196,193,44,89,193                   ; vmulps        %ymm9,%ymm10,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
  DB  196,227,249,22,205,1                ; vpextrq       $0x1,%xmm1,%rbp
  DB  65,137,233                          ; mov           %ebp,%r9d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,203                 ; vmovq         %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,57,201,1                ; vextracti128  $0x1,%ymm1,%xmm1
  DB  196,195,249,22,203,1                ; vpextrq       $0x1,%xmm1,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  196,193,249,126,207                 ; vmovq         %xmm1,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,12,16,0              ; vpinsrb       $0x0,(%r8,%r10,1),%xmm0,%xmm1
  DB  196,195,113,32,12,24,1              ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm1,%xmm1
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,113,32,203,2                ; vpinsrb       $0x2,%ebx,%xmm1,%xmm1
  DB  65,15,182,44,40                     ; movzbl        (%r8,%rbp,1),%ebp
  DB  196,227,113,32,205,3                ; vpinsrb       $0x3,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,32                     ; movzbl        (%r8,%r12,1),%ebp
  DB  196,227,113,32,205,4                ; vpinsrb       $0x4,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,56                     ; movzbl        (%r8,%r15,1),%ebp
  DB  196,227,113,32,205,5                ; vpinsrb       $0x5,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
  DB  196,227,113,32,205,6                ; vpinsrb       $0x6,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
  DB  196,227,113,32,205,7                ; vpinsrb       $0x7,%ebp,%xmm1,%xmm1
  DB  196,226,125,49,201                  ; vpmovzxbd     %xmm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
  DB  196,227,249,22,213,1                ; vpextrq       $0x1,%xmm2,%rbp
  DB  65,137,233                          ; mov           %ebp,%r9d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,211                 ; vmovq         %xmm2,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,57,210,1                ; vextracti128  $0x1,%ymm2,%xmm2
  DB  196,195,249,22,211,1                ; vpextrq       $0x1,%xmm2,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  196,193,249,126,215                 ; vmovq         %xmm2,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,20,16,0              ; vpinsrb       $0x0,(%r8,%r10,1),%xmm0,%xmm2
  DB  196,195,105,32,20,24,1              ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm2,%xmm2
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,105,32,211,2                ; vpinsrb       $0x2,%ebx,%xmm2,%xmm2
  DB  65,15,182,44,40                     ; movzbl        (%r8,%rbp,1),%ebp
  DB  196,227,105,32,213,3                ; vpinsrb       $0x3,%ebp,%xmm2,%xmm2
  DB  67,15,182,44,32                     ; movzbl        (%r8,%r12,1),%ebp
  DB  196,227,105,32,213,4                ; vpinsrb       $0x4,%ebp,%xmm2,%xmm2
  DB  67,15,182,44,56                     ; movzbl        (%r8,%r15,1),%ebp
  DB  196,227,105,32,213,5                ; vpinsrb       $0x5,%ebp,%xmm2,%xmm2
  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
  DB  196,227,105,32,213,6                ; vpinsrb       $0x6,%ebp,%xmm2,%xmm2
  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
  DB  196,227,105,32,213,7                ; vpinsrb       $0x7,%ebp,%xmm2,%xmm2
  DB  196,226,125,49,210                  ; vpmovzxbd     %xmm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  197,253,91,219                      ; vcvtps2dq     %ymm3,%ymm3
  DB  196,227,249,22,221,1                ; vpextrq       $0x1,%xmm3,%rbp
  DB  65,137,232                          ; mov           %ebp,%r8d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,219                 ; vmovq         %xmm3,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,57,219,1                ; vextracti128  $0x1,%ymm3,%xmm3
  DB  196,195,249,22,218,1                ; vpextrq       $0x1,%xmm3,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,222                 ; vmovq         %xmm3,%r14
  DB  69,137,247                          ; mov           %r14d,%r15d
  DB  73,193,238,32                       ; shr           $0x20,%r14
  DB  196,163,121,32,28,8,0               ; vpinsrb       $0x0,(%rax,%r9,1),%xmm0,%xmm3
  DB  196,227,97,32,28,24,1               ; vpinsrb       $0x1,(%rax,%rbx,1),%xmm3,%xmm3
  DB  66,15,182,28,0                      ; movzbl        (%rax,%r8,1),%ebx
  DB  196,227,97,32,219,2                 ; vpinsrb       $0x2,%ebx,%xmm3,%xmm3
  DB  15,182,44,40                        ; movzbl        (%rax,%rbp,1),%ebp
  DB  196,227,97,32,221,3                 ; vpinsrb       $0x3,%ebp,%xmm3,%xmm3
  DB  66,15,182,44,56                     ; movzbl        (%rax,%r15,1),%ebp
  DB  196,227,97,32,221,4                 ; vpinsrb       $0x4,%ebp,%xmm3,%xmm3
  DB  66,15,182,44,48                     ; movzbl        (%rax,%r14,1),%ebp
  DB  196,227,97,32,221,5                 ; vpinsrb       $0x5,%ebp,%xmm3,%xmm3
  DB  66,15,182,44,24                     ; movzbl        (%rax,%r11,1),%ebp
  DB  196,227,97,32,221,6                 ; vpinsrb       $0x6,%ebp,%xmm3,%xmm3
  DB  66,15,182,4,16                      ; movzbl        (%rax,%r10,1),%eax
  DB  196,227,97,32,216,7                 ; vpinsrb       $0x7,%eax,%xmm3,%xmm3
  DB  196,226,125,49,219                  ; vpmovzxbd     %xmm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  197,180,89,219                      ; vmulps        %ymm3,%ymm9,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_byte_tables_rgb_hsw
_sk_byte_tables_rgb_hsw LABEL PROC
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  68,139,64,24                        ; mov           0x18(%rax),%r8d
  DB  65,255,200                          ; dec           %r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  68,137,197                          ; mov           %r8d,%ebp
  DB  77,137,194                          ; mov           %r8,%r10
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,192                 ; vmovq         %xmm0,%r8
  DB  69,137,195                          ; mov           %r8d,%r11d
  DB  77,137,199                          ; mov           %r8,%r15
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  69,137,198                          ; mov           %r8d,%r14d
  DB  77,137,196                          ; mov           %r8,%r12
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,221                          ; mov           %ebx,%r13d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  196,131,121,32,4,25,0               ; vpinsrb       $0x0,(%r9,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,57,1               ; vpinsrb       $0x1,(%r9,%r15,1),%xmm0,%xmm0
  DB  65,15,182,44,41                     ; movzbl        (%r9,%rbp,1),%ebp
  DB  196,227,121,32,197,2                ; vpinsrb       $0x2,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,17                     ; movzbl        (%r9,%r10,1),%ebp
  DB  196,227,121,32,197,3                ; vpinsrb       $0x3,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,41                     ; movzbl        (%r9,%r13,1),%ebp
  DB  196,227,121,32,197,4                ; vpinsrb       $0x4,%ebp,%xmm0,%xmm0
  DB  65,15,182,44,25                     ; movzbl        (%r9,%rbx,1),%ebp
  DB  196,227,121,32,197,5                ; vpinsrb       $0x5,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,49                     ; movzbl        (%r9,%r14,1),%ebp
  DB  196,227,121,32,197,6                ; vpinsrb       $0x6,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,33                     ; movzbl        (%r9,%r12,1),%ebp
  DB  196,227,121,32,197,7                ; vpinsrb       $0x7,%ebp,%xmm0,%xmm0
  DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
  DB  197,124,91,208                      ; vcvtdq2ps     %ymm0,%ymm10
  DB  189,129,128,128,59                  ; mov           $0x3b808081,%ebp
  DB  197,249,110,197                     ; vmovd         %ebp,%xmm0
  DB  196,98,125,88,200                   ; vpbroadcastd  %xmm0,%ymm9
  DB  196,193,44,89,193                   ; vmulps        %ymm9,%ymm10,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
  DB  196,227,249,22,205,1                ; vpextrq       $0x1,%xmm1,%rbp
  DB  65,137,233                          ; mov           %ebp,%r9d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,203                 ; vmovq         %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,57,201,1                ; vextracti128  $0x1,%ymm1,%xmm1
  DB  196,195,249,22,203,1                ; vpextrq       $0x1,%xmm1,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  196,193,249,126,207                 ; vmovq         %xmm1,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,12,16,0              ; vpinsrb       $0x0,(%r8,%r10,1),%xmm0,%xmm1
  DB  196,195,113,32,12,24,1              ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm1,%xmm1
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,113,32,203,2                ; vpinsrb       $0x2,%ebx,%xmm1,%xmm1
  DB  65,15,182,44,40                     ; movzbl        (%r8,%rbp,1),%ebp
  DB  196,227,113,32,205,3                ; vpinsrb       $0x3,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,32                     ; movzbl        (%r8,%r12,1),%ebp
  DB  196,227,113,32,205,4                ; vpinsrb       $0x4,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,56                     ; movzbl        (%r8,%r15,1),%ebp
  DB  196,227,113,32,205,5                ; vpinsrb       $0x5,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
  DB  196,227,113,32,205,6                ; vpinsrb       $0x6,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
  DB  196,227,113,32,205,7                ; vpinsrb       $0x7,%ebp,%xmm1,%xmm1
  DB  196,226,125,49,201                  ; vpmovzxbd     %xmm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  72,139,64,16                        ; mov           0x10(%rax),%rax
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
  DB  196,227,249,22,213,1                ; vpextrq       $0x1,%xmm2,%rbp
  DB  65,137,232                          ; mov           %ebp,%r8d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,211                 ; vmovq         %xmm2,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,57,210,1                ; vextracti128  $0x1,%ymm2,%xmm2
  DB  196,195,249,22,210,1                ; vpextrq       $0x1,%xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,214                 ; vmovq         %xmm2,%r14
  DB  69,137,247                          ; mov           %r14d,%r15d
  DB  73,193,238,32                       ; shr           $0x20,%r14
  DB  196,163,121,32,20,8,0               ; vpinsrb       $0x0,(%rax,%r9,1),%xmm0,%xmm2
  DB  196,227,105,32,20,24,1              ; vpinsrb       $0x1,(%rax,%rbx,1),%xmm2,%xmm2
  DB  66,15,182,28,0                      ; movzbl        (%rax,%r8,1),%ebx
  DB  196,227,105,32,211,2                ; vpinsrb       $0x2,%ebx,%xmm2,%xmm2
  DB  15,182,44,40                        ; movzbl        (%rax,%rbp,1),%ebp
  DB  196,227,105,32,213,3                ; vpinsrb       $0x3,%ebp,%xmm2,%xmm2
  DB  66,15,182,44,56                     ; movzbl        (%rax,%r15,1),%ebp
  DB  196,227,105,32,213,4                ; vpinsrb       $0x4,%ebp,%xmm2,%xmm2
  DB  66,15,182,44,48                     ; movzbl        (%rax,%r14,1),%ebp
  DB  196,227,105,32,213,5                ; vpinsrb       $0x5,%ebp,%xmm2,%xmm2
  DB  66,15,182,44,24                     ; movzbl        (%rax,%r11,1),%ebp
  DB  196,227,105,32,213,6                ; vpinsrb       $0x6,%ebp,%xmm2,%xmm2
  DB  66,15,182,4,16                      ; movzbl        (%rax,%r10,1),%eax
  DB  196,227,105,32,208,7                ; vpinsrb       $0x7,%eax,%xmm2,%xmm2
  DB  196,226,125,49,210                  ; vpmovzxbd     %xmm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_r_hsw
_sk_table_r_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,125,91,192                      ; vcvtps2dq     %ymm0,%ymm8
  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
  DB  196,130,53,146,4,128                ; vgatherdps    %ymm9,(%r8,%ymm8,4),%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_g_hsw
_sk_table_g_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,125,91,193                      ; vcvtps2dq     %ymm1,%ymm8
  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
  DB  196,130,53,146,12,128               ; vgatherdps    %ymm9,(%r8,%ymm8,4),%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_b_hsw
_sk_table_b_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,125,91,194                      ; vcvtps2dq     %ymm2,%ymm8
  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
  DB  196,130,53,146,20,128               ; vgatherdps    %ymm9,(%r8,%ymm8,4),%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_a_hsw
_sk_table_a_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  197,125,91,195                      ; vcvtps2dq     %ymm3,%ymm8
  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
  DB  196,130,53,146,28,128               ; vgatherdps    %ymm9,(%r8,%ymm8,4),%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_r_hsw
_sk_parametric_r_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,124,194,192,2                ; vcmpleps      %ymm8,%ymm0,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  196,66,125,168,202                  ; vfmadd213ps   %ymm10,%ymm0,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  196,66,125,168,211                  ; vfmadd213ps   %ymm11,%ymm0,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,232                   ; vpbroadcastd  %xmm0,%ymm13
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  196,65,125,219,210                  ; vpand         %ymm10,%ymm0,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,45,235,208                      ; vpor          %ymm0,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,240                   ; vpbroadcastd  %xmm0,%ymm14
  DB  196,66,37,186,245                   ; vfmsub231ps   %ymm13,%ymm11,%ymm14
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,216                   ; vpbroadcastd  %xmm0,%ymm11
  DB  196,66,45,172,222                   ; vfnmadd213ps  %ymm14,%ymm10,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,232                   ; vpbroadcastd  %xmm0,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,172,88,192                      ; vaddps        %ymm0,%ymm10,%ymm0
  DB  197,148,94,192                      ; vdivps        %ymm0,%ymm13,%ymm0
  DB  197,164,92,192                      ; vsubps        %ymm0,%ymm11,%ymm0
  DB  197,28,89,216                       ; vmulps        %ymm0,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,224                   ; vpbroadcastd  %xmm0,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  196,65,124,88,219                   ; vaddps        %ymm11,%ymm0,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,232                   ; vpbroadcastd  %xmm0,%ymm13
  DB  196,66,45,172,235                   ; vfnmadd213ps  %ymm11,%ymm10,%ymm13
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,216                   ; vpbroadcastd  %xmm0,%ymm11
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  196,193,124,92,194                  ; vsubps        %ymm10,%ymm0,%ymm0
  DB  197,164,94,192                      ; vdivps        %ymm0,%ymm11,%ymm0
  DB  197,148,88,192                      ; vaddps        %ymm0,%ymm13,%ymm0
  DB  197,156,89,192                      ; vmulps        %ymm0,%ymm12,%ymm0
  DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
  DB  196,195,125,74,193,128              ; vblendvps     %ymm8,%ymm9,%ymm0,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,124,95,192                   ; vmaxps        %ymm8,%ymm0,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,188,93,192                      ; vminps        %ymm0,%ymm8,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_g_hsw
_sk_parametric_g_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,116,194,192,2                ; vcmpleps      %ymm8,%ymm1,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  196,66,117,168,202                  ; vfmadd213ps   %ymm10,%ymm1,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  196,66,117,168,211                  ; vfmadd213ps   %ymm11,%ymm1,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,233                   ; vpbroadcastd  %xmm1,%ymm13
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  196,65,117,219,210                  ; vpand         %ymm10,%ymm1,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,45,235,209                      ; vpor          %ymm1,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,241                   ; vpbroadcastd  %xmm1,%ymm14
  DB  196,66,37,186,245                   ; vfmsub231ps   %ymm13,%ymm11,%ymm14
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,217                   ; vpbroadcastd  %xmm1,%ymm11
  DB  196,66,45,172,222                   ; vfnmadd213ps  %ymm14,%ymm10,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,233                   ; vpbroadcastd  %xmm1,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,172,88,201                      ; vaddps        %ymm1,%ymm10,%ymm1
  DB  197,148,94,201                      ; vdivps        %ymm1,%ymm13,%ymm1
  DB  197,164,92,201                      ; vsubps        %ymm1,%ymm11,%ymm1
  DB  197,28,89,217                       ; vmulps        %ymm1,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,225                   ; vpbroadcastd  %xmm1,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  196,65,116,88,219                   ; vaddps        %ymm11,%ymm1,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,233                   ; vpbroadcastd  %xmm1,%ymm13
  DB  196,66,45,172,235                   ; vfnmadd213ps  %ymm11,%ymm10,%ymm13
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,217                   ; vpbroadcastd  %xmm1,%ymm11
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  196,193,116,92,202                  ; vsubps        %ymm10,%ymm1,%ymm1
  DB  197,164,94,201                      ; vdivps        %ymm1,%ymm11,%ymm1
  DB  197,148,88,201                      ; vaddps        %ymm1,%ymm13,%ymm1
  DB  197,156,89,201                      ; vmulps        %ymm1,%ymm12,%ymm1
  DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,116,88,202                  ; vaddps        %ymm10,%ymm1,%ymm1
  DB  196,195,117,74,201,128              ; vblendvps     %ymm8,%ymm9,%ymm1,%ymm1
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,116,95,192                   ; vmaxps        %ymm8,%ymm1,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,188,93,201                      ; vminps        %ymm1,%ymm8,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_b_hsw
_sk_parametric_b_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,108,194,192,2                ; vcmpleps      %ymm8,%ymm2,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  196,66,109,168,202                  ; vfmadd213ps   %ymm10,%ymm2,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  196,66,109,168,211                  ; vfmadd213ps   %ymm11,%ymm2,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,98,125,88,234                   ; vpbroadcastd  %xmm2,%ymm13
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  196,65,109,219,210                  ; vpand         %ymm10,%ymm2,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  197,45,235,210                      ; vpor          %ymm2,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,98,125,88,242                   ; vpbroadcastd  %xmm2,%ymm14
  DB  196,66,37,186,245                   ; vfmsub231ps   %ymm13,%ymm11,%ymm14
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,98,125,88,218                   ; vpbroadcastd  %xmm2,%ymm11
  DB  196,66,45,172,222                   ; vfnmadd213ps  %ymm14,%ymm10,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,98,125,88,234                   ; vpbroadcastd  %xmm2,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  197,172,88,210                      ; vaddps        %ymm2,%ymm10,%ymm2
  DB  197,148,94,210                      ; vdivps        %ymm2,%ymm13,%ymm2
  DB  197,164,92,210                      ; vsubps        %ymm2,%ymm11,%ymm2
  DB  197,28,89,218                       ; vmulps        %ymm2,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,98,125,88,226                   ; vpbroadcastd  %xmm2,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  196,65,108,88,219                   ; vaddps        %ymm11,%ymm2,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,98,125,88,234                   ; vpbroadcastd  %xmm2,%ymm13
  DB  196,66,45,172,235                   ; vfnmadd213ps  %ymm11,%ymm10,%ymm13
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,98,125,88,218                   ; vpbroadcastd  %xmm2,%ymm11
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  196,193,108,92,210                  ; vsubps        %ymm10,%ymm2,%ymm2
  DB  197,164,94,210                      ; vdivps        %ymm2,%ymm11,%ymm2
  DB  197,148,88,210                      ; vaddps        %ymm2,%ymm13,%ymm2
  DB  197,156,89,210                      ; vmulps        %ymm2,%ymm12,%ymm2
  DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
  DB  196,195,109,74,209,128              ; vblendvps     %ymm8,%ymm9,%ymm2,%ymm2
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,108,95,192                   ; vmaxps        %ymm8,%ymm2,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  197,188,93,210                      ; vminps        %ymm2,%ymm8,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_a_hsw
_sk_parametric_a_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,100,194,192,2                ; vcmpleps      %ymm8,%ymm3,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  196,66,101,168,202                  ; vfmadd213ps   %ymm10,%ymm3,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  196,66,101,168,211                  ; vfmadd213ps   %ymm11,%ymm3,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,235                   ; vpbroadcastd  %xmm3,%ymm13
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,65,101,219,210                  ; vpand         %ymm10,%ymm3,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,45,235,211                      ; vpor          %ymm3,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,243                   ; vpbroadcastd  %xmm3,%ymm14
  DB  196,66,37,186,245                   ; vfmsub231ps   %ymm13,%ymm11,%ymm14
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,219                   ; vpbroadcastd  %xmm3,%ymm11
  DB  196,66,45,172,222                   ; vfnmadd213ps  %ymm14,%ymm10,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,235                   ; vpbroadcastd  %xmm3,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,172,88,219                      ; vaddps        %ymm3,%ymm10,%ymm3
  DB  197,148,94,219                      ; vdivps        %ymm3,%ymm13,%ymm3
  DB  197,164,92,219                      ; vsubps        %ymm3,%ymm11,%ymm3
  DB  197,28,89,219                       ; vmulps        %ymm3,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,227                   ; vpbroadcastd  %xmm3,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,65,100,88,219                   ; vaddps        %ymm11,%ymm3,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,235                   ; vpbroadcastd  %xmm3,%ymm13
  DB  196,66,45,172,235                   ; vfnmadd213ps  %ymm11,%ymm10,%ymm13
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,98,125,88,219                   ; vpbroadcastd  %xmm3,%ymm11
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,193,100,92,218                  ; vsubps        %ymm10,%ymm3,%ymm3
  DB  197,164,94,219                      ; vdivps        %ymm3,%ymm11,%ymm3
  DB  197,148,88,219                      ; vaddps        %ymm3,%ymm13,%ymm3
  DB  197,156,89,219                      ; vmulps        %ymm3,%ymm12,%ymm3
  DB  197,253,91,219                      ; vcvtps2dq     %ymm3,%ymm3
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,100,88,218                  ; vaddps        %ymm10,%ymm3,%ymm3
  DB  196,195,101,74,217,128              ; vblendvps     %ymm8,%ymm9,%ymm3,%ymm3
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,100,95,192                   ; vmaxps        %ymm8,%ymm3,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,188,93,219                      ; vminps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_a8_hsw
_sk_load_a8_hsw LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,50                              ; jne           2408 <_sk_load_a8_hsw+0x42>
  DB  197,250,126,0                       ; vmovq         (%rax),%xmm0
  DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           2410 <_sk_load_a8_hsw+0x4a>
  DB  196,193,249,110,193                 ; vmovq         %r9,%xmm0
  DB  235,173                             ; jmp           23da <_sk_load_a8_hsw+0x14>

PUBLIC _sk_gather_a8_hsw
_sk_gather_a8_hsw LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  196,226,125,88,80,16                ; vpbroadcastd  0x10(%rax),%ymm2
  DB  196,226,109,64,201                  ; vpmulld       %ymm1,%ymm2,%ymm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
  DB  196,227,249,22,192,1                ; vpextrq       $0x1,%xmm0,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,194                 ; vmovq         %xmm0,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,199                 ; vmovq         %xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,4,24,0               ; vpinsrb       $0x0,(%r8,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,16,1               ; vpinsrb       $0x1,(%r8,%r10,1),%xmm0,%xmm0
  DB  71,15,182,12,8                      ; movzbl        (%r8,%r9,1),%r9d
  DB  196,195,121,32,193,2                ; vpinsrb       $0x2,%r9d,%xmm0,%xmm0
  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
  DB  196,227,121,32,192,3                ; vpinsrb       $0x3,%eax,%xmm0,%xmm0
  DB  67,15,182,4,32                      ; movzbl        (%r8,%r12,1),%eax
  DB  196,227,121,32,192,4                ; vpinsrb       $0x4,%eax,%xmm0,%xmm0
  DB  67,15,182,4,56                      ; movzbl        (%r8,%r15,1),%eax
  DB  196,227,121,32,192,5                ; vpinsrb       $0x5,%eax,%xmm0,%xmm0
  DB  67,15,182,4,48                      ; movzbl        (%r8,%r14,1),%eax
  DB  196,227,121,32,192,6                ; vpinsrb       $0x6,%eax,%xmm0,%xmm0
  DB  65,15,182,4,24                      ; movzbl        (%r8,%rbx,1),%eax
  DB  196,227,121,32,192,7                ; vpinsrb       $0x7,%eax,%xmm0,%xmm0
  DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,237,239,210                     ; vpxor         %ymm2,%ymm2,%ymm2
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_a8_hsw
_sk_store_a8_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  184,0,0,127,67                      ; mov           $0x437f0000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
  DB  196,66,57,43,193                    ; vpackusdw     %xmm9,%xmm8,%xmm8
  DB  196,65,57,103,192                   ; vpackuswb     %xmm8,%xmm8,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,10                              ; jne           2545 <_sk_store_a8_hsw+0x3b>
  DB  196,65,123,17,4,57                  ; vmovsd        %xmm8,(%r9,%rdi,1)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  119,236                             ; ja            2541 <_sk_store_a8_hsw+0x37>
  DB  196,66,121,48,192                   ; vpmovzxbw     %xmm8,%xmm8
  DB  65,15,182,192                       ; movzbl        %r8b,%eax
  DB  76,141,5,67,0,0,0                   ; lea           0x43(%rip),%r8        # 25a8 <_sk_store_a8_hsw+0x9e>
  DB  73,99,4,128                         ; movslq        (%r8,%rax,4),%rax
  DB  76,1,192                            ; add           %r8,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,67,121,20,68,57,6,12            ; vpextrb       $0xc,%xmm8,0x6(%r9,%rdi,1)
  DB  196,67,121,20,68,57,5,10            ; vpextrb       $0xa,%xmm8,0x5(%r9,%rdi,1)
  DB  196,67,121,20,68,57,4,8             ; vpextrb       $0x8,%xmm8,0x4(%r9,%rdi,1)
  DB  196,67,121,20,68,57,3,6             ; vpextrb       $0x6,%xmm8,0x3(%r9,%rdi,1)
  DB  196,67,121,20,68,57,2,4             ; vpextrb       $0x4,%xmm8,0x2(%r9,%rdi,1)
  DB  196,67,121,20,68,57,1,2             ; vpextrb       $0x2,%xmm8,0x1(%r9,%rdi,1)
  DB  196,67,121,20,4,57,0                ; vpextrb       $0x0,%xmm8,(%r9,%rdi,1)
  DB  235,154                             ; jmp           2541 <_sk_store_a8_hsw+0x37>
  DB  144                                 ; nop
  DB  246,255                             ; idiv          %bh
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  238                                 ; out           %al,(%dx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,230                             ; jmpq          *%rsi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  222,255                             ; fdivrp        %st,%st(7)
  DB  255                                 ; (bad)
  DB  255,214                             ; callq         *%rsi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,206                             ; dec           %esi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,198                             ; inc           %esi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_g8_hsw
_sk_load_g8_hsw LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,60                              ; jne           2610 <_sk_load_g8_hsw+0x4c>
  DB  197,250,126,0                       ; vmovq         (%rax),%xmm0
  DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,217                  ; vpbroadcastd  %xmm1,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
  DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           2618 <_sk_load_g8_hsw+0x54>
  DB  196,193,249,110,193                 ; vmovq         %r9,%xmm0
  DB  235,163                             ; jmp           25d8 <_sk_load_g8_hsw+0x14>

PUBLIC _sk_gather_g8_hsw
_sk_gather_g8_hsw LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  196,226,125,88,80,16                ; vpbroadcastd  0x10(%rax),%ymm2
  DB  196,226,109,64,201                  ; vpmulld       %ymm1,%ymm2,%ymm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
  DB  196,227,249,22,192,1                ; vpextrq       $0x1,%xmm0,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,194                 ; vmovq         %xmm0,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,199                 ; vmovq         %xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,4,24,0               ; vpinsrb       $0x0,(%r8,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,16,1               ; vpinsrb       $0x1,(%r8,%r10,1),%xmm0,%xmm0
  DB  71,15,182,12,8                      ; movzbl        (%r8,%r9,1),%r9d
  DB  196,195,121,32,193,2                ; vpinsrb       $0x2,%r9d,%xmm0,%xmm0
  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
  DB  196,227,121,32,192,3                ; vpinsrb       $0x3,%eax,%xmm0,%xmm0
  DB  67,15,182,4,32                      ; movzbl        (%r8,%r12,1),%eax
  DB  196,227,121,32,192,4                ; vpinsrb       $0x4,%eax,%xmm0,%xmm0
  DB  67,15,182,4,56                      ; movzbl        (%r8,%r15,1),%eax
  DB  196,227,121,32,192,5                ; vpinsrb       $0x5,%eax,%xmm0,%xmm0
  DB  67,15,182,4,48                      ; movzbl        (%r8,%r14,1),%eax
  DB  196,227,121,32,192,6                ; vpinsrb       $0x6,%eax,%xmm0,%xmm0
  DB  65,15,182,4,24                      ; movzbl        (%r8,%rbx,1),%eax
  DB  196,227,121,32,192,7                ; vpinsrb       $0x7,%eax,%xmm0,%xmm0
  DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,217                  ; vpbroadcastd  %xmm1,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
  DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_i8_hsw
_sk_gather_i8_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,192                          ; mov           %rax,%r8
  DB  77,133,192                          ; test          %r8,%r8
  DB  116,5                               ; je            272b <_sk_gather_i8_hsw+0xf>
  DB  76,137,192                          ; mov           %r8,%rax
  DB  235,2                               ; jmp           272d <_sk_gather_i8_hsw+0x11>
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  196,226,125,88,80,16                ; vpbroadcastd  0x10(%rax),%ymm2
  DB  196,226,109,64,201                  ; vpmulld       %ymm1,%ymm2,%ymm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
  DB  196,227,249,22,192,1                ; vpextrq       $0x1,%xmm0,%rax
  DB  65,137,194                          ; mov           %eax,%r10d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,195                 ; vmovq         %xmm0,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,223                          ; mov           %ebx,%r15d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,196                 ; vmovq         %xmm0,%r12
  DB  69,137,229                          ; mov           %r12d,%r13d
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  196,131,121,32,4,49,0               ; vpinsrb       $0x0,(%r9,%r14,1),%xmm0,%xmm0
  DB  196,131,121,32,4,25,1               ; vpinsrb       $0x1,(%r9,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,17,2               ; vpinsrb       $0x2,(%r9,%r10,1),%xmm0,%xmm0
  DB  196,195,121,32,4,1,3                ; vpinsrb       $0x3,(%r9,%rax,1),%xmm0,%xmm0
  DB  196,131,121,32,4,41,4               ; vpinsrb       $0x4,(%r9,%r13,1),%xmm0,%xmm0
  DB  196,131,121,32,4,33,5               ; vpinsrb       $0x5,(%r9,%r12,1),%xmm0,%xmm0
  DB  196,131,121,32,4,57,6               ; vpinsrb       $0x6,(%r9,%r15,1),%xmm0,%xmm0
  DB  196,195,121,32,4,25,7               ; vpinsrb       $0x7,(%r9,%rbx,1),%xmm0,%xmm0
  DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
  DB  73,139,64,8                         ; mov           0x8(%r8),%rax
  DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
  DB  196,226,117,144,28,128              ; vpgatherdd    %ymm1,(%rax,%ymm0,4),%ymm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,208                  ; vpbroadcastd  %xmm0,%ymm2
  DB  197,237,219,195                     ; vpand         %ymm3,%ymm2,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,98,125,88,193                   ; vpbroadcastd  %xmm1,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  197,245,114,211,8                   ; vpsrld        $0x8,%ymm3,%ymm1
  DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
  DB  197,181,114,211,16                  ; vpsrld        $0x10,%ymm3,%ymm9
  DB  196,193,109,219,209                 ; vpand         %ymm9,%ymm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
  DB  197,229,114,211,24                  ; vpsrld        $0x18,%ymm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_565_hsw
_sk_load_565_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,149,0,0,0                    ; jne           28df <_sk_load_565_hsw+0xa3>
  DB  196,193,122,111,4,122               ; vmovdqu       (%r10,%rdi,2),%xmm0
  DB  196,226,125,51,208                  ; vpmovzxwd     %xmm0,%ymm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,253,219,194                     ; vpand         %ymm2,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,245,219,202                     ; vpand         %ymm2,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  15,135,89,255,255,255               ; ja            2850 <_sk_load_565_hsw+0x14>
  DB  69,15,182,192                       ; movzbl        %r8b,%r8d
  DB  76,141,13,74,0,0,0                  ; lea           0x4a(%rip),%r9        # 294c <_sk_load_565_hsw+0x110>
  DB  75,99,4,129                         ; movslq        (%r9,%r8,4),%rax
  DB  76,1,200                            ; add           %r9,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  196,193,121,196,68,122,12,6         ; vpinsrw       $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,10,5         ; vpinsrw       $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,8,4          ; vpinsrw       $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,6,3          ; vpinsrw       $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,4,2          ; vpinsrw       $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,2,1          ; vpinsrw       $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,4,122,0             ; vpinsrw       $0x0,(%r10,%rdi,2),%xmm0,%xmm0
  DB  233,5,255,255,255                   ; jmpq          2850 <_sk_load_565_hsw+0x14>
  DB  144                                 ; nop
  DB  243,255                             ; repz          (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  235,255                             ; jmp           2951 <_sk_load_565_hsw+0x115>
  DB  255                                 ; (bad)
  DB  255,227                             ; jmpq          *%rbx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  219,255                             ; (bad)
  DB  255                                 ; (bad)
  DB  255,211                             ; callq         *%rbx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,203                             ; dec           %ebx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  191                                 ; .byte         0xbf
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_gather_565_hsw
_sk_gather_565_hsw LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  196,226,125,88,80,16                ; vpbroadcastd  0x10(%rax),%ymm2
  DB  196,226,109,64,201                  ; vpmulld       %ymm1,%ymm2,%ymm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
  DB  196,227,249,22,192,1                ; vpextrq       $0x1,%xmm0,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,194                 ; vmovq         %xmm0,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,199                 ; vmovq         %xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  71,15,183,20,80                     ; movzwl        (%r8,%r10,2),%r10d
  DB  71,15,183,28,88                     ; movzwl        (%r8,%r11,2),%r11d
  DB  196,193,121,110,195                 ; vmovd         %r11d,%xmm0
  DB  196,193,121,196,194,1               ; vpinsrw       $0x1,%r10d,%xmm0,%xmm0
  DB  71,15,183,12,72                     ; movzwl        (%r8,%r9,2),%r9d
  DB  196,193,121,196,193,2               ; vpinsrw       $0x2,%r9d,%xmm0,%xmm0
  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
  DB  197,249,196,192,3                   ; vpinsrw       $0x3,%eax,%xmm0,%xmm0
  DB  67,15,183,4,96                      ; movzwl        (%r8,%r12,2),%eax
  DB  197,249,196,192,4                   ; vpinsrw       $0x4,%eax,%xmm0,%xmm0
  DB  67,15,183,4,120                     ; movzwl        (%r8,%r15,2),%eax
  DB  197,249,196,192,5                   ; vpinsrw       $0x5,%eax,%xmm0,%xmm0
  DB  67,15,183,4,112                     ; movzwl        (%r8,%r14,2),%eax
  DB  197,249,196,192,6                   ; vpinsrw       $0x6,%eax,%xmm0,%xmm0
  DB  65,15,183,4,88                      ; movzwl        (%r8,%rbx,2),%eax
  DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
  DB  196,226,125,51,208                  ; vpmovzxwd     %xmm0,%ymm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,253,219,194                     ; vpand         %ymm2,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,245,219,202                     ; vpand         %ymm2,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_565_hsw
_sk_store_565_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  184,0,0,248,65                      ; mov           $0x41f80000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  196,193,53,114,241,11               ; vpslld        $0xb,%ymm9,%ymm9
  DB  184,0,0,124,66                      ; mov           $0x427c0000,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  197,44,89,209                       ; vmulps        %ymm1,%ymm10,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,45,114,242,5                ; vpslld        $0x5,%ymm10,%ymm10
  DB  196,65,45,235,201                   ; vpor          %ymm9,%ymm10,%ymm9
  DB  197,60,89,194                       ; vmulps        %ymm2,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,65,53,235,192                   ; vpor          %ymm8,%ymm9,%ymm8
  DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
  DB  196,66,57,43,193                    ; vpackusdw     %xmm9,%xmm8,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,10                              ; jne           2b17 <_sk_store_565_hsw+0x6c>
  DB  196,65,122,127,4,121                ; vmovdqu       %xmm8,(%r9,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  119,236                             ; ja            2b13 <_sk_store_565_hsw+0x68>
  DB  65,15,182,192                       ; movzbl        %r8b,%eax
  DB  76,141,5,66,0,0,0                   ; lea           0x42(%rip),%r8        # 2b74 <_sk_store_565_hsw+0xc9>
  DB  73,99,4,128                         ; movslq        (%r8,%rax,4),%rax
  DB  76,1,192                            ; add           %r8,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,67,121,21,68,121,12,6           ; vpextrw       $0x6,%xmm8,0xc(%r9,%rdi,2)
  DB  196,67,121,21,68,121,10,5           ; vpextrw       $0x5,%xmm8,0xa(%r9,%rdi,2)
  DB  196,67,121,21,68,121,8,4            ; vpextrw       $0x4,%xmm8,0x8(%r9,%rdi,2)
  DB  196,67,121,21,68,121,6,3            ; vpextrw       $0x3,%xmm8,0x6(%r9,%rdi,2)
  DB  196,67,121,21,68,121,4,2            ; vpextrw       $0x2,%xmm8,0x4(%r9,%rdi,2)
  DB  196,67,121,21,68,121,2,1            ; vpextrw       $0x1,%xmm8,0x2(%r9,%rdi,2)
  DB  196,67,121,21,4,121,0               ; vpextrw       $0x0,%xmm8,(%r9,%rdi,2)
  DB  235,159                             ; jmp           2b13 <_sk_store_565_hsw+0x68>
  DB  247,255                             ; idiv          %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  239                                 ; out           %eax,(%dx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,231                             ; jmpq          *%rdi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  223,255                             ; (bad)
  DB  255                                 ; (bad)
  DB  255,215                             ; callq         *%rdi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,207                             ; dec           %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,199                             ; inc           %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_4444_hsw
_sk_load_4444_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,179,0,0,0                    ; jne           2c51 <_sk_load_4444_hsw+0xc1>
  DB  196,193,122,111,4,122               ; vmovdqu       (%r10,%rdi,2),%xmm0
  DB  196,98,125,51,200                   ; vpmovzxwd     %xmm0,%ymm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  196,193,125,219,193                 ; vpand         %ymm9,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  196,193,117,219,201                 ; vpand         %ymm9,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  196,193,109,219,209                 ; vpand         %ymm9,%ymm2,%ymm2
  DB  197,124,91,194                      ; vcvtdq2ps     %ymm2,%ymm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,193,101,219,217                 ; vpand         %ymm9,%ymm3,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  15,135,59,255,255,255               ; ja            2ba4 <_sk_load_4444_hsw+0x14>
  DB  69,15,182,192                       ; movzbl        %r8b,%r8d
  DB  76,141,13,76,0,0,0                  ; lea           0x4c(%rip),%r9        # 2cc0 <_sk_load_4444_hsw+0x130>
  DB  75,99,4,129                         ; movslq        (%r9,%r8,4),%rax
  DB  76,1,200                            ; add           %r9,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  196,193,121,196,68,122,12,6         ; vpinsrw       $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,10,5         ; vpinsrw       $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,8,4          ; vpinsrw       $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,6,3          ; vpinsrw       $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,4,2          ; vpinsrw       $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,2,1          ; vpinsrw       $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,4,122,0             ; vpinsrw       $0x0,(%r10,%rdi,2),%xmm0,%xmm0
  DB  233,231,254,255,255                 ; jmpq          2ba4 <_sk_load_4444_hsw+0x14>
  DB  15,31,0                             ; nopl          (%rax)
  DB  241                                 ; icebp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  233,255,255,255,225                 ; jmpq          ffffffffe2002cc8 <_sk_callback_hsw+0xffffffffe1ffeace>
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  217,255                             ; fcos
  DB  255                                 ; (bad)
  DB  255,209                             ; callq         *%rcx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,201                             ; dec           %ecx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  189                                 ; .byte         0xbd
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_gather_4444_hsw
_sk_gather_4444_hsw LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  196,226,125,88,80,16                ; vpbroadcastd  0x10(%rax),%ymm2
  DB  196,226,109,64,201                  ; vpmulld       %ymm1,%ymm2,%ymm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
  DB  196,227,249,22,192,1                ; vpextrq       $0x1,%xmm0,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,194                 ; vmovq         %xmm0,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,199                 ; vmovq         %xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  71,15,183,20,80                     ; movzwl        (%r8,%r10,2),%r10d
  DB  71,15,183,28,88                     ; movzwl        (%r8,%r11,2),%r11d
  DB  196,193,121,110,195                 ; vmovd         %r11d,%xmm0
  DB  196,193,121,196,194,1               ; vpinsrw       $0x1,%r10d,%xmm0,%xmm0
  DB  71,15,183,12,72                     ; movzwl        (%r8,%r9,2),%r9d
  DB  196,193,121,196,193,2               ; vpinsrw       $0x2,%r9d,%xmm0,%xmm0
  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
  DB  197,249,196,192,3                   ; vpinsrw       $0x3,%eax,%xmm0,%xmm0
  DB  67,15,183,4,96                      ; movzwl        (%r8,%r12,2),%eax
  DB  197,249,196,192,4                   ; vpinsrw       $0x4,%eax,%xmm0,%xmm0
  DB  67,15,183,4,120                     ; movzwl        (%r8,%r15,2),%eax
  DB  197,249,196,192,5                   ; vpinsrw       $0x5,%eax,%xmm0,%xmm0
  DB  67,15,183,4,112                     ; movzwl        (%r8,%r14,2),%eax
  DB  197,249,196,192,6                   ; vpinsrw       $0x6,%eax,%xmm0,%xmm0
  DB  65,15,183,4,88                      ; movzwl        (%r8,%rbx,2),%eax
  DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
  DB  196,98,125,51,200                   ; vpmovzxwd     %xmm0,%ymm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  196,193,125,219,193                 ; vpand         %ymm9,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  196,193,117,219,201                 ; vpand         %ymm9,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  196,193,109,219,209                 ; vpand         %ymm9,%ymm2,%ymm2
  DB  197,124,91,194                      ; vcvtdq2ps     %ymm2,%ymm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  196,193,101,219,217                 ; vpand         %ymm9,%ymm3,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_4444_hsw
_sk_store_4444_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  184,0,0,112,65                      ; mov           $0x41700000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  196,193,53,114,241,12               ; vpslld        $0xc,%ymm9,%ymm9
  DB  197,60,89,209                       ; vmulps        %ymm1,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,45,114,242,8                ; vpslld        $0x8,%ymm10,%ymm10
  DB  196,65,45,235,201                   ; vpor          %ymm9,%ymm10,%ymm9
  DB  197,60,89,210                       ; vmulps        %ymm2,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,45,114,242,4                ; vpslld        $0x4,%ymm10,%ymm10
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,65,45,235,192                   ; vpor          %ymm8,%ymm10,%ymm8
  DB  196,65,53,235,192                   ; vpor          %ymm8,%ymm9,%ymm8
  DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
  DB  196,66,57,43,193                    ; vpackusdw     %xmm9,%xmm8,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,10                              ; jne           2eaf <_sk_store_4444_hsw+0x72>
  DB  196,65,122,127,4,121                ; vmovdqu       %xmm8,(%r9,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  119,236                             ; ja            2eab <_sk_store_4444_hsw+0x6e>
  DB  65,15,182,192                       ; movzbl        %r8b,%eax
  DB  76,141,5,66,0,0,0                   ; lea           0x42(%rip),%r8        # 2f0c <_sk_store_4444_hsw+0xcf>
  DB  73,99,4,128                         ; movslq        (%r8,%rax,4),%rax
  DB  76,1,192                            ; add           %r8,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,67,121,21,68,121,12,6           ; vpextrw       $0x6,%xmm8,0xc(%r9,%rdi,2)
  DB  196,67,121,21,68,121,10,5           ; vpextrw       $0x5,%xmm8,0xa(%r9,%rdi,2)
  DB  196,67,121,21,68,121,8,4            ; vpextrw       $0x4,%xmm8,0x8(%r9,%rdi,2)
  DB  196,67,121,21,68,121,6,3            ; vpextrw       $0x3,%xmm8,0x6(%r9,%rdi,2)
  DB  196,67,121,21,68,121,4,2            ; vpextrw       $0x2,%xmm8,0x4(%r9,%rdi,2)
  DB  196,67,121,21,68,121,2,1            ; vpextrw       $0x1,%xmm8,0x2(%r9,%rdi,2)
  DB  196,67,121,21,4,121,0               ; vpextrw       $0x0,%xmm8,(%r9,%rdi,2)
  DB  235,159                             ; jmp           2eab <_sk_store_4444_hsw+0x6e>
  DB  247,255                             ; idiv          %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  239                                 ; out           %eax,(%dx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,231                             ; jmpq          *%rdi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  223,255                             ; (bad)
  DB  255                                 ; (bad)
  DB  255,215                             ; callq         *%rdi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,207                             ; dec           %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,199                             ; inc           %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_8888_hsw
_sk_load_8888_hsw LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  76,3,8                              ; add           (%rax),%r9
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,104                             ; jne           2fa5 <_sk_load_8888_hsw+0x7d>
  DB  196,193,126,111,25                  ; vmovdqu       (%r9),%ymm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,208                  ; vpbroadcastd  %xmm0,%ymm2
  DB  197,237,219,195                     ; vpand         %ymm3,%ymm2,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,98,125,88,193                   ; vpbroadcastd  %xmm1,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  197,245,114,211,8                   ; vpsrld        $0x8,%ymm3,%ymm1
  DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
  DB  197,181,114,211,16                  ; vpsrld        $0x10,%ymm3,%ymm9
  DB  196,193,109,219,209                 ; vpand         %ymm9,%ymm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
  DB  197,229,114,211,24                  ; vpsrld        $0x18,%ymm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  185,8,0,0,0                         ; mov           $0x8,%ecx
  DB  68,41,193                           ; sub           %r8d,%ecx
  DB  192,225,3                           ; shl           $0x3,%cl
  DB  72,199,192,255,255,255,255          ; mov           $0xffffffffffffffff,%rax
  DB  72,211,232                          ; shr           %cl,%rax
  DB  196,225,249,110,192                 ; vmovq         %rax,%xmm0
  DB  196,226,125,33,192                  ; vpmovsxbd     %xmm0,%ymm0
  DB  196,194,125,140,25                  ; vpmaskmovd    (%r9),%ymm0,%ymm3
  DB  233,116,255,255,255                 ; jmpq          2f42 <_sk_load_8888_hsw+0x1a>

PUBLIC _sk_gather_8888_hsw
_sk_gather_8888_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  196,226,125,88,80,16                ; vpbroadcastd  0x10(%rax),%ymm2
  DB  196,226,109,64,201                  ; vpmulld       %ymm1,%ymm2,%ymm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
  DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
  DB  196,194,117,144,28,128              ; vpgatherdd    %ymm1,(%r8,%ymm0,4),%ymm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,208                  ; vpbroadcastd  %xmm0,%ymm2
  DB  197,237,219,195                     ; vpand         %ymm3,%ymm2,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,98,125,88,193                   ; vpbroadcastd  %xmm1,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  197,245,114,211,8                   ; vpsrld        $0x8,%ymm3,%ymm1
  DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
  DB  197,181,114,211,16                  ; vpsrld        $0x10,%ymm3,%ymm9
  DB  196,193,109,219,209                 ; vpand         %ymm9,%ymm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
  DB  197,229,114,211,24                  ; vpsrld        $0x18,%ymm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_8888_hsw
_sk_store_8888_hsw LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  76,3,8                              ; add           (%rax),%r9
  DB  184,0,0,127,67                      ; mov           $0x437f0000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  197,60,89,209                       ; vmulps        %ymm1,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,45,114,242,8                ; vpslld        $0x8,%ymm10,%ymm10
  DB  196,65,45,235,201                   ; vpor          %ymm9,%ymm10,%ymm9
  DB  197,60,89,210                       ; vmulps        %ymm2,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,45,114,242,16               ; vpslld        $0x10,%ymm10,%ymm10
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,193,61,114,240,24               ; vpslld        $0x18,%ymm8,%ymm8
  DB  196,65,45,235,192                   ; vpor          %ymm8,%ymm10,%ymm8
  DB  196,65,53,235,192                   ; vpor          %ymm8,%ymm9,%ymm8
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,12                              ; jne           30c8 <_sk_store_8888_hsw+0x74>
  DB  196,65,126,127,1                    ; vmovdqu       %ymm8,(%r9)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  185,8,0,0,0                         ; mov           $0x8,%ecx
  DB  68,41,193                           ; sub           %r8d,%ecx
  DB  192,225,3                           ; shl           $0x3,%cl
  DB  72,199,192,255,255,255,255          ; mov           $0xffffffffffffffff,%rax
  DB  72,211,232                          ; shr           %cl,%rax
  DB  196,97,249,110,200                  ; vmovq         %rax,%xmm9
  DB  196,66,125,33,201                   ; vpmovsxbd     %xmm9,%ymm9
  DB  196,66,53,142,1                     ; vpmaskmovd    %ymm8,%ymm9,(%r9)
  DB  235,211                             ; jmp           30c1 <_sk_store_8888_hsw+0x6d>

PUBLIC _sk_load_f16_hsw
_sk_load_f16_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,97                              ; jne           3159 <_sk_load_f16_hsw+0x6b>
  DB  197,121,16,4,248                    ; vmovupd       (%rax,%rdi,8),%xmm8
  DB  197,249,16,84,248,16                ; vmovupd       0x10(%rax,%rdi,8),%xmm2
  DB  197,249,16,92,248,32                ; vmovupd       0x20(%rax,%rdi,8),%xmm3
  DB  197,122,111,76,248,48               ; vmovdqu       0x30(%rax,%rdi,8),%xmm9
  DB  197,185,97,194                      ; vpunpcklwd    %xmm2,%xmm8,%xmm0
  DB  197,185,105,210                     ; vpunpckhwd    %xmm2,%xmm8,%xmm2
  DB  196,193,97,97,201                   ; vpunpcklwd    %xmm9,%xmm3,%xmm1
  DB  196,193,97,105,217                  ; vpunpckhwd    %xmm9,%xmm3,%xmm3
  DB  197,121,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm8
  DB  197,121,105,202                     ; vpunpckhwd    %xmm2,%xmm0,%xmm9
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,241,105,219                     ; vpunpckhwd    %xmm3,%xmm1,%xmm3
  DB  197,185,108,194                     ; vpunpcklqdq   %xmm2,%xmm8,%xmm0
  DB  196,226,125,19,192                  ; vcvtph2ps     %xmm0,%ymm0
  DB  197,185,109,202                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm1
  DB  196,226,125,19,201                  ; vcvtph2ps     %xmm1,%ymm1
  DB  197,177,108,211                     ; vpunpcklqdq   %xmm3,%xmm9,%xmm2
  DB  196,226,125,19,210                  ; vcvtph2ps     %xmm2,%ymm2
  DB  197,177,109,219                     ; vpunpckhqdq   %xmm3,%xmm9,%xmm3
  DB  196,226,125,19,219                  ; vcvtph2ps     %xmm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,123,16,4,248                    ; vmovsd        (%rax,%rdi,8),%xmm8
  DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,79                              ; je            31b8 <_sk_load_f16_hsw+0xca>
  DB  197,57,22,68,248,8                  ; vmovhpd       0x8(%rax,%rdi,8),%xmm8,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,67                              ; jb            31b8 <_sk_load_f16_hsw+0xca>
  DB  197,251,16,84,248,16                ; vmovsd        0x10(%rax,%rdi,8),%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  116,68                              ; je            31c5 <_sk_load_f16_hsw+0xd7>
  DB  197,233,22,84,248,24                ; vmovhpd       0x18(%rax,%rdi,8),%xmm2,%xmm2
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,56                              ; jb            31c5 <_sk_load_f16_hsw+0xd7>
  DB  197,251,16,92,248,32                ; vmovsd        0x20(%rax,%rdi,8),%xmm3
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  15,132,114,255,255,255              ; je            310f <_sk_load_f16_hsw+0x21>
  DB  197,225,22,92,248,40                ; vmovhpd       0x28(%rax,%rdi,8),%xmm3,%xmm3
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  15,130,98,255,255,255               ; jb            310f <_sk_load_f16_hsw+0x21>
  DB  197,122,126,76,248,48               ; vmovq         0x30(%rax,%rdi,8),%xmm9
  DB  233,87,255,255,255                  ; jmpq          310f <_sk_load_f16_hsw+0x21>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  197,233,87,210                      ; vxorpd        %xmm2,%xmm2,%xmm2
  DB  233,74,255,255,255                  ; jmpq          310f <_sk_load_f16_hsw+0x21>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  233,65,255,255,255                  ; jmpq          310f <_sk_load_f16_hsw+0x21>

PUBLIC _sk_gather_f16_hsw
_sk_gather_f16_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  196,226,125,88,80,16                ; vpbroadcastd  0x10(%rax),%ymm2
  DB  196,226,109,64,201                  ; vpmulld       %ymm1,%ymm2,%ymm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
  DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
  DB  197,237,118,210                     ; vpcmpeqd      %ymm2,%ymm2,%ymm2
  DB  196,194,237,144,28,192              ; vpgatherdq    %ymm2,(%r8,%xmm0,8),%ymm3
  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
  DB  196,194,245,144,20,192              ; vpgatherdq    %ymm1,(%r8,%xmm0,8),%ymm2
  DB  196,227,125,57,216,1                ; vextracti128  $0x1,%ymm3,%xmm0
  DB  196,227,125,57,209,1                ; vextracti128  $0x1,%ymm2,%xmm1
  DB  197,97,97,192                       ; vpunpcklwd    %xmm0,%xmm3,%xmm8
  DB  197,225,105,192                     ; vpunpckhwd    %xmm0,%xmm3,%xmm0
  DB  197,233,97,217                      ; vpunpcklwd    %xmm1,%xmm2,%xmm3
  DB  197,233,105,201                     ; vpunpckhwd    %xmm1,%xmm2,%xmm1
  DB  197,57,97,200                       ; vpunpcklwd    %xmm0,%xmm8,%xmm9
  DB  197,57,105,192                      ; vpunpckhwd    %xmm0,%xmm8,%xmm8
  DB  197,225,97,209                      ; vpunpcklwd    %xmm1,%xmm3,%xmm2
  DB  197,225,105,217                     ; vpunpckhwd    %xmm1,%xmm3,%xmm3
  DB  197,177,108,194                     ; vpunpcklqdq   %xmm2,%xmm9,%xmm0
  DB  196,226,125,19,192                  ; vcvtph2ps     %xmm0,%ymm0
  DB  197,177,109,202                     ; vpunpckhqdq   %xmm2,%xmm9,%xmm1
  DB  196,226,125,19,201                  ; vcvtph2ps     %xmm1,%ymm1
  DB  197,185,108,211                     ; vpunpcklqdq   %xmm3,%xmm8,%xmm2
  DB  196,226,125,19,210                  ; vcvtph2ps     %xmm2,%ymm2
  DB  197,185,109,219                     ; vpunpckhqdq   %xmm3,%xmm8,%xmm3
  DB  196,226,125,19,219                  ; vcvtph2ps     %xmm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_f16_hsw
_sk_store_f16_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  196,195,125,29,192,4                ; vcvtps2ph     $0x4,%ymm0,%xmm8
  DB  196,195,125,29,201,4                ; vcvtps2ph     $0x4,%ymm1,%xmm9
  DB  196,195,125,29,210,4                ; vcvtps2ph     $0x4,%ymm2,%xmm10
  DB  196,195,125,29,219,4                ; vcvtps2ph     $0x4,%ymm3,%xmm11
  DB  196,65,57,97,225                    ; vpunpcklwd    %xmm9,%xmm8,%xmm12
  DB  196,65,57,105,193                   ; vpunpckhwd    %xmm9,%xmm8,%xmm8
  DB  196,65,41,97,203                    ; vpunpcklwd    %xmm11,%xmm10,%xmm9
  DB  196,65,41,105,235                   ; vpunpckhwd    %xmm11,%xmm10,%xmm13
  DB  196,65,25,98,217                    ; vpunpckldq    %xmm9,%xmm12,%xmm11
  DB  196,65,25,106,209                   ; vpunpckhdq    %xmm9,%xmm12,%xmm10
  DB  196,65,57,98,205                    ; vpunpckldq    %xmm13,%xmm8,%xmm9
  DB  196,65,57,106,197                   ; vpunpckhdq    %xmm13,%xmm8,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,27                              ; jne           32bd <_sk_store_f16_hsw+0x65>
  DB  197,120,17,28,248                   ; vmovups       %xmm11,(%rax,%rdi,8)
  DB  197,120,17,84,248,16                ; vmovups       %xmm10,0x10(%rax,%rdi,8)
  DB  197,120,17,76,248,32                ; vmovups       %xmm9,0x20(%rax,%rdi,8)
  DB  197,122,127,68,248,48               ; vmovdqu       %xmm8,0x30(%rax,%rdi,8)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,121,214,28,248                  ; vmovq         %xmm11,(%rax,%rdi,8)
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,241                             ; je            32b9 <_sk_store_f16_hsw+0x61>
  DB  197,121,23,92,248,8                 ; vmovhpd       %xmm11,0x8(%rax,%rdi,8)
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,229                             ; jb            32b9 <_sk_store_f16_hsw+0x61>
  DB  197,121,214,84,248,16               ; vmovq         %xmm10,0x10(%rax,%rdi,8)
  DB  116,221                             ; je            32b9 <_sk_store_f16_hsw+0x61>
  DB  197,121,23,84,248,24                ; vmovhpd       %xmm10,0x18(%rax,%rdi,8)
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,209                             ; jb            32b9 <_sk_store_f16_hsw+0x61>
  DB  197,121,214,76,248,32               ; vmovq         %xmm9,0x20(%rax,%rdi,8)
  DB  116,201                             ; je            32b9 <_sk_store_f16_hsw+0x61>
  DB  197,121,23,76,248,40                ; vmovhpd       %xmm9,0x28(%rax,%rdi,8)
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,189                             ; jb            32b9 <_sk_store_f16_hsw+0x61>
  DB  197,121,214,68,248,48               ; vmovq         %xmm8,0x30(%rax,%rdi,8)
  DB  235,181                             ; jmp           32b9 <_sk_store_f16_hsw+0x61>

PUBLIC _sk_load_u16_be_hsw
_sk_load_u16_be_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  72,141,4,189,0,0,0,0                ; lea           0x0(,%rdi,4),%rax
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,205,0,0,0                    ; jne           33e7 <_sk_load_u16_be_hsw+0xe3>
  DB  196,65,121,16,4,64                  ; vmovupd       (%r8,%rax,2),%xmm8
  DB  196,193,121,16,84,64,16             ; vmovupd       0x10(%r8,%rax,2),%xmm2
  DB  196,193,121,16,92,64,32             ; vmovupd       0x20(%r8,%rax,2),%xmm3
  DB  196,65,122,111,76,64,48             ; vmovdqu       0x30(%r8,%rax,2),%xmm9
  DB  197,185,97,194                      ; vpunpcklwd    %xmm2,%xmm8,%xmm0
  DB  197,185,105,210                     ; vpunpckhwd    %xmm2,%xmm8,%xmm2
  DB  196,193,97,97,201                   ; vpunpcklwd    %xmm9,%xmm3,%xmm1
  DB  196,193,97,105,217                  ; vpunpckhwd    %xmm9,%xmm3,%xmm3
  DB  197,121,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm8
  DB  197,121,105,202                     ; vpunpckhwd    %xmm2,%xmm0,%xmm9
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,113,105,219                     ; vpunpckhwd    %xmm3,%xmm1,%xmm11
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,98,125,88,208                   ; vpbroadcastd  %xmm0,%ymm10
  DB  197,185,108,194                     ; vpunpcklqdq   %xmm2,%xmm8,%xmm0
  DB  197,241,113,240,8                   ; vpsllw        $0x8,%xmm0,%xmm1
  DB  197,249,113,208,8                   ; vpsrlw        $0x8,%xmm0,%xmm0
  DB  197,241,235,192                     ; vpor          %xmm0,%xmm1,%xmm0
  DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  197,185,109,202                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm1
  DB  197,233,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm2
  DB  197,241,113,209,8                   ; vpsrlw        $0x8,%xmm1,%xmm1
  DB  197,233,235,201                     ; vpor          %xmm1,%xmm2,%xmm1
  DB  196,226,125,51,201                  ; vpmovzxwd     %xmm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  196,193,49,108,211                  ; vpunpcklqdq   %xmm11,%xmm9,%xmm2
  DB  197,225,113,242,8                   ; vpsllw        $0x8,%xmm2,%xmm3
  DB  197,233,113,210,8                   ; vpsrlw        $0x8,%xmm2,%xmm2
  DB  197,225,235,210                     ; vpor          %xmm2,%xmm3,%xmm2
  DB  196,226,125,51,210                  ; vpmovzxwd     %xmm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
  DB  196,193,49,109,219                  ; vpunpckhqdq   %xmm11,%xmm9,%xmm3
  DB  197,185,113,243,8                   ; vpsllw        $0x8,%xmm3,%xmm8
  DB  197,225,113,211,8                   ; vpsrlw        $0x8,%xmm3,%xmm3
  DB  197,185,235,219                     ; vpor          %xmm3,%xmm8,%xmm3
  DB  196,226,125,51,219                  ; vpmovzxwd     %xmm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  197,172,89,219                      ; vmulps        %ymm3,%ymm10,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,65,123,16,4,64                  ; vmovsd        (%r8,%rax,2),%xmm8
  DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,85                              ; je            344d <_sk_load_u16_be_hsw+0x149>
  DB  196,65,57,22,68,64,8                ; vmovhpd       0x8(%r8,%rax,2),%xmm8,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,72                              ; jb            344d <_sk_load_u16_be_hsw+0x149>
  DB  196,193,123,16,84,64,16             ; vmovsd        0x10(%r8,%rax,2),%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  116,72                              ; je            345a <_sk_load_u16_be_hsw+0x156>
  DB  196,193,105,22,84,64,24             ; vmovhpd       0x18(%r8,%rax,2),%xmm2,%xmm2
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,59                              ; jb            345a <_sk_load_u16_be_hsw+0x156>
  DB  196,193,123,16,92,64,32             ; vmovsd        0x20(%r8,%rax,2),%xmm3
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  15,132,5,255,255,255                ; je            3335 <_sk_load_u16_be_hsw+0x31>
  DB  196,193,97,22,92,64,40              ; vmovhpd       0x28(%r8,%rax,2),%xmm3,%xmm3
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  15,130,244,254,255,255              ; jb            3335 <_sk_load_u16_be_hsw+0x31>
  DB  196,65,122,126,76,64,48             ; vmovq         0x30(%r8,%rax,2),%xmm9
  DB  233,232,254,255,255                 ; jmpq          3335 <_sk_load_u16_be_hsw+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  197,233,87,210                      ; vxorpd        %xmm2,%xmm2,%xmm2
  DB  233,219,254,255,255                 ; jmpq          3335 <_sk_load_u16_be_hsw+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  233,210,254,255,255                 ; jmpq          3335 <_sk_load_u16_be_hsw+0x31>

PUBLIC _sk_load_rgb_u16_be_hsw
_sk_load_rgb_u16_be_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  72,141,4,127                        ; lea           (%rdi,%rdi,2),%rax
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,211,0,0,0                    ; jne           3548 <_sk_load_rgb_u16_be_hsw+0xe5>
  DB  196,193,122,111,4,64                ; vmovdqu       (%r8,%rax,2),%xmm0
  DB  196,193,122,111,84,64,12            ; vmovdqu       0xc(%r8,%rax,2),%xmm2
  DB  196,193,122,111,76,64,24            ; vmovdqu       0x18(%r8,%rax,2),%xmm1
  DB  196,193,122,111,92,64,32            ; vmovdqu       0x20(%r8,%rax,2),%xmm3
  DB  197,225,115,219,4                   ; vpsrldq       $0x4,%xmm3,%xmm3
  DB  197,185,115,216,6                   ; vpsrldq       $0x6,%xmm0,%xmm8
  DB  197,177,115,218,6                   ; vpsrldq       $0x6,%xmm2,%xmm9
  DB  197,161,115,217,6                   ; vpsrldq       $0x6,%xmm1,%xmm11
  DB  197,169,115,219,6                   ; vpsrldq       $0x6,%xmm3,%xmm10
  DB  197,249,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm0
  DB  196,193,57,97,209                   ; vpunpcklwd    %xmm9,%xmm8,%xmm2
  DB  197,241,97,203                      ; vpunpcklwd    %xmm3,%xmm1,%xmm1
  DB  196,193,33,97,218                   ; vpunpcklwd    %xmm10,%xmm11,%xmm3
  DB  197,121,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm8
  DB  197,121,105,202                     ; vpunpckhwd    %xmm2,%xmm0,%xmm9
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,241,105,219                     ; vpunpckhwd    %xmm3,%xmm1,%xmm3
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,98,125,88,208                   ; vpbroadcastd  %xmm0,%ymm10
  DB  197,185,108,194                     ; vpunpcklqdq   %xmm2,%xmm8,%xmm0
  DB  197,241,113,240,8                   ; vpsllw        $0x8,%xmm0,%xmm1
  DB  197,249,113,208,8                   ; vpsrlw        $0x8,%xmm0,%xmm0
  DB  197,241,235,192                     ; vpor          %xmm0,%xmm1,%xmm0
  DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  197,185,109,202                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm1
  DB  197,233,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm2
  DB  197,241,113,209,8                   ; vpsrlw        $0x8,%xmm1,%xmm1
  DB  197,233,235,201                     ; vpor          %xmm1,%xmm2,%xmm1
  DB  196,226,125,51,201                  ; vpmovzxwd     %xmm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  197,177,108,211                     ; vpunpcklqdq   %xmm3,%xmm9,%xmm2
  DB  197,225,113,242,8                   ; vpsllw        $0x8,%xmm2,%xmm3
  DB  197,233,113,210,8                   ; vpsrlw        $0x8,%xmm2,%xmm2
  DB  197,225,235,210                     ; vpor          %xmm2,%xmm3,%xmm2
  DB  196,226,125,51,210                  ; vpmovzxwd     %xmm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,193,121,110,4,64                ; vmovd         (%r8,%rax,2),%xmm0
  DB  196,193,121,196,68,64,4,2           ; vpinsrw       $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  117,5                               ; jne           3561 <_sk_load_rgb_u16_be_hsw+0xfe>
  DB  233,72,255,255,255                  ; jmpq          34a9 <_sk_load_rgb_u16_be_hsw+0x46>
  DB  196,193,121,110,76,64,6             ; vmovd         0x6(%r8,%rax,2),%xmm1
  DB  196,65,113,196,68,64,10,2           ; vpinsrw       $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,26                              ; jb            3590 <_sk_load_rgb_u16_be_hsw+0x12d>
  DB  196,193,121,110,76,64,12            ; vmovd         0xc(%r8,%rax,2),%xmm1
  DB  196,193,113,196,84,64,16,2          ; vpinsrw       $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  117,10                              ; jne           3595 <_sk_load_rgb_u16_be_hsw+0x132>
  DB  233,25,255,255,255                  ; jmpq          34a9 <_sk_load_rgb_u16_be_hsw+0x46>
  DB  233,20,255,255,255                  ; jmpq          34a9 <_sk_load_rgb_u16_be_hsw+0x46>
  DB  196,193,121,110,76,64,18            ; vmovd         0x12(%r8,%rax,2),%xmm1
  DB  196,65,113,196,76,64,22,2           ; vpinsrw       $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,26                              ; jb            35c4 <_sk_load_rgb_u16_be_hsw+0x161>
  DB  196,193,121,110,76,64,24            ; vmovd         0x18(%r8,%rax,2),%xmm1
  DB  196,193,113,196,76,64,28,2          ; vpinsrw       $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  117,10                              ; jne           35c9 <_sk_load_rgb_u16_be_hsw+0x166>
  DB  233,229,254,255,255                 ; jmpq          34a9 <_sk_load_rgb_u16_be_hsw+0x46>
  DB  233,224,254,255,255                 ; jmpq          34a9 <_sk_load_rgb_u16_be_hsw+0x46>
  DB  196,193,121,110,92,64,30            ; vmovd         0x1e(%r8,%rax,2),%xmm3
  DB  196,65,97,196,92,64,34,2            ; vpinsrw       $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,20                              ; jb            35f2 <_sk_load_rgb_u16_be_hsw+0x18f>
  DB  196,193,121,110,92,64,36            ; vmovd         0x24(%r8,%rax,2),%xmm3
  DB  196,193,97,196,92,64,40,2           ; vpinsrw       $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
  DB  233,183,254,255,255                 ; jmpq          34a9 <_sk_load_rgb_u16_be_hsw+0x46>
  DB  233,178,254,255,255                 ; jmpq          34a9 <_sk_load_rgb_u16_be_hsw+0x46>

PUBLIC _sk_store_u16_be_hsw
_sk_store_u16_be_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  184,0,255,127,71                    ; mov           $0x477fff00,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  196,67,125,25,202,1                 ; vextractf128  $0x1,%ymm9,%xmm10
  DB  196,66,49,43,202                    ; vpackusdw     %xmm10,%xmm9,%xmm9
  DB  196,193,41,113,241,8                ; vpsllw        $0x8,%xmm9,%xmm10
  DB  196,193,49,113,209,8                ; vpsrlw        $0x8,%xmm9,%xmm9
  DB  196,65,41,235,201                   ; vpor          %xmm9,%xmm10,%xmm9
  DB  197,60,89,209                       ; vmulps        %ymm1,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,67,125,25,211,1                 ; vextractf128  $0x1,%ymm10,%xmm11
  DB  196,66,41,43,211                    ; vpackusdw     %xmm11,%xmm10,%xmm10
  DB  196,193,33,113,242,8                ; vpsllw        $0x8,%xmm10,%xmm11
  DB  196,193,41,113,210,8                ; vpsrlw        $0x8,%xmm10,%xmm10
  DB  196,65,33,235,210                   ; vpor          %xmm10,%xmm11,%xmm10
  DB  197,60,89,218                       ; vmulps        %ymm2,%ymm8,%ymm11
  DB  196,65,125,91,219                   ; vcvtps2dq     %ymm11,%ymm11
  DB  196,67,125,25,220,1                 ; vextractf128  $0x1,%ymm11,%xmm12
  DB  196,66,33,43,220                    ; vpackusdw     %xmm12,%xmm11,%xmm11
  DB  196,193,25,113,243,8                ; vpsllw        $0x8,%xmm11,%xmm12
  DB  196,193,33,113,211,8                ; vpsrlw        $0x8,%xmm11,%xmm11
  DB  196,65,25,235,219                   ; vpor          %xmm11,%xmm12,%xmm11
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,67,125,25,196,1                 ; vextractf128  $0x1,%ymm8,%xmm12
  DB  196,66,57,43,196                    ; vpackusdw     %xmm12,%xmm8,%xmm8
  DB  196,193,25,113,240,8                ; vpsllw        $0x8,%xmm8,%xmm12
  DB  196,193,57,113,208,8                ; vpsrlw        $0x8,%xmm8,%xmm8
  DB  196,65,25,235,192                   ; vpor          %xmm8,%xmm12,%xmm8
  DB  196,65,49,97,226                    ; vpunpcklwd    %xmm10,%xmm9,%xmm12
  DB  196,65,49,105,234                   ; vpunpckhwd    %xmm10,%xmm9,%xmm13
  DB  196,65,33,97,200                    ; vpunpcklwd    %xmm8,%xmm11,%xmm9
  DB  196,65,33,105,192                   ; vpunpckhwd    %xmm8,%xmm11,%xmm8
  DB  196,65,25,98,217                    ; vpunpckldq    %xmm9,%xmm12,%xmm11
  DB  196,65,25,106,209                   ; vpunpckhdq    %xmm9,%xmm12,%xmm10
  DB  196,65,17,98,200                    ; vpunpckldq    %xmm8,%xmm13,%xmm9
  DB  196,65,17,106,192                   ; vpunpckhdq    %xmm8,%xmm13,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,31                              ; jne           36f2 <_sk_store_u16_be_hsw+0xfb>
  DB  196,1,120,17,28,72                  ; vmovups       %xmm11,(%r8,%r9,2)
  DB  196,1,120,17,84,72,16               ; vmovups       %xmm10,0x10(%r8,%r9,2)
  DB  196,1,120,17,76,72,32               ; vmovups       %xmm9,0x20(%r8,%r9,2)
  DB  196,1,122,127,68,72,48              ; vmovdqu       %xmm8,0x30(%r8,%r9,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,1,121,214,28,72                 ; vmovq         %xmm11,(%r8,%r9,2)
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,240                             ; je            36ee <_sk_store_u16_be_hsw+0xf7>
  DB  196,1,121,23,92,72,8                ; vmovhpd       %xmm11,0x8(%r8,%r9,2)
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,227                             ; jb            36ee <_sk_store_u16_be_hsw+0xf7>
  DB  196,1,121,214,84,72,16              ; vmovq         %xmm10,0x10(%r8,%r9,2)
  DB  116,218                             ; je            36ee <_sk_store_u16_be_hsw+0xf7>
  DB  196,1,121,23,84,72,24               ; vmovhpd       %xmm10,0x18(%r8,%r9,2)
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,205                             ; jb            36ee <_sk_store_u16_be_hsw+0xf7>
  DB  196,1,121,214,76,72,32              ; vmovq         %xmm9,0x20(%r8,%r9,2)
  DB  116,196                             ; je            36ee <_sk_store_u16_be_hsw+0xf7>
  DB  196,1,121,23,76,72,40               ; vmovhpd       %xmm9,0x28(%r8,%r9,2)
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,183                             ; jb            36ee <_sk_store_u16_be_hsw+0xf7>
  DB  196,1,121,214,68,72,48              ; vmovq         %xmm8,0x30(%r8,%r9,2)
  DB  235,174                             ; jmp           36ee <_sk_store_u16_be_hsw+0xf7>

PUBLIC _sk_load_f32_hsw
_sk_load_f32_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  119,110                             ; ja            37b6 <_sk_load_f32_hsw+0x76>
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  76,141,21,134,0,0,0                 ; lea           0x86(%rip),%r10        # 37e0 <_sk_load_f32_hsw+0xa0>
  DB  73,99,4,138                         ; movslq        (%r10,%rcx,4),%rax
  DB  76,1,208                            ; add           %r10,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,3,125,24,68,136,112,1           ; vinsertf128   $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8
  DB  196,131,125,24,92,136,96,1          ; vinsertf128   $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3
  DB  196,131,125,24,76,136,80,1          ; vinsertf128   $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1
  DB  196,131,125,24,84,136,64,1          ; vinsertf128   $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2
  DB  196,129,121,16,68,136,48            ; vmovupd       0x30(%r8,%r9,4),%xmm0
  DB  196,195,125,13,192,12               ; vblendpd      $0xc,%ymm8,%ymm0,%ymm0
  DB  196,1,121,16,68,136,32              ; vmovupd       0x20(%r8,%r9,4),%xmm8
  DB  196,99,61,13,203,12                 ; vblendpd      $0xc,%ymm3,%ymm8,%ymm9
  DB  196,129,121,16,92,136,16            ; vmovupd       0x10(%r8,%r9,4),%xmm3
  DB  196,99,101,13,209,12                ; vblendpd      $0xc,%ymm1,%ymm3,%ymm10
  DB  196,129,121,16,12,136               ; vmovupd       (%r8,%r9,4),%xmm1
  DB  196,227,117,13,202,12               ; vblendpd      $0xc,%ymm2,%ymm1,%ymm1
  DB  196,193,116,20,210                  ; vunpcklps     %ymm10,%ymm1,%ymm2
  DB  196,193,116,21,218                  ; vunpckhps     %ymm10,%ymm1,%ymm3
  DB  197,180,20,200                      ; vunpcklps     %ymm0,%ymm9,%ymm1
  DB  197,52,21,192                       ; vunpckhps     %ymm0,%ymm9,%ymm8
  DB  197,237,20,193                      ; vunpcklpd     %ymm1,%ymm2,%ymm0
  DB  197,237,21,201                      ; vunpckhpd     %ymm1,%ymm2,%ymm1
  DB  196,193,101,20,208                  ; vunpcklpd     %ymm8,%ymm3,%ymm2
  DB  196,193,101,21,216                  ; vunpckhpd     %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  102,144                             ; xchg          %ax,%ax
  DB  131,255,255                         ; cmp           $0xffffffff,%edi
  DB  255,202                             ; dec           %edx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  189,255,255,255,176                 ; mov           $0xb0ffffff,%ebp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,163,255,255,255,155             ; jmpq          *-0x64000001(%rbx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,147,255,255,255,139             ; callq         *-0x74000001(%rbx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_store_f32_hsw
_sk_store_f32_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  72,141,4,189,0,0,0,0                ; lea           0x0(,%rdi,4),%rax
  DB  197,124,20,193                      ; vunpcklps     %ymm1,%ymm0,%ymm8
  DB  197,124,21,217                      ; vunpckhps     %ymm1,%ymm0,%ymm11
  DB  197,108,20,203                      ; vunpcklps     %ymm3,%ymm2,%ymm9
  DB  197,108,21,227                      ; vunpckhps     %ymm3,%ymm2,%ymm12
  DB  196,65,61,20,209                    ; vunpcklpd     %ymm9,%ymm8,%ymm10
  DB  196,65,61,21,201                    ; vunpckhpd     %ymm9,%ymm8,%ymm9
  DB  196,65,37,20,196                    ; vunpcklpd     %ymm12,%ymm11,%ymm8
  DB  196,65,37,21,220                    ; vunpckhpd     %ymm12,%ymm11,%ymm11
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,55                              ; jne           386d <_sk_store_f32_hsw+0x6d>
  DB  196,67,45,24,225,1                  ; vinsertf128   $0x1,%xmm9,%ymm10,%ymm12
  DB  196,67,61,24,235,1                  ; vinsertf128   $0x1,%xmm11,%ymm8,%ymm13
  DB  196,67,45,6,201,49                  ; vperm2f128    $0x31,%ymm9,%ymm10,%ymm9
  DB  196,67,61,6,195,49                  ; vperm2f128    $0x31,%ymm11,%ymm8,%ymm8
  DB  196,65,125,17,36,128                ; vmovupd       %ymm12,(%r8,%rax,4)
  DB  196,65,125,17,108,128,32            ; vmovupd       %ymm13,0x20(%r8,%rax,4)
  DB  196,65,125,17,76,128,64             ; vmovupd       %ymm9,0x40(%r8,%rax,4)
  DB  196,65,125,17,68,128,96             ; vmovupd       %ymm8,0x60(%r8,%rax,4)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,65,121,17,20,128                ; vmovupd       %xmm10,(%r8,%rax,4)
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,240                             ; je            3869 <_sk_store_f32_hsw+0x69>
  DB  196,65,121,17,76,128,16             ; vmovupd       %xmm9,0x10(%r8,%rax,4)
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,227                             ; jb            3869 <_sk_store_f32_hsw+0x69>
  DB  196,65,121,17,68,128,32             ; vmovupd       %xmm8,0x20(%r8,%rax,4)
  DB  116,218                             ; je            3869 <_sk_store_f32_hsw+0x69>
  DB  196,65,121,17,92,128,48             ; vmovupd       %xmm11,0x30(%r8,%rax,4)
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,205                             ; jb            3869 <_sk_store_f32_hsw+0x69>
  DB  196,67,125,25,84,128,64,1           ; vextractf128  $0x1,%ymm10,0x40(%r8,%rax,4)
  DB  116,195                             ; je            3869 <_sk_store_f32_hsw+0x69>
  DB  196,67,125,25,76,128,80,1           ; vextractf128  $0x1,%ymm9,0x50(%r8,%rax,4)
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,181                             ; jb            3869 <_sk_store_f32_hsw+0x69>
  DB  196,67,125,25,68,128,96,1           ; vextractf128  $0x1,%ymm8,0x60(%r8,%rax,4)
  DB  235,171                             ; jmp           3869 <_sk_store_f32_hsw+0x69>

PUBLIC _sk_clamp_x_hsw
_sk_clamp_x_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
  DB  196,98,125,88,0                     ; vpbroadcastd  (%rax),%ymm8
  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
  DB  196,65,61,254,193                   ; vpaddd        %ymm9,%ymm8,%ymm8
  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_y_hsw
_sk_clamp_y_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,188,95,201                      ; vmaxps        %ymm1,%ymm8,%ymm1
  DB  196,98,125,88,0                     ; vpbroadcastd  (%rax),%ymm8
  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
  DB  196,65,61,254,193                   ; vpaddd        %ymm9,%ymm8,%ymm8
  DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_x_hsw
_sk_repeat_x_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,65,124,94,200                   ; vdivps        %ymm8,%ymm0,%ymm9
  DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
  DB  196,98,61,172,200                   ; vfnmadd213ps  %ymm0,%ymm8,%ymm9
  DB  197,253,118,192                     ; vpcmpeqd      %ymm0,%ymm0,%ymm0
  DB  197,189,254,192                     ; vpaddd        %ymm0,%ymm8,%ymm0
  DB  197,180,93,192                      ; vminps        %ymm0,%ymm9,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_y_hsw
_sk_repeat_y_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,65,116,94,200                   ; vdivps        %ymm8,%ymm1,%ymm9
  DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
  DB  196,98,61,172,201                   ; vfnmadd213ps  %ymm1,%ymm8,%ymm9
  DB  197,245,118,201                     ; vpcmpeqd      %ymm1,%ymm1,%ymm1
  DB  197,189,254,201                     ; vpaddd        %ymm1,%ymm8,%ymm1
  DB  197,180,93,201                      ; vminps        %ymm1,%ymm9,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_x_hsw
_sk_mirror_x_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,122,16,0                        ; vmovss        (%rax),%xmm8
  DB  196,66,125,24,200                   ; vbroadcastss  %xmm8,%ymm9
  DB  196,65,124,92,209                   ; vsubps        %ymm9,%ymm0,%ymm10
  DB  196,193,58,88,192                   ; vaddss        %xmm8,%xmm8,%xmm0
  DB  196,226,125,24,192                  ; vbroadcastss  %xmm0,%ymm0
  DB  197,44,94,192                       ; vdivps        %ymm0,%ymm10,%ymm8
  DB  196,67,125,8,192,1                  ; vroundps      $0x1,%ymm8,%ymm8
  DB  196,66,125,172,194                  ; vfnmadd213ps  %ymm10,%ymm0,%ymm8
  DB  196,193,60,92,193                   ; vsubps        %ymm9,%ymm8,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,60,92,192                       ; vsubps        %ymm0,%ymm8,%ymm8
  DB  197,188,84,192                      ; vandps        %ymm0,%ymm8,%ymm0
  DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
  DB  196,65,53,254,192                   ; vpaddd        %ymm8,%ymm9,%ymm8
  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_y_hsw
_sk_mirror_y_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,122,16,0                        ; vmovss        (%rax),%xmm8
  DB  196,66,125,24,200                   ; vbroadcastss  %xmm8,%ymm9
  DB  196,65,116,92,209                   ; vsubps        %ymm9,%ymm1,%ymm10
  DB  196,193,58,88,200                   ; vaddss        %xmm8,%xmm8,%xmm1
  DB  196,226,125,24,201                  ; vbroadcastss  %xmm1,%ymm1
  DB  197,44,94,193                       ; vdivps        %ymm1,%ymm10,%ymm8
  DB  196,67,125,8,192,1                  ; vroundps      $0x1,%ymm8,%ymm8
  DB  196,66,117,172,194                  ; vfnmadd213ps  %ymm10,%ymm1,%ymm8
  DB  196,193,60,92,201                   ; vsubps        %ymm9,%ymm8,%ymm1
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,60,92,193                       ; vsubps        %ymm1,%ymm8,%ymm8
  DB  197,188,84,201                      ; vandps        %ymm1,%ymm8,%ymm1
  DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
  DB  196,65,53,254,192                   ; vpaddd        %ymm8,%ymm9,%ymm8
  DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_luminance_to_alpha_hsw
_sk_luminance_to_alpha_hsw LABEL PROC
  DB  184,208,179,89,62                   ; mov           $0x3e59b3d0,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,98,125,88,195                   ; vpbroadcastd  %xmm3,%ymm8
  DB  184,89,23,55,63                     ; mov           $0x3f371759,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,226,125,88,219                  ; vpbroadcastd  %xmm3,%ymm3
  DB  197,228,89,201                      ; vmulps        %ymm1,%ymm3,%ymm1
  DB  196,98,125,168,193                  ; vfmadd213ps   %ymm1,%ymm0,%ymm8
  DB  184,152,221,147,61                  ; mov           $0x3d93dd98,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,226,125,88,216                  ; vpbroadcastd  %xmm0,%ymm3
  DB  196,194,109,168,216                 ; vfmadd213ps   %ymm8,%ymm2,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_2x3_hsw
_sk_matrix_2x3_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,8                     ; vbroadcastss  (%rax),%ymm9
  DB  196,98,125,24,80,8                  ; vbroadcastss  0x8(%rax),%ymm10
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,66,117,184,194                  ; vfmadd231ps   %ymm10,%ymm1,%ymm8
  DB  196,66,125,184,193                  ; vfmadd231ps   %ymm9,%ymm0,%ymm8
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,12                 ; vbroadcastss  0xc(%rax),%ymm11
  DB  196,98,125,24,72,20                 ; vbroadcastss  0x14(%rax),%ymm9
  DB  196,66,117,184,203                  ; vfmadd231ps   %ymm11,%ymm1,%ymm9
  DB  196,66,125,184,202                  ; vfmadd231ps   %ymm10,%ymm0,%ymm9
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  197,124,41,201                      ; vmovaps       %ymm9,%ymm1
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_3x4_hsw
_sk_matrix_3x4_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,8                     ; vbroadcastss  (%rax),%ymm9
  DB  196,98,125,24,80,12                 ; vbroadcastss  0xc(%rax),%ymm10
  DB  196,98,125,24,88,24                 ; vbroadcastss  0x18(%rax),%ymm11
  DB  196,98,125,24,64,36                 ; vbroadcastss  0x24(%rax),%ymm8
  DB  196,66,109,184,195                  ; vfmadd231ps   %ymm11,%ymm2,%ymm8
  DB  196,66,117,184,194                  ; vfmadd231ps   %ymm10,%ymm1,%ymm8
  DB  196,66,125,184,193                  ; vfmadd231ps   %ymm9,%ymm0,%ymm8
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,16                 ; vbroadcastss  0x10(%rax),%ymm11
  DB  196,98,125,24,96,28                 ; vbroadcastss  0x1c(%rax),%ymm12
  DB  196,98,125,24,72,40                 ; vbroadcastss  0x28(%rax),%ymm9
  DB  196,66,109,184,204                  ; vfmadd231ps   %ymm12,%ymm2,%ymm9
  DB  196,66,117,184,203                  ; vfmadd231ps   %ymm11,%ymm1,%ymm9
  DB  196,66,125,184,202                  ; vfmadd231ps   %ymm10,%ymm0,%ymm9
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  196,98,125,24,96,20                 ; vbroadcastss  0x14(%rax),%ymm12
  DB  196,98,125,24,104,32                ; vbroadcastss  0x20(%rax),%ymm13
  DB  196,98,125,24,80,44                 ; vbroadcastss  0x2c(%rax),%ymm10
  DB  196,66,109,184,213                  ; vfmadd231ps   %ymm13,%ymm2,%ymm10
  DB  196,66,117,184,212                  ; vfmadd231ps   %ymm12,%ymm1,%ymm10
  DB  196,66,125,184,211                  ; vfmadd231ps   %ymm11,%ymm0,%ymm10
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  197,124,41,201                      ; vmovaps       %ymm9,%ymm1
  DB  197,124,41,210                      ; vmovaps       %ymm10,%ymm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_4x5_hsw
_sk_matrix_4x5_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,8                     ; vbroadcastss  (%rax),%ymm9
  DB  196,98,125,24,80,16                 ; vbroadcastss  0x10(%rax),%ymm10
  DB  196,98,125,24,88,32                 ; vbroadcastss  0x20(%rax),%ymm11
  DB  196,98,125,24,96,48                 ; vbroadcastss  0x30(%rax),%ymm12
  DB  196,98,125,24,64,64                 ; vbroadcastss  0x40(%rax),%ymm8
  DB  196,66,101,184,196                  ; vfmadd231ps   %ymm12,%ymm3,%ymm8
  DB  196,66,109,184,195                  ; vfmadd231ps   %ymm11,%ymm2,%ymm8
  DB  196,66,117,184,194                  ; vfmadd231ps   %ymm10,%ymm1,%ymm8
  DB  196,66,125,184,193                  ; vfmadd231ps   %ymm9,%ymm0,%ymm8
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,20                 ; vbroadcastss  0x14(%rax),%ymm11
  DB  196,98,125,24,96,36                 ; vbroadcastss  0x24(%rax),%ymm12
  DB  196,98,125,24,104,52                ; vbroadcastss  0x34(%rax),%ymm13
  DB  196,98,125,24,72,68                 ; vbroadcastss  0x44(%rax),%ymm9
  DB  196,66,101,184,205                  ; vfmadd231ps   %ymm13,%ymm3,%ymm9
  DB  196,66,109,184,204                  ; vfmadd231ps   %ymm12,%ymm2,%ymm9
  DB  196,66,117,184,203                  ; vfmadd231ps   %ymm11,%ymm1,%ymm9
  DB  196,66,125,184,202                  ; vfmadd231ps   %ymm10,%ymm0,%ymm9
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  196,98,125,24,96,24                 ; vbroadcastss  0x18(%rax),%ymm12
  DB  196,98,125,24,104,40                ; vbroadcastss  0x28(%rax),%ymm13
  DB  196,98,125,24,112,56                ; vbroadcastss  0x38(%rax),%ymm14
  DB  196,98,125,24,80,72                 ; vbroadcastss  0x48(%rax),%ymm10
  DB  196,66,101,184,214                  ; vfmadd231ps   %ymm14,%ymm3,%ymm10
  DB  196,66,109,184,213                  ; vfmadd231ps   %ymm13,%ymm2,%ymm10
  DB  196,66,117,184,212                  ; vfmadd231ps   %ymm12,%ymm1,%ymm10
  DB  196,66,125,184,211                  ; vfmadd231ps   %ymm11,%ymm0,%ymm10
  DB  196,98,125,24,96,12                 ; vbroadcastss  0xc(%rax),%ymm12
  DB  196,98,125,24,104,28                ; vbroadcastss  0x1c(%rax),%ymm13
  DB  196,98,125,24,112,44                ; vbroadcastss  0x2c(%rax),%ymm14
  DB  196,98,125,24,120,60                ; vbroadcastss  0x3c(%rax),%ymm15
  DB  196,98,125,24,88,76                 ; vbroadcastss  0x4c(%rax),%ymm11
  DB  196,66,101,184,223                  ; vfmadd231ps   %ymm15,%ymm3,%ymm11
  DB  196,66,109,184,222                  ; vfmadd231ps   %ymm14,%ymm2,%ymm11
  DB  196,66,117,184,221                  ; vfmadd231ps   %ymm13,%ymm1,%ymm11
  DB  196,66,125,184,220                  ; vfmadd231ps   %ymm12,%ymm0,%ymm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  197,124,41,201                      ; vmovaps       %ymm9,%ymm1
  DB  197,124,41,210                      ; vmovaps       %ymm10,%ymm2
  DB  197,124,41,219                      ; vmovaps       %ymm11,%ymm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_perspective_hsw
_sk_matrix_perspective_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,98,125,24,72,4                  ; vbroadcastss  0x4(%rax),%ymm9
  DB  196,98,125,24,80,8                  ; vbroadcastss  0x8(%rax),%ymm10
  DB  196,66,117,184,209                  ; vfmadd231ps   %ymm9,%ymm1,%ymm10
  DB  196,66,125,184,208                  ; vfmadd231ps   %ymm8,%ymm0,%ymm10
  DB  196,98,125,24,64,12                 ; vbroadcastss  0xc(%rax),%ymm8
  DB  196,98,125,24,72,16                 ; vbroadcastss  0x10(%rax),%ymm9
  DB  196,98,125,24,88,20                 ; vbroadcastss  0x14(%rax),%ymm11
  DB  196,66,117,184,217                  ; vfmadd231ps   %ymm9,%ymm1,%ymm11
  DB  196,66,125,184,216                  ; vfmadd231ps   %ymm8,%ymm0,%ymm11
  DB  196,98,125,24,64,24                 ; vbroadcastss  0x18(%rax),%ymm8
  DB  196,98,125,24,72,28                 ; vbroadcastss  0x1c(%rax),%ymm9
  DB  196,98,125,24,96,32                 ; vbroadcastss  0x20(%rax),%ymm12
  DB  196,66,117,184,225                  ; vfmadd231ps   %ymm9,%ymm1,%ymm12
  DB  196,66,125,184,224                  ; vfmadd231ps   %ymm8,%ymm0,%ymm12
  DB  196,193,124,83,204                  ; vrcpps        %ymm12,%ymm1
  DB  197,172,89,193                      ; vmulps        %ymm1,%ymm10,%ymm0
  DB  197,164,89,201                      ; vmulps        %ymm1,%ymm11,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_hsw
_sk_linear_gradient_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,98,125,24,88,20                 ; vbroadcastss  0x14(%rax),%ymm11
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  196,98,125,24,72,28                 ; vbroadcastss  0x1c(%rax),%ymm9
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  77,133,192                          ; test          %r8,%r8
  DB  15,132,143,0,0,0                    ; je            3cf9 <_sk_linear_gradient_hsw+0xb5>
  DB  72,139,64,8                         ; mov           0x8(%rax),%rax
  DB  72,131,192,32                       ; add           $0x20,%rax
  DB  196,65,28,87,228                    ; vxorps        %ymm12,%ymm12,%ymm12
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  196,98,125,24,104,224               ; vbroadcastss  -0x20(%rax),%ymm13
  DB  196,65,124,194,237,1                ; vcmpltps      %ymm13,%ymm0,%ymm13
  DB  196,98,125,24,112,228               ; vbroadcastss  -0x1c(%rax),%ymm14
  DB  196,67,13,74,228,208                ; vblendvps     %ymm13,%ymm12,%ymm14,%ymm12
  DB  196,98,125,24,112,232               ; vbroadcastss  -0x18(%rax),%ymm14
  DB  196,227,13,74,201,208               ; vblendvps     %ymm13,%ymm1,%ymm14,%ymm1
  DB  196,98,125,24,112,236               ; vbroadcastss  -0x14(%rax),%ymm14
  DB  196,227,13,74,210,208               ; vblendvps     %ymm13,%ymm2,%ymm14,%ymm2
  DB  196,98,125,24,112,240               ; vbroadcastss  -0x10(%rax),%ymm14
  DB  196,227,13,74,219,208               ; vblendvps     %ymm13,%ymm3,%ymm14,%ymm3
  DB  196,98,125,24,112,244               ; vbroadcastss  -0xc(%rax),%ymm14
  DB  196,67,13,74,192,208                ; vblendvps     %ymm13,%ymm8,%ymm14,%ymm8
  DB  196,98,125,24,112,248               ; vbroadcastss  -0x8(%rax),%ymm14
  DB  196,67,13,74,219,208                ; vblendvps     %ymm13,%ymm11,%ymm14,%ymm11
  DB  196,98,125,24,112,252               ; vbroadcastss  -0x4(%rax),%ymm14
  DB  196,67,13,74,210,208                ; vblendvps     %ymm13,%ymm10,%ymm14,%ymm10
  DB  196,98,125,24,48                    ; vbroadcastss  (%rax),%ymm14
  DB  196,67,13,74,201,208                ; vblendvps     %ymm13,%ymm9,%ymm14,%ymm9
  DB  72,131,192,36                       ; add           $0x24,%rax
  DB  73,255,200                          ; dec           %r8
  DB  117,140                             ; jne           3c83 <_sk_linear_gradient_hsw+0x3f>
  DB  235,17                              ; jmp           3d0a <_sk_linear_gradient_hsw+0xc6>
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  196,65,28,87,228                    ; vxorps        %ymm12,%ymm12,%ymm12
  DB  196,66,125,184,196                  ; vfmadd231ps   %ymm12,%ymm0,%ymm8
  DB  196,194,125,168,203                 ; vfmadd213ps   %ymm11,%ymm0,%ymm1
  DB  196,194,125,168,210                 ; vfmadd213ps   %ymm10,%ymm0,%ymm2
  DB  196,194,125,168,217                 ; vfmadd213ps   %ymm9,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_2stops_hsw
_sk_linear_gradient_2stops_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,226,125,24,8                    ; vbroadcastss  (%rax),%ymm1
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,98,125,184,193                  ; vfmadd231ps   %ymm1,%ymm0,%ymm8
  DB  196,226,125,24,80,4                 ; vbroadcastss  0x4(%rax),%ymm2
  DB  196,226,125,24,72,20                ; vbroadcastss  0x14(%rax),%ymm1
  DB  196,226,125,184,202                 ; vfmadd231ps   %ymm2,%ymm0,%ymm1
  DB  196,226,125,24,88,8                 ; vbroadcastss  0x8(%rax),%ymm3
  DB  196,226,125,24,80,24                ; vbroadcastss  0x18(%rax),%ymm2
  DB  196,226,125,184,211                 ; vfmadd231ps   %ymm3,%ymm0,%ymm2
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,226,125,24,88,28                ; vbroadcastss  0x1c(%rax),%ymm3
  DB  196,194,125,184,217                 ; vfmadd231ps   %ymm9,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_save_xy_hsw
_sk_save_xy_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,88,200                       ; vaddps        %ymm0,%ymm8,%ymm9
  DB  196,67,125,8,209,1                  ; vroundps      $0x1,%ymm9,%ymm10
  DB  196,65,52,92,202                    ; vsubps        %ymm10,%ymm9,%ymm9
  DB  197,60,88,193                       ; vaddps        %ymm1,%ymm8,%ymm8
  DB  196,67,125,8,208,1                  ; vroundps      $0x1,%ymm8,%ymm10
  DB  196,65,60,92,194                    ; vsubps        %ymm10,%ymm8,%ymm8
  DB  197,252,17,0                        ; vmovups       %ymm0,(%rax)
  DB  197,252,17,72,32                    ; vmovups       %ymm1,0x20(%rax)
  DB  197,124,17,72,64                    ; vmovups       %ymm9,0x40(%rax)
  DB  197,124,17,64,96                    ; vmovups       %ymm8,0x60(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_accumulate_hsw
_sk_accumulate_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,16,128,128,0,0,0            ; vmovups       0x80(%rax),%ymm8
  DB  197,60,89,128,160,0,0,0             ; vmulps        0xa0(%rax),%ymm8,%ymm8
  DB  196,226,61,184,224                  ; vfmadd231ps   %ymm0,%ymm8,%ymm4
  DB  196,226,61,184,233                  ; vfmadd231ps   %ymm1,%ymm8,%ymm5
  DB  196,226,61,184,242                  ; vfmadd231ps   %ymm2,%ymm8,%ymm6
  DB  196,98,101,168,199                  ; vfmadd213ps   %ymm7,%ymm3,%ymm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,199                      ; vmovaps       %ymm8,%ymm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_nx_hsw
_sk_bilinear_nx_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_px_hsw
_sk_bilinear_px_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  197,124,16,64,64                    ; vmovups       0x40(%rax),%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_ny_hsw
_sk_bilinear_ny_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_py_hsw
_sk_bilinear_py_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  197,124,16,64,96                    ; vmovups       0x60(%rax),%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3x_hsw
_sk_bicubic_n3x_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,191                  ; mov           $0xbfc00000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,66,125,88,219                   ; vpbroadcastd  %xmm11,%ymm11
  DB  196,66,61,168,211                   ; vfmadd213ps   %ymm11,%ymm8,%ymm10
  DB  196,65,44,89,193                    ; vmulps        %ymm9,%ymm10,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1x_hsw
_sk_bicubic_n1x_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,200                  ; vmovd         %r8d,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  196,66,61,168,202                   ; vfmadd213ps   %ymm10,%ymm8,%ymm9
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  196,66,61,184,209                   ; vfmadd231ps   %ymm9,%ymm8,%ymm10
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,200                  ; vmovd         %r8d,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
  DB  197,124,17,136,128,0,0,0            ; vmovups       %ymm9,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1x_hsw
_sk_bicubic_p1x_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,98,125,88,192                   ; vpbroadcastd  %xmm0,%ymm8
  DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
  DB  197,124,16,72,64                    ; vmovups       0x40(%rax),%ymm9
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,66,125,88,219                   ; vpbroadcastd  %xmm11,%ymm11
  DB  196,66,53,168,211                   ; vfmadd213ps   %ymm11,%ymm9,%ymm10
  DB  196,66,53,168,208                   ; vfmadd213ps   %ymm8,%ymm9,%ymm10
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,66,53,184,194                   ; vfmadd231ps   %ymm10,%ymm9,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3x_hsw
_sk_bicubic_p3x_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,226,125,88,192                  ; vpbroadcastd  %xmm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  197,124,16,64,64                    ; vmovups       0x40(%rax),%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,66,125,88,219                   ; vpbroadcastd  %xmm11,%ymm11
  DB  196,66,61,168,211                   ; vfmadd213ps   %ymm11,%ymm8,%ymm10
  DB  196,65,52,89,194                    ; vmulps        %ymm10,%ymm9,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3y_hsw
_sk_bicubic_n3y_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,191                  ; mov           $0xbfc00000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,66,125,88,219                   ; vpbroadcastd  %xmm11,%ymm11
  DB  196,66,61,168,211                   ; vfmadd213ps   %ymm11,%ymm8,%ymm10
  DB  196,65,44,89,193                    ; vmulps        %ymm9,%ymm10,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1y_hsw
_sk_bicubic_n1y_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,200                  ; vmovd         %r8d,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  196,66,61,168,202                   ; vfmadd213ps   %ymm10,%ymm8,%ymm9
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  196,66,61,184,209                   ; vfmadd231ps   %ymm9,%ymm8,%ymm10
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,200                  ; vmovd         %r8d,%xmm9
  DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
  DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
  DB  197,124,17,136,160,0,0,0            ; vmovups       %ymm9,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1y_hsw
_sk_bicubic_p1y_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,98,125,88,193                   ; vpbroadcastd  %xmm1,%ymm8
  DB  197,188,88,72,32                    ; vaddps        0x20(%rax),%ymm8,%ymm1
  DB  197,124,16,72,96                    ; vmovups       0x60(%rax),%ymm9
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,66,125,88,219                   ; vpbroadcastd  %xmm11,%ymm11
  DB  196,66,53,168,211                   ; vfmadd213ps   %ymm11,%ymm9,%ymm10
  DB  196,66,53,168,208                   ; vfmadd213ps   %ymm8,%ymm9,%ymm10
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
  DB  196,66,53,184,194                   ; vfmadd231ps   %ymm10,%ymm9,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3y_hsw
_sk_bicubic_p3y_hsw LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  197,124,16,64,96                    ; vmovups       0x60(%rax),%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,66,125,88,210                   ; vpbroadcastd  %xmm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,66,125,88,219                   ; vpbroadcastd  %xmm11,%ymm11
  DB  196,66,61,168,211                   ; vfmadd213ps   %ymm11,%ymm8,%ymm10
  DB  196,65,52,89,194                    ; vmulps        %ymm10,%ymm9,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_callback_hsw
_sk_callback_hsw LABEL PROC
  DB  65,86                               ; push          %r14
  DB  83                                  ; push          %rbx
  DB  72,129,236,40,1,0,0                 ; sub           $0x128,%rsp
  DB  197,252,17,188,36,0,1,0,0           ; vmovups       %ymm7,0x100(%rsp)
  DB  197,252,17,180,36,224,0,0,0         ; vmovups       %ymm6,0xe0(%rsp)
  DB  197,252,17,172,36,192,0,0,0         ; vmovups       %ymm5,0xc0(%rsp)
  DB  197,252,17,164,36,160,0,0,0         ; vmovups       %ymm4,0xa0(%rsp)
  DB  197,252,17,156,36,128,0,0,0         ; vmovups       %ymm3,0x80(%rsp)
  DB  197,252,17,84,36,96                 ; vmovups       %ymm2,0x60(%rsp)
  DB  197,252,17,76,36,64                 ; vmovups       %ymm1,0x40(%rsp)
  DB  197,252,17,68,36,32                 ; vmovups       %ymm0,0x20(%rsp)
  DB  72,137,203                          ; mov           %rcx,%rbx
  DB  73,137,214                          ; mov           %rdx,%r14
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,72,8                         ; mov           0x8(%rax),%rcx
  DB  72,133,219                          ; test          %rbx,%rbx
  DB  186,8,0,0,0                         ; mov           $0x8,%edx
  DB  15,69,211                           ; cmovne        %ebx,%edx
  DB  197,248,119                         ; vzeroupper
  DB  255,16                              ; callq         *(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  72,137,217                          ; mov           %rbx,%rcx
  DB  197,252,16,68,36,32                 ; vmovups       0x20(%rsp),%ymm0
  DB  197,252,16,76,36,64                 ; vmovups       0x40(%rsp),%ymm1
  DB  197,252,16,84,36,96                 ; vmovups       0x60(%rsp),%ymm2
  DB  197,252,16,156,36,128,0,0,0         ; vmovups       0x80(%rsp),%ymm3
  DB  197,252,16,164,36,160,0,0,0         ; vmovups       0xa0(%rsp),%ymm4
  DB  197,252,16,172,36,192,0,0,0         ; vmovups       0xc0(%rsp),%ymm5
  DB  197,252,16,180,36,224,0,0,0         ; vmovups       0xe0(%rsp),%ymm6
  DB  197,252,16,188,36,0,1,0,0           ; vmovups       0x100(%rsp),%ymm7
  DB  72,129,196,40,1,0,0                 ; add           $0x128,%rsp
  DB  91                                  ; pop           %rbx
  DB  65,94                               ; pop           %r14
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_start_pipeline_avx
_sk_start_pipeline_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  86                                  ; push          %rsi
  DB  87                                  ; push          %rdi
  DB  83                                  ; push          %rbx
  DB  72,129,236,160,0,0,0                ; sub           $0xa0,%rsp
  DB  197,120,41,188,36,144,0,0,0         ; vmovaps       %xmm15,0x90(%rsp)
  DB  197,120,41,180,36,128,0,0,0         ; vmovaps       %xmm14,0x80(%rsp)
  DB  197,120,41,108,36,112               ; vmovaps       %xmm13,0x70(%rsp)
  DB  197,120,41,100,36,96                ; vmovaps       %xmm12,0x60(%rsp)
  DB  197,120,41,92,36,80                 ; vmovaps       %xmm11,0x50(%rsp)
  DB  197,120,41,84,36,64                 ; vmovaps       %xmm10,0x40(%rsp)
  DB  197,120,41,76,36,48                 ; vmovaps       %xmm9,0x30(%rsp)
  DB  197,120,41,68,36,32                 ; vmovaps       %xmm8,0x20(%rsp)
  DB  197,248,41,124,36,16                ; vmovaps       %xmm7,0x10(%rsp)
  DB  197,248,41,52,36                    ; vmovaps       %xmm6,(%rsp)
  DB  77,137,205                          ; mov           %r9,%r13
  DB  77,137,198                          ; mov           %r8,%r14
  DB  72,137,203                          ; mov           %rcx,%rbx
  DB  72,137,214                          ; mov           %rdx,%rsi
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,199                          ; mov           %rax,%r15
  DB  73,137,244                          ; mov           %rsi,%r12
  DB  72,141,67,8                         ; lea           0x8(%rbx),%rax
  DB  76,57,232                           ; cmp           %r13,%rax
  DB  118,5                               ; jbe           75 <_sk_start_pipeline_avx+0x75>
  DB  72,137,223                          ; mov           %rbx,%rdi
  DB  235,65                              ; jmp           b6 <_sk_start_pipeline_avx+0xb6>
  DB  185,0,0,0,0                         ; mov           $0x0,%ecx
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
  DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
  DB  197,204,87,246                      ; vxorps        %ymm6,%ymm6,%ymm6
  DB  197,196,87,255                      ; vxorps        %ymm7,%ymm7,%ymm7
  DB  72,137,223                          ; mov           %rbx,%rdi
  DB  76,137,230                          ; mov           %r12,%rsi
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  65,255,215                          ; callq         *%r15
  DB  72,141,123,8                        ; lea           0x8(%rbx),%rdi
  DB  72,131,195,16                       ; add           $0x10,%rbx
  DB  76,57,235                           ; cmp           %r13,%rbx
  DB  72,137,251                          ; mov           %rdi,%rbx
  DB  118,191                             ; jbe           75 <_sk_start_pipeline_avx+0x75>
  DB  76,137,233                          ; mov           %r13,%rcx
  DB  72,41,249                           ; sub           %rdi,%rcx
  DB  116,41                              ; je            e7 <_sk_start_pipeline_avx+0xe7>
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
  DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
  DB  197,204,87,246                      ; vxorps        %ymm6,%ymm6,%ymm6
  DB  197,196,87,255                      ; vxorps        %ymm7,%ymm7,%ymm7
  DB  76,137,230                          ; mov           %r12,%rsi
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  65,255,215                          ; callq         *%r15
  DB  76,137,232                          ; mov           %r13,%rax
  DB  197,248,40,52,36                    ; vmovaps       (%rsp),%xmm6
  DB  197,248,40,124,36,16                ; vmovaps       0x10(%rsp),%xmm7
  DB  197,120,40,68,36,32                 ; vmovaps       0x20(%rsp),%xmm8
  DB  197,120,40,76,36,48                 ; vmovaps       0x30(%rsp),%xmm9
  DB  197,120,40,84,36,64                 ; vmovaps       0x40(%rsp),%xmm10
  DB  197,120,40,92,36,80                 ; vmovaps       0x50(%rsp),%xmm11
  DB  197,120,40,100,36,96                ; vmovaps       0x60(%rsp),%xmm12
  DB  197,120,40,108,36,112               ; vmovaps       0x70(%rsp),%xmm13
  DB  197,120,40,180,36,128,0,0,0         ; vmovaps       0x80(%rsp),%xmm14
  DB  197,120,40,188,36,144,0,0,0         ; vmovaps       0x90(%rsp),%xmm15
  DB  72,129,196,160,0,0,0                ; add           $0xa0,%rsp
  DB  91                                  ; pop           %rbx
  DB  95                                  ; pop           %rdi
  DB  94                                  ; pop           %rsi
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  197,248,119                         ; vzeroupper
  DB  195                                 ; retq

PUBLIC _sk_just_return_avx
_sk_just_return_avx LABEL PROC
  DB  195                                 ; retq

PUBLIC _sk_seed_shader_avx
_sk_seed_shader_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,249,110,199                     ; vmovd         %edi,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
  DB  197,252,88,2                        ; vaddps        (%rdx),%ymm0,%ymm0
  DB  196,226,125,24,16                   ; vbroadcastss  (%rax),%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,236,88,201                      ; vaddps        %ymm1,%ymm2,%ymm1
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
  DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
  DB  197,204,87,246                      ; vxorps        %ymm6,%ymm6,%ymm6
  DB  197,196,87,255                      ; vxorps        %ymm7,%ymm7,%ymm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_constant_color_avx
_sk_constant_color_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,226,125,24,0                    ; vbroadcastss  (%rax),%ymm0
  DB  196,226,125,24,72,4                 ; vbroadcastss  0x4(%rax),%ymm1
  DB  196,226,125,24,80,8                 ; vbroadcastss  0x8(%rax),%ymm2
  DB  196,226,125,24,88,12                ; vbroadcastss  0xc(%rax),%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clear_avx
_sk_clear_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcatop_avx
_sk_srcatop_avx LABEL PROC
  DB  197,124,89,199                      ; vmulps        %ymm7,%ymm0,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,124,92,203                      ; vsubps        %ymm3,%ymm0,%ymm9
  DB  197,180,89,196                      ; vmulps        %ymm4,%ymm9,%ymm0
  DB  197,188,88,192                      ; vaddps        %ymm0,%ymm8,%ymm0
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,52,89,197                       ; vmulps        %ymm5,%ymm9,%ymm8
  DB  196,193,116,88,200                  ; vaddps        %ymm8,%ymm1,%ymm1
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,52,89,198                       ; vmulps        %ymm6,%ymm9,%ymm8
  DB  196,193,108,88,208                  ; vaddps        %ymm8,%ymm2,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  197,52,89,199                       ; vmulps        %ymm7,%ymm9,%ymm8
  DB  196,193,100,88,216                  ; vaddps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstatop_avx
_sk_dstatop_avx LABEL PROC
  DB  197,100,89,196                      ; vmulps        %ymm4,%ymm3,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  197,52,92,207                       ; vsubps        %ymm7,%ymm9,%ymm9
  DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
  DB  197,188,88,192                      ; vaddps        %ymm0,%ymm8,%ymm0
  DB  197,100,89,197                      ; vmulps        %ymm5,%ymm3,%ymm8
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  197,188,88,201                      ; vaddps        %ymm1,%ymm8,%ymm1
  DB  197,100,89,198                      ; vmulps        %ymm6,%ymm3,%ymm8
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  197,188,88,210                      ; vaddps        %ymm2,%ymm8,%ymm2
  DB  197,100,89,199                      ; vmulps        %ymm7,%ymm3,%ymm8
  DB  197,180,89,219                      ; vmulps        %ymm3,%ymm9,%ymm3
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcin_avx
_sk_srcin_avx LABEL PROC
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstin_avx
_sk_dstin_avx LABEL PROC
  DB  197,228,89,196                      ; vmulps        %ymm4,%ymm3,%ymm0
  DB  197,228,89,205                      ; vmulps        %ymm5,%ymm3,%ymm1
  DB  197,228,89,214                      ; vmulps        %ymm6,%ymm3,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcout_avx
_sk_srcout_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstout_avx
_sk_dstout_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,92,219                      ; vsubps        %ymm3,%ymm0,%ymm3
  DB  197,228,89,196                      ; vmulps        %ymm4,%ymm3,%ymm0
  DB  197,228,89,205                      ; vmulps        %ymm5,%ymm3,%ymm1
  DB  197,228,89,214                      ; vmulps        %ymm6,%ymm3,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcover_avx
_sk_srcover_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,204                       ; vmulps        %ymm4,%ymm8,%ymm9
  DB  197,180,88,192                      ; vaddps        %ymm0,%ymm9,%ymm0
  DB  197,60,89,205                       ; vmulps        %ymm5,%ymm8,%ymm9
  DB  197,180,88,201                      ; vaddps        %ymm1,%ymm9,%ymm1
  DB  197,60,89,206                       ; vmulps        %ymm6,%ymm8,%ymm9
  DB  197,180,88,210                      ; vaddps        %ymm2,%ymm9,%ymm2
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstover_avx
_sk_dstover_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,236,88,214                      ; vaddps        %ymm6,%ymm2,%ymm2
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  197,228,88,223                      ; vaddps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_modulate_avx
_sk_modulate_avx LABEL PROC
  DB  197,252,89,196                      ; vmulps        %ymm4,%ymm0,%ymm0
  DB  197,244,89,205                      ; vmulps        %ymm5,%ymm1,%ymm1
  DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_multiply_avx
_sk_multiply_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
  DB  197,52,89,208                       ; vmulps        %ymm0,%ymm9,%ymm10
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,220                       ; vmulps        %ymm4,%ymm8,%ymm11
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  197,252,89,196                      ; vmulps        %ymm4,%ymm0,%ymm0
  DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
  DB  197,52,89,209                       ; vmulps        %ymm1,%ymm9,%ymm10
  DB  197,60,89,221                       ; vmulps        %ymm5,%ymm8,%ymm11
  DB  196,65,36,88,210                    ; vaddps        %ymm10,%ymm11,%ymm10
  DB  197,244,89,205                      ; vmulps        %ymm5,%ymm1,%ymm1
  DB  196,193,116,88,202                  ; vaddps        %ymm10,%ymm1,%ymm1
  DB  197,52,89,210                       ; vmulps        %ymm2,%ymm9,%ymm10
  DB  197,60,89,222                       ; vmulps        %ymm6,%ymm8,%ymm11
  DB  196,65,36,88,210                    ; vaddps        %ymm10,%ymm11,%ymm10
  DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
  DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
  DB  197,52,89,203                       ; vmulps        %ymm3,%ymm9,%ymm9
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  196,193,100,88,216                  ; vaddps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_plus__avx
_sk_plus__avx LABEL PROC
  DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
  DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
  DB  197,236,88,214                      ; vaddps        %ymm6,%ymm2,%ymm2
  DB  197,228,88,223                      ; vaddps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_screen_avx
_sk_screen_avx LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,196                      ; vmulps        %ymm4,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,205                      ; vmulps        %ymm5,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  197,100,88,199                      ; vaddps        %ymm7,%ymm3,%ymm8
  DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
  DB  197,188,92,219                      ; vsubps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_xor__avx
_sk_xor__avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
  DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,212                       ; vmulps        %ymm4,%ymm8,%ymm10
  DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  197,60,89,213                       ; vmulps        %ymm5,%ymm8,%ymm10
  DB  197,172,88,201                      ; vaddps        %ymm1,%ymm10,%ymm1
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  197,60,89,214                       ; vmulps        %ymm6,%ymm8,%ymm10
  DB  197,172,88,210                      ; vaddps        %ymm2,%ymm10,%ymm2
  DB  197,180,89,219                      ; vmulps        %ymm3,%ymm9,%ymm3
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_darken_avx
_sk_darken_avx LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,100,89,204                      ; vmulps        %ymm4,%ymm3,%ymm9
  DB  196,193,124,95,193                  ; vmaxps        %ymm9,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,100,89,205                      ; vmulps        %ymm5,%ymm3,%ymm9
  DB  196,193,116,95,201                  ; vmaxps        %ymm9,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
  DB  196,193,108,95,209                  ; vmaxps        %ymm9,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lighten_avx
_sk_lighten_avx LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,100,89,204                      ; vmulps        %ymm4,%ymm3,%ymm9
  DB  196,193,124,93,193                  ; vminps        %ymm9,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,100,89,205                      ; vmulps        %ymm5,%ymm3,%ymm9
  DB  196,193,116,93,201                  ; vminps        %ymm9,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
  DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_difference_avx
_sk_difference_avx LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
  DB  197,100,89,204                      ; vmulps        %ymm4,%ymm3,%ymm9
  DB  196,193,124,93,193                  ; vminps        %ymm9,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
  DB  197,100,89,205                      ; vmulps        %ymm5,%ymm3,%ymm9
  DB  196,193,116,93,201                  ; vminps        %ymm9,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,215                      ; vmulps        %ymm7,%ymm2,%ymm2
  DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
  DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_exclusion_avx
_sk_exclusion_avx LABEL PROC
  DB  197,124,88,196                      ; vaddps        %ymm4,%ymm0,%ymm8
  DB  197,252,89,196                      ; vmulps        %ymm4,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
  DB  197,116,88,197                      ; vaddps        %ymm5,%ymm1,%ymm8
  DB  197,244,89,205                      ; vmulps        %ymm5,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
  DB  197,108,88,198                      ; vaddps        %ymm6,%ymm2,%ymm8
  DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colorburn_avx
_sk_colorburn_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
  DB  197,52,89,216                       ; vmulps        %ymm0,%ymm9,%ymm11
  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,228                       ; vmulps        %ymm4,%ymm8,%ymm12
  DB  197,68,92,236                       ; vsubps        %ymm4,%ymm7,%ymm13
  DB  197,20,89,235                       ; vmulps        %ymm3,%ymm13,%ymm13
  DB  197,20,94,232                       ; vdivps        %ymm0,%ymm13,%ymm13
  DB  196,65,68,93,237                    ; vminps        %ymm13,%ymm7,%ymm13
  DB  196,65,68,92,237                    ; vsubps        %ymm13,%ymm7,%ymm13
  DB  197,20,89,235                       ; vmulps        %ymm3,%ymm13,%ymm13
  DB  196,65,20,88,235                    ; vaddps        %ymm11,%ymm13,%ymm13
  DB  196,65,28,88,237                    ; vaddps        %ymm13,%ymm12,%ymm13
  DB  197,28,88,224                       ; vaddps        %ymm0,%ymm12,%ymm12
  DB  196,193,124,194,194,0               ; vcmpeqps      %ymm10,%ymm0,%ymm0
  DB  196,195,21,74,196,0                 ; vblendvps     %ymm0,%ymm12,%ymm13,%ymm0
  DB  197,92,194,231,0                    ; vcmpeqps      %ymm7,%ymm4,%ymm12
  DB  197,36,88,220                       ; vaddps        %ymm4,%ymm11,%ymm11
  DB  196,195,125,74,195,192              ; vblendvps     %ymm12,%ymm11,%ymm0,%ymm0
  DB  197,52,89,217                       ; vmulps        %ymm1,%ymm9,%ymm11
  DB  197,60,89,229                       ; vmulps        %ymm5,%ymm8,%ymm12
  DB  197,68,92,237                       ; vsubps        %ymm5,%ymm7,%ymm13
  DB  197,20,89,235                       ; vmulps        %ymm3,%ymm13,%ymm13
  DB  197,20,94,233                       ; vdivps        %ymm1,%ymm13,%ymm13
  DB  196,65,68,93,237                    ; vminps        %ymm13,%ymm7,%ymm13
  DB  196,65,68,92,237                    ; vsubps        %ymm13,%ymm7,%ymm13
  DB  197,20,89,235                       ; vmulps        %ymm3,%ymm13,%ymm13
  DB  196,65,36,88,237                    ; vaddps        %ymm13,%ymm11,%ymm13
  DB  196,65,28,88,237                    ; vaddps        %ymm13,%ymm12,%ymm13
  DB  197,28,88,225                       ; vaddps        %ymm1,%ymm12,%ymm12
  DB  196,193,116,194,202,0               ; vcmpeqps      %ymm10,%ymm1,%ymm1
  DB  196,195,21,74,204,16                ; vblendvps     %ymm1,%ymm12,%ymm13,%ymm1
  DB  197,84,194,231,0                    ; vcmpeqps      %ymm7,%ymm5,%ymm12
  DB  197,36,88,221                       ; vaddps        %ymm5,%ymm11,%ymm11
  DB  196,195,117,74,203,192              ; vblendvps     %ymm12,%ymm11,%ymm1,%ymm1
  DB  197,52,89,202                       ; vmulps        %ymm2,%ymm9,%ymm9
  DB  196,65,108,194,210,0                ; vcmpeqps      %ymm10,%ymm2,%ymm10
  DB  197,60,89,222                       ; vmulps        %ymm6,%ymm8,%ymm11
  DB  197,68,92,230                       ; vsubps        %ymm6,%ymm7,%ymm12
  DB  197,28,89,227                       ; vmulps        %ymm3,%ymm12,%ymm12
  DB  197,28,94,226                       ; vdivps        %ymm2,%ymm12,%ymm12
  DB  197,164,88,210                      ; vaddps        %ymm2,%ymm11,%ymm2
  DB  196,65,68,93,228                    ; vminps        %ymm12,%ymm7,%ymm12
  DB  196,65,68,92,228                    ; vsubps        %ymm12,%ymm7,%ymm12
  DB  197,28,89,227                       ; vmulps        %ymm3,%ymm12,%ymm12
  DB  196,65,52,88,228                    ; vaddps        %ymm12,%ymm9,%ymm12
  DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
  DB  196,227,37,74,210,160               ; vblendvps     %ymm10,%ymm2,%ymm11,%ymm2
  DB  197,76,194,215,0                    ; vcmpeqps      %ymm7,%ymm6,%ymm10
  DB  197,52,88,206                       ; vaddps        %ymm6,%ymm9,%ymm9
  DB  196,195,109,74,209,160              ; vblendvps     %ymm10,%ymm9,%ymm2,%ymm2
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colordodge_avx
_sk_colordodge_avx LABEL PROC
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  197,52,92,215                       ; vsubps        %ymm7,%ymm9,%ymm10
  DB  197,44,89,216                       ; vmulps        %ymm0,%ymm10,%ymm11
  DB  197,52,92,203                       ; vsubps        %ymm3,%ymm9,%ymm9
  DB  197,100,89,228                      ; vmulps        %ymm4,%ymm3,%ymm12
  DB  197,100,92,232                      ; vsubps        %ymm0,%ymm3,%ymm13
  DB  196,65,28,94,229                    ; vdivps        %ymm13,%ymm12,%ymm12
  DB  197,52,89,236                       ; vmulps        %ymm4,%ymm9,%ymm13
  DB  196,65,68,93,228                    ; vminps        %ymm12,%ymm7,%ymm12
  DB  197,28,89,227                       ; vmulps        %ymm3,%ymm12,%ymm12
  DB  196,65,28,88,227                    ; vaddps        %ymm11,%ymm12,%ymm12
  DB  196,65,20,88,228                    ; vaddps        %ymm12,%ymm13,%ymm12
  DB  197,20,88,232                       ; vaddps        %ymm0,%ymm13,%ymm13
  DB  197,252,194,195,0                   ; vcmpeqps      %ymm3,%ymm0,%ymm0
  DB  196,195,29,74,197,0                 ; vblendvps     %ymm0,%ymm13,%ymm12,%ymm0
  DB  196,65,92,194,224,0                 ; vcmpeqps      %ymm8,%ymm4,%ymm12
  DB  197,36,88,220                       ; vaddps        %ymm4,%ymm11,%ymm11
  DB  196,195,125,74,195,192              ; vblendvps     %ymm12,%ymm11,%ymm0,%ymm0
  DB  197,44,89,217                       ; vmulps        %ymm1,%ymm10,%ymm11
  DB  197,100,89,229                      ; vmulps        %ymm5,%ymm3,%ymm12
  DB  197,100,92,233                      ; vsubps        %ymm1,%ymm3,%ymm13
  DB  196,65,28,94,229                    ; vdivps        %ymm13,%ymm12,%ymm12
  DB  197,52,89,237                       ; vmulps        %ymm5,%ymm9,%ymm13
  DB  196,65,68,93,228                    ; vminps        %ymm12,%ymm7,%ymm12
  DB  197,28,89,227                       ; vmulps        %ymm3,%ymm12,%ymm12
  DB  196,65,28,88,227                    ; vaddps        %ymm11,%ymm12,%ymm12
  DB  196,65,20,88,228                    ; vaddps        %ymm12,%ymm13,%ymm12
  DB  197,20,88,233                       ; vaddps        %ymm1,%ymm13,%ymm13
  DB  197,244,194,203,0                   ; vcmpeqps      %ymm3,%ymm1,%ymm1
  DB  196,195,29,74,205,16                ; vblendvps     %ymm1,%ymm13,%ymm12,%ymm1
  DB  196,65,84,194,224,0                 ; vcmpeqps      %ymm8,%ymm5,%ymm12
  DB  197,36,88,221                       ; vaddps        %ymm5,%ymm11,%ymm11
  DB  196,195,117,74,203,192              ; vblendvps     %ymm12,%ymm11,%ymm1,%ymm1
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  197,100,89,222                      ; vmulps        %ymm6,%ymm3,%ymm11
  DB  197,100,92,226                      ; vsubps        %ymm2,%ymm3,%ymm12
  DB  196,65,36,94,220                    ; vdivps        %ymm12,%ymm11,%ymm11
  DB  197,52,89,230                       ; vmulps        %ymm6,%ymm9,%ymm12
  DB  196,65,68,93,219                    ; vminps        %ymm11,%ymm7,%ymm11
  DB  197,36,89,219                       ; vmulps        %ymm3,%ymm11,%ymm11
  DB  196,65,44,88,219                    ; vaddps        %ymm11,%ymm10,%ymm11
  DB  196,65,28,88,219                    ; vaddps        %ymm11,%ymm12,%ymm11
  DB  197,28,88,226                       ; vaddps        %ymm2,%ymm12,%ymm12
  DB  197,236,194,211,0                   ; vcmpeqps      %ymm3,%ymm2,%ymm2
  DB  196,195,37,74,212,32                ; vblendvps     %ymm2,%ymm12,%ymm11,%ymm2
  DB  196,65,76,194,192,0                 ; vcmpeqps      %ymm8,%ymm6,%ymm8
  DB  197,44,88,214                       ; vaddps        %ymm6,%ymm10,%ymm10
  DB  196,195,109,74,210,128              ; vblendvps     %ymm8,%ymm10,%ymm2,%ymm2
  DB  197,52,89,199                       ; vmulps        %ymm7,%ymm9,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hardlight_avx
_sk_hardlight_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
  DB  197,44,89,200                       ; vmulps        %ymm0,%ymm10,%ymm9
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,220                       ; vmulps        %ymm4,%ymm8,%ymm11
  DB  196,65,52,88,219                    ; vaddps        %ymm11,%ymm9,%ymm11
  DB  197,124,88,200                      ; vaddps        %ymm0,%ymm0,%ymm9
  DB  197,52,194,227,2                    ; vcmpleps      %ymm3,%ymm9,%ymm12
  DB  197,124,89,204                      ; vmulps        %ymm4,%ymm0,%ymm9
  DB  196,65,52,88,233                    ; vaddps        %ymm9,%ymm9,%ymm13
  DB  197,100,89,207                      ; vmulps        %ymm7,%ymm3,%ymm9
  DB  197,68,92,244                       ; vsubps        %ymm4,%ymm7,%ymm14
  DB  197,228,92,192                      ; vsubps        %ymm0,%ymm3,%ymm0
  DB  196,193,124,89,198                  ; vmulps        %ymm14,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
  DB  196,195,125,74,197,192              ; vblendvps     %ymm12,%ymm13,%ymm0,%ymm0
  DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
  DB  197,44,89,217                       ; vmulps        %ymm1,%ymm10,%ymm11
  DB  197,60,89,229                       ; vmulps        %ymm5,%ymm8,%ymm12
  DB  196,65,28,88,219                    ; vaddps        %ymm11,%ymm12,%ymm11
  DB  197,116,88,225                      ; vaddps        %ymm1,%ymm1,%ymm12
  DB  197,28,194,227,2                    ; vcmpleps      %ymm3,%ymm12,%ymm12
  DB  197,116,89,237                      ; vmulps        %ymm5,%ymm1,%ymm13
  DB  196,65,20,88,237                    ; vaddps        %ymm13,%ymm13,%ymm13
  DB  197,68,92,245                       ; vsubps        %ymm5,%ymm7,%ymm14
  DB  197,228,92,201                      ; vsubps        %ymm1,%ymm3,%ymm1
  DB  196,193,116,89,206                  ; vmulps        %ymm14,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,180,92,201                      ; vsubps        %ymm1,%ymm9,%ymm1
  DB  196,195,117,74,205,192              ; vblendvps     %ymm12,%ymm13,%ymm1,%ymm1
  DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  197,60,89,222                       ; vmulps        %ymm6,%ymm8,%ymm11
  DB  196,65,36,88,210                    ; vaddps        %ymm10,%ymm11,%ymm10
  DB  197,108,88,218                      ; vaddps        %ymm2,%ymm2,%ymm11
  DB  197,36,194,219,2                    ; vcmpleps      %ymm3,%ymm11,%ymm11
  DB  197,108,89,230                      ; vmulps        %ymm6,%ymm2,%ymm12
  DB  196,65,28,88,228                    ; vaddps        %ymm12,%ymm12,%ymm12
  DB  197,68,92,238                       ; vsubps        %ymm6,%ymm7,%ymm13
  DB  197,228,92,210                      ; vsubps        %ymm2,%ymm3,%ymm2
  DB  196,193,108,89,213                  ; vmulps        %ymm13,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,180,92,210                      ; vsubps        %ymm2,%ymm9,%ymm2
  DB  196,195,109,74,212,176              ; vblendvps     %ymm11,%ymm12,%ymm2,%ymm2
  DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_overlay_avx
_sk_overlay_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
  DB  197,44,89,200                       ; vmulps        %ymm0,%ymm10,%ymm9
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,220                       ; vmulps        %ymm4,%ymm8,%ymm11
  DB  196,65,52,88,219                    ; vaddps        %ymm11,%ymm9,%ymm11
  DB  197,92,88,204                       ; vaddps        %ymm4,%ymm4,%ymm9
  DB  197,52,194,231,2                    ; vcmpleps      %ymm7,%ymm9,%ymm12
  DB  197,124,89,204                      ; vmulps        %ymm4,%ymm0,%ymm9
  DB  196,65,52,88,233                    ; vaddps        %ymm9,%ymm9,%ymm13
  DB  197,100,89,207                      ; vmulps        %ymm7,%ymm3,%ymm9
  DB  197,68,92,244                       ; vsubps        %ymm4,%ymm7,%ymm14
  DB  197,228,92,192                      ; vsubps        %ymm0,%ymm3,%ymm0
  DB  196,193,124,89,198                  ; vmulps        %ymm14,%ymm0,%ymm0
  DB  197,252,88,192                      ; vaddps        %ymm0,%ymm0,%ymm0
  DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
  DB  196,195,125,74,197,192              ; vblendvps     %ymm12,%ymm13,%ymm0,%ymm0
  DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
  DB  197,44,89,217                       ; vmulps        %ymm1,%ymm10,%ymm11
  DB  197,60,89,229                       ; vmulps        %ymm5,%ymm8,%ymm12
  DB  196,65,28,88,219                    ; vaddps        %ymm11,%ymm12,%ymm11
  DB  197,84,88,229                       ; vaddps        %ymm5,%ymm5,%ymm12
  DB  197,28,194,231,2                    ; vcmpleps      %ymm7,%ymm12,%ymm12
  DB  197,116,89,237                      ; vmulps        %ymm5,%ymm1,%ymm13
  DB  196,65,20,88,237                    ; vaddps        %ymm13,%ymm13,%ymm13
  DB  197,68,92,245                       ; vsubps        %ymm5,%ymm7,%ymm14
  DB  197,228,92,201                      ; vsubps        %ymm1,%ymm3,%ymm1
  DB  196,193,116,89,206                  ; vmulps        %ymm14,%ymm1,%ymm1
  DB  197,244,88,201                      ; vaddps        %ymm1,%ymm1,%ymm1
  DB  197,180,92,201                      ; vsubps        %ymm1,%ymm9,%ymm1
  DB  196,195,117,74,205,192              ; vblendvps     %ymm12,%ymm13,%ymm1,%ymm1
  DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  197,60,89,222                       ; vmulps        %ymm6,%ymm8,%ymm11
  DB  196,65,36,88,210                    ; vaddps        %ymm10,%ymm11,%ymm10
  DB  197,76,88,222                       ; vaddps        %ymm6,%ymm6,%ymm11
  DB  197,36,194,223,2                    ; vcmpleps      %ymm7,%ymm11,%ymm11
  DB  197,108,89,230                      ; vmulps        %ymm6,%ymm2,%ymm12
  DB  196,65,28,88,228                    ; vaddps        %ymm12,%ymm12,%ymm12
  DB  197,68,92,238                       ; vsubps        %ymm6,%ymm7,%ymm13
  DB  197,228,92,210                      ; vsubps        %ymm2,%ymm3,%ymm2
  DB  196,193,108,89,213                  ; vmulps        %ymm13,%ymm2,%ymm2
  DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
  DB  197,180,92,210                      ; vsubps        %ymm2,%ymm9,%ymm2
  DB  196,195,109,74,212,176              ; vblendvps     %ymm11,%ymm12,%ymm2,%ymm2
  DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_softlight_avx
_sk_softlight_avx LABEL PROC
  DB  72,131,236,56                       ; sub           $0x38,%rsp
  DB  197,252,17,20,36                    ; vmovups       %ymm2,(%rsp)
  DB  197,252,40,209                      ; vmovaps       %ymm1,%ymm2
  DB  196,65,52,87,201                    ; vxorps        %ymm9,%ymm9,%ymm9
  DB  197,52,194,215,1                    ; vcmpltps      %ymm7,%ymm9,%ymm10
  DB  197,92,94,199                       ; vdivps        %ymm7,%ymm4,%ymm8
  DB  196,67,53,74,216,160                ; vblendvps     %ymm10,%ymm8,%ymm9,%ymm11
  DB  196,65,36,88,195                    ; vaddps        %ymm11,%ymm11,%ymm8
  DB  196,65,60,88,224                    ; vaddps        %ymm8,%ymm8,%ymm12
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,65,28,89,236                    ; vmulps        %ymm12,%ymm12,%ymm13
  DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
  DB  196,65,36,92,232                    ; vsubps        %ymm8,%ymm11,%ymm13
  DB  196,65,28,89,237                    ; vmulps        %ymm13,%ymm12,%ymm13
  DB  184,0,0,224,64                      ; mov           $0x40e00000,%eax
  DB  197,121,110,224                     ; vmovd         %eax,%xmm12
  DB  196,67,121,4,228,0                  ; vpermilps     $0x0,%xmm12,%xmm12
  DB  196,67,29,24,228,1                  ; vinsertf128   $0x1,%xmm12,%ymm12,%ymm12
  DB  196,65,36,89,244                    ; vmulps        %ymm12,%ymm11,%ymm14
  DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
  DB  196,65,124,82,243                   ; vrsqrtps      %ymm11,%ymm14
  DB  196,65,124,83,246                   ; vrcpps        %ymm14,%ymm14
  DB  196,65,12,92,243                    ; vsubps        %ymm11,%ymm14,%ymm14
  DB  197,92,88,252                       ; vaddps        %ymm4,%ymm4,%ymm15
  DB  196,65,4,88,255                     ; vaddps        %ymm15,%ymm15,%ymm15
  DB  197,4,194,255,2                     ; vcmpleps      %ymm7,%ymm15,%ymm15
  DB  196,67,13,74,237,240                ; vblendvps     %ymm15,%ymm13,%ymm14,%ymm13
  DB  197,124,88,240                      ; vaddps        %ymm0,%ymm0,%ymm14
  DB  197,12,92,251                       ; vsubps        %ymm3,%ymm14,%ymm15
  DB  196,65,60,92,219                    ; vsubps        %ymm11,%ymm8,%ymm11
  DB  196,65,4,89,219                     ; vmulps        %ymm11,%ymm15,%ymm11
  DB  197,36,88,219                       ; vaddps        %ymm3,%ymm11,%ymm11
  DB  197,36,89,220                       ; vmulps        %ymm4,%ymm11,%ymm11
  DB  197,4,89,255                        ; vmulps        %ymm7,%ymm15,%ymm15
  DB  196,65,4,89,237                     ; vmulps        %ymm13,%ymm15,%ymm13
  DB  197,100,89,252                      ; vmulps        %ymm4,%ymm3,%ymm15
  DB  196,65,4,88,237                     ; vaddps        %ymm13,%ymm15,%ymm13
  DB  197,12,194,243,2                    ; vcmpleps      %ymm3,%ymm14,%ymm14
  DB  196,195,21,74,203,224               ; vblendvps     %ymm14,%ymm11,%ymm13,%ymm1
  DB  197,84,94,239                       ; vdivps        %ymm7,%ymm5,%ymm13
  DB  196,67,53,74,237,160                ; vblendvps     %ymm10,%ymm13,%ymm9,%ymm13
  DB  196,65,20,88,245                    ; vaddps        %ymm13,%ymm13,%ymm14
  DB  196,65,12,88,246                    ; vaddps        %ymm14,%ymm14,%ymm14
  DB  196,65,12,89,254                    ; vmulps        %ymm14,%ymm14,%ymm15
  DB  196,65,12,88,247                    ; vaddps        %ymm15,%ymm14,%ymm14
  DB  196,65,20,92,248                    ; vsubps        %ymm8,%ymm13,%ymm15
  DB  196,65,4,89,246                     ; vmulps        %ymm14,%ymm15,%ymm14
  DB  196,65,28,89,253                    ; vmulps        %ymm13,%ymm12,%ymm15
  DB  196,65,4,88,246                     ; vaddps        %ymm14,%ymm15,%ymm14
  DB  196,65,124,82,253                   ; vrsqrtps      %ymm13,%ymm15
  DB  196,65,124,83,255                   ; vrcpps        %ymm15,%ymm15
  DB  196,65,4,92,253                     ; vsubps        %ymm13,%ymm15,%ymm15
  DB  197,84,88,221                       ; vaddps        %ymm5,%ymm5,%ymm11
  DB  196,65,36,88,219                    ; vaddps        %ymm11,%ymm11,%ymm11
  DB  197,36,194,223,2                    ; vcmpleps      %ymm7,%ymm11,%ymm11
  DB  196,67,5,74,222,176                 ; vblendvps     %ymm11,%ymm14,%ymm15,%ymm11
  DB  197,108,88,242                      ; vaddps        %ymm2,%ymm2,%ymm14
  DB  196,65,60,92,237                    ; vsubps        %ymm13,%ymm8,%ymm13
  DB  197,12,92,251                       ; vsubps        %ymm3,%ymm14,%ymm15
  DB  196,65,4,89,237                     ; vmulps        %ymm13,%ymm15,%ymm13
  DB  197,4,89,255                        ; vmulps        %ymm7,%ymm15,%ymm15
  DB  196,65,4,89,219                     ; vmulps        %ymm11,%ymm15,%ymm11
  DB  197,100,89,253                      ; vmulps        %ymm5,%ymm3,%ymm15
  DB  196,65,4,88,219                     ; vaddps        %ymm11,%ymm15,%ymm11
  DB  197,20,88,235                       ; vaddps        %ymm3,%ymm13,%ymm13
  DB  197,20,89,237                       ; vmulps        %ymm5,%ymm13,%ymm13
  DB  197,12,194,243,2                    ; vcmpleps      %ymm3,%ymm14,%ymm14
  DB  196,67,37,74,237,224                ; vblendvps     %ymm14,%ymm13,%ymm11,%ymm13
  DB  197,76,94,223                       ; vdivps        %ymm7,%ymm6,%ymm11
  DB  196,67,53,74,203,160                ; vblendvps     %ymm10,%ymm11,%ymm9,%ymm9
  DB  196,65,52,88,209                    ; vaddps        %ymm9,%ymm9,%ymm10
  DB  196,65,44,88,210                    ; vaddps        %ymm10,%ymm10,%ymm10
  DB  196,65,44,89,218                    ; vmulps        %ymm10,%ymm10,%ymm11
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  196,65,52,92,216                    ; vsubps        %ymm8,%ymm9,%ymm11
  DB  196,65,36,89,210                    ; vmulps        %ymm10,%ymm11,%ymm10
  DB  196,65,28,89,217                    ; vmulps        %ymm9,%ymm12,%ymm11
  DB  196,65,36,88,210                    ; vaddps        %ymm10,%ymm11,%ymm10
  DB  196,65,124,82,217                   ; vrsqrtps      %ymm9,%ymm11
  DB  196,65,124,83,219                   ; vrcpps        %ymm11,%ymm11
  DB  196,65,36,92,217                    ; vsubps        %ymm9,%ymm11,%ymm11
  DB  197,76,88,230                       ; vaddps        %ymm6,%ymm6,%ymm12
  DB  196,65,28,88,228                    ; vaddps        %ymm12,%ymm12,%ymm12
  DB  197,28,194,231,2                    ; vcmpleps      %ymm7,%ymm12,%ymm12
  DB  196,67,37,74,210,192                ; vblendvps     %ymm12,%ymm10,%ymm11,%ymm10
  DB  197,124,16,52,36                    ; vmovups       (%rsp),%ymm14
  DB  196,65,12,88,222                    ; vaddps        %ymm14,%ymm14,%ymm11
  DB  197,36,92,227                       ; vsubps        %ymm3,%ymm11,%ymm12
  DB  196,65,60,92,201                    ; vsubps        %ymm9,%ymm8,%ymm9
  DB  196,65,28,89,201                    ; vmulps        %ymm9,%ymm12,%ymm9
  DB  197,28,89,231                       ; vmulps        %ymm7,%ymm12,%ymm12
  DB  196,65,28,89,210                    ; vmulps        %ymm10,%ymm12,%ymm10
  DB  197,100,89,230                      ; vmulps        %ymm6,%ymm3,%ymm12
  DB  196,65,28,88,210                    ; vaddps        %ymm10,%ymm12,%ymm10
  DB  197,52,88,203                       ; vaddps        %ymm3,%ymm9,%ymm9
  DB  197,52,89,206                       ; vmulps        %ymm6,%ymm9,%ymm9
  DB  197,36,194,219,2                    ; vcmpleps      %ymm3,%ymm11,%ymm11
  DB  196,67,45,74,201,176                ; vblendvps     %ymm11,%ymm9,%ymm10,%ymm9
  DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
  DB  197,60,89,220                       ; vmulps        %ymm4,%ymm8,%ymm11
  DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
  DB  197,244,88,192                      ; vaddps        %ymm0,%ymm1,%ymm0
  DB  197,172,89,202                      ; vmulps        %ymm2,%ymm10,%ymm1
  DB  197,188,89,213                      ; vmulps        %ymm5,%ymm8,%ymm2
  DB  197,236,88,201                      ; vaddps        %ymm1,%ymm2,%ymm1
  DB  196,193,116,88,205                  ; vaddps        %ymm13,%ymm1,%ymm1
  DB  196,193,44,89,214                   ; vmulps        %ymm14,%ymm10,%ymm2
  DB  197,60,89,214                       ; vmulps        %ymm6,%ymm8,%ymm10
  DB  197,172,88,210                      ; vaddps        %ymm2,%ymm10,%ymm2
  DB  196,193,108,88,209                  ; vaddps        %ymm9,%ymm2,%ymm2
  DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
  DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,131,196,56                       ; add           $0x38,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_0_avx
_sk_clamp_0_avx LABEL PROC
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
  DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
  DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
  DB  196,193,100,95,216                  ; vmaxps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_1_avx
_sk_clamp_1_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
  DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
  DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
  DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_a_avx
_sk_clamp_a_avx LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
  DB  197,252,93,195                      ; vminps        %ymm3,%ymm0,%ymm0
  DB  197,244,93,203                      ; vminps        %ymm3,%ymm1,%ymm1
  DB  197,236,93,211                      ; vminps        %ymm3,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_set_rgb_avx
_sk_set_rgb_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,226,125,24,0                    ; vbroadcastss  (%rax),%ymm0
  DB  196,226,125,24,72,4                 ; vbroadcastss  0x4(%rax),%ymm1
  DB  196,226,125,24,80,8                 ; vbroadcastss  0x8(%rax),%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_rb_avx
_sk_swap_rb_avx LABEL PROC
  DB  197,124,40,192                      ; vmovaps       %ymm0,%ymm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,194                      ; vmovaps       %ymm2,%ymm0
  DB  197,124,41,194                      ; vmovaps       %ymm8,%ymm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_avx
_sk_swap_avx LABEL PROC
  DB  197,124,40,195                      ; vmovaps       %ymm3,%ymm8
  DB  197,124,40,202                      ; vmovaps       %ymm2,%ymm9
  DB  197,124,40,209                      ; vmovaps       %ymm1,%ymm10
  DB  197,124,40,216                      ; vmovaps       %ymm0,%ymm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,196                      ; vmovaps       %ymm4,%ymm0
  DB  197,252,40,205                      ; vmovaps       %ymm5,%ymm1
  DB  197,252,40,214                      ; vmovaps       %ymm6,%ymm2
  DB  197,252,40,223                      ; vmovaps       %ymm7,%ymm3
  DB  197,124,41,220                      ; vmovaps       %ymm11,%ymm4
  DB  197,124,41,213                      ; vmovaps       %ymm10,%ymm5
  DB  197,124,41,206                      ; vmovaps       %ymm9,%ymm6
  DB  197,124,41,199                      ; vmovaps       %ymm8,%ymm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_src_dst_avx
_sk_move_src_dst_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,224                      ; vmovaps       %ymm0,%ymm4
  DB  197,252,40,233                      ; vmovaps       %ymm1,%ymm5
  DB  197,252,40,242                      ; vmovaps       %ymm2,%ymm6
  DB  197,252,40,251                      ; vmovaps       %ymm3,%ymm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_dst_src_avx
_sk_move_dst_src_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,196                      ; vmovaps       %ymm4,%ymm0
  DB  197,252,40,205                      ; vmovaps       %ymm5,%ymm1
  DB  197,252,40,214                      ; vmovaps       %ymm6,%ymm2
  DB  197,252,40,223                      ; vmovaps       %ymm7,%ymm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_premul_avx
_sk_premul_avx LABEL PROC
  DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
  DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
  DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_unpremul_avx
_sk_unpremul_avx LABEL PROC
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,100,194,200,0                ; vcmpeqps      %ymm8,%ymm3,%ymm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  197,44,94,211                       ; vdivps        %ymm3,%ymm10,%ymm10
  DB  196,67,45,74,192,144                ; vblendvps     %ymm9,%ymm8,%ymm10,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_srgb_avx
_sk_from_srgb_avx LABEL PROC
  DB  184,145,131,158,61                  ; mov           $0x3d9e8391,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
  DB  184,154,153,153,62                  ; mov           $0x3e99999a,%eax
  DB  197,121,110,216                     ; vmovd         %eax,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  184,92,143,50,63                    ; mov           $0x3f328f5c,%eax
  DB  197,121,110,224                     ; vmovd         %eax,%xmm12
  DB  196,67,121,4,228,0                  ; vpermilps     $0x0,%xmm12,%xmm12
  DB  196,67,29,24,228,1                  ; vinsertf128   $0x1,%xmm12,%ymm12,%ymm12
  DB  197,36,89,232                       ; vmulps        %ymm0,%ymm11,%ymm13
  DB  196,65,20,88,236                    ; vaddps        %ymm12,%ymm13,%ymm13
  DB  184,10,215,35,59                    ; mov           $0x3b23d70a,%eax
  DB  197,121,110,240                     ; vmovd         %eax,%xmm14
  DB  196,67,121,4,246,0                  ; vpermilps     $0x0,%xmm14,%xmm14
  DB  196,67,13,24,246,1                  ; vinsertf128   $0x1,%xmm14,%ymm14,%ymm14
  DB  196,65,44,89,213                    ; vmulps        %ymm13,%ymm10,%ymm10
  DB  196,65,12,88,210                    ; vaddps        %ymm10,%ymm14,%ymm10
  DB  184,174,71,97,61                    ; mov           $0x3d6147ae,%eax
  DB  197,121,110,232                     ; vmovd         %eax,%xmm13
  DB  196,67,121,4,237,0                  ; vpermilps     $0x0,%xmm13,%xmm13
  DB  196,67,21,24,237,1                  ; vinsertf128   $0x1,%xmm13,%ymm13,%ymm13
  DB  196,193,124,194,197,1               ; vcmpltps      %ymm13,%ymm0,%ymm0
  DB  196,195,45,74,193,0                 ; vblendvps     %ymm0,%ymm9,%ymm10,%ymm0
  DB  197,60,89,201                       ; vmulps        %ymm1,%ymm8,%ymm9
  DB  197,116,89,209                      ; vmulps        %ymm1,%ymm1,%ymm10
  DB  197,36,89,249                       ; vmulps        %ymm1,%ymm11,%ymm15
  DB  196,65,28,88,255                    ; vaddps        %ymm15,%ymm12,%ymm15
  DB  196,65,44,89,215                    ; vmulps        %ymm15,%ymm10,%ymm10
  DB  196,65,12,88,210                    ; vaddps        %ymm10,%ymm14,%ymm10
  DB  196,193,116,194,205,1               ; vcmpltps      %ymm13,%ymm1,%ymm1
  DB  196,195,45,74,201,16                ; vblendvps     %ymm1,%ymm9,%ymm10,%ymm1
  DB  197,60,89,194                       ; vmulps        %ymm2,%ymm8,%ymm8
  DB  197,108,89,202                      ; vmulps        %ymm2,%ymm2,%ymm9
  DB  197,36,89,210                       ; vmulps        %ymm2,%ymm11,%ymm10
  DB  196,65,28,88,210                    ; vaddps        %ymm10,%ymm12,%ymm10
  DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
  DB  196,65,12,88,201                    ; vaddps        %ymm9,%ymm14,%ymm9
  DB  196,193,108,194,213,1               ; vcmpltps      %ymm13,%ymm2,%ymm2
  DB  196,195,53,74,208,32                ; vblendvps     %ymm2,%ymm8,%ymm9,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_srgb_avx
_sk_to_srgb_avx LABEL PROC
  DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
  DB  196,65,124,83,232                   ; vrcpps        %ymm8,%ymm13
  DB  196,65,124,82,240                   ; vrsqrtps      %ymm8,%ymm14
  DB  184,41,92,71,65                     ; mov           $0x41475c29,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,89,224                       ; vmulps        %ymm0,%ymm8,%ymm12
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  184,194,135,210,62                  ; mov           $0x3ed287c2,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  184,206,111,48,63                   ; mov           $0x3f306fce,%eax
  DB  197,121,110,216                     ; vmovd         %eax,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  184,168,87,202,61                   ; mov           $0x3dca57a8,%eax
  DB  53,0,0,0,128                        ; xor           $0x80000000,%eax
  DB  197,121,110,248                     ; vmovd         %eax,%xmm15
  DB  196,67,121,4,255,0                  ; vpermilps     $0x0,%xmm15,%xmm15
  DB  196,67,5,24,255,1                   ; vinsertf128   $0x1,%xmm15,%ymm15,%ymm15
  DB  196,65,20,89,235                    ; vmulps        %ymm11,%ymm13,%ymm13
  DB  196,65,20,88,239                    ; vaddps        %ymm15,%ymm13,%ymm13
  DB  196,65,12,89,242                    ; vmulps        %ymm10,%ymm14,%ymm14
  DB  196,65,12,88,237                    ; vaddps        %ymm13,%ymm14,%ymm13
  DB  196,65,52,93,237                    ; vminps        %ymm13,%ymm9,%ymm13
  DB  184,4,231,140,59                    ; mov           $0x3b8ce704,%eax
  DB  197,121,110,240                     ; vmovd         %eax,%xmm14
  DB  196,67,121,4,246,0                  ; vpermilps     $0x0,%xmm14,%xmm14
  DB  196,67,13,24,246,1                  ; vinsertf128   $0x1,%xmm14,%ymm14,%ymm14
  DB  196,193,124,194,198,1               ; vcmpltps      %ymm14,%ymm0,%ymm0
  DB  196,195,21,74,196,0                 ; vblendvps     %ymm0,%ymm12,%ymm13,%ymm0
  DB  197,124,82,225                      ; vrsqrtps      %ymm1,%ymm12
  DB  196,65,124,83,236                   ; vrcpps        %ymm12,%ymm13
  DB  196,65,124,82,228                   ; vrsqrtps      %ymm12,%ymm12
  DB  196,65,36,89,237                    ; vmulps        %ymm13,%ymm11,%ymm13
  DB  196,65,4,88,237                     ; vaddps        %ymm13,%ymm15,%ymm13
  DB  196,65,44,89,228                    ; vmulps        %ymm12,%ymm10,%ymm12
  DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
  DB  197,60,89,233                       ; vmulps        %ymm1,%ymm8,%ymm13
  DB  196,65,52,93,228                    ; vminps        %ymm12,%ymm9,%ymm12
  DB  196,193,116,194,206,1               ; vcmpltps      %ymm14,%ymm1,%ymm1
  DB  196,195,29,74,205,16                ; vblendvps     %ymm1,%ymm13,%ymm12,%ymm1
  DB  197,124,82,226                      ; vrsqrtps      %ymm2,%ymm12
  DB  196,65,124,83,236                   ; vrcpps        %ymm12,%ymm13
  DB  196,65,36,89,221                    ; vmulps        %ymm13,%ymm11,%ymm11
  DB  196,65,4,88,219                     ; vaddps        %ymm11,%ymm15,%ymm11
  DB  196,65,124,82,228                   ; vrsqrtps      %ymm12,%ymm12
  DB  196,65,44,89,212                    ; vmulps        %ymm12,%ymm10,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  196,65,52,93,202                    ; vminps        %ymm10,%ymm9,%ymm9
  DB  197,60,89,194                       ; vmulps        %ymm2,%ymm8,%ymm8
  DB  196,193,108,194,214,1               ; vcmpltps      %ymm14,%ymm2,%ymm2
  DB  196,195,53,74,208,32                ; vblendvps     %ymm2,%ymm8,%ymm9,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_2dot2_avx
_sk_from_2dot2_avx LABEL PROC
  DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,200                   ; vrsqrtps      %ymm8,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  197,252,89,192                      ; vmulps        %ymm0,%ymm0,%ymm0
  DB  196,65,60,89,208                    ; vmulps        %ymm8,%ymm8,%ymm10
  DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
  DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  196,65,124,82,210                   ; vrsqrtps      %ymm10,%ymm10
  DB  197,244,89,201                      ; vmulps        %ymm1,%ymm1,%ymm1
  DB  196,65,52,89,217                    ; vmulps        %ymm9,%ymm9,%ymm11
  DB  196,65,52,89,203                    ; vmulps        %ymm11,%ymm9,%ymm9
  DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
  DB  197,124,82,202                      ; vrsqrtps      %ymm2,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  196,65,124,82,210                   ; vrsqrtps      %ymm10,%ymm10
  DB  197,236,89,210                      ; vmulps        %ymm2,%ymm2,%ymm2
  DB  196,65,52,89,217                    ; vmulps        %ymm9,%ymm9,%ymm11
  DB  196,65,52,89,203                    ; vmulps        %ymm11,%ymm9,%ymm9
  DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
  DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
  DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_2dot2_avx
_sk_to_2dot2_avx LABEL PROC
  DB  197,252,82,192                      ; vrsqrtps      %ymm0,%ymm0
  DB  197,124,82,192                      ; vrsqrtps      %ymm0,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,192                   ; vrsqrtps      %ymm8,%ymm8
  DB  196,65,124,82,200                   ; vrsqrtps      %ymm8,%ymm9
  DB  197,252,83,192                      ; vrcpps        %ymm0,%ymm0
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  196,65,124,83,193                   ; vrcpps        %ymm9,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
  DB  197,252,82,201                      ; vrsqrtps      %ymm1,%ymm1
  DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  197,252,83,201                      ; vrcpps        %ymm1,%ymm1
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  196,65,124,83,202                   ; vrcpps        %ymm10,%ymm9
  DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
  DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
  DB  197,252,82,210                      ; vrsqrtps      %ymm2,%ymm2
  DB  197,124,82,202                      ; vrsqrtps      %ymm2,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,201                   ; vrsqrtps      %ymm9,%ymm9
  DB  196,65,124,82,209                   ; vrsqrtps      %ymm9,%ymm10
  DB  197,252,83,210                      ; vrcpps        %ymm2,%ymm2
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  196,65,124,83,202                   ; vrcpps        %ymm10,%ymm9
  DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
  DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_rgb_to_hsl_avx
_sk_rgb_to_hsl_avx LABEL PROC
  DB  197,124,95,193                      ; vmaxps        %ymm1,%ymm0,%ymm8
  DB  197,60,95,194                       ; vmaxps        %ymm2,%ymm8,%ymm8
  DB  197,124,93,201                      ; vminps        %ymm1,%ymm0,%ymm9
  DB  197,52,93,202                       ; vminps        %ymm2,%ymm9,%ymm9
  DB  196,65,60,92,209                    ; vsubps        %ymm9,%ymm8,%ymm10
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,121,110,216                     ; vmovd         %eax,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  196,65,36,94,218                    ; vdivps        %ymm10,%ymm11,%ymm11
  DB  65,184,171,170,42,62                ; mov           $0x3e2aaaab,%r8d
  DB  197,116,92,226                      ; vsubps        %ymm2,%ymm1,%ymm12
  DB  196,65,28,89,227                    ; vmulps        %ymm11,%ymm12,%ymm12
  DB  65,185,0,0,192,64                   ; mov           $0x40c00000,%r9d
  DB  197,108,92,232                      ; vsubps        %ymm0,%ymm2,%ymm13
  DB  196,65,20,89,235                    ; vmulps        %ymm11,%ymm13,%ymm13
  DB  65,186,0,0,0,64                     ; mov           $0x40000000,%r10d
  DB  197,124,92,241                      ; vsubps        %ymm1,%ymm0,%ymm14
  DB  196,65,12,89,219                    ; vmulps        %ymm11,%ymm14,%ymm11
  DB  184,0,0,128,64                      ; mov           $0x40800000,%eax
  DB  197,121,110,240                     ; vmovd         %eax,%xmm14
  DB  196,67,121,4,246,0                  ; vpermilps     $0x0,%xmm14,%xmm14
  DB  196,67,13,24,246,1                  ; vinsertf128   $0x1,%xmm14,%ymm14,%ymm14
  DB  196,65,36,88,222                    ; vaddps        %ymm14,%ymm11,%ymm11
  DB  196,65,121,110,242                  ; vmovd         %r10d,%xmm14
  DB  197,244,194,210,1                   ; vcmpltps      %ymm2,%ymm1,%ymm2
  DB  197,188,194,201,0                   ; vcmpeqps      %ymm1,%ymm8,%ymm1
  DB  196,67,121,4,246,0                  ; vpermilps     $0x0,%xmm14,%xmm14
  DB  196,67,13,24,246,1                  ; vinsertf128   $0x1,%xmm14,%ymm14,%ymm14
  DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
  DB  196,67,37,74,221,16                 ; vblendvps     %ymm1,%ymm13,%ymm11,%ymm11
  DB  196,193,121,110,201                 ; vmovd         %r9d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  196,65,20,87,237                    ; vxorps        %ymm13,%ymm13,%ymm13
  DB  196,227,21,74,201,32                ; vblendvps     %ymm2,%ymm1,%ymm13,%ymm1
  DB  196,193,116,88,204                  ; vaddps        %ymm12,%ymm1,%ymm1
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  197,188,194,192,0                   ; vcmpeqps      %ymm0,%ymm8,%ymm0
  DB  196,227,37,74,193,0                 ; vblendvps     %ymm0,%ymm1,%ymm11,%ymm0
  DB  196,193,60,88,201                   ; vaddps        %ymm9,%ymm8,%ymm1
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,99,109,24,218,1                 ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm11
  DB  196,193,116,89,211                  ; vmulps        %ymm11,%ymm1,%ymm2
  DB  197,36,194,218,1                    ; vcmpltps      %ymm2,%ymm11,%ymm11
  DB  196,65,12,92,224                    ; vsubps        %ymm8,%ymm14,%ymm12
  DB  196,65,28,92,225                    ; vsubps        %ymm9,%ymm12,%ymm12
  DB  196,195,117,74,204,176              ; vblendvps     %ymm11,%ymm12,%ymm1,%ymm1
  DB  196,65,60,194,193,0                 ; vcmpeqps      %ymm9,%ymm8,%ymm8
  DB  197,172,94,201                      ; vdivps        %ymm1,%ymm10,%ymm1
  DB  196,67,125,74,205,128               ; vblendvps     %ymm8,%ymm13,%ymm0,%ymm9
  DB  196,195,117,74,205,128              ; vblendvps     %ymm8,%ymm13,%ymm1,%ymm1
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hsl_to_rgb_avx
_sk_hsl_to_rgb_avx LABEL PROC
  DB  72,129,236,184,0,0,0                ; sub           $0xb8,%rsp
  DB  197,252,17,188,36,128,0,0,0         ; vmovups       %ymm7,0x80(%rsp)
  DB  197,252,17,116,36,96                ; vmovups       %ymm6,0x60(%rsp)
  DB  197,252,17,108,36,64                ; vmovups       %ymm5,0x40(%rsp)
  DB  197,252,17,100,36,32                ; vmovups       %ymm4,0x20(%rsp)
  DB  197,252,17,28,36                    ; vmovups       %ymm3,(%rsp)
  DB  197,252,40,226                      ; vmovaps       %ymm2,%ymm4
  DB  197,252,40,233                      ; vmovaps       %ymm1,%ymm5
  DB  197,252,40,216                      ; vmovaps       %ymm0,%ymm3
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm8
  DB  196,193,92,194,192,1                ; vcmpltps      %ymm8,%ymm4,%ymm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,99,109,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm10
  DB  197,172,88,213                      ; vaddps        %ymm5,%ymm10,%ymm2
  DB  197,236,89,212                      ; vmulps        %ymm4,%ymm2,%ymm2
  DB  197,84,88,204                       ; vaddps        %ymm4,%ymm5,%ymm9
  DB  197,84,89,220                       ; vmulps        %ymm4,%ymm5,%ymm11
  DB  196,65,52,92,203                    ; vsubps        %ymm11,%ymm9,%ymm9
  DB  196,99,53,74,202,0                  ; vblendvps     %ymm0,%ymm2,%ymm9,%ymm9
  DB  65,184,0,0,0,64                     ; mov           $0x40000000,%r8d
  DB  184,171,170,170,62                  ; mov           $0x3eaaaaab,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,224,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm12
  DB  197,28,88,251                       ; vaddps        %ymm3,%ymm12,%ymm15
  DB  184,0,0,0,0                         ; mov           $0x0,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,232,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm13
  DB  196,193,44,194,199,1                ; vcmpltps      %ymm15,%ymm10,%ymm0
  DB  196,193,4,92,210                    ; vsubps        %ymm10,%ymm15,%ymm2
  DB  196,227,5,74,194,0                  ; vblendvps     %ymm0,%ymm2,%ymm15,%ymm0
  DB  196,193,4,194,213,1                 ; vcmpltps      %ymm13,%ymm15,%ymm2
  DB  196,65,44,88,223                    ; vaddps        %ymm15,%ymm10,%ymm11
  DB  196,195,125,74,203,32               ; vblendvps     %ymm2,%ymm11,%ymm0,%ymm1
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,89,196                      ; vmulps        %ymm4,%ymm0,%ymm0
  DB  196,65,124,92,217                   ; vsubps        %ymm9,%ymm0,%ymm11
  DB  65,184,171,170,42,62                ; mov           $0x3e2aaaab,%r8d
  DB  184,0,0,192,64                      ; mov           $0x40c00000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,193,52,92,211                   ; vsubps        %ymm11,%ymm9,%ymm2
  DB  197,108,89,240                      ; vmulps        %ymm0,%ymm2,%ymm14
  DB  184,171,170,42,63                   ; mov           $0x3f2aaaab,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,92,209                      ; vsubps        %ymm1,%ymm0,%ymm2
  DB  197,140,89,210                      ; vmulps        %ymm2,%ymm14,%ymm2
  DB  197,164,88,210                      ; vaddps        %ymm2,%ymm11,%ymm2
  DB  197,244,194,248,1                   ; vcmpltps      %ymm0,%ymm1,%ymm7
  DB  196,227,37,74,210,112               ; vblendvps     %ymm7,%ymm2,%ymm11,%ymm2
  DB  196,193,116,194,248,1               ; vcmpltps      %ymm8,%ymm1,%ymm7
  DB  196,195,109,74,249,112              ; vblendvps     %ymm7,%ymm9,%ymm2,%ymm7
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,244,194,202,1                   ; vcmpltps      %ymm2,%ymm1,%ymm1
  DB  196,65,4,89,254                     ; vmulps        %ymm14,%ymm15,%ymm15
  DB  196,65,36,88,255                    ; vaddps        %ymm15,%ymm11,%ymm15
  DB  196,67,69,74,255,16                 ; vblendvps     %ymm1,%ymm15,%ymm7,%ymm15
  DB  197,172,194,203,1                   ; vcmpltps      %ymm3,%ymm10,%ymm1
  DB  196,193,100,92,250                  ; vsubps        %ymm10,%ymm3,%ymm7
  DB  196,227,101,74,207,16               ; vblendvps     %ymm1,%ymm7,%ymm3,%ymm1
  DB  196,193,100,194,253,1               ; vcmpltps      %ymm13,%ymm3,%ymm7
  DB  197,172,88,243                      ; vaddps        %ymm3,%ymm10,%ymm6
  DB  196,227,117,74,206,112              ; vblendvps     %ymm7,%ymm6,%ymm1,%ymm1
  DB  197,252,92,241                      ; vsubps        %ymm1,%ymm0,%ymm6
  DB  197,140,89,246                      ; vmulps        %ymm6,%ymm14,%ymm6
  DB  197,164,88,246                      ; vaddps        %ymm6,%ymm11,%ymm6
  DB  197,244,194,248,1                   ; vcmpltps      %ymm0,%ymm1,%ymm7
  DB  196,227,37,74,246,112               ; vblendvps     %ymm7,%ymm6,%ymm11,%ymm6
  DB  196,193,116,194,248,1               ; vcmpltps      %ymm8,%ymm1,%ymm7
  DB  196,195,77,74,241,112               ; vblendvps     %ymm7,%ymm9,%ymm6,%ymm6
  DB  197,244,194,202,1                   ; vcmpltps      %ymm2,%ymm1,%ymm1
  DB  197,140,89,251                      ; vmulps        %ymm3,%ymm14,%ymm7
  DB  197,164,88,255                      ; vaddps        %ymm7,%ymm11,%ymm7
  DB  196,227,77,74,207,16                ; vblendvps     %ymm1,%ymm7,%ymm6,%ymm1
  DB  196,193,100,92,220                  ; vsubps        %ymm12,%ymm3,%ymm3
  DB  197,172,194,243,1                   ; vcmpltps      %ymm3,%ymm10,%ymm6
  DB  196,193,100,92,250                  ; vsubps        %ymm10,%ymm3,%ymm7
  DB  196,227,101,74,247,96               ; vblendvps     %ymm6,%ymm7,%ymm3,%ymm6
  DB  196,193,100,194,253,1               ; vcmpltps      %ymm13,%ymm3,%ymm7
  DB  197,44,88,211                       ; vaddps        %ymm3,%ymm10,%ymm10
  DB  196,195,77,74,242,112               ; vblendvps     %ymm7,%ymm10,%ymm6,%ymm6
  DB  197,204,194,248,1                   ; vcmpltps      %ymm0,%ymm6,%ymm7
  DB  197,252,92,198                      ; vsubps        %ymm6,%ymm0,%ymm0
  DB  197,140,89,192                      ; vmulps        %ymm0,%ymm14,%ymm0
  DB  197,164,88,192                      ; vaddps        %ymm0,%ymm11,%ymm0
  DB  196,227,37,74,192,112               ; vblendvps     %ymm7,%ymm0,%ymm11,%ymm0
  DB  196,193,76,194,248,1                ; vcmpltps      %ymm8,%ymm6,%ymm7
  DB  196,195,125,74,193,112              ; vblendvps     %ymm7,%ymm9,%ymm0,%ymm0
  DB  197,204,194,210,1                   ; vcmpltps      %ymm2,%ymm6,%ymm2
  DB  196,193,100,89,222                  ; vmulps        %ymm14,%ymm3,%ymm3
  DB  197,164,88,219                      ; vaddps        %ymm3,%ymm11,%ymm3
  DB  196,227,125,74,211,32               ; vblendvps     %ymm2,%ymm3,%ymm0,%ymm2
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,212,194,216,0                   ; vcmpeqps      %ymm0,%ymm5,%ymm3
  DB  196,227,5,74,196,48                 ; vblendvps     %ymm3,%ymm4,%ymm15,%ymm0
  DB  196,227,117,74,204,48               ; vblendvps     %ymm3,%ymm4,%ymm1,%ymm1
  DB  196,227,109,74,212,48               ; vblendvps     %ymm3,%ymm4,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,16,28,36                    ; vmovups       (%rsp),%ymm3
  DB  197,252,16,100,36,32                ; vmovups       0x20(%rsp),%ymm4
  DB  197,252,16,108,36,64                ; vmovups       0x40(%rsp),%ymm5
  DB  197,252,16,116,36,96                ; vmovups       0x60(%rsp),%ymm6
  DB  197,252,16,188,36,128,0,0,0         ; vmovups       0x80(%rsp),%ymm7
  DB  72,129,196,184,0,0,0                ; add           $0xb8,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_1_float_avx
_sk_scale_1_float_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_u8_avx
_sk_scale_u8_avx LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,80                              ; jne           14ff <_sk_scale_u8_avx+0x60>
  DB  197,122,126,0                       ; vmovq         (%rax),%xmm8
  DB  196,66,121,49,200                   ; vpmovzxbd     %xmm8,%xmm9
  DB  196,67,121,4,192,229                ; vpermilps     $0xe5,%xmm8,%xmm8
  DB  196,66,121,49,192                   ; vpmovzxbd     %xmm8,%xmm8
  DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           1507 <_sk_scale_u8_avx+0x68>
  DB  196,65,249,110,193                  ; vmovq         %r9,%xmm8
  DB  235,143                             ; jmp           14b3 <_sk_scale_u8_avx+0x14>

PUBLIC _sk_lerp_1_float_avx
_sk_lerp_1_float_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
  DB  197,244,92,205                      ; vsubps        %ymm5,%ymm1,%ymm1
  DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
  DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
  DB  197,236,92,214                      ; vsubps        %ymm6,%ymm2,%ymm2
  DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
  DB  197,236,88,214                      ; vaddps        %ymm6,%ymm2,%ymm2
  DB  197,228,92,223                      ; vsubps        %ymm7,%ymm3,%ymm3
  DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
  DB  197,228,88,223                      ; vaddps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_u8_avx
_sk_lerp_u8_avx LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,116                             ; jne           15e7 <_sk_lerp_u8_avx+0x84>
  DB  197,122,126,0                       ; vmovq         (%rax),%xmm8
  DB  196,66,121,49,200                   ; vpmovzxbd     %xmm8,%xmm9
  DB  196,67,121,4,192,229                ; vpermilps     $0xe5,%xmm8,%xmm8
  DB  196,66,121,49,192                   ; vpmovzxbd     %xmm8,%xmm8
  DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,121,110,200                     ; vmovd         %eax,%xmm9
  DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
  DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
  DB  197,244,92,205                      ; vsubps        %ymm5,%ymm1,%ymm1
  DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
  DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
  DB  197,236,92,214                      ; vsubps        %ymm6,%ymm2,%ymm2
  DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
  DB  197,236,88,214                      ; vaddps        %ymm6,%ymm2,%ymm2
  DB  197,228,92,223                      ; vsubps        %ymm7,%ymm3,%ymm3
  DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
  DB  197,228,88,223                      ; vaddps        %ymm7,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           15ef <_sk_lerp_u8_avx+0x8c>
  DB  196,65,249,110,193                  ; vmovq         %r9,%xmm8
  DB  233,104,255,255,255                 ; jmpq          1577 <_sk_lerp_u8_avx+0x14>

PUBLIC _sk_lerp_565_avx
_sk_lerp_565_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,250,0,0,0                    ; jne           1717 <_sk_lerp_565_avx+0x108>
  DB  196,65,122,111,4,122                ; vmovdqu       (%r10,%rdi,2),%xmm8
  DB  197,225,239,219                     ; vpxor         %xmm3,%xmm3,%xmm3
  DB  197,185,105,219                     ; vpunpckhwd    %xmm3,%xmm8,%xmm3
  DB  196,66,121,51,192                   ; vpmovzxwd     %xmm8,%xmm8
  DB  196,99,61,24,195,1                  ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm8
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
  DB  197,124,91,203                      ; vcvtdq2ps     %ymm3,%ymm9
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,52,89,203                       ; vmulps        %ymm3,%ymm9,%ymm9
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
  DB  197,124,91,211                      ; vcvtdq2ps     %ymm3,%ymm10
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,44,89,211                       ; vmulps        %ymm3,%ymm10,%ymm10
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
  DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
  DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
  DB  197,244,92,205                      ; vsubps        %ymm5,%ymm1,%ymm1
  DB  196,193,116,89,202                  ; vmulps        %ymm10,%ymm1,%ymm1
  DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
  DB  197,236,92,214                      ; vsubps        %ymm6,%ymm2,%ymm2
  DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
  DB  197,236,88,214                      ; vaddps        %ymm6,%ymm2,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  15,135,243,254,255,255              ; ja            1623 <_sk_lerp_565_avx+0x14>
  DB  69,15,182,192                       ; movzbl        %r8b,%r8d
  DB  76,141,13,73,0,0,0                  ; lea           0x49(%rip),%r9        # 1784 <_sk_lerp_565_avx+0x175>
  DB  75,99,4,129                         ; movslq        (%r9,%r8,4),%rax
  DB  76,1,200                            ; add           %r9,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,225,239,219                     ; vpxor         %xmm3,%xmm3,%xmm3
  DB  196,65,97,196,68,122,12,6           ; vpinsrw       $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm8
  DB  196,65,57,196,68,122,10,5           ; vpinsrw       $0x5,0xa(%r10,%rdi,2),%xmm8,%xmm8
  DB  196,65,57,196,68,122,8,4            ; vpinsrw       $0x4,0x8(%r10,%rdi,2),%xmm8,%xmm8
  DB  196,65,57,196,68,122,6,3            ; vpinsrw       $0x3,0x6(%r10,%rdi,2),%xmm8,%xmm8
  DB  196,65,57,196,68,122,4,2            ; vpinsrw       $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8
  DB  196,65,57,196,68,122,2,1            ; vpinsrw       $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8
  DB  196,65,57,196,4,122,0               ; vpinsrw       $0x0,(%r10,%rdi,2),%xmm8,%xmm8
  DB  233,159,254,255,255                 ; jmpq          1623 <_sk_lerp_565_avx+0x14>
  DB  244                                 ; hlt
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  236                                 ; in            (%dx),%al
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,228                             ; jmpq          *%rsp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  220,255                             ; fdivr         %st,%st(7)
  DB  255                                 ; (bad)
  DB  255,212                             ; callq         *%rsp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,204                             ; dec           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,192                             ; inc           %eax
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_tables_avx
_sk_load_tables_avx LABEL PROC
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,56,2,0,0                     ; jne           19f0 <_sk_load_tables_avx+0x250>
  DB  196,65,124,16,4,184                 ; vmovups       (%r8,%rdi,4),%ymm8
  DB  187,255,0,0,0                       ; mov           $0xff,%ebx
  DB  197,249,110,195                     ; vmovd         %ebx,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,99,125,24,200,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm9
  DB  196,193,52,84,192                   ; vandps        %ymm8,%ymm9,%ymm0
  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
  DB  69,137,214                          ; mov           %r10d,%r14d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  196,227,125,25,192,1                ; vextractf128  $0x1,%ymm0,%xmm0
  DB  196,193,249,126,196                 ; vmovq         %xmm0,%r12
  DB  69,137,231                          ; mov           %r12d,%r15d
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,221                          ; mov           %ebx,%r13d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  72,139,104,8                        ; mov           0x8(%rax),%rbp
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  196,161,122,16,68,189,0             ; vmovss        0x0(%rbp,%r15,4),%xmm0
  DB  196,163,121,33,68,165,0,16          ; vinsertps     $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
  DB  196,161,122,16,76,173,0             ; vmovss        0x0(%rbp,%r13,4),%xmm1
  DB  196,227,121,33,193,32               ; vinsertps     $0x20,%xmm1,%xmm0,%xmm0
  DB  197,250,16,76,157,0                 ; vmovss        0x0(%rbp,%rbx,4),%xmm1
  DB  196,227,121,33,193,48               ; vinsertps     $0x30,%xmm1,%xmm0,%xmm0
  DB  196,161,122,16,76,157,0             ; vmovss        0x0(%rbp,%r11,4),%xmm1
  DB  196,163,113,33,76,141,0,16          ; vinsertps     $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
  DB  196,161,122,16,92,181,0             ; vmovss        0x0(%rbp,%r14,4),%xmm3
  DB  196,227,113,33,203,32               ; vinsertps     $0x20,%xmm3,%xmm1,%xmm1
  DB  196,161,122,16,92,149,0             ; vmovss        0x0(%rbp,%r10,4),%xmm3
  DB  196,227,113,33,203,48               ; vinsertps     $0x30,%xmm3,%xmm1,%xmm1
  DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
  DB  196,193,113,114,208,8               ; vpsrld        $0x8,%xmm8,%xmm1
  DB  196,67,125,25,194,1                 ; vextractf128  $0x1,%ymm8,%xmm10
  DB  196,193,105,114,210,8               ; vpsrld        $0x8,%xmm10,%xmm2
  DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
  DB  197,180,84,201                      ; vandps        %ymm1,%ymm9,%ymm1
  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
  DB  69,137,214                          ; mov           %r10d,%r14d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  196,227,125,25,201,1                ; vextractf128  $0x1,%ymm1,%xmm1
  DB  196,225,249,126,205                 ; vmovq         %xmm1,%rbp
  DB  65,137,239                          ; mov           %ebp,%r15d
  DB  196,227,249,22,203,1                ; vpextrq       $0x1,%xmm1,%rbx
  DB  65,137,220                          ; mov           %ebx,%r12d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,129,122,16,12,184               ; vmovss        (%r8,%r15,4),%xmm1
  DB  196,195,113,33,12,168,16            ; vinsertps     $0x10,(%r8,%rbp,4),%xmm1,%xmm1
  DB  196,129,122,16,20,160               ; vmovss        (%r8,%r12,4),%xmm2
  DB  196,227,113,33,202,32               ; vinsertps     $0x20,%xmm2,%xmm1,%xmm1
  DB  196,193,122,16,20,152               ; vmovss        (%r8,%rbx,4),%xmm2
  DB  196,227,113,33,202,48               ; vinsertps     $0x30,%xmm2,%xmm1,%xmm1
  DB  196,129,122,16,20,152               ; vmovss        (%r8,%r11,4),%xmm2
  DB  196,131,105,33,20,136,16            ; vinsertps     $0x10,(%r8,%r9,4),%xmm2,%xmm2
  DB  196,129,122,16,28,176               ; vmovss        (%r8,%r14,4),%xmm3
  DB  196,227,105,33,211,32               ; vinsertps     $0x20,%xmm3,%xmm2,%xmm2
  DB  196,129,122,16,28,144               ; vmovss        (%r8,%r10,4),%xmm3
  DB  196,227,105,33,211,48               ; vinsertps     $0x30,%xmm3,%xmm2,%xmm2
  DB  196,227,109,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm2,%ymm1
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  196,193,105,114,208,16              ; vpsrld        $0x10,%xmm8,%xmm2
  DB  196,193,97,114,210,16               ; vpsrld        $0x10,%xmm10,%xmm3
  DB  196,227,109,24,211,1                ; vinsertf128   $0x1,%xmm3,%ymm2,%ymm2
  DB  197,180,84,210                      ; vandps        %ymm2,%ymm9,%ymm2
  DB  196,193,249,126,208                 ; vmovq         %xmm2,%r8
  DB  69,137,194                          ; mov           %r8d,%r10d
  DB  196,195,249,22,209,1                ; vpextrq       $0x1,%xmm2,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  73,193,232,32                       ; shr           $0x20,%r8
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,225,249,126,213                 ; vmovq         %xmm2,%rbp
  DB  65,137,238                          ; mov           %ebp,%r14d
  DB  196,227,249,22,211,1                ; vpextrq       $0x1,%xmm2,%rbx
  DB  65,137,223                          ; mov           %ebx,%r15d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,161,122,16,20,176               ; vmovss        (%rax,%r14,4),%xmm2
  DB  196,227,105,33,20,168,16            ; vinsertps     $0x10,(%rax,%rbp,4),%xmm2,%xmm2
  DB  196,161,122,16,28,184               ; vmovss        (%rax,%r15,4),%xmm3
  DB  196,227,105,33,211,32               ; vinsertps     $0x20,%xmm3,%xmm2,%xmm2
  DB  197,250,16,28,152                   ; vmovss        (%rax,%rbx,4),%xmm3
  DB  196,99,105,33,203,48                ; vinsertps     $0x30,%xmm3,%xmm2,%xmm9
  DB  196,161,122,16,28,144               ; vmovss        (%rax,%r10,4),%xmm3
  DB  196,163,97,33,28,128,16             ; vinsertps     $0x10,(%rax,%r8,4),%xmm3,%xmm3
  DB  196,161,122,16,20,152               ; vmovss        (%rax,%r11,4),%xmm2
  DB  196,227,97,33,210,32                ; vinsertps     $0x20,%xmm2,%xmm3,%xmm2
  DB  196,161,122,16,28,136               ; vmovss        (%rax,%r9,4),%xmm3
  DB  196,227,105,33,211,48               ; vinsertps     $0x30,%xmm3,%xmm2,%xmm2
  DB  196,195,109,24,209,1                ; vinsertf128   $0x1,%xmm9,%ymm2,%ymm2
  DB  196,193,57,114,208,24               ; vpsrld        $0x18,%xmm8,%xmm8
  DB  196,193,97,114,210,24               ; vpsrld        $0x18,%xmm10,%xmm3
  DB  196,227,61,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax
  DB  137,203                             ; mov           %ecx,%ebx
  DB  128,227,7                           ; and           $0x7,%bl
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  254,203                             ; dec           %bl
  DB  128,251,6                           ; cmp           $0x6,%bl
  DB  15,135,185,253,255,255              ; ja            17be <_sk_load_tables_avx+0x1e>
  DB  15,182,219                          ; movzbl        %bl,%ebx
  DB  76,141,13,137,0,0,0                 ; lea           0x89(%rip),%r9        # 1a98 <_sk_load_tables_avx+0x2f8>
  DB  73,99,28,153                        ; movslq        (%r9,%rbx,4),%rbx
  DB  76,1,203                            ; add           %r9,%rbx
  DB  255,227                             ; jmpq          *%rbx
  DB  196,193,121,110,68,184,24           ; vmovd         0x18(%r8,%rdi,4),%xmm0
  DB  197,249,112,192,68                  ; vpshufd       $0x44,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  196,99,117,12,192,64                ; vblendps      $0x40,%ymm0,%ymm1,%ymm8
  DB  196,99,125,25,192,1                 ; vextractf128  $0x1,%ymm8,%xmm0
  DB  196,195,121,34,68,184,20,1          ; vpinsrd       $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0
  DB  196,99,61,24,192,1                  ; vinsertf128   $0x1,%xmm0,%ymm8,%ymm8
  DB  196,99,125,25,192,1                 ; vextractf128  $0x1,%ymm8,%xmm0
  DB  196,195,121,34,68,184,16,0          ; vpinsrd       $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0
  DB  196,99,61,24,192,1                  ; vinsertf128   $0x1,%xmm0,%ymm8,%ymm8
  DB  196,195,57,34,68,184,12,3           ; vpinsrd       $0x3,0xc(%r8,%rdi,4),%xmm8,%xmm0
  DB  196,99,61,12,192,15                 ; vblendps      $0xf,%ymm0,%ymm8,%ymm8
  DB  196,195,57,34,68,184,8,2            ; vpinsrd       $0x2,0x8(%r8,%rdi,4),%xmm8,%xmm0
  DB  196,99,61,12,192,15                 ; vblendps      $0xf,%ymm0,%ymm8,%ymm8
  DB  196,195,57,34,68,184,4,1            ; vpinsrd       $0x1,0x4(%r8,%rdi,4),%xmm8,%xmm0
  DB  196,99,61,12,192,15                 ; vblendps      $0xf,%ymm0,%ymm8,%ymm8
  DB  196,195,57,34,4,184,0               ; vpinsrd       $0x0,(%r8,%rdi,4),%xmm8,%xmm0
  DB  196,99,61,12,192,15                 ; vblendps      $0xf,%ymm0,%ymm8,%ymm8
  DB  233,38,253,255,255                  ; jmpq          17be <_sk_load_tables_avx+0x1e>
  DB  238                                 ; out           %al,(%dx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,224                             ; jmpq          *%rax
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,210                             ; callq         *%rdx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,196                             ; inc           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,176,255,255,255,156             ; pushq         -0x63000001(%rax)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff
  DB  128,255,255                         ; cmp           $0xff,%bh
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_tables_u16_be_avx
_sk_load_tables_u16_be_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,165,2,0,0                    ; jne           1d6f <_sk_load_tables_u16_be_avx+0x2bb>
  DB  196,1,121,16,4,72                   ; vmovupd       (%r8,%r9,2),%xmm8
  DB  196,129,121,16,84,72,16             ; vmovupd       0x10(%r8,%r9,2),%xmm2
  DB  196,129,121,16,92,72,32             ; vmovupd       0x20(%r8,%r9,2),%xmm3
  DB  196,1,122,111,76,72,48              ; vmovdqu       0x30(%r8,%r9,2),%xmm9
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  197,185,97,194                      ; vpunpcklwd    %xmm2,%xmm8,%xmm0
  DB  197,185,105,210                     ; vpunpckhwd    %xmm2,%xmm8,%xmm2
  DB  196,193,97,97,201                   ; vpunpcklwd    %xmm9,%xmm3,%xmm1
  DB  196,193,97,105,217                  ; vpunpckhwd    %xmm9,%xmm3,%xmm3
  DB  197,121,97,218                      ; vpunpcklwd    %xmm2,%xmm0,%xmm11
  DB  197,121,105,202                     ; vpunpckhwd    %xmm2,%xmm0,%xmm9
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,113,105,211                     ; vpunpckhwd    %xmm3,%xmm1,%xmm10
  DB  197,161,108,194                     ; vpunpcklqdq   %xmm2,%xmm11,%xmm0
  DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
  DB  196,193,121,105,200                 ; vpunpckhwd    %xmm8,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  187,255,0,0,0                       ; mov           $0xff,%ebx
  DB  197,249,110,203                     ; vmovd         %ebx,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,99,117,24,225,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm12
  DB  196,193,124,84,196                  ; vandps        %ymm12,%ymm0,%ymm0
  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
  DB  69,137,214                          ; mov           %r10d,%r14d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  196,227,125,25,192,1                ; vextractf128  $0x1,%ymm0,%xmm0
  DB  196,193,249,126,196                 ; vmovq         %xmm0,%r12
  DB  69,137,231                          ; mov           %r12d,%r15d
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,221                          ; mov           %ebx,%r13d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  72,139,104,8                        ; mov           0x8(%rax),%rbp
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  196,161,122,16,68,189,0             ; vmovss        0x0(%rbp,%r15,4),%xmm0
  DB  196,163,121,33,68,165,0,16          ; vinsertps     $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
  DB  196,161,122,16,76,173,0             ; vmovss        0x0(%rbp,%r13,4),%xmm1
  DB  196,227,121,33,193,32               ; vinsertps     $0x20,%xmm1,%xmm0,%xmm0
  DB  197,250,16,76,157,0                 ; vmovss        0x0(%rbp,%rbx,4),%xmm1
  DB  196,227,121,33,193,48               ; vinsertps     $0x30,%xmm1,%xmm0,%xmm0
  DB  196,161,122,16,76,157,0             ; vmovss        0x0(%rbp,%r11,4),%xmm1
  DB  196,163,113,33,76,141,0,16          ; vinsertps     $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
  DB  196,161,122,16,92,181,0             ; vmovss        0x0(%rbp,%r14,4),%xmm3
  DB  196,227,113,33,203,32               ; vinsertps     $0x20,%xmm3,%xmm1,%xmm1
  DB  196,161,122,16,92,149,0             ; vmovss        0x0(%rbp,%r10,4),%xmm3
  DB  196,227,113,33,203,48               ; vinsertps     $0x30,%xmm3,%xmm1,%xmm1
  DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
  DB  197,161,109,202                     ; vpunpckhqdq   %xmm2,%xmm11,%xmm1
  DB  196,193,113,105,208                 ; vpunpckhwd    %xmm8,%xmm1,%xmm2
  DB  196,226,121,51,201                  ; vpmovzxwd     %xmm1,%xmm1
  DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
  DB  196,193,116,84,204                  ; vandps        %ymm12,%ymm1,%ymm1
  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
  DB  69,137,214                          ; mov           %r10d,%r14d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  196,227,125,25,201,1                ; vextractf128  $0x1,%ymm1,%xmm1
  DB  196,225,249,126,205                 ; vmovq         %xmm1,%rbp
  DB  65,137,239                          ; mov           %ebp,%r15d
  DB  196,227,249,22,203,1                ; vpextrq       $0x1,%xmm1,%rbx
  DB  65,137,220                          ; mov           %ebx,%r12d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,129,122,16,12,184               ; vmovss        (%r8,%r15,4),%xmm1
  DB  196,195,113,33,12,168,16            ; vinsertps     $0x10,(%r8,%rbp,4),%xmm1,%xmm1
  DB  196,129,122,16,20,160               ; vmovss        (%r8,%r12,4),%xmm2
  DB  196,227,113,33,202,32               ; vinsertps     $0x20,%xmm2,%xmm1,%xmm1
  DB  196,193,122,16,20,152               ; vmovss        (%r8,%rbx,4),%xmm2
  DB  196,227,113,33,202,48               ; vinsertps     $0x30,%xmm2,%xmm1,%xmm1
  DB  196,129,122,16,20,152               ; vmovss        (%r8,%r11,4),%xmm2
  DB  196,131,105,33,20,136,16            ; vinsertps     $0x10,(%r8,%r9,4),%xmm2,%xmm2
  DB  196,129,122,16,28,176               ; vmovss        (%r8,%r14,4),%xmm3
  DB  196,227,105,33,211,32               ; vinsertps     $0x20,%xmm3,%xmm2,%xmm2
  DB  196,129,122,16,28,144               ; vmovss        (%r8,%r10,4),%xmm3
  DB  196,227,105,33,211,48               ; vinsertps     $0x30,%xmm3,%xmm2,%xmm2
  DB  196,227,109,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm2,%ymm1
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  196,193,49,108,210                  ; vpunpcklqdq   %xmm10,%xmm9,%xmm2
  DB  196,193,105,105,216                 ; vpunpckhwd    %xmm8,%xmm2,%xmm3
  DB  196,226,121,51,210                  ; vpmovzxwd     %xmm2,%xmm2
  DB  196,227,109,24,211,1                ; vinsertf128   $0x1,%xmm3,%ymm2,%ymm2
  DB  196,193,108,84,212                  ; vandps        %ymm12,%ymm2,%ymm2
  DB  196,193,249,126,208                 ; vmovq         %xmm2,%r8
  DB  69,137,194                          ; mov           %r8d,%r10d
  DB  196,195,249,22,209,1                ; vpextrq       $0x1,%xmm2,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  73,193,232,32                       ; shr           $0x20,%r8
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,225,249,126,213                 ; vmovq         %xmm2,%rbp
  DB  65,137,238                          ; mov           %ebp,%r14d
  DB  196,227,249,22,211,1                ; vpextrq       $0x1,%xmm2,%rbx
  DB  65,137,223                          ; mov           %ebx,%r15d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,161,122,16,20,176               ; vmovss        (%rax,%r14,4),%xmm2
  DB  196,227,105,33,20,168,16            ; vinsertps     $0x10,(%rax,%rbp,4),%xmm2,%xmm2
  DB  196,161,122,16,28,184               ; vmovss        (%rax,%r15,4),%xmm3
  DB  196,227,105,33,211,32               ; vinsertps     $0x20,%xmm3,%xmm2,%xmm2
  DB  197,250,16,28,152                   ; vmovss        (%rax,%rbx,4),%xmm3
  DB  196,99,105,33,219,48                ; vinsertps     $0x30,%xmm3,%xmm2,%xmm11
  DB  196,161,122,16,28,144               ; vmovss        (%rax,%r10,4),%xmm3
  DB  196,163,97,33,28,128,16             ; vinsertps     $0x10,(%rax,%r8,4),%xmm3,%xmm3
  DB  196,161,122,16,20,152               ; vmovss        (%rax,%r11,4),%xmm2
  DB  196,227,97,33,210,32                ; vinsertps     $0x20,%xmm2,%xmm3,%xmm2
  DB  196,161,122,16,28,136               ; vmovss        (%rax,%r9,4),%xmm3
  DB  196,227,105,33,211,48               ; vinsertps     $0x30,%xmm3,%xmm2,%xmm2
  DB  196,195,109,24,211,1                ; vinsertf128   $0x1,%xmm11,%ymm2,%ymm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,99,101,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm11
  DB  196,193,49,109,218                  ; vpunpckhqdq   %xmm10,%xmm9,%xmm3
  DB  197,177,113,243,8                   ; vpsllw        $0x8,%xmm3,%xmm9
  DB  197,225,113,211,8                   ; vpsrlw        $0x8,%xmm3,%xmm3
  DB  197,177,235,219                     ; vpor          %xmm3,%xmm9,%xmm3
  DB  196,65,97,105,192                   ; vpunpckhwd    %xmm8,%xmm3,%xmm8
  DB  196,226,121,51,219                  ; vpmovzxwd     %xmm3,%xmm3
  DB  196,195,101,24,216,1                ; vinsertf128   $0x1,%xmm8,%ymm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,219                  ; vmulps        %ymm11,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax
  DB  196,1,123,16,4,72                   ; vmovsd        (%r8,%r9,2),%xmm8
  DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,85                              ; je            1dd5 <_sk_load_tables_u16_be_avx+0x321>
  DB  196,1,57,22,68,72,8                 ; vmovhpd       0x8(%r8,%r9,2),%xmm8,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,72                              ; jb            1dd5 <_sk_load_tables_u16_be_avx+0x321>
  DB  196,129,123,16,84,72,16             ; vmovsd        0x10(%r8,%r9,2),%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  116,72                              ; je            1de2 <_sk_load_tables_u16_be_avx+0x32e>
  DB  196,129,105,22,84,72,24             ; vmovhpd       0x18(%r8,%r9,2),%xmm2,%xmm2
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,59                              ; jb            1de2 <_sk_load_tables_u16_be_avx+0x32e>
  DB  196,129,123,16,92,72,32             ; vmovsd        0x20(%r8,%r9,2),%xmm3
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  15,132,45,253,255,255               ; je            1ae5 <_sk_load_tables_u16_be_avx+0x31>
  DB  196,129,97,22,92,72,40              ; vmovhpd       0x28(%r8,%r9,2),%xmm3,%xmm3
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  15,130,28,253,255,255               ; jb            1ae5 <_sk_load_tables_u16_be_avx+0x31>
  DB  196,1,122,126,76,72,48              ; vmovq         0x30(%r8,%r9,2),%xmm9
  DB  233,16,253,255,255                  ; jmpq          1ae5 <_sk_load_tables_u16_be_avx+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  197,233,87,210                      ; vxorpd        %xmm2,%xmm2,%xmm2
  DB  233,3,253,255,255                   ; jmpq          1ae5 <_sk_load_tables_u16_be_avx+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  233,250,252,255,255                 ; jmpq          1ae5 <_sk_load_tables_u16_be_avx+0x31>

PUBLIC _sk_load_tables_rgb_u16_be_avx
_sk_load_tables_rgb_u16_be_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,127                       ; lea           (%rdi,%rdi,2),%r9
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,146,2,0,0                    ; jne           208f <_sk_load_tables_rgb_u16_be_avx+0x2a4>
  DB  196,129,122,111,4,72                ; vmovdqu       (%r8,%r9,2),%xmm0
  DB  196,129,122,111,84,72,12            ; vmovdqu       0xc(%r8,%r9,2),%xmm2
  DB  196,129,122,111,76,72,24            ; vmovdqu       0x18(%r8,%r9,2),%xmm1
  DB  196,129,122,111,92,72,32            ; vmovdqu       0x20(%r8,%r9,2),%xmm3
  DB  197,225,115,219,4                   ; vpsrldq       $0x4,%xmm3,%xmm3
  DB  197,185,115,216,6                   ; vpsrldq       $0x6,%xmm0,%xmm8
  DB  197,177,115,218,6                   ; vpsrldq       $0x6,%xmm2,%xmm9
  DB  197,161,115,217,6                   ; vpsrldq       $0x6,%xmm1,%xmm11
  DB  197,169,115,219,6                   ; vpsrldq       $0x6,%xmm3,%xmm10
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  197,249,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm0
  DB  196,193,57,97,209                   ; vpunpcklwd    %xmm9,%xmm8,%xmm2
  DB  197,241,97,203                      ; vpunpcklwd    %xmm3,%xmm1,%xmm1
  DB  196,193,33,97,218                   ; vpunpcklwd    %xmm10,%xmm11,%xmm3
  DB  197,121,97,218                      ; vpunpcklwd    %xmm2,%xmm0,%xmm11
  DB  197,121,105,194                     ; vpunpckhwd    %xmm2,%xmm0,%xmm8
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,113,105,211                     ; vpunpckhwd    %xmm3,%xmm1,%xmm10
  DB  197,161,108,194                     ; vpunpcklqdq   %xmm2,%xmm11,%xmm0
  DB  196,65,25,239,228                   ; vpxor         %xmm12,%xmm12,%xmm12
  DB  196,193,121,105,204                 ; vpunpckhwd    %xmm12,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  187,255,0,0,0                       ; mov           $0xff,%ebx
  DB  197,249,110,203                     ; vmovd         %ebx,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,99,117,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm9
  DB  196,193,124,84,193                  ; vandps        %ymm9,%ymm0,%ymm0
  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
  DB  69,137,214                          ; mov           %r10d,%r14d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  196,227,125,25,192,1                ; vextractf128  $0x1,%ymm0,%xmm0
  DB  196,193,249,126,196                 ; vmovq         %xmm0,%r12
  DB  69,137,231                          ; mov           %r12d,%r15d
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,221                          ; mov           %ebx,%r13d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  72,139,104,8                        ; mov           0x8(%rax),%rbp
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  196,161,122,16,68,189,0             ; vmovss        0x0(%rbp,%r15,4),%xmm0
  DB  196,163,121,33,68,165,0,16          ; vinsertps     $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
  DB  196,161,122,16,76,173,0             ; vmovss        0x0(%rbp,%r13,4),%xmm1
  DB  196,227,121,33,193,32               ; vinsertps     $0x20,%xmm1,%xmm0,%xmm0
  DB  197,250,16,76,157,0                 ; vmovss        0x0(%rbp,%rbx,4),%xmm1
  DB  196,227,121,33,193,48               ; vinsertps     $0x30,%xmm1,%xmm0,%xmm0
  DB  196,161,122,16,76,157,0             ; vmovss        0x0(%rbp,%r11,4),%xmm1
  DB  196,163,113,33,76,141,0,16          ; vinsertps     $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
  DB  196,161,122,16,92,181,0             ; vmovss        0x0(%rbp,%r14,4),%xmm3
  DB  196,227,113,33,203,32               ; vinsertps     $0x20,%xmm3,%xmm1,%xmm1
  DB  196,161,122,16,92,149,0             ; vmovss        0x0(%rbp,%r10,4),%xmm3
  DB  196,227,113,33,203,48               ; vinsertps     $0x30,%xmm3,%xmm1,%xmm1
  DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
  DB  197,161,109,202                     ; vpunpckhqdq   %xmm2,%xmm11,%xmm1
  DB  196,193,113,105,212                 ; vpunpckhwd    %xmm12,%xmm1,%xmm2
  DB  196,226,121,51,201                  ; vpmovzxwd     %xmm1,%xmm1
  DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
  DB  196,193,116,84,201                  ; vandps        %ymm9,%ymm1,%ymm1
  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
  DB  69,137,214                          ; mov           %r10d,%r14d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  196,227,125,25,201,1                ; vextractf128  $0x1,%ymm1,%xmm1
  DB  196,225,249,126,205                 ; vmovq         %xmm1,%rbp
  DB  65,137,239                          ; mov           %ebp,%r15d
  DB  196,227,249,22,203,1                ; vpextrq       $0x1,%xmm1,%rbx
  DB  65,137,220                          ; mov           %ebx,%r12d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,129,122,16,12,184               ; vmovss        (%r8,%r15,4),%xmm1
  DB  196,195,113,33,12,168,16            ; vinsertps     $0x10,(%r8,%rbp,4),%xmm1,%xmm1
  DB  196,129,122,16,20,160               ; vmovss        (%r8,%r12,4),%xmm2
  DB  196,227,113,33,202,32               ; vinsertps     $0x20,%xmm2,%xmm1,%xmm1
  DB  196,193,122,16,20,152               ; vmovss        (%r8,%rbx,4),%xmm2
  DB  196,227,113,33,202,48               ; vinsertps     $0x30,%xmm2,%xmm1,%xmm1
  DB  196,129,122,16,20,152               ; vmovss        (%r8,%r11,4),%xmm2
  DB  196,131,105,33,20,136,16            ; vinsertps     $0x10,(%r8,%r9,4),%xmm2,%xmm2
  DB  196,129,122,16,28,176               ; vmovss        (%r8,%r14,4),%xmm3
  DB  196,227,105,33,211,32               ; vinsertps     $0x20,%xmm3,%xmm2,%xmm2
  DB  196,129,122,16,28,144               ; vmovss        (%r8,%r10,4),%xmm3
  DB  196,227,105,33,211,48               ; vinsertps     $0x30,%xmm3,%xmm2,%xmm2
  DB  196,227,109,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm2,%ymm1
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  196,193,57,108,210                  ; vpunpcklqdq   %xmm10,%xmm8,%xmm2
  DB  196,193,105,105,220                 ; vpunpckhwd    %xmm12,%xmm2,%xmm3
  DB  196,226,121,51,210                  ; vpmovzxwd     %xmm2,%xmm2
  DB  196,227,109,24,211,1                ; vinsertf128   $0x1,%xmm3,%ymm2,%ymm2
  DB  196,193,108,84,209                  ; vandps        %ymm9,%ymm2,%ymm2
  DB  196,193,249,126,208                 ; vmovq         %xmm2,%r8
  DB  69,137,194                          ; mov           %r8d,%r10d
  DB  196,195,249,22,209,1                ; vpextrq       $0x1,%xmm2,%r9
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  73,193,232,32                       ; shr           $0x20,%r8
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,225,249,126,213                 ; vmovq         %xmm2,%rbp
  DB  65,137,238                          ; mov           %ebp,%r14d
  DB  196,227,249,22,211,1                ; vpextrq       $0x1,%xmm2,%rbx
  DB  65,137,223                          ; mov           %ebx,%r15d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,161,122,16,20,176               ; vmovss        (%rax,%r14,4),%xmm2
  DB  196,227,105,33,20,168,16            ; vinsertps     $0x10,(%rax,%rbp,4),%xmm2,%xmm2
  DB  196,161,122,16,28,184               ; vmovss        (%rax,%r15,4),%xmm3
  DB  196,227,105,33,211,32               ; vinsertps     $0x20,%xmm3,%xmm2,%xmm2
  DB  197,250,16,28,152                   ; vmovss        (%rax,%rbx,4),%xmm3
  DB  196,99,105,33,195,48                ; vinsertps     $0x30,%xmm3,%xmm2,%xmm8
  DB  196,161,122,16,28,144               ; vmovss        (%rax,%r10,4),%xmm3
  DB  196,163,97,33,28,128,16             ; vinsertps     $0x10,(%rax,%r8,4),%xmm3,%xmm3
  DB  196,161,122,16,20,152               ; vmovss        (%rax,%r11,4),%xmm2
  DB  196,227,97,33,210,32                ; vinsertps     $0x20,%xmm2,%xmm3,%xmm2
  DB  196,161,122,16,28,136               ; vmovss        (%rax,%r9,4),%xmm3
  DB  196,227,105,33,211,48               ; vinsertps     $0x30,%xmm3,%xmm2,%xmm2
  DB  196,195,109,24,208,1                ; vinsertf128   $0x1,%xmm8,%ymm2,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax
  DB  196,129,121,110,4,72                ; vmovd         (%r8,%r9,2),%xmm0
  DB  196,129,121,196,68,72,4,2           ; vpinsrw       $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  117,5                               ; jne           20a8 <_sk_load_tables_rgb_u16_be_avx+0x2bd>
  DB  233,137,253,255,255                 ; jmpq          1e31 <_sk_load_tables_rgb_u16_be_avx+0x46>
  DB  196,129,121,110,76,72,6             ; vmovd         0x6(%r8,%r9,2),%xmm1
  DB  196,1,113,196,68,72,10,2            ; vpinsrw       $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,26                              ; jb            20d7 <_sk_load_tables_rgb_u16_be_avx+0x2ec>
  DB  196,129,121,110,76,72,12            ; vmovd         0xc(%r8,%r9,2),%xmm1
  DB  196,129,113,196,84,72,16,2          ; vpinsrw       $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  117,10                              ; jne           20dc <_sk_load_tables_rgb_u16_be_avx+0x2f1>
  DB  233,90,253,255,255                  ; jmpq          1e31 <_sk_load_tables_rgb_u16_be_avx+0x46>
  DB  233,85,253,255,255                  ; jmpq          1e31 <_sk_load_tables_rgb_u16_be_avx+0x46>
  DB  196,129,121,110,76,72,18            ; vmovd         0x12(%r8,%r9,2),%xmm1
  DB  196,1,113,196,76,72,22,2            ; vpinsrw       $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,26                              ; jb            210b <_sk_load_tables_rgb_u16_be_avx+0x320>
  DB  196,129,121,110,76,72,24            ; vmovd         0x18(%r8,%r9,2),%xmm1
  DB  196,129,113,196,76,72,28,2          ; vpinsrw       $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  117,10                              ; jne           2110 <_sk_load_tables_rgb_u16_be_avx+0x325>
  DB  233,38,253,255,255                  ; jmpq          1e31 <_sk_load_tables_rgb_u16_be_avx+0x46>
  DB  233,33,253,255,255                  ; jmpq          1e31 <_sk_load_tables_rgb_u16_be_avx+0x46>
  DB  196,129,121,110,92,72,30            ; vmovd         0x1e(%r8,%r9,2),%xmm3
  DB  196,1,97,196,92,72,34,2             ; vpinsrw       $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,20                              ; jb            2139 <_sk_load_tables_rgb_u16_be_avx+0x34e>
  DB  196,129,121,110,92,72,36            ; vmovd         0x24(%r8,%r9,2),%xmm3
  DB  196,129,97,196,92,72,40,2           ; vpinsrw       $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
  DB  233,248,252,255,255                 ; jmpq          1e31 <_sk_load_tables_rgb_u16_be_avx+0x46>
  DB  233,243,252,255,255                 ; jmpq          1e31 <_sk_load_tables_rgb_u16_be_avx+0x46>

PUBLIC _sk_byte_tables_avx
_sk_byte_tables_avx LABEL PROC
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,127,67                   ; mov           $0x437f0000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  68,137,197                          ; mov           %r8d,%ebp
  DB  77,137,194                          ; mov           %r8,%r10
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,192                 ; vmovq         %xmm0,%r8
  DB  69,137,195                          ; mov           %r8d,%r11d
  DB  77,137,199                          ; mov           %r8,%r15
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,227,125,25,192,1                ; vextractf128  $0x1,%ymm0,%xmm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  69,137,198                          ; mov           %r8d,%r14d
  DB  77,137,196                          ; mov           %r8,%r12
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,221                          ; mov           %ebx,%r13d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  196,131,121,32,4,25,0               ; vpinsrb       $0x0,(%r9,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,57,1               ; vpinsrb       $0x1,(%r9,%r15,1),%xmm0,%xmm0
  DB  65,15,182,44,41                     ; movzbl        (%r9,%rbp,1),%ebp
  DB  196,227,121,32,197,2                ; vpinsrb       $0x2,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,17                     ; movzbl        (%r9,%r10,1),%ebp
  DB  196,227,121,32,197,3                ; vpinsrb       $0x3,%ebp,%xmm0,%xmm0
  DB  196,98,121,49,200                   ; vpmovzxbd     %xmm0,%xmm9
  DB  196,131,121,32,4,41,0               ; vpinsrb       $0x0,(%r9,%r13,1),%xmm0,%xmm0
  DB  196,195,121,32,4,25,1               ; vpinsrb       $0x1,(%r9,%rbx,1),%xmm0,%xmm0
  DB  67,15,182,44,49                     ; movzbl        (%r9,%r14,1),%ebp
  DB  196,227,121,32,197,2                ; vpinsrb       $0x2,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,33                     ; movzbl        (%r9,%r12,1),%ebp
  DB  196,227,121,32,197,3                ; vpinsrb       $0x3,%ebp,%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  196,227,53,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm9,%ymm0
  DB  197,124,91,208                      ; vcvtdq2ps     %ymm0,%ymm10
  DB  189,129,128,128,59                  ; mov           $0x3b808081,%ebp
  DB  197,249,110,197                     ; vmovd         %ebp,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,200,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm9
  DB  196,193,44,89,193                   ; vmulps        %ymm9,%ymm10,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
  DB  196,227,249,22,205,1                ; vpextrq       $0x1,%xmm1,%rbp
  DB  65,137,233                          ; mov           %ebp,%r9d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,203                 ; vmovq         %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,25,201,1                ; vextractf128  $0x1,%ymm1,%xmm1
  DB  196,195,249,22,203,1                ; vpextrq       $0x1,%xmm1,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  196,193,249,126,207                 ; vmovq         %xmm1,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,12,16,0              ; vpinsrb       $0x0,(%r8,%r10,1),%xmm0,%xmm1
  DB  196,195,113,32,12,24,1              ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm1,%xmm1
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,113,32,203,2                ; vpinsrb       $0x2,%ebx,%xmm1,%xmm1
  DB  65,15,182,44,40                     ; movzbl        (%r8,%rbp,1),%ebp
  DB  196,227,113,32,205,3                ; vpinsrb       $0x3,%ebp,%xmm1,%xmm1
  DB  196,98,121,49,209                   ; vpmovzxbd     %xmm1,%xmm10
  DB  196,131,121,32,12,32,0              ; vpinsrb       $0x0,(%r8,%r12,1),%xmm0,%xmm1
  DB  196,131,113,32,12,56,1              ; vpinsrb       $0x1,(%r8,%r15,1),%xmm1,%xmm1
  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
  DB  196,227,113,32,205,2                ; vpinsrb       $0x2,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
  DB  196,227,113,32,205,3                ; vpinsrb       $0x3,%ebp,%xmm1,%xmm1
  DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
  DB  196,227,45,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm10,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
  DB  196,227,249,22,213,1                ; vpextrq       $0x1,%xmm2,%rbp
  DB  65,137,233                          ; mov           %ebp,%r9d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,211                 ; vmovq         %xmm2,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,195,249,22,211,1                ; vpextrq       $0x1,%xmm2,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  196,193,249,126,215                 ; vmovq         %xmm2,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,20,16,0              ; vpinsrb       $0x0,(%r8,%r10,1),%xmm0,%xmm2
  DB  196,195,105,32,20,24,1              ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm2,%xmm2
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,105,32,211,2                ; vpinsrb       $0x2,%ebx,%xmm2,%xmm2
  DB  65,15,182,44,40                     ; movzbl        (%r8,%rbp,1),%ebp
  DB  196,227,105,32,213,3                ; vpinsrb       $0x3,%ebp,%xmm2,%xmm2
  DB  196,98,121,49,210                   ; vpmovzxbd     %xmm2,%xmm10
  DB  196,131,121,32,20,32,0              ; vpinsrb       $0x0,(%r8,%r12,1),%xmm0,%xmm2
  DB  196,131,105,32,20,56,1              ; vpinsrb       $0x1,(%r8,%r15,1),%xmm2,%xmm2
  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
  DB  196,227,105,32,213,2                ; vpinsrb       $0x2,%ebp,%xmm2,%xmm2
  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
  DB  196,227,105,32,213,3                ; vpinsrb       $0x3,%ebp,%xmm2,%xmm2
  DB  196,226,121,49,210                  ; vpmovzxbd     %xmm2,%xmm2
  DB  196,227,45,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm10,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  197,253,91,219                      ; vcvtps2dq     %ymm3,%ymm3
  DB  196,227,249,22,221,1                ; vpextrq       $0x1,%xmm3,%rbp
  DB  65,137,232                          ; mov           %ebp,%r8d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,219                 ; vmovq         %xmm3,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,25,219,1                ; vextractf128  $0x1,%ymm3,%xmm3
  DB  196,195,249,22,218,1                ; vpextrq       $0x1,%xmm3,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,222                 ; vmovq         %xmm3,%r14
  DB  69,137,247                          ; mov           %r14d,%r15d
  DB  73,193,238,32                       ; shr           $0x20,%r14
  DB  196,163,121,32,28,8,0               ; vpinsrb       $0x0,(%rax,%r9,1),%xmm0,%xmm3
  DB  196,227,97,32,28,24,1               ; vpinsrb       $0x1,(%rax,%rbx,1),%xmm3,%xmm3
  DB  66,15,182,28,0                      ; movzbl        (%rax,%r8,1),%ebx
  DB  196,227,97,32,219,2                 ; vpinsrb       $0x2,%ebx,%xmm3,%xmm3
  DB  15,182,44,40                        ; movzbl        (%rax,%rbp,1),%ebp
  DB  196,227,97,32,221,3                 ; vpinsrb       $0x3,%ebp,%xmm3,%xmm3
  DB  196,98,121,49,195                   ; vpmovzxbd     %xmm3,%xmm8
  DB  196,163,121,32,28,56,0              ; vpinsrb       $0x0,(%rax,%r15,1),%xmm0,%xmm3
  DB  196,163,97,32,28,48,1               ; vpinsrb       $0x1,(%rax,%r14,1),%xmm3,%xmm3
  DB  66,15,182,44,24                     ; movzbl        (%rax,%r11,1),%ebp
  DB  196,227,97,32,221,2                 ; vpinsrb       $0x2,%ebp,%xmm3,%xmm3
  DB  66,15,182,4,16                      ; movzbl        (%rax,%r10,1),%eax
  DB  196,227,97,32,216,3                 ; vpinsrb       $0x3,%eax,%xmm3,%xmm3
  DB  196,226,121,49,219                  ; vpmovzxbd     %xmm3,%xmm3
  DB  196,227,61,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  197,180,89,219                      ; vmulps        %ymm3,%ymm9,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_byte_tables_rgb_avx
_sk_byte_tables_rgb_avx LABEL PROC
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  68,139,64,24                        ; mov           0x18(%rax),%r8d
  DB  65,255,200                          ; dec           %r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,65,121,112,192,0                ; vpshufd       $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  68,137,197                          ; mov           %r8d,%ebp
  DB  77,137,194                          ; mov           %r8,%r10
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,192                 ; vmovq         %xmm0,%r8
  DB  69,137,195                          ; mov           %r8d,%r11d
  DB  77,137,199                          ; mov           %r8,%r15
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,227,125,25,192,1                ; vextractf128  $0x1,%ymm0,%xmm0
  DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
  DB  69,137,198                          ; mov           %r8d,%r14d
  DB  77,137,196                          ; mov           %r8,%r12
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,221                          ; mov           %ebx,%r13d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  196,131,121,32,4,25,0               ; vpinsrb       $0x0,(%r9,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,57,1               ; vpinsrb       $0x1,(%r9,%r15,1),%xmm0,%xmm0
  DB  65,15,182,44,41                     ; movzbl        (%r9,%rbp,1),%ebp
  DB  196,227,121,32,197,2                ; vpinsrb       $0x2,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,17                     ; movzbl        (%r9,%r10,1),%ebp
  DB  196,227,121,32,197,3                ; vpinsrb       $0x3,%ebp,%xmm0,%xmm0
  DB  196,98,121,49,200                   ; vpmovzxbd     %xmm0,%xmm9
  DB  196,131,121,32,4,41,0               ; vpinsrb       $0x0,(%r9,%r13,1),%xmm0,%xmm0
  DB  196,195,121,32,4,25,1               ; vpinsrb       $0x1,(%r9,%rbx,1),%xmm0,%xmm0
  DB  67,15,182,44,49                     ; movzbl        (%r9,%r14,1),%ebp
  DB  196,227,121,32,197,2                ; vpinsrb       $0x2,%ebp,%xmm0,%xmm0
  DB  67,15,182,44,33                     ; movzbl        (%r9,%r12,1),%ebp
  DB  196,227,121,32,197,3                ; vpinsrb       $0x3,%ebp,%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  196,227,53,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm9,%ymm0
  DB  197,124,91,208                      ; vcvtdq2ps     %ymm0,%ymm10
  DB  189,129,128,128,59                  ; mov           $0x3b808081,%ebp
  DB  197,249,110,197                     ; vmovd         %ebp,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,200,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm9
  DB  196,193,44,89,193                   ; vmulps        %ymm9,%ymm10,%ymm0
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
  DB  196,227,249,22,205,1                ; vpextrq       $0x1,%xmm1,%rbp
  DB  65,137,233                          ; mov           %ebp,%r9d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,203                 ; vmovq         %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,25,201,1                ; vextractf128  $0x1,%ymm1,%xmm1
  DB  196,195,249,22,203,1                ; vpextrq       $0x1,%xmm1,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  196,193,249,126,207                 ; vmovq         %xmm1,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,131,121,32,12,16,0              ; vpinsrb       $0x0,(%r8,%r10,1),%xmm0,%xmm1
  DB  196,195,113,32,12,24,1              ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm1,%xmm1
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,113,32,203,2                ; vpinsrb       $0x2,%ebx,%xmm1,%xmm1
  DB  65,15,182,44,40                     ; movzbl        (%r8,%rbp,1),%ebp
  DB  196,227,113,32,205,3                ; vpinsrb       $0x3,%ebp,%xmm1,%xmm1
  DB  196,98,121,49,209                   ; vpmovzxbd     %xmm1,%xmm10
  DB  196,131,121,32,12,32,0              ; vpinsrb       $0x0,(%r8,%r12,1),%xmm0,%xmm1
  DB  196,131,113,32,12,56,1              ; vpinsrb       $0x1,(%r8,%r15,1),%xmm1,%xmm1
  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
  DB  196,227,113,32,205,2                ; vpinsrb       $0x2,%ebp,%xmm1,%xmm1
  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
  DB  196,227,113,32,205,3                ; vpinsrb       $0x3,%ebp,%xmm1,%xmm1
  DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
  DB  196,227,45,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm10,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  72,139,64,16                        ; mov           0x10(%rax),%rax
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
  DB  196,227,249,22,213,1                ; vpextrq       $0x1,%xmm2,%rbp
  DB  65,137,232                          ; mov           %ebp,%r8d
  DB  72,193,237,32                       ; shr           $0x20,%rbp
  DB  196,225,249,126,211                 ; vmovq         %xmm2,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,195,249,22,210,1                ; vpextrq       $0x1,%xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,193,249,126,214                 ; vmovq         %xmm2,%r14
  DB  69,137,247                          ; mov           %r14d,%r15d
  DB  73,193,238,32                       ; shr           $0x20,%r14
  DB  196,163,121,32,20,8,0               ; vpinsrb       $0x0,(%rax,%r9,1),%xmm0,%xmm2
  DB  196,227,105,32,20,24,1              ; vpinsrb       $0x1,(%rax,%rbx,1),%xmm2,%xmm2
  DB  66,15,182,28,0                      ; movzbl        (%rax,%r8,1),%ebx
  DB  196,227,105,32,211,2                ; vpinsrb       $0x2,%ebx,%xmm2,%xmm2
  DB  15,182,44,40                        ; movzbl        (%rax,%rbp,1),%ebp
  DB  196,227,105,32,213,3                ; vpinsrb       $0x3,%ebp,%xmm2,%xmm2
  DB  196,98,121,49,194                   ; vpmovzxbd     %xmm2,%xmm8
  DB  196,163,121,32,20,56,0              ; vpinsrb       $0x0,(%rax,%r15,1),%xmm0,%xmm2
  DB  196,163,105,32,20,48,1              ; vpinsrb       $0x1,(%rax,%r14,1),%xmm2,%xmm2
  DB  66,15,182,44,24                     ; movzbl        (%rax,%r11,1),%ebp
  DB  196,227,105,32,213,2                ; vpinsrb       $0x2,%ebp,%xmm2,%xmm2
  DB  66,15,182,4,16                      ; movzbl        (%rax,%r10,1),%eax
  DB  196,227,105,32,208,3                ; vpinsrb       $0x3,%eax,%xmm2,%xmm2
  DB  196,226,121,49,210                  ; vpmovzxbd     %xmm2,%xmm2
  DB  196,227,61,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm8,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,180,89,210                      ; vmulps        %ymm2,%ymm9,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_r_avx
_sk_table_r_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,65,121,112,192,0                ; vpshufd       $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
  DB  196,227,249,22,192,1                ; vpextrq       $0x1,%xmm0,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,194                 ; vmovq         %xmm0,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,25,192,1                ; vextractf128  $0x1,%ymm0,%xmm0
  DB  196,227,249,22,195,1                ; vpextrq       $0x1,%xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,199                 ; vmovq         %xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,129,122,16,4,160                ; vmovss        (%r8,%r12,4),%xmm0
  DB  196,3,121,33,4,184,16               ; vinsertps     $0x10,(%r8,%r15,4),%xmm0,%xmm8
  DB  196,129,122,16,4,176                ; vmovss        (%r8,%r14,4),%xmm0
  DB  196,99,57,33,192,32                 ; vinsertps     $0x20,%xmm0,%xmm8,%xmm8
  DB  196,193,122,16,4,152                ; vmovss        (%r8,%rbx,4),%xmm0
  DB  196,99,57,33,192,48                 ; vinsertps     $0x30,%xmm0,%xmm8,%xmm8
  DB  196,129,122,16,4,152                ; vmovss        (%r8,%r11,4),%xmm0
  DB  196,3,121,33,12,144,16              ; vinsertps     $0x10,(%r8,%r10,4),%xmm0,%xmm9
  DB  196,129,122,16,4,136                ; vmovss        (%r8,%r9,4),%xmm0
  DB  196,99,49,33,200,32                 ; vinsertps     $0x20,%xmm0,%xmm9,%xmm9
  DB  196,193,122,16,4,128                ; vmovss        (%r8,%rax,4),%xmm0
  DB  196,227,49,33,192,48                ; vinsertps     $0x30,%xmm0,%xmm9,%xmm0
  DB  196,195,125,24,192,1                ; vinsertf128   $0x1,%xmm8,%ymm0,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_g_avx
_sk_table_g_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,65,121,112,192,0                ; vpshufd       $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
  DB  196,227,249,22,200,1                ; vpextrq       $0x1,%xmm1,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,202                 ; vmovq         %xmm1,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,25,201,1                ; vextractf128  $0x1,%ymm1,%xmm1
  DB  196,227,249,22,203,1                ; vpextrq       $0x1,%xmm1,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,207                 ; vmovq         %xmm1,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,129,122,16,12,160               ; vmovss        (%r8,%r12,4),%xmm1
  DB  196,3,113,33,4,184,16               ; vinsertps     $0x10,(%r8,%r15,4),%xmm1,%xmm8
  DB  196,129,122,16,12,176               ; vmovss        (%r8,%r14,4),%xmm1
  DB  196,99,57,33,193,32                 ; vinsertps     $0x20,%xmm1,%xmm8,%xmm8
  DB  196,193,122,16,12,152               ; vmovss        (%r8,%rbx,4),%xmm1
  DB  196,99,57,33,193,48                 ; vinsertps     $0x30,%xmm1,%xmm8,%xmm8
  DB  196,129,122,16,12,152               ; vmovss        (%r8,%r11,4),%xmm1
  DB  196,3,113,33,12,144,16              ; vinsertps     $0x10,(%r8,%r10,4),%xmm1,%xmm9
  DB  196,129,122,16,12,136               ; vmovss        (%r8,%r9,4),%xmm1
  DB  196,99,49,33,201,32                 ; vinsertps     $0x20,%xmm1,%xmm9,%xmm9
  DB  196,193,122,16,12,128               ; vmovss        (%r8,%rax,4),%xmm1
  DB  196,227,49,33,201,48                ; vinsertps     $0x30,%xmm1,%xmm9,%xmm1
  DB  196,195,117,24,200,1                ; vinsertf128   $0x1,%xmm8,%ymm1,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_b_avx
_sk_table_b_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,65,121,112,192,0                ; vpshufd       $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
  DB  196,227,249,22,208,1                ; vpextrq       $0x1,%xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,210                 ; vmovq         %xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,227,249,22,211,1                ; vpextrq       $0x1,%xmm2,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,215                 ; vmovq         %xmm2,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,129,122,16,20,160               ; vmovss        (%r8,%r12,4),%xmm2
  DB  196,3,105,33,4,184,16               ; vinsertps     $0x10,(%r8,%r15,4),%xmm2,%xmm8
  DB  196,129,122,16,20,176               ; vmovss        (%r8,%r14,4),%xmm2
  DB  196,99,57,33,194,32                 ; vinsertps     $0x20,%xmm2,%xmm8,%xmm8
  DB  196,193,122,16,20,152               ; vmovss        (%r8,%rbx,4),%xmm2
  DB  196,99,57,33,194,48                 ; vinsertps     $0x30,%xmm2,%xmm8,%xmm8
  DB  196,129,122,16,20,152               ; vmovss        (%r8,%r11,4),%xmm2
  DB  196,3,105,33,12,144,16              ; vinsertps     $0x10,(%r8,%r10,4),%xmm2,%xmm9
  DB  196,129,122,16,20,136               ; vmovss        (%r8,%r9,4),%xmm2
  DB  196,99,49,33,202,32                 ; vinsertps     $0x20,%xmm2,%xmm9,%xmm9
  DB  196,193,122,16,20,128               ; vmovss        (%r8,%rax,4),%xmm2
  DB  196,227,49,33,210,48                ; vinsertps     $0x30,%xmm2,%xmm9,%xmm2
  DB  196,195,109,24,208,1                ; vinsertf128   $0x1,%xmm8,%ymm2,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_a_avx
_sk_table_a_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,65,121,112,192,0                ; vpshufd       $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  197,253,91,219                      ; vcvtps2dq     %ymm3,%ymm3
  DB  196,227,249,22,216,1                ; vpextrq       $0x1,%xmm3,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,218                 ; vmovq         %xmm3,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  196,227,125,25,219,1                ; vextractf128  $0x1,%ymm3,%xmm3
  DB  196,227,249,22,219,1                ; vpextrq       $0x1,%xmm3,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,249,126,223                 ; vmovq         %xmm3,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,129,122,16,28,160               ; vmovss        (%r8,%r12,4),%xmm3
  DB  196,3,97,33,4,184,16                ; vinsertps     $0x10,(%r8,%r15,4),%xmm3,%xmm8
  DB  196,129,122,16,28,176               ; vmovss        (%r8,%r14,4),%xmm3
  DB  196,99,57,33,195,32                 ; vinsertps     $0x20,%xmm3,%xmm8,%xmm8
  DB  196,193,122,16,28,152               ; vmovss        (%r8,%rbx,4),%xmm3
  DB  196,99,57,33,195,48                 ; vinsertps     $0x30,%xmm3,%xmm8,%xmm8
  DB  196,129,122,16,28,152               ; vmovss        (%r8,%r11,4),%xmm3
  DB  196,3,97,33,12,144,16               ; vinsertps     $0x10,(%r8,%r10,4),%xmm3,%xmm9
  DB  196,129,122,16,28,136               ; vmovss        (%r8,%r9,4),%xmm3
  DB  196,99,49,33,203,32                 ; vinsertps     $0x20,%xmm3,%xmm9,%xmm9
  DB  196,193,122,16,28,128               ; vmovss        (%r8,%rax,4),%xmm3
  DB  196,227,49,33,219,48                ; vinsertps     $0x30,%xmm3,%xmm9,%xmm3
  DB  196,195,101,24,216,1                ; vinsertf128   $0x1,%xmm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_r_avx
_sk_parametric_r_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,124,194,192,2                ; vcmpleps      %ymm8,%ymm0,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  197,52,89,200                       ; vmulps        %ymm0,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  196,65,124,88,211                   ; vaddps        %ymm11,%ymm0,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,36,89,216                       ; vmulps        %ymm0,%ymm11,%ymm11
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,65,124,84,210                   ; vandps        %ymm10,%ymm0,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,44,86,208                       ; vorps         %ymm0,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,36,92,216                       ; vsubps        %ymm0,%ymm11,%ymm11
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  197,36,92,216                       ; vsubps        %ymm0,%ymm11,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,232,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,172,88,192                      ; vaddps        %ymm0,%ymm10,%ymm0
  DB  197,148,94,192                      ; vdivps        %ymm0,%ymm13,%ymm0
  DB  197,164,92,192                      ; vsubps        %ymm0,%ymm11,%ymm0
  DB  197,28,89,216                       ; vmulps        %ymm0,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,224,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,65,124,88,219                   ; vaddps        %ymm11,%ymm0,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,193,124,89,194                  ; vmulps        %ymm10,%ymm0,%ymm0
  DB  197,36,92,216                       ; vsubps        %ymm0,%ymm11,%ymm11
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,232,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm13
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,193,124,92,194                  ; vsubps        %ymm10,%ymm0,%ymm0
  DB  197,148,94,192                      ; vdivps        %ymm0,%ymm13,%ymm0
  DB  197,164,88,192                      ; vaddps        %ymm0,%ymm11,%ymm0
  DB  197,156,89,192                      ; vmulps        %ymm0,%ymm12,%ymm0
  DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
  DB  196,195,125,74,193,128              ; vblendvps     %ymm8,%ymm9,%ymm0,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,124,95,192                   ; vmaxps        %ymm8,%ymm0,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,188,93,192                      ; vminps        %ymm0,%ymm8,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_g_avx
_sk_parametric_g_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,116,194,192,2                ; vcmpleps      %ymm8,%ymm1,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  197,52,89,201                       ; vmulps        %ymm1,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  196,65,116,88,211                   ; vaddps        %ymm11,%ymm1,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,36,89,217                       ; vmulps        %ymm1,%ymm11,%ymm11
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  196,65,116,84,210                   ; vandps        %ymm10,%ymm1,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,44,86,209                       ; vorps         %ymm1,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,36,92,217                       ; vsubps        %ymm1,%ymm11,%ymm11
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  197,36,92,217                       ; vsubps        %ymm1,%ymm11,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,99,117,24,233,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,172,88,201                      ; vaddps        %ymm1,%ymm10,%ymm1
  DB  197,148,94,201                      ; vdivps        %ymm1,%ymm13,%ymm1
  DB  197,164,92,201                      ; vsubps        %ymm1,%ymm11,%ymm1
  DB  197,28,89,217                       ; vmulps        %ymm1,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,99,117,24,225,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  196,65,116,88,219                   ; vaddps        %ymm11,%ymm1,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  196,193,116,89,202                  ; vmulps        %ymm10,%ymm1,%ymm1
  DB  197,36,92,217                       ; vsubps        %ymm1,%ymm11,%ymm11
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,99,117,24,233,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm13
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  196,193,116,92,202                  ; vsubps        %ymm10,%ymm1,%ymm1
  DB  197,148,94,201                      ; vdivps        %ymm1,%ymm13,%ymm1
  DB  197,164,88,201                      ; vaddps        %ymm1,%ymm11,%ymm1
  DB  197,156,89,201                      ; vmulps        %ymm1,%ymm12,%ymm1
  DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,116,88,202                  ; vaddps        %ymm10,%ymm1,%ymm1
  DB  196,195,117,74,201,128              ; vblendvps     %ymm8,%ymm9,%ymm1,%ymm1
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,116,95,192                   ; vmaxps        %ymm8,%ymm1,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,188,93,201                      ; vminps        %ymm1,%ymm8,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_b_avx
_sk_parametric_b_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,108,194,192,2                ; vcmpleps      %ymm8,%ymm2,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  197,52,89,202                       ; vmulps        %ymm2,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
  DB  196,65,108,88,211                   ; vaddps        %ymm11,%ymm2,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,36,89,218                       ; vmulps        %ymm2,%ymm11,%ymm11
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  197,249,112,210,0                   ; vpshufd       $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  196,65,108,84,210                   ; vandps        %ymm10,%ymm2,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  197,249,112,210,0                   ; vpshufd       $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,44,86,210                       ; vorps         %ymm2,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,36,92,218                       ; vsubps        %ymm2,%ymm11,%ymm11
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
  DB  197,36,92,218                       ; vsubps        %ymm2,%ymm11,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,99,109,24,234,1                 ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,172,88,210                      ; vaddps        %ymm2,%ymm10,%ymm2
  DB  197,148,94,210                      ; vdivps        %ymm2,%ymm13,%ymm2
  DB  197,164,92,210                      ; vsubps        %ymm2,%ymm11,%ymm2
  DB  197,28,89,218                       ; vmulps        %ymm2,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,99,109,24,226,1                 ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  196,65,108,88,219                   ; vaddps        %ymm11,%ymm2,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  196,193,108,89,210                  ; vmulps        %ymm10,%ymm2,%ymm2
  DB  197,36,92,218                       ; vsubps        %ymm2,%ymm11,%ymm11
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,99,109,24,234,1                 ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm13
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,208                 ; vmovd         %r8d,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  196,193,108,92,210                  ; vsubps        %ymm10,%ymm2,%ymm2
  DB  197,148,94,210                      ; vdivps        %ymm2,%ymm13,%ymm2
  DB  197,164,88,210                      ; vaddps        %ymm2,%ymm11,%ymm2
  DB  197,156,89,210                      ; vmulps        %ymm2,%ymm12,%ymm2
  DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
  DB  196,195,109,74,209,128              ; vblendvps     %ymm8,%ymm9,%ymm2,%ymm2
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,108,95,192                   ; vmaxps        %ymm8,%ymm2,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,188,93,210                      ; vminps        %ymm2,%ymm8,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_a_avx
_sk_parametric_a_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,65,100,194,192,2                ; vcmpleps      %ymm8,%ymm3,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  197,52,89,203                       ; vmulps        %ymm3,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  196,98,125,24,80,4                  ; vbroadcastss  0x4(%rax),%ymm10
  DB  196,98,125,24,88,8                  ; vbroadcastss  0x8(%rax),%ymm11
  DB  197,172,89,219                      ; vmulps        %ymm3,%ymm10,%ymm3
  DB  196,65,100,88,211                   ; vaddps        %ymm11,%ymm3,%ymm10
  DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
  DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
  DB  65,184,0,0,0,52                     ; mov           $0x34000000,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,36,89,219                       ; vmulps        %ymm3,%ymm11,%ymm11
  DB  65,184,255,255,127,0                ; mov           $0x7fffff,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,65,100,84,210                   ; vandps        %ymm10,%ymm3,%ymm10
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,44,86,211                       ; vorps         %ymm3,%ymm10,%ymm10
  DB  65,184,119,115,248,66               ; mov           $0x42f87377,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,36,92,219                       ; vsubps        %ymm3,%ymm11,%ymm11
  DB  65,184,117,191,191,63               ; mov           $0x3fbfbf75,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,172,89,219                      ; vmulps        %ymm3,%ymm10,%ymm3
  DB  197,36,92,219                       ; vsubps        %ymm3,%ymm11,%ymm11
  DB  65,184,163,233,220,63               ; mov           $0x3fdce9a3,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,99,101,24,235,1                 ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm13
  DB  65,184,249,68,180,62                ; mov           $0x3eb444f9,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,172,88,219                      ; vaddps        %ymm3,%ymm10,%ymm3
  DB  197,148,94,219                      ; vdivps        %ymm3,%ymm13,%ymm3
  DB  197,164,92,219                      ; vsubps        %ymm3,%ymm11,%ymm3
  DB  197,28,89,219                       ; vmulps        %ymm3,%ymm12,%ymm11
  DB  196,67,125,8,211,1                  ; vroundps      $0x1,%ymm11,%ymm10
  DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,99,101,24,227,1                 ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm12
  DB  65,184,81,140,242,66                ; mov           $0x42f28c51,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,65,100,88,219                   ; vaddps        %ymm11,%ymm3,%ymm11
  DB  65,184,141,188,190,63               ; mov           $0x3fbebc8d,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,193,100,89,218                  ; vmulps        %ymm10,%ymm3,%ymm3
  DB  197,36,92,219                       ; vsubps        %ymm3,%ymm11,%ymm11
  DB  65,184,254,210,221,65               ; mov           $0x41ddd2fe,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,99,101,24,235,1                 ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm13
  DB  65,184,248,245,154,64               ; mov           $0x409af5f8,%r8d
  DB  196,193,121,110,216                 ; vmovd         %r8d,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,193,100,92,218                  ; vsubps        %ymm10,%ymm3,%ymm3
  DB  197,148,94,219                      ; vdivps        %ymm3,%ymm13,%ymm3
  DB  197,164,88,219                      ; vaddps        %ymm3,%ymm11,%ymm3
  DB  197,156,89,219                      ; vmulps        %ymm3,%ymm12,%ymm3
  DB  197,253,91,219                      ; vcvtps2dq     %ymm3,%ymm3
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,193,100,88,218                  ; vaddps        %ymm10,%ymm3,%ymm3
  DB  196,195,101,74,217,128              ; vblendvps     %ymm8,%ymm9,%ymm3,%ymm3
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  196,65,100,95,192                   ; vmaxps        %ymm8,%ymm3,%ymm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,188,93,219                      ; vminps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_a8_avx
_sk_load_a8_avx LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,74                              ; jne           3146 <_sk_load_a8_avx+0x5a>
  DB  197,250,126,0                       ; vmovq         (%rax),%xmm0
  DB  196,226,121,49,200                  ; vpmovzxbd     %xmm0,%xmm1
  DB  196,227,121,4,192,229               ; vpermilps     $0xe5,%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           314e <_sk_load_a8_avx+0x62>
  DB  196,193,249,110,193                 ; vmovq         %r9,%xmm0
  DB  235,149                             ; jmp           3100 <_sk_load_a8_avx+0x14>

PUBLIC _sk_gather_a8_avx
_sk_gather_a8_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,209                      ; vcvttps2dq    %ymm1,%ymm2
  DB  197,249,110,72,16                   ; vmovd         0x10(%rax),%xmm1
  DB  197,249,112,217,0                   ; vpshufd       $0x0,%xmm1,%xmm3
  DB  196,226,97,64,202                   ; vpmulld       %xmm2,%xmm3,%xmm1
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,226,97,64,210                   ; vpmulld       %xmm2,%xmm3,%xmm2
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  196,227,249,22,208,1                ; vpextrq       $0x1,%xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,210                 ; vmovq         %xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  197,241,254,192                     ; vpaddd        %xmm0,%xmm1,%xmm0
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  196,195,249,22,199,1                ; vpextrq       $0x1,%xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,131,121,32,4,48,0               ; vpinsrb       $0x0,(%r8,%r14,1),%xmm0,%xmm0
  DB  196,195,121,32,4,24,1               ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm0,%xmm0
  DB  67,15,182,28,32                     ; movzbl        (%r8,%r12,1),%ebx
  DB  196,227,121,32,195,2                ; vpinsrb       $0x2,%ebx,%xmm0,%xmm0
  DB  67,15,182,28,56                     ; movzbl        (%r8,%r15,1),%ebx
  DB  196,227,121,32,195,3                ; vpinsrb       $0x3,%ebx,%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  196,131,121,32,12,24,0              ; vpinsrb       $0x0,(%r8,%r11,1),%xmm0,%xmm1
  DB  196,131,113,32,12,16,1              ; vpinsrb       $0x1,(%r8,%r10,1),%xmm1,%xmm1
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,113,32,203,2                ; vpinsrb       $0x2,%ebx,%xmm1,%xmm1
  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
  DB  196,227,113,32,200,3                ; vpinsrb       $0x3,%eax,%xmm1,%xmm1
  DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_a8_avx
_sk_store_a8_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  184,0,0,127,67                      ; mov           $0x437f0000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
  DB  196,66,57,43,193                    ; vpackusdw     %xmm9,%xmm8,%xmm8
  DB  196,65,57,103,192                   ; vpackuswb     %xmm8,%xmm8,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,10                              ; jne           32a7 <_sk_store_a8_avx+0x42>
  DB  196,65,123,17,4,57                  ; vmovsd        %xmm8,(%r9,%rdi,1)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  119,236                             ; ja            32a3 <_sk_store_a8_avx+0x3e>
  DB  196,66,121,48,192                   ; vpmovzxbw     %xmm8,%xmm8
  DB  65,15,182,192                       ; movzbl        %r8b,%eax
  DB  76,141,5,69,0,0,0                   ; lea           0x45(%rip),%r8        # 330c <_sk_store_a8_avx+0xa7>
  DB  73,99,4,128                         ; movslq        (%r8,%rax,4),%rax
  DB  76,1,192                            ; add           %r8,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,67,121,20,68,57,6,12            ; vpextrb       $0xc,%xmm8,0x6(%r9,%rdi,1)
  DB  196,67,121,20,68,57,5,10            ; vpextrb       $0xa,%xmm8,0x5(%r9,%rdi,1)
  DB  196,67,121,20,68,57,4,8             ; vpextrb       $0x8,%xmm8,0x4(%r9,%rdi,1)
  DB  196,67,121,20,68,57,3,6             ; vpextrb       $0x6,%xmm8,0x3(%r9,%rdi,1)
  DB  196,67,121,20,68,57,2,4             ; vpextrb       $0x4,%xmm8,0x2(%r9,%rdi,1)
  DB  196,67,121,20,68,57,1,2             ; vpextrb       $0x2,%xmm8,0x1(%r9,%rdi,1)
  DB  196,67,121,20,4,57,0                ; vpextrb       $0x0,%xmm8,(%r9,%rdi,1)
  DB  235,154                             ; jmp           32a3 <_sk_store_a8_avx+0x3e>
  DB  15,31,0                             ; nopl          (%rax)
  DB  244                                 ; hlt
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  236                                 ; in            (%dx),%al
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,228                             ; jmpq          *%rsp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  220,255                             ; fdivr         %st,%st(7)
  DB  255                                 ; (bad)
  DB  255,212                             ; callq         *%rsp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,204                             ; dec           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,196                             ; inc           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_g8_avx
_sk_load_g8_avx LABEL PROC
  DB  73,137,200                          ; mov           %rcx,%r8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,1,248                            ; add           %rdi,%rax
  DB  77,133,192                          ; test          %r8,%r8
  DB  117,91                              ; jne           3393 <_sk_load_g8_avx+0x6b>
  DB  197,250,126,0                       ; vmovq         (%rax),%xmm0
  DB  196,226,121,49,200                  ; vpmovzxbd     %xmm0,%xmm1
  DB  196,227,121,4,192,229               ; vpermilps     $0xe5,%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,217,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,193                          ; mov           %r8,%rcx
  DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
  DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
  DB  255,224                             ; jmpq          *%rax
  DB  49,201                              ; xor           %ecx,%ecx
  DB  77,137,194                          ; mov           %r8,%r10
  DB  69,49,201                           ; xor           %r9d,%r9d
  DB  68,15,182,24                        ; movzbl        (%rax),%r11d
  DB  72,255,192                          ; inc           %rax
  DB  73,211,227                          ; shl           %cl,%r11
  DB  77,9,217                            ; or            %r11,%r9
  DB  72,131,193,8                        ; add           $0x8,%rcx
  DB  73,255,202                          ; dec           %r10
  DB  117,234                             ; jne           339b <_sk_load_g8_avx+0x73>
  DB  196,193,249,110,193                 ; vmovq         %r9,%xmm0
  DB  235,132                             ; jmp           333c <_sk_load_g8_avx+0x14>

PUBLIC _sk_gather_g8_avx
_sk_gather_g8_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,209                      ; vcvttps2dq    %ymm1,%ymm2
  DB  197,249,110,72,16                   ; vmovd         0x10(%rax),%xmm1
  DB  197,249,112,217,0                   ; vpshufd       $0x0,%xmm1,%xmm3
  DB  196,226,97,64,202                   ; vpmulld       %xmm2,%xmm3,%xmm1
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,226,97,64,210                   ; vpmulld       %xmm2,%xmm3,%xmm2
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  196,227,249,22,208,1                ; vpextrq       $0x1,%xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,210                 ; vmovq         %xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  197,241,254,192                     ; vpaddd        %xmm0,%xmm1,%xmm0
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  196,195,249,22,199,1                ; vpextrq       $0x1,%xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,131,121,32,4,48,0               ; vpinsrb       $0x0,(%r8,%r14,1),%xmm0,%xmm0
  DB  196,195,121,32,4,24,1               ; vpinsrb       $0x1,(%r8,%rbx,1),%xmm0,%xmm0
  DB  67,15,182,28,32                     ; movzbl        (%r8,%r12,1),%ebx
  DB  196,227,121,32,195,2                ; vpinsrb       $0x2,%ebx,%xmm0,%xmm0
  DB  67,15,182,28,56                     ; movzbl        (%r8,%r15,1),%ebx
  DB  196,227,121,32,195,3                ; vpinsrb       $0x3,%ebx,%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  196,131,121,32,12,24,0              ; vpinsrb       $0x0,(%r8,%r11,1),%xmm0,%xmm1
  DB  196,131,113,32,12,16,1              ; vpinsrb       $0x1,(%r8,%r10,1),%xmm1,%xmm1
  DB  67,15,182,28,8                      ; movzbl        (%r8,%r9,1),%ebx
  DB  196,227,113,32,203,2                ; vpinsrb       $0x2,%ebx,%xmm1,%xmm1
  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
  DB  196,227,113,32,200,3                ; vpinsrb       $0x3,%eax,%xmm1,%xmm1
  DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,217,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
  DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_i8_avx
_sk_gather_i8_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,192                          ; mov           %rax,%r8
  DB  77,133,192                          ; test          %r8,%r8
  DB  116,5                               ; je            34d2 <_sk_gather_i8_avx+0xf>
  DB  76,137,192                          ; mov           %r8,%rax
  DB  235,2                               ; jmp           34d4 <_sk_gather_i8_avx+0x11>
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  197,254,91,209                      ; vcvttps2dq    %ymm1,%ymm2
  DB  197,249,110,72,16                   ; vmovd         0x10(%rax),%xmm1
  DB  197,249,112,217,0                   ; vpshufd       $0x0,%xmm1,%xmm3
  DB  196,226,97,64,202                   ; vpmulld       %xmm2,%xmm3,%xmm1
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,226,97,64,210                   ; vpmulld       %xmm2,%xmm3,%xmm2
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  196,227,249,22,208,1                ; vpextrq       $0x1,%xmm2,%rax
  DB  65,137,194                          ; mov           %eax,%r10d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,211                 ; vmovq         %xmm2,%r11
  DB  69,137,222                          ; mov           %r11d,%r14d
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  197,241,254,192                     ; vpaddd        %xmm0,%xmm1,%xmm0
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,223                          ; mov           %ebx,%r15d
  DB  196,195,249,22,196,1                ; vpextrq       $0x1,%xmm0,%r12
  DB  69,137,229                          ; mov           %r12d,%r13d
  DB  73,193,236,32                       ; shr           $0x20,%r12
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,131,121,32,4,49,0               ; vpinsrb       $0x0,(%r9,%r14,1),%xmm0,%xmm0
  DB  196,131,121,32,4,25,1               ; vpinsrb       $0x1,(%r9,%r11,1),%xmm0,%xmm0
  DB  196,131,121,32,4,17,2               ; vpinsrb       $0x2,(%r9,%r10,1),%xmm0,%xmm0
  DB  196,195,121,32,4,1,3                ; vpinsrb       $0x3,(%r9,%rax,1),%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
  DB  196,193,249,126,195                 ; vmovq         %xmm0,%r11
  DB  196,131,121,32,4,57,0               ; vpinsrb       $0x0,(%r9,%r15,1),%xmm0,%xmm0
  DB  196,195,121,32,4,25,1               ; vpinsrb       $0x1,(%r9,%rbx,1),%xmm0,%xmm0
  DB  196,131,121,32,4,41,2               ; vpinsrb       $0x2,(%r9,%r13,1),%xmm0,%xmm0
  DB  196,131,121,32,4,33,3               ; vpinsrb       $0x3,(%r9,%r12,1),%xmm0,%xmm0
  DB  196,226,121,49,192                  ; vpmovzxbd     %xmm0,%xmm0
  DB  73,139,88,8                         ; mov           0x8(%r8),%rbx
  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
  DB  69,137,200                          ; mov           %r9d,%r8d
  DB  73,193,233,30                       ; shr           $0x1e,%r9
  DB  196,227,249,22,192,1                ; vpextrq       $0x1,%xmm0,%rax
  DB  65,137,198                          ; mov           %eax,%r14d
  DB  72,193,232,30                       ; shr           $0x1e,%rax
  DB  69,137,223                          ; mov           %r11d,%r15d
  DB  73,193,235,30                       ; shr           $0x1e,%r11
  DB  69,137,212                          ; mov           %r10d,%r12d
  DB  73,193,234,30                       ; shr           $0x1e,%r10
  DB  196,161,121,110,4,131               ; vmovd         (%rbx,%r8,4),%xmm0
  DB  196,163,121,34,4,11,1               ; vpinsrd       $0x1,(%rbx,%r9,1),%xmm0,%xmm0
  DB  196,163,121,34,4,179,2              ; vpinsrd       $0x2,(%rbx,%r14,4),%xmm0,%xmm0
  DB  196,99,121,34,4,3,3                 ; vpinsrd       $0x3,(%rbx,%rax,1),%xmm0,%xmm8
  DB  196,161,121,110,4,187               ; vmovd         (%rbx,%r15,4),%xmm0
  DB  196,163,121,34,4,27,1               ; vpinsrd       $0x1,(%rbx,%r11,1),%xmm0,%xmm0
  DB  196,163,121,34,4,163,2              ; vpinsrd       $0x2,(%rbx,%r12,4),%xmm0,%xmm0
  DB  196,163,121,34,28,19,3              ; vpinsrd       $0x3,(%rbx,%r10,1),%xmm0,%xmm3
  DB  196,227,61,24,195,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,99,117,24,217,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm11
  DB  197,164,84,192                      ; vandps        %ymm0,%ymm11,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,99,117,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm9
  DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
  DB  196,193,41,114,208,8                ; vpsrld        $0x8,%xmm8,%xmm10
  DB  197,241,114,211,8                   ; vpsrld        $0x8,%xmm3,%xmm1
  DB  196,227,45,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm10,%ymm1
  DB  197,164,84,201                      ; vandps        %ymm1,%ymm11,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
  DB  196,193,41,114,208,16               ; vpsrld        $0x10,%xmm8,%xmm10
  DB  197,233,114,211,16                  ; vpsrld        $0x10,%xmm3,%xmm2
  DB  196,227,45,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm10,%ymm2
  DB  197,164,84,210                      ; vandps        %ymm2,%ymm11,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
  DB  196,193,57,114,208,24               ; vpsrld        $0x18,%xmm8,%xmm8
  DB  197,225,114,211,24                  ; vpsrld        $0x18,%xmm3,%xmm3
  DB  196,227,61,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,217                  ; vmulps        %ymm9,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_565_avx
_sk_load_565_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,209,0,0,0                    ; jne           376e <_sk_load_565_avx+0xdf>
  DB  196,193,122,111,4,122               ; vmovdqu       (%r10,%rdi,2),%xmm0
  DB  197,241,239,201                     ; vpxor         %xmm1,%xmm1,%xmm1
  DB  197,249,105,201                     ; vpunpckhwd    %xmm1,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,209,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,84,194                      ; vandps        %ymm2,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,244,84,202                      ; vandps        %ymm2,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  15,135,29,255,255,255               ; ja            36a3 <_sk_load_565_avx+0x14>
  DB  69,15,182,192                       ; movzbl        %r8b,%r8d
  DB  76,141,13,75,0,0,0                  ; lea           0x4b(%rip),%r9        # 37dc <_sk_load_565_avx+0x14d>
  DB  75,99,4,129                         ; movslq        (%r9,%r8,4),%rax
  DB  76,1,200                            ; add           %r9,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  196,193,121,196,68,122,12,6         ; vpinsrw       $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,10,5         ; vpinsrw       $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,8,4          ; vpinsrw       $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,6,3          ; vpinsrw       $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,4,2          ; vpinsrw       $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,2,1          ; vpinsrw       $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,4,122,0             ; vpinsrw       $0x0,(%r10,%rdi,2),%xmm0,%xmm0
  DB  233,201,254,255,255                 ; jmpq          36a3 <_sk_load_565_avx+0x14>
  DB  102,144                             ; xchg          %ax,%ax
  DB  242,255                             ; repnz         (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  234                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,226                             ; jmpq          *%rdx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  218,255                             ; (bad)
  DB  255                                 ; (bad)
  DB  255,210                             ; callq         *%rdx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,202                             ; dec           %edx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  190                                 ; .byte         0xbe
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_gather_565_avx
_sk_gather_565_avx LABEL PROC
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,209                      ; vcvttps2dq    %ymm1,%ymm2
  DB  197,249,110,72,16                   ; vmovd         0x10(%rax),%xmm1
  DB  197,249,112,217,0                   ; vpshufd       $0x0,%xmm1,%xmm3
  DB  196,226,97,64,202                   ; vpmulld       %xmm2,%xmm3,%xmm1
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,226,97,64,210                   ; vpmulld       %xmm2,%xmm3,%xmm2
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  196,227,249,22,208,1                ; vpextrq       $0x1,%xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,210                 ; vmovq         %xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  197,241,254,192                     ; vpaddd        %xmm0,%xmm1,%xmm0
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  196,195,249,22,199,1                ; vpextrq       $0x1,%xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  65,15,183,28,88                     ; movzwl        (%r8,%rbx,2),%ebx
  DB  67,15,183,44,112                    ; movzwl        (%r8,%r14,2),%ebp
  DB  197,249,110,197                     ; vmovd         %ebp,%xmm0
  DB  197,249,196,195,1                   ; vpinsrw       $0x1,%ebx,%xmm0,%xmm0
  DB  67,15,183,28,96                     ; movzwl        (%r8,%r12,2),%ebx
  DB  197,249,196,195,2                   ; vpinsrw       $0x2,%ebx,%xmm0,%xmm0
  DB  67,15,183,28,120                    ; movzwl        (%r8,%r15,2),%ebx
  DB  197,249,196,195,3                   ; vpinsrw       $0x3,%ebx,%xmm0,%xmm0
  DB  67,15,183,44,88                     ; movzwl        (%r8,%r11,2),%ebp
  DB  197,249,196,197,4                   ; vpinsrw       $0x4,%ebp,%xmm0,%xmm0
  DB  67,15,183,44,80                     ; movzwl        (%r8,%r10,2),%ebp
  DB  197,249,196,197,5                   ; vpinsrw       $0x5,%ebp,%xmm0,%xmm0
  DB  67,15,183,44,72                     ; movzwl        (%r8,%r9,2),%ebp
  DB  197,249,196,197,6                   ; vpinsrw       $0x6,%ebp,%xmm0,%xmm0
  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
  DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
  DB  197,241,239,201                     ; vpxor         %xmm1,%xmm1,%xmm1
  DB  197,249,105,201                     ; vpunpckhwd    %xmm1,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,209,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,84,194                      ; vandps        %ymm2,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,244,84,202                      ; vandps        %ymm2,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_565_avx
_sk_store_565_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  184,0,0,248,65                      ; mov           $0x41f80000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  196,193,41,114,241,11               ; vpslld        $0xb,%xmm9,%xmm10
  DB  196,67,125,25,201,1                 ; vextractf128  $0x1,%ymm9,%xmm9
  DB  196,193,49,114,241,11               ; vpslld        $0xb,%xmm9,%xmm9
  DB  196,67,45,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm10,%ymm9
  DB  184,0,0,124,66                      ; mov           $0x427c0000,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  197,44,89,209                       ; vmulps        %ymm1,%ymm10,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,33,114,242,5                ; vpslld        $0x5,%xmm10,%xmm11
  DB  196,67,125,25,210,1                 ; vextractf128  $0x1,%ymm10,%xmm10
  DB  196,193,41,114,242,5                ; vpslld        $0x5,%xmm10,%xmm10
  DB  196,67,37,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm11,%ymm10
  DB  196,65,45,86,201                    ; vorpd         %ymm9,%ymm10,%ymm9
  DB  197,60,89,194                       ; vmulps        %ymm2,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,65,53,86,192                    ; vorpd         %ymm8,%ymm9,%ymm8
  DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
  DB  196,66,57,43,193                    ; vpackusdw     %xmm9,%xmm8,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,10                              ; jne           3a27 <_sk_store_565_avx+0x9e>
  DB  196,65,122,127,4,121                ; vmovdqu       %xmm8,(%r9,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  119,236                             ; ja            3a23 <_sk_store_565_avx+0x9a>
  DB  65,15,182,192                       ; movzbl        %r8b,%eax
  DB  76,141,5,66,0,0,0                   ; lea           0x42(%rip),%r8        # 3a84 <_sk_store_565_avx+0xfb>
  DB  73,99,4,128                         ; movslq        (%r8,%rax,4),%rax
  DB  76,1,192                            ; add           %r8,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,67,121,21,68,121,12,6           ; vpextrw       $0x6,%xmm8,0xc(%r9,%rdi,2)
  DB  196,67,121,21,68,121,10,5           ; vpextrw       $0x5,%xmm8,0xa(%r9,%rdi,2)
  DB  196,67,121,21,68,121,8,4            ; vpextrw       $0x4,%xmm8,0x8(%r9,%rdi,2)
  DB  196,67,121,21,68,121,6,3            ; vpextrw       $0x3,%xmm8,0x6(%r9,%rdi,2)
  DB  196,67,121,21,68,121,4,2            ; vpextrw       $0x2,%xmm8,0x4(%r9,%rdi,2)
  DB  196,67,121,21,68,121,2,1            ; vpextrw       $0x1,%xmm8,0x2(%r9,%rdi,2)
  DB  196,67,121,21,4,121,0               ; vpextrw       $0x0,%xmm8,(%r9,%rdi,2)
  DB  235,159                             ; jmp           3a23 <_sk_store_565_avx+0x9a>
  DB  247,255                             ; idiv          %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  239                                 ; out           %eax,(%dx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,231                             ; jmpq          *%rdi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  223,255                             ; (bad)
  DB  255                                 ; (bad)
  DB  255,215                             ; callq         *%rdi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,207                             ; dec           %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,199                             ; inc           %edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_4444_avx
_sk_load_4444_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,245,0,0,0                    ; jne           3ba3 <_sk_load_4444_avx+0x103>
  DB  196,193,122,111,4,122               ; vmovdqu       (%r10,%rdi,2),%xmm0
  DB  197,241,239,201                     ; vpxor         %xmm1,%xmm1,%xmm1
  DB  197,249,105,201                     ; vpunpckhwd    %xmm1,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,99,125,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,193,124,84,193                  ; vandps        %ymm9,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  196,193,116,84,201                  ; vandps        %ymm9,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  197,249,112,210,0                   ; vpshufd       $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  196,193,108,84,209                  ; vandps        %ymm9,%ymm2,%ymm2
  DB  197,124,91,194                      ; vcvtdq2ps     %ymm2,%ymm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,193,100,84,217                  ; vandps        %ymm9,%ymm3,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  15,135,249,254,255,255              ; ja            3ab4 <_sk_load_4444_avx+0x14>
  DB  69,15,182,192                       ; movzbl        %r8b,%r8d
  DB  76,141,13,74,0,0,0                  ; lea           0x4a(%rip),%r9        # 3c10 <_sk_load_4444_avx+0x170>
  DB  75,99,4,129                         ; movslq        (%r9,%r8,4),%rax
  DB  76,1,200                            ; add           %r9,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
  DB  196,193,121,196,68,122,12,6         ; vpinsrw       $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,10,5         ; vpinsrw       $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,8,4          ; vpinsrw       $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,6,3          ; vpinsrw       $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,4,2          ; vpinsrw       $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,68,122,2,1          ; vpinsrw       $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
  DB  196,193,121,196,4,122,0             ; vpinsrw       $0x0,(%r10,%rdi,2),%xmm0,%xmm0
  DB  233,165,254,255,255                 ; jmpq          3ab4 <_sk_load_4444_avx+0x14>
  DB  144                                 ; nop
  DB  243,255                             ; repz          (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  235,255                             ; jmp           3c15 <_sk_load_4444_avx+0x175>
  DB  255                                 ; (bad)
  DB  255,227                             ; jmpq          *%rbx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  219,255                             ; (bad)
  DB  255                                 ; (bad)
  DB  255,211                             ; callq         *%rbx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,203                             ; dec           %ebx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  191                                 ; .byte         0xbf
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_gather_4444_avx
_sk_gather_4444_avx LABEL PROC
  DB  85                                  ; push          %rbp
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,209                      ; vcvttps2dq    %ymm1,%ymm2
  DB  197,249,110,72,16                   ; vmovd         0x10(%rax),%xmm1
  DB  197,249,112,217,0                   ; vpshufd       $0x0,%xmm1,%xmm3
  DB  196,226,97,64,202                   ; vpmulld       %xmm2,%xmm3,%xmm1
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,226,97,64,210                   ; vpmulld       %xmm2,%xmm3,%xmm2
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  196,227,249,22,208,1                ; vpextrq       $0x1,%xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,210                 ; vmovq         %xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  197,241,254,192                     ; vpaddd        %xmm0,%xmm1,%xmm0
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  196,195,249,22,199,1                ; vpextrq       $0x1,%xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  65,15,183,28,88                     ; movzwl        (%r8,%rbx,2),%ebx
  DB  67,15,183,44,112                    ; movzwl        (%r8,%r14,2),%ebp
  DB  197,249,110,197                     ; vmovd         %ebp,%xmm0
  DB  197,249,196,195,1                   ; vpinsrw       $0x1,%ebx,%xmm0,%xmm0
  DB  67,15,183,28,96                     ; movzwl        (%r8,%r12,2),%ebx
  DB  197,249,196,195,2                   ; vpinsrw       $0x2,%ebx,%xmm0,%xmm0
  DB  67,15,183,28,120                    ; movzwl        (%r8,%r15,2),%ebx
  DB  197,249,196,195,3                   ; vpinsrw       $0x3,%ebx,%xmm0,%xmm0
  DB  67,15,183,44,88                     ; movzwl        (%r8,%r11,2),%ebp
  DB  197,249,196,197,4                   ; vpinsrw       $0x4,%ebp,%xmm0,%xmm0
  DB  67,15,183,44,80                     ; movzwl        (%r8,%r10,2),%ebp
  DB  197,249,196,197,5                   ; vpinsrw       $0x5,%ebp,%xmm0,%xmm0
  DB  67,15,183,44,72                     ; movzwl        (%r8,%r9,2),%ebp
  DB  197,249,196,197,6                   ; vpinsrw       $0x6,%ebp,%xmm0,%xmm0
  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
  DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
  DB  197,241,239,201                     ; vpxor         %xmm1,%xmm1,%xmm1
  DB  197,249,105,201                     ; vpunpckhwd    %xmm1,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,99,125,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  196,193,124,84,193                  ; vandps        %ymm9,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  196,193,116,84,201                  ; vandps        %ymm9,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  197,249,112,210,0                   ; vpshufd       $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  196,193,108,84,209                  ; vandps        %ymm9,%ymm2,%ymm2
  DB  197,124,91,194                      ; vcvtdq2ps     %ymm2,%ymm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  196,227,121,4,210,0                 ; vpermilps     $0x0,%xmm2,%xmm2
  DB  196,227,109,24,210,1                ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm2
  DB  197,188,89,210                      ; vmulps        %ymm2,%ymm8,%ymm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  196,193,100,84,217                  ; vandps        %ymm9,%ymm3,%ymm3
  DB  197,124,91,195                      ; vcvtdq2ps     %ymm3,%ymm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,188,89,219                      ; vmulps        %ymm3,%ymm8,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  93                                  ; pop           %rbp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_4444_avx
_sk_store_4444_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  184,0,0,112,65                      ; mov           $0x41700000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  196,193,41,114,241,12               ; vpslld        $0xc,%xmm9,%xmm10
  DB  196,67,125,25,201,1                 ; vextractf128  $0x1,%ymm9,%xmm9
  DB  196,193,49,114,241,12               ; vpslld        $0xc,%xmm9,%xmm9
  DB  196,67,45,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm10,%ymm9
  DB  197,60,89,209                       ; vmulps        %ymm1,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,33,114,242,8                ; vpslld        $0x8,%xmm10,%xmm11
  DB  196,67,125,25,210,1                 ; vextractf128  $0x1,%ymm10,%xmm10
  DB  196,193,41,114,242,8                ; vpslld        $0x8,%xmm10,%xmm10
  DB  196,67,37,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm11,%ymm10
  DB  196,65,45,86,201                    ; vorpd         %ymm9,%ymm10,%ymm9
  DB  197,60,89,210                       ; vmulps        %ymm2,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,33,114,242,4                ; vpslld        $0x4,%xmm10,%xmm11
  DB  196,67,125,25,210,1                 ; vextractf128  $0x1,%ymm10,%xmm10
  DB  196,193,41,114,242,4                ; vpslld        $0x4,%xmm10,%xmm10
  DB  196,67,37,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm11,%ymm10
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,65,45,86,192                    ; vorpd         %ymm8,%ymm10,%ymm8
  DB  196,65,53,86,192                    ; vorpd         %ymm8,%ymm9,%ymm8
  DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
  DB  196,66,57,43,193                    ; vpackusdw     %xmm9,%xmm8,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,10                              ; jne           3e90 <_sk_store_4444_avx+0xaf>
  DB  196,65,122,127,4,121                ; vmovdqu       %xmm8,(%r9,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  119,236                             ; ja            3e8c <_sk_store_4444_avx+0xab>
  DB  65,15,182,192                       ; movzbl        %r8b,%eax
  DB  76,141,5,69,0,0,0                   ; lea           0x45(%rip),%r8        # 3ef0 <_sk_store_4444_avx+0x10f>
  DB  73,99,4,128                         ; movslq        (%r8,%rax,4),%rax
  DB  76,1,192                            ; add           %r8,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,67,121,21,68,121,12,6           ; vpextrw       $0x6,%xmm8,0xc(%r9,%rdi,2)
  DB  196,67,121,21,68,121,10,5           ; vpextrw       $0x5,%xmm8,0xa(%r9,%rdi,2)
  DB  196,67,121,21,68,121,8,4            ; vpextrw       $0x4,%xmm8,0x8(%r9,%rdi,2)
  DB  196,67,121,21,68,121,6,3            ; vpextrw       $0x3,%xmm8,0x6(%r9,%rdi,2)
  DB  196,67,121,21,68,121,4,2            ; vpextrw       $0x2,%xmm8,0x4(%r9,%rdi,2)
  DB  196,67,121,21,68,121,2,1            ; vpextrw       $0x1,%xmm8,0x2(%r9,%rdi,2)
  DB  196,67,121,21,4,121,0               ; vpextrw       $0x0,%xmm8,(%r9,%rdi,2)
  DB  235,159                             ; jmp           3e8c <_sk_store_4444_avx+0xab>
  DB  15,31,0                             ; nopl          (%rax)
  DB  244                                 ; hlt
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  236                                 ; in            (%dx),%al
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,228                             ; jmpq          *%rsp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  220,255                             ; fdivr         %st,%st(7)
  DB  255                                 ; (bad)
  DB  255,212                             ; callq         *%rsp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,204                             ; dec           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,196                             ; inc           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_8888_avx
_sk_load_8888_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,157,0,0,0                    ; jne           3fb7 <_sk_load_8888_avx+0xab>
  DB  196,65,124,16,12,186                ; vmovups       (%r10,%rdi,4),%ymm9
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  197,249,112,192,0                   ; vpshufd       $0x0,%xmm0,%xmm0
  DB  196,99,125,24,216,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm11
  DB  196,193,36,84,193                   ; vandps        %ymm9,%ymm11,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,99,117,24,193,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm8
  DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
  DB  196,193,41,114,209,8                ; vpsrld        $0x8,%xmm9,%xmm10
  DB  196,99,125,25,203,1                 ; vextractf128  $0x1,%ymm9,%xmm3
  DB  197,241,114,211,8                   ; vpsrld        $0x8,%xmm3,%xmm1
  DB  196,227,45,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm10,%ymm1
  DB  197,164,84,201                      ; vandps        %ymm1,%ymm11,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
  DB  196,193,41,114,209,16               ; vpsrld        $0x10,%xmm9,%xmm10
  DB  197,233,114,211,16                  ; vpsrld        $0x10,%xmm3,%xmm2
  DB  196,227,45,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm10,%ymm2
  DB  197,164,84,210                      ; vandps        %ymm2,%ymm11,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
  DB  196,193,49,114,209,24               ; vpsrld        $0x18,%xmm9,%xmm9
  DB  197,225,114,211,24                  ; vpsrld        $0x18,%xmm3,%xmm3
  DB  196,227,53,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm9,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  196,65,52,87,201                    ; vxorps        %ymm9,%ymm9,%ymm9
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  15,135,80,255,255,255               ; ja            3f20 <_sk_load_8888_avx+0x14>
  DB  69,15,182,192                       ; movzbl        %r8b,%r8d
  DB  76,141,13,137,0,0,0                 ; lea           0x89(%rip),%r9        # 4064 <_sk_load_8888_avx+0x158>
  DB  75,99,4,129                         ; movslq        (%r9,%r8,4),%rax
  DB  76,1,200                            ; add           %r9,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,193,121,110,68,186,24           ; vmovd         0x18(%r10,%rdi,4),%xmm0
  DB  197,249,112,192,68                  ; vpshufd       $0x44,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  196,99,117,12,200,64                ; vblendps      $0x40,%ymm0,%ymm1,%ymm9
  DB  196,99,125,25,200,1                 ; vextractf128  $0x1,%ymm9,%xmm0
  DB  196,195,121,34,68,186,20,1          ; vpinsrd       $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0
  DB  196,99,53,24,200,1                  ; vinsertf128   $0x1,%xmm0,%ymm9,%ymm9
  DB  196,99,125,25,200,1                 ; vextractf128  $0x1,%ymm9,%xmm0
  DB  196,195,121,34,68,186,16,0          ; vpinsrd       $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0
  DB  196,99,53,24,200,1                  ; vinsertf128   $0x1,%xmm0,%ymm9,%ymm9
  DB  196,195,49,34,68,186,12,3           ; vpinsrd       $0x3,0xc(%r10,%rdi,4),%xmm9,%xmm0
  DB  196,99,53,12,200,15                 ; vblendps      $0xf,%ymm0,%ymm9,%ymm9
  DB  196,195,49,34,68,186,8,2            ; vpinsrd       $0x2,0x8(%r10,%rdi,4),%xmm9,%xmm0
  DB  196,99,53,12,200,15                 ; vblendps      $0xf,%ymm0,%ymm9,%ymm9
  DB  196,195,49,34,68,186,4,1            ; vpinsrd       $0x1,0x4(%r10,%rdi,4),%xmm9,%xmm0
  DB  196,99,53,12,200,15                 ; vblendps      $0xf,%ymm0,%ymm9,%ymm9
  DB  196,195,49,34,4,186,0               ; vpinsrd       $0x0,(%r10,%rdi,4),%xmm9,%xmm0
  DB  196,99,53,12,200,15                 ; vblendps      $0xf,%ymm0,%ymm9,%ymm9
  DB  233,188,254,255,255                 ; jmpq          3f20 <_sk_load_8888_avx+0x14>
  DB  238                                 ; out           %al,(%dx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,224                             ; jmpq          *%rax
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,210                             ; callq         *%rdx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,196                             ; inc           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,176,255,255,255,156             ; pushq         -0x63000001(%rax)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff
  DB  128,255,255                         ; cmp           $0xff,%bh
  DB  255                                 ; .byte         0xff

PUBLIC _sk_gather_8888_avx
_sk_gather_8888_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
  DB  197,249,110,80,16                   ; vmovd         0x10(%rax),%xmm2
  DB  197,249,112,210,0                   ; vpshufd       $0x0,%xmm2,%xmm2
  DB  196,226,105,64,217                  ; vpmulld       %xmm1,%xmm2,%xmm3
  DB  196,227,125,25,201,1                ; vextractf128  $0x1,%ymm1,%xmm1
  DB  196,226,105,64,201                  ; vpmulld       %xmm1,%xmm2,%xmm1
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  196,227,125,25,194,1                ; vextractf128  $0x1,%ymm0,%xmm2
  DB  197,241,254,202                     ; vpaddd        %xmm2,%xmm1,%xmm1
  DB  196,225,249,126,200                 ; vmovq         %xmm1,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  197,225,254,192                     ; vpaddd        %xmm0,%xmm3,%xmm0
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  196,195,249,22,199,1                ; vpextrq       $0x1,%xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  196,129,121,110,4,176               ; vmovd         (%r8,%r14,4),%xmm0
  DB  196,195,121,34,4,152,1              ; vpinsrd       $0x1,(%r8,%rbx,4),%xmm0,%xmm0
  DB  196,131,121,34,4,160,2              ; vpinsrd       $0x2,(%r8,%r12,4),%xmm0,%xmm0
  DB  196,3,121,34,4,184,3                ; vpinsrd       $0x3,(%r8,%r15,4),%xmm0,%xmm8
  DB  196,129,121,110,4,136               ; vmovd         (%r8,%r9,4),%xmm0
  DB  196,195,121,34,4,128,1              ; vpinsrd       $0x1,(%r8,%rax,4),%xmm0,%xmm0
  DB  196,131,121,34,4,152,2              ; vpinsrd       $0x2,(%r8,%r11,4),%xmm0,%xmm0
  DB  196,131,121,34,28,144,3             ; vpinsrd       $0x3,(%r8,%r10,4),%xmm0,%xmm3
  DB  196,227,61,24,195,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,99,117,24,217,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm11
  DB  197,164,84,192                      ; vandps        %ymm0,%ymm11,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,99,117,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm9
  DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
  DB  196,193,41,114,208,8                ; vpsrld        $0x8,%xmm8,%xmm10
  DB  197,241,114,211,8                   ; vpsrld        $0x8,%xmm3,%xmm1
  DB  196,227,45,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm10,%ymm1
  DB  197,164,84,201                      ; vandps        %ymm1,%ymm11,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
  DB  196,193,41,114,208,16               ; vpsrld        $0x10,%xmm8,%xmm10
  DB  197,233,114,211,16                  ; vpsrld        $0x10,%xmm3,%xmm2
  DB  196,227,45,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm10,%ymm2
  DB  197,164,84,210                      ; vandps        %ymm2,%ymm11,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
  DB  196,193,57,114,208,24               ; vpsrld        $0x18,%xmm8,%xmm8
  DB  197,225,114,211,24                  ; vpsrld        $0x18,%xmm3,%xmm3
  DB  196,227,61,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  196,193,100,89,217                  ; vmulps        %ymm9,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_8888_avx
_sk_store_8888_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  184,0,0,127,67                      ; mov           $0x437f0000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  197,60,89,209                       ; vmulps        %ymm1,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,33,114,242,8                ; vpslld        $0x8,%xmm10,%xmm11
  DB  196,67,125,25,210,1                 ; vextractf128  $0x1,%ymm10,%xmm10
  DB  196,193,41,114,242,8                ; vpslld        $0x8,%xmm10,%xmm10
  DB  196,67,37,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm11,%ymm10
  DB  196,65,45,86,201                    ; vorpd         %ymm9,%ymm10,%ymm9
  DB  197,60,89,210                       ; vmulps        %ymm2,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,193,33,114,242,16               ; vpslld        $0x10,%xmm10,%xmm11
  DB  196,67,125,25,210,1                 ; vextractf128  $0x1,%ymm10,%xmm10
  DB  196,193,41,114,242,16               ; vpslld        $0x10,%xmm10,%xmm10
  DB  196,67,37,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm11,%ymm10
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,193,33,114,240,24               ; vpslld        $0x18,%xmm8,%xmm11
  DB  196,67,125,25,192,1                 ; vextractf128  $0x1,%ymm8,%xmm8
  DB  196,193,57,114,240,24               ; vpslld        $0x18,%xmm8,%xmm8
  DB  196,67,37,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm11,%ymm8
  DB  196,65,45,86,192                    ; vorpd         %ymm8,%ymm10,%ymm8
  DB  196,65,53,86,192                    ; vorpd         %ymm8,%ymm9,%ymm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,10                              ; jne           4265 <_sk_store_8888_avx+0xa4>
  DB  196,65,124,17,4,185                 ; vmovups       %ymm8,(%r9,%rdi,4)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  65,128,224,7                        ; and           $0x7,%r8b
  DB  65,254,200                          ; dec           %r8b
  DB  65,128,248,6                        ; cmp           $0x6,%r8b
  DB  119,236                             ; ja            4261 <_sk_store_8888_avx+0xa0>
  DB  65,15,182,192                       ; movzbl        %r8b,%eax
  DB  76,141,5,84,0,0,0                   ; lea           0x54(%rip),%r8        # 42d4 <_sk_store_8888_avx+0x113>
  DB  73,99,4,128                         ; movslq        (%r8,%rax,4),%rax
  DB  76,1,192                            ; add           %r8,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
  DB  196,67,121,22,76,185,24,2           ; vpextrd       $0x2,%xmm9,0x18(%r9,%rdi,4)
  DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
  DB  196,67,121,22,76,185,20,1           ; vpextrd       $0x1,%xmm9,0x14(%r9,%rdi,4)
  DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
  DB  196,65,122,17,76,185,16             ; vmovss        %xmm9,0x10(%r9,%rdi,4)
  DB  196,67,121,22,68,185,12,3           ; vpextrd       $0x3,%xmm8,0xc(%r9,%rdi,4)
  DB  196,67,121,22,68,185,8,2            ; vpextrd       $0x2,%xmm8,0x8(%r9,%rdi,4)
  DB  196,67,121,22,68,185,4,1            ; vpextrd       $0x1,%xmm8,0x4(%r9,%rdi,4)
  DB  196,65,121,126,4,185                ; vmovd         %xmm8,(%r9,%rdi,4)
  DB  235,143                             ; jmp           4261 <_sk_store_8888_avx+0xa0>
  DB  102,144                             ; xchg          %ax,%ax
  DB  246,255                             ; idiv          %bh
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  238                                 ; out           %al,(%dx)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,230                             ; jmpq          *%rsi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  222,255                             ; fdivrp        %st,%st(7)
  DB  255                                 ; (bad)
  DB  255,209                             ; callq         *%rcx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,195                             ; inc           %ebx
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff
  DB  181,255                             ; mov           $0xff,%ch
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_load_f16_avx
_sk_load_f16_avx LABEL PROC
  DB  72,131,236,120                      ; sub           $0x78,%rsp
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  197,252,17,124,36,64                ; vmovups       %ymm7,0x40(%rsp)
  DB  197,252,17,116,36,32                ; vmovups       %ymm6,0x20(%rsp)
  DB  197,252,17,44,36                    ; vmovups       %ymm5,(%rsp)
  DB  15,133,104,2,0,0                    ; jne           457b <_sk_load_f16_avx+0x28b>
  DB  197,121,16,4,248                    ; vmovupd       (%rax,%rdi,8),%xmm8
  DB  197,249,16,84,248,16                ; vmovupd       0x10(%rax,%rdi,8),%xmm2
  DB  197,249,16,76,248,32                ; vmovupd       0x20(%rax,%rdi,8),%xmm1
  DB  197,122,111,76,248,48               ; vmovdqu       0x30(%rax,%rdi,8),%xmm9
  DB  197,185,97,194                      ; vpunpcklwd    %xmm2,%xmm8,%xmm0
  DB  197,185,105,210                     ; vpunpckhwd    %xmm2,%xmm8,%xmm2
  DB  196,193,113,97,217                  ; vpunpcklwd    %xmm9,%xmm1,%xmm3
  DB  196,193,113,105,201                 ; vpunpckhwd    %xmm9,%xmm1,%xmm1
  DB  197,121,97,242                      ; vpunpcklwd    %xmm2,%xmm0,%xmm14
  DB  197,121,105,194                     ; vpunpckhwd    %xmm2,%xmm0,%xmm8
  DB  197,97,97,249                       ; vpunpcklwd    %xmm1,%xmm3,%xmm15
  DB  197,97,105,209                      ; vpunpckhwd    %xmm1,%xmm3,%xmm10
  DB  196,193,9,108,199                   ; vpunpcklqdq   %xmm15,%xmm14,%xmm0
  DB  196,65,25,239,228                   ; vpxor         %xmm12,%xmm12,%xmm12
  DB  196,193,121,105,204                 ; vpunpckhwd    %xmm12,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  184,0,128,0,0                       ; mov           $0x8000,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
  DB  196,99,117,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm9
  DB  196,193,124,84,201                  ; vandps        %ymm9,%ymm0,%ymm1
  DB  184,0,124,0,0                       ; mov           $0x7c00,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,99,101,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm11
  DB  196,193,124,84,219                  ; vandps        %ymm11,%ymm0,%ymm3
  DB  197,252,87,193                      ; vxorps        %ymm1,%ymm0,%ymm0
  DB  196,227,125,25,218,1                ; vextractf128  $0x1,%ymm3,%xmm2
  DB  196,193,105,118,212                 ; vpcmpeqd      %xmm12,%xmm2,%xmm2
  DB  196,193,97,118,220                  ; vpcmpeqd      %xmm12,%xmm3,%xmm3
  DB  196,227,101,24,242,1                ; vinsertf128   $0x1,%xmm2,%ymm3,%ymm6
  DB  196,227,125,25,203,1                ; vextractf128  $0x1,%ymm1,%xmm3
  DB  197,145,114,243,16                  ; vpslld        $0x10,%xmm3,%xmm13
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,233,114,243,13                  ; vpslld        $0xd,%xmm3,%xmm2
  DB  184,0,0,0,56                        ; mov           $0x38000000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  197,145,254,251                     ; vpaddd        %xmm3,%xmm13,%xmm7
  DB  197,193,254,210                     ; vpaddd        %xmm2,%xmm7,%xmm2
  DB  197,241,114,241,16                  ; vpslld        $0x10,%xmm1,%xmm1
  DB  197,249,114,240,13                  ; vpslld        $0xd,%xmm0,%xmm0
  DB  197,241,254,203                     ; vpaddd        %xmm3,%xmm1,%xmm1
  DB  197,241,254,192                     ; vpaddd        %xmm0,%xmm1,%xmm0
  DB  196,227,125,24,194,1                ; vinsertf128   $0x1,%xmm2,%ymm0,%ymm0
  DB  196,65,20,87,237                    ; vxorps        %ymm13,%ymm13,%ymm13
  DB  196,195,125,74,197,96               ; vblendvps     %ymm6,%ymm13,%ymm0,%ymm0
  DB  196,193,9,109,207                   ; vpunpckhqdq   %xmm15,%xmm14,%xmm1
  DB  196,193,113,105,212                 ; vpunpckhwd    %xmm12,%xmm1,%xmm2
  DB  196,226,121,51,201                  ; vpmovzxwd     %xmm1,%xmm1
  DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
  DB  196,193,116,84,209                  ; vandps        %ymm9,%ymm1,%ymm2
  DB  196,193,116,84,243                  ; vandps        %ymm11,%ymm1,%ymm6
  DB  197,244,87,202                      ; vxorps        %ymm2,%ymm1,%ymm1
  DB  196,227,125,25,247,1                ; vextractf128  $0x1,%ymm6,%xmm7
  DB  196,193,65,118,252                  ; vpcmpeqd      %xmm12,%xmm7,%xmm7
  DB  196,193,73,118,244                  ; vpcmpeqd      %xmm12,%xmm6,%xmm6
  DB  196,99,77,24,247,1                  ; vinsertf128   $0x1,%xmm7,%ymm6,%ymm14
  DB  196,227,125,25,215,1                ; vextractf128  $0x1,%ymm2,%xmm7
  DB  197,193,114,247,16                  ; vpslld        $0x10,%xmm7,%xmm7
  DB  196,227,125,25,206,1                ; vextractf128  $0x1,%ymm1,%xmm6
  DB  197,201,114,246,13                  ; vpslld        $0xd,%xmm6,%xmm6
  DB  197,193,254,251                     ; vpaddd        %xmm3,%xmm7,%xmm7
  DB  197,193,254,246                     ; vpaddd        %xmm6,%xmm7,%xmm6
  DB  197,233,114,242,16                  ; vpslld        $0x10,%xmm2,%xmm2
  DB  197,241,114,241,13                  ; vpslld        $0xd,%xmm1,%xmm1
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  197,233,254,201                     ; vpaddd        %xmm1,%xmm2,%xmm1
  DB  196,227,117,24,206,1                ; vinsertf128   $0x1,%xmm6,%ymm1,%ymm1
  DB  196,195,117,74,205,224              ; vblendvps     %ymm14,%ymm13,%ymm1,%ymm1
  DB  196,193,57,108,210                  ; vpunpcklqdq   %xmm10,%xmm8,%xmm2
  DB  196,193,105,105,244                 ; vpunpckhwd    %xmm12,%xmm2,%xmm6
  DB  196,226,121,51,210                  ; vpmovzxwd     %xmm2,%xmm2
  DB  196,227,109,24,214,1                ; vinsertf128   $0x1,%xmm6,%ymm2,%ymm2
  DB  196,193,108,84,243                  ; vandps        %ymm11,%ymm2,%ymm6
  DB  196,227,125,25,247,1                ; vextractf128  $0x1,%ymm6,%xmm7
  DB  196,193,65,118,252                  ; vpcmpeqd      %xmm12,%xmm7,%xmm7
  DB  196,193,73,118,244                  ; vpcmpeqd      %xmm12,%xmm6,%xmm6
  DB  196,99,77,24,247,1                  ; vinsertf128   $0x1,%xmm7,%ymm6,%ymm14
  DB  196,193,108,84,249                  ; vandps        %ymm9,%ymm2,%ymm7
  DB  197,236,87,215                      ; vxorps        %ymm7,%ymm2,%ymm2
  DB  196,227,125,25,254,1                ; vextractf128  $0x1,%ymm7,%xmm6
  DB  197,129,114,246,16                  ; vpslld        $0x10,%xmm6,%xmm15
  DB  196,227,125,25,214,1                ; vextractf128  $0x1,%ymm2,%xmm6
  DB  197,209,114,246,13                  ; vpslld        $0xd,%xmm6,%xmm5
  DB  197,129,254,243                     ; vpaddd        %xmm3,%xmm15,%xmm6
  DB  197,201,254,237                     ; vpaddd        %xmm5,%xmm6,%xmm5
  DB  197,201,114,247,16                  ; vpslld        $0x10,%xmm7,%xmm6
  DB  197,233,114,242,13                  ; vpslld        $0xd,%xmm2,%xmm2
  DB  197,201,254,243                     ; vpaddd        %xmm3,%xmm6,%xmm6
  DB  197,201,254,210                     ; vpaddd        %xmm2,%xmm6,%xmm2
  DB  196,227,109,24,213,1                ; vinsertf128   $0x1,%xmm5,%ymm2,%ymm2
  DB  196,195,109,74,213,224              ; vblendvps     %ymm14,%ymm13,%ymm2,%ymm2
  DB  196,193,57,109,234                  ; vpunpckhqdq   %xmm10,%xmm8,%xmm5
  DB  196,193,81,105,244                  ; vpunpckhwd    %xmm12,%xmm5,%xmm6
  DB  196,226,121,51,237                  ; vpmovzxwd     %xmm5,%xmm5
  DB  196,227,85,24,238,1                 ; vinsertf128   $0x1,%xmm6,%ymm5,%ymm5
  DB  196,193,84,84,243                   ; vandps        %ymm11,%ymm5,%ymm6
  DB  196,227,125,25,247,1                ; vextractf128  $0x1,%ymm6,%xmm7
  DB  196,193,65,118,252                  ; vpcmpeqd      %xmm12,%xmm7,%xmm7
  DB  196,193,73,118,244                  ; vpcmpeqd      %xmm12,%xmm6,%xmm6
  DB  196,65,84,84,193                    ; vandps        %ymm9,%ymm5,%ymm8
  DB  196,193,84,87,232                   ; vxorps        %ymm8,%ymm5,%ymm5
  DB  196,99,77,24,207,1                  ; vinsertf128   $0x1,%xmm7,%ymm6,%ymm9
  DB  196,99,125,25,199,1                 ; vextractf128  $0x1,%ymm8,%xmm7
  DB  197,193,114,247,16                  ; vpslld        $0x10,%xmm7,%xmm7
  DB  196,193,73,114,240,16               ; vpslld        $0x10,%xmm8,%xmm6
  DB  197,201,254,243                     ; vpaddd        %xmm3,%xmm6,%xmm6
  DB  197,193,254,219                     ; vpaddd        %xmm3,%xmm7,%xmm3
  DB  196,227,125,25,239,1                ; vextractf128  $0x1,%ymm5,%xmm7
  DB  197,193,114,247,13                  ; vpslld        $0xd,%xmm7,%xmm7
  DB  197,225,254,223                     ; vpaddd        %xmm7,%xmm3,%xmm3
  DB  197,209,114,245,13                  ; vpslld        $0xd,%xmm5,%xmm5
  DB  197,201,254,237                     ; vpaddd        %xmm5,%xmm6,%xmm5
  DB  196,227,85,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm5,%ymm3
  DB  196,195,101,74,221,144              ; vblendvps     %ymm9,%ymm13,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,16,44,36                    ; vmovups       (%rsp),%ymm5
  DB  197,252,16,116,36,32                ; vmovups       0x20(%rsp),%ymm6
  DB  197,252,16,124,36,64                ; vmovups       0x40(%rsp),%ymm7
  DB  72,131,196,120                      ; add           $0x78,%rsp
  DB  255,224                             ; jmpq          *%rax
  DB  197,123,16,4,248                    ; vmovsd        (%rax,%rdi,8),%xmm8
  DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,79                              ; je            45da <_sk_load_f16_avx+0x2ea>
  DB  197,57,22,68,248,8                  ; vmovhpd       0x8(%rax,%rdi,8),%xmm8,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,67                              ; jb            45da <_sk_load_f16_avx+0x2ea>
  DB  197,251,16,84,248,16                ; vmovsd        0x10(%rax,%rdi,8),%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  116,68                              ; je            45e7 <_sk_load_f16_avx+0x2f7>
  DB  197,233,22,84,248,24                ; vmovhpd       0x18(%rax,%rdi,8),%xmm2,%xmm2
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,56                              ; jb            45e7 <_sk_load_f16_avx+0x2f7>
  DB  197,251,16,76,248,32                ; vmovsd        0x20(%rax,%rdi,8),%xmm1
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  15,132,107,253,255,255              ; je            432a <_sk_load_f16_avx+0x3a>
  DB  197,241,22,76,248,40                ; vmovhpd       0x28(%rax,%rdi,8),%xmm1,%xmm1
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  15,130,91,253,255,255               ; jb            432a <_sk_load_f16_avx+0x3a>
  DB  197,122,126,76,248,48               ; vmovq         0x30(%rax,%rdi,8),%xmm9
  DB  233,80,253,255,255                  ; jmpq          432a <_sk_load_f16_avx+0x3a>
  DB  197,241,87,201                      ; vxorpd        %xmm1,%xmm1,%xmm1
  DB  197,233,87,210                      ; vxorpd        %xmm2,%xmm2,%xmm2
  DB  233,67,253,255,255                  ; jmpq          432a <_sk_load_f16_avx+0x3a>
  DB  197,241,87,201                      ; vxorpd        %xmm1,%xmm1,%xmm1
  DB  233,58,253,255,255                  ; jmpq          432a <_sk_load_f16_avx+0x3a>

PUBLIC _sk_gather_f16_avx
_sk_gather_f16_avx LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,84                               ; push          %r12
  DB  83                                  ; push          %rbx
  DB  72,129,236,152,0,0,0                ; sub           $0x98,%rsp
  DB  197,252,17,124,36,96                ; vmovups       %ymm7,0x60(%rsp)
  DB  197,252,17,116,36,64                ; vmovups       %ymm6,0x40(%rsp)
  DB  197,252,17,108,36,32                ; vmovups       %ymm5,0x20(%rsp)
  DB  197,252,17,36,36                    ; vmovups       %ymm4,(%rsp)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  197,254,91,209                      ; vcvttps2dq    %ymm1,%ymm2
  DB  197,249,110,72,16                   ; vmovd         0x10(%rax),%xmm1
  DB  197,249,112,217,0                   ; vpshufd       $0x0,%xmm1,%xmm3
  DB  196,226,97,64,202                   ; vpmulld       %xmm2,%xmm3,%xmm1
  DB  196,227,125,25,210,1                ; vextractf128  $0x1,%ymm2,%xmm2
  DB  196,226,97,64,210                   ; vpmulld       %xmm2,%xmm3,%xmm2
  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  196,227,249,22,208,1                ; vpextrq       $0x1,%xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  196,193,249,126,210                 ; vmovq         %xmm2,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  197,241,254,192                     ; vpaddd        %xmm0,%xmm1,%xmm0
  DB  196,225,249,126,195                 ; vmovq         %xmm0,%rbx
  DB  65,137,222                          ; mov           %ebx,%r14d
  DB  196,195,249,22,199,1                ; vpextrq       $0x1,%xmm0,%r15
  DB  69,137,252                          ; mov           %r15d,%r12d
  DB  73,193,239,32                       ; shr           $0x20,%r15
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  196,193,122,126,4,216               ; vmovq         (%r8,%rbx,8),%xmm0
  DB  196,129,122,126,12,240              ; vmovq         (%r8,%r14,8),%xmm1
  DB  197,113,108,200                     ; vpunpcklqdq   %xmm0,%xmm1,%xmm9
  DB  196,129,122,126,12,248              ; vmovq         (%r8,%r15,8),%xmm1
  DB  196,129,122,126,20,224              ; vmovq         (%r8,%r12,8),%xmm2
  DB  197,233,108,201                     ; vpunpcklqdq   %xmm1,%xmm2,%xmm1
  DB  196,129,122,126,20,208              ; vmovq         (%r8,%r10,8),%xmm2
  DB  196,129,122,126,28,216              ; vmovq         (%r8,%r11,8),%xmm3
  DB  197,97,108,210                      ; vpunpcklqdq   %xmm2,%xmm3,%xmm10
  DB  196,65,122,126,4,192                ; vmovq         (%r8,%rax,8),%xmm8
  DB  196,129,122,126,28,200              ; vmovq         (%r8,%r9,8),%xmm3
  DB  196,193,97,108,216                  ; vpunpcklqdq   %xmm8,%xmm3,%xmm3
  DB  197,177,97,193                      ; vpunpcklwd    %xmm1,%xmm9,%xmm0
  DB  197,177,105,201                     ; vpunpckhwd    %xmm1,%xmm9,%xmm1
  DB  197,169,97,211                      ; vpunpcklwd    %xmm3,%xmm10,%xmm2
  DB  197,169,105,219                     ; vpunpckhwd    %xmm3,%xmm10,%xmm3
  DB  197,121,97,241                      ; vpunpcklwd    %xmm1,%xmm0,%xmm14
  DB  197,121,105,193                     ; vpunpckhwd    %xmm1,%xmm0,%xmm8
  DB  197,105,97,251                      ; vpunpcklwd    %xmm3,%xmm2,%xmm15
  DB  197,105,105,211                     ; vpunpckhwd    %xmm3,%xmm2,%xmm10
  DB  196,193,9,108,199                   ; vpunpcklqdq   %xmm15,%xmm14,%xmm0
  DB  196,65,25,239,228                   ; vpxor         %xmm12,%xmm12,%xmm12
  DB  196,193,121,105,212                 ; vpunpckhwd    %xmm12,%xmm0,%xmm2
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,194,1                ; vinsertf128   $0x1,%xmm2,%ymm0,%ymm0
  DB  184,0,128,0,0                       ; mov           $0x8000,%eax
  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
  DB  197,249,112,210,0                   ; vpshufd       $0x0,%xmm2,%xmm2
  DB  196,99,109,24,202,1                 ; vinsertf128   $0x1,%xmm2,%ymm2,%ymm9
  DB  196,193,124,84,209                  ; vandps        %ymm9,%ymm0,%ymm2
  DB  184,0,124,0,0                       ; mov           $0x7c00,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  196,99,101,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm11
  DB  196,193,124,84,219                  ; vandps        %ymm11,%ymm0,%ymm3
  DB  197,252,87,194                      ; vxorps        %ymm2,%ymm0,%ymm0
  DB  196,227,125,25,217,1                ; vextractf128  $0x1,%ymm3,%xmm1
  DB  196,193,113,118,204                 ; vpcmpeqd      %xmm12,%xmm1,%xmm1
  DB  196,193,97,118,220                  ; vpcmpeqd      %xmm12,%xmm3,%xmm3
  DB  196,227,101,24,225,1                ; vinsertf128   $0x1,%xmm1,%ymm3,%ymm4
  DB  196,227,125,25,211,1                ; vextractf128  $0x1,%ymm2,%xmm3
  DB  197,145,114,243,16                  ; vpslld        $0x10,%xmm3,%xmm13
  DB  196,227,125,25,195,1                ; vextractf128  $0x1,%ymm0,%xmm3
  DB  197,241,114,243,13                  ; vpslld        $0xd,%xmm3,%xmm1
  DB  184,0,0,0,56                        ; mov           $0x38000000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  197,249,112,219,0                   ; vpshufd       $0x0,%xmm3,%xmm3
  DB  197,145,254,251                     ; vpaddd        %xmm3,%xmm13,%xmm7
  DB  197,193,254,201                     ; vpaddd        %xmm1,%xmm7,%xmm1
  DB  197,233,114,242,16                  ; vpslld        $0x10,%xmm2,%xmm2
  DB  197,249,114,240,13                  ; vpslld        $0xd,%xmm0,%xmm0
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  197,233,254,192                     ; vpaddd        %xmm0,%xmm2,%xmm0
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  196,65,20,87,237                    ; vxorps        %ymm13,%ymm13,%ymm13
  DB  196,195,125,74,197,64               ; vblendvps     %ymm4,%ymm13,%ymm0,%ymm0
  DB  196,193,9,109,207                   ; vpunpckhqdq   %xmm15,%xmm14,%xmm1
  DB  196,193,113,105,212                 ; vpunpckhwd    %xmm12,%xmm1,%xmm2
  DB  196,226,121,51,201                  ; vpmovzxwd     %xmm1,%xmm1
  DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
  DB  196,193,116,84,209                  ; vandps        %ymm9,%ymm1,%ymm2
  DB  196,193,116,84,227                  ; vandps        %ymm11,%ymm1,%ymm4
  DB  197,244,87,202                      ; vxorps        %ymm2,%ymm1,%ymm1
  DB  196,227,125,25,231,1                ; vextractf128  $0x1,%ymm4,%xmm7
  DB  196,193,65,118,252                  ; vpcmpeqd      %xmm12,%xmm7,%xmm7
  DB  196,193,89,118,228                  ; vpcmpeqd      %xmm12,%xmm4,%xmm4
  DB  196,227,93,24,231,1                 ; vinsertf128   $0x1,%xmm7,%ymm4,%ymm4
  DB  196,227,125,25,215,1                ; vextractf128  $0x1,%ymm2,%xmm7
  DB  197,193,114,247,16                  ; vpslld        $0x10,%xmm7,%xmm7
  DB  196,227,125,25,206,1                ; vextractf128  $0x1,%ymm1,%xmm6
  DB  197,201,114,246,13                  ; vpslld        $0xd,%xmm6,%xmm6
  DB  197,193,254,251                     ; vpaddd        %xmm3,%xmm7,%xmm7
  DB  197,193,254,246                     ; vpaddd        %xmm6,%xmm7,%xmm6
  DB  197,233,114,242,16                  ; vpslld        $0x10,%xmm2,%xmm2
  DB  197,241,114,241,13                  ; vpslld        $0xd,%xmm1,%xmm1
  DB  197,233,254,211                     ; vpaddd        %xmm3,%xmm2,%xmm2
  DB  197,233,254,201                     ; vpaddd        %xmm1,%xmm2,%xmm1
  DB  196,227,117,24,206,1                ; vinsertf128   $0x1,%xmm6,%ymm1,%ymm1
  DB  196,195,117,74,205,64               ; vblendvps     %ymm4,%ymm13,%ymm1,%ymm1
  DB  196,193,57,108,210                  ; vpunpcklqdq   %xmm10,%xmm8,%xmm2
  DB  196,193,105,105,228                 ; vpunpckhwd    %xmm12,%xmm2,%xmm4
  DB  196,226,121,51,210                  ; vpmovzxwd     %xmm2,%xmm2
  DB  196,227,109,24,212,1                ; vinsertf128   $0x1,%xmm4,%ymm2,%ymm2
  DB  196,193,108,84,227                  ; vandps        %ymm11,%ymm2,%ymm4
  DB  196,227,125,25,230,1                ; vextractf128  $0x1,%ymm4,%xmm6
  DB  196,193,73,118,244                  ; vpcmpeqd      %xmm12,%xmm6,%xmm6
  DB  196,193,89,118,228                  ; vpcmpeqd      %xmm12,%xmm4,%xmm4
  DB  196,227,93,24,230,1                 ; vinsertf128   $0x1,%xmm6,%ymm4,%ymm4
  DB  196,193,108,84,241                  ; vandps        %ymm9,%ymm2,%ymm6
  DB  197,236,87,214                      ; vxorps        %ymm6,%ymm2,%ymm2
  DB  196,227,125,25,247,1                ; vextractf128  $0x1,%ymm6,%xmm7
  DB  197,193,114,247,16                  ; vpslld        $0x10,%xmm7,%xmm7
  DB  196,227,125,25,213,1                ; vextractf128  $0x1,%ymm2,%xmm5
  DB  197,209,114,245,13                  ; vpslld        $0xd,%xmm5,%xmm5
  DB  197,193,254,251                     ; vpaddd        %xmm3,%xmm7,%xmm7
  DB  197,193,254,237                     ; vpaddd        %xmm5,%xmm7,%xmm5
  DB  197,201,114,246,16                  ; vpslld        $0x10,%xmm6,%xmm6
  DB  197,233,114,242,13                  ; vpslld        $0xd,%xmm2,%xmm2
  DB  197,201,254,243                     ; vpaddd        %xmm3,%xmm6,%xmm6
  DB  197,201,254,210                     ; vpaddd        %xmm2,%xmm6,%xmm2
  DB  196,227,109,24,213,1                ; vinsertf128   $0x1,%xmm5,%ymm2,%ymm2
  DB  196,195,109,74,213,64               ; vblendvps     %ymm4,%ymm13,%ymm2,%ymm2
  DB  196,193,57,109,226                  ; vpunpckhqdq   %xmm10,%xmm8,%xmm4
  DB  196,193,89,105,236                  ; vpunpckhwd    %xmm12,%xmm4,%xmm5
  DB  196,226,121,51,228                  ; vpmovzxwd     %xmm4,%xmm4
  DB  196,227,93,24,229,1                 ; vinsertf128   $0x1,%xmm5,%ymm4,%ymm4
  DB  196,193,92,84,235                   ; vandps        %ymm11,%ymm4,%ymm5
  DB  196,227,125,25,238,1                ; vextractf128  $0x1,%ymm5,%xmm6
  DB  196,193,73,118,244                  ; vpcmpeqd      %xmm12,%xmm6,%xmm6
  DB  196,193,81,118,236                  ; vpcmpeqd      %xmm12,%xmm5,%xmm5
  DB  196,193,92,84,249                   ; vandps        %ymm9,%ymm4,%ymm7
  DB  197,220,87,231                      ; vxorps        %ymm7,%ymm4,%ymm4
  DB  196,227,85,24,238,1                 ; vinsertf128   $0x1,%xmm6,%ymm5,%ymm5
  DB  196,227,125,25,254,1                ; vextractf128  $0x1,%ymm7,%xmm6
  DB  197,201,114,246,16                  ; vpslld        $0x10,%xmm6,%xmm6
  DB  197,193,114,247,16                  ; vpslld        $0x10,%xmm7,%xmm7
  DB  197,193,254,251                     ; vpaddd        %xmm3,%xmm7,%xmm7
  DB  197,201,254,219                     ; vpaddd        %xmm3,%xmm6,%xmm3
  DB  196,227,125,25,230,1                ; vextractf128  $0x1,%ymm4,%xmm6
  DB  197,201,114,246,13                  ; vpslld        $0xd,%xmm6,%xmm6
  DB  197,225,254,222                     ; vpaddd        %xmm6,%xmm3,%xmm3
  DB  197,217,114,244,13                  ; vpslld        $0xd,%xmm4,%xmm4
  DB  197,193,254,228                     ; vpaddd        %xmm4,%xmm7,%xmm4
  DB  196,227,93,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm4,%ymm3
  DB  196,195,101,74,221,80               ; vblendvps     %ymm5,%ymm13,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,16,36,36                    ; vmovups       (%rsp),%ymm4
  DB  197,252,16,108,36,32                ; vmovups       0x20(%rsp),%ymm5
  DB  197,252,16,116,36,64                ; vmovups       0x40(%rsp),%ymm6
  DB  197,252,16,124,36,96                ; vmovups       0x60(%rsp),%ymm7
  DB  72,129,196,152,0,0,0                ; add           $0x98,%rsp
  DB  91                                  ; pop           %rbx
  DB  65,92                               ; pop           %r12
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_f16_avx
_sk_store_f16_avx LABEL PROC
  DB  72,129,236,216,0,0,0                ; sub           $0xd8,%rsp
  DB  197,252,17,188,36,160,0,0,0         ; vmovups       %ymm7,0xa0(%rsp)
  DB  197,252,17,180,36,128,0,0,0         ; vmovups       %ymm6,0x80(%rsp)
  DB  197,252,17,108,36,96                ; vmovups       %ymm5,0x60(%rsp)
  DB  197,252,17,100,36,64                ; vmovups       %ymm4,0x40(%rsp)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  184,0,0,0,128                       ; mov           $0x80000000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,65,121,112,192,0                ; vpshufd       $0x0,%xmm8,%xmm8
  DB  196,67,61,24,200,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm9
  DB  197,52,84,208                       ; vandps        %ymm0,%ymm9,%ymm10
  DB  197,252,17,4,36                     ; vmovups       %ymm0,(%rsp)
  DB  196,65,124,87,218                   ; vxorps        %ymm10,%ymm0,%ymm11
  DB  184,0,0,128,56                      ; mov           $0x38800000,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  196,65,36,194,224,1                 ; vcmpltps      %ymm8,%ymm11,%ymm12
  DB  196,67,125,25,213,1                 ; vextractf128  $0x1,%ymm10,%xmm13
  DB  196,193,17,114,213,16               ; vpsrld        $0x10,%xmm13,%xmm13
  DB  196,193,9,114,210,16                ; vpsrld        $0x10,%xmm10,%xmm14
  DB  196,193,1,114,211,13                ; vpsrld        $0xd,%xmm11,%xmm15
  DB  196,67,125,25,218,1                 ; vextractf128  $0x1,%ymm11,%xmm10
  DB  196,193,33,114,210,13               ; vpsrld        $0xd,%xmm10,%xmm11
  DB  184,0,192,1,0                       ; mov           $0x1c000,%eax
  DB  197,121,110,208                     ; vmovd         %eax,%xmm10
  DB  196,65,121,112,210,0                ; vpshufd       $0x0,%xmm10,%xmm10
  DB  196,65,9,250,242                    ; vpsubd        %xmm10,%xmm14,%xmm14
  DB  196,65,17,250,234                   ; vpsubd        %xmm10,%xmm13,%xmm13
  DB  196,65,17,254,219                   ; vpaddd        %xmm11,%xmm13,%xmm11
  DB  196,65,9,254,239                    ; vpaddd        %xmm15,%xmm14,%xmm13
  DB  196,67,21,24,235,1                  ; vinsertf128   $0x1,%xmm11,%ymm13,%ymm13
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  196,99,21,74,224,192                ; vblendvps     %ymm12,%ymm0,%ymm13,%ymm12
  DB  197,52,84,233                       ; vandps        %ymm1,%ymm9,%ymm13
  DB  197,252,17,76,36,32                 ; vmovups       %ymm1,0x20(%rsp)
  DB  196,65,116,87,245                   ; vxorps        %ymm13,%ymm1,%ymm14
  DB  196,67,125,25,239,1                 ; vextractf128  $0x1,%ymm13,%xmm15
  DB  196,193,1,114,215,16                ; vpsrld        $0x10,%xmm15,%xmm15
  DB  196,67,125,25,243,1                 ; vextractf128  $0x1,%ymm14,%xmm11
  DB  196,193,33,114,211,13               ; vpsrld        $0xd,%xmm11,%xmm11
  DB  196,193,1,250,250                   ; vpsubd        %xmm10,%xmm15,%xmm7
  DB  196,193,65,254,251                  ; vpaddd        %xmm11,%xmm7,%xmm7
  DB  196,193,73,114,213,16               ; vpsrld        $0x10,%xmm13,%xmm6
  DB  196,193,73,250,242                  ; vpsubd        %xmm10,%xmm6,%xmm6
  DB  196,193,81,114,214,13               ; vpsrld        $0xd,%xmm14,%xmm5
  DB  197,201,254,237                     ; vpaddd        %xmm5,%xmm6,%xmm5
  DB  196,193,12,194,240,1                ; vcmpltps      %ymm8,%ymm14,%ymm6
  DB  196,227,85,24,239,1                 ; vinsertf128   $0x1,%xmm7,%ymm5,%ymm5
  DB  196,99,85,74,232,96                 ; vblendvps     %ymm6,%ymm0,%ymm5,%ymm13
  DB  197,180,84,234                      ; vandps        %ymm2,%ymm9,%ymm5
  DB  196,227,125,25,238,1                ; vextractf128  $0x1,%ymm5,%xmm6
  DB  197,201,114,214,16                  ; vpsrld        $0x10,%xmm6,%xmm6
  DB  197,236,87,253                      ; vxorps        %ymm5,%ymm2,%ymm7
  DB  196,227,125,25,252,1                ; vextractf128  $0x1,%ymm7,%xmm4
  DB  197,217,114,212,13                  ; vpsrld        $0xd,%xmm4,%xmm4
  DB  196,193,73,250,242                  ; vpsubd        %xmm10,%xmm6,%xmm6
  DB  197,201,254,228                     ; vpaddd        %xmm4,%xmm6,%xmm4
  DB  197,209,114,213,16                  ; vpsrld        $0x10,%xmm5,%xmm5
  DB  196,193,81,250,234                  ; vpsubd        %xmm10,%xmm5,%xmm5
  DB  197,201,114,215,13                  ; vpsrld        $0xd,%xmm7,%xmm6
  DB  197,209,254,238                     ; vpaddd        %xmm6,%xmm5,%xmm5
  DB  196,227,85,24,228,1                 ; vinsertf128   $0x1,%xmm4,%ymm5,%ymm4
  DB  196,193,68,194,232,1                ; vcmpltps      %ymm8,%ymm7,%ymm5
  DB  196,227,93,74,224,80                ; vblendvps     %ymm5,%ymm0,%ymm4,%ymm4
  DB  197,180,84,235                      ; vandps        %ymm3,%ymm9,%ymm5
  DB  196,227,125,25,238,1                ; vextractf128  $0x1,%ymm5,%xmm6
  DB  197,201,114,214,16                  ; vpsrld        $0x10,%xmm6,%xmm6
  DB  197,193,114,213,16                  ; vpsrld        $0x10,%xmm5,%xmm7
  DB  196,193,65,250,250                  ; vpsubd        %xmm10,%xmm7,%xmm7
  DB  196,193,73,250,242                  ; vpsubd        %xmm10,%xmm6,%xmm6
  DB  197,228,87,237                      ; vxorps        %ymm5,%ymm3,%ymm5
  DB  196,227,125,25,233,1                ; vextractf128  $0x1,%ymm5,%xmm1
  DB  197,241,114,209,13                  ; vpsrld        $0xd,%xmm1,%xmm1
  DB  197,201,254,201                     ; vpaddd        %xmm1,%xmm6,%xmm1
  DB  196,193,84,194,240,1                ; vcmpltps      %ymm8,%ymm5,%ymm6
  DB  197,209,114,213,13                  ; vpsrld        $0xd,%xmm5,%xmm5
  DB  197,193,254,237                     ; vpaddd        %xmm5,%xmm7,%xmm5
  DB  196,227,85,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm5,%ymm1
  DB  196,227,117,74,192,96               ; vblendvps     %ymm6,%ymm0,%ymm1,%ymm0
  DB  196,99,125,25,225,1                 ; vextractf128  $0x1,%ymm12,%xmm1
  DB  196,226,25,43,201                   ; vpackusdw     %xmm1,%xmm12,%xmm1
  DB  196,99,125,25,237,1                 ; vextractf128  $0x1,%ymm13,%xmm5
  DB  196,226,17,43,237                   ; vpackusdw     %xmm5,%xmm13,%xmm5
  DB  196,227,125,25,230,1                ; vextractf128  $0x1,%ymm4,%xmm6
  DB  196,226,89,43,230                   ; vpackusdw     %xmm6,%xmm4,%xmm4
  DB  196,227,125,25,198,1                ; vextractf128  $0x1,%ymm0,%xmm6
  DB  196,226,121,43,198                  ; vpackusdw     %xmm6,%xmm0,%xmm0
  DB  197,241,97,245                      ; vpunpcklwd    %xmm5,%xmm1,%xmm6
  DB  197,241,105,205                     ; vpunpckhwd    %xmm5,%xmm1,%xmm1
  DB  197,217,97,232                      ; vpunpcklwd    %xmm0,%xmm4,%xmm5
  DB  197,217,105,192                     ; vpunpckhwd    %xmm0,%xmm4,%xmm0
  DB  197,73,98,221                       ; vpunpckldq    %xmm5,%xmm6,%xmm11
  DB  197,73,106,213                      ; vpunpckhdq    %xmm5,%xmm6,%xmm10
  DB  197,113,98,200                      ; vpunpckldq    %xmm0,%xmm1,%xmm9
  DB  197,113,106,192                     ; vpunpckhdq    %xmm0,%xmm1,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,79                              ; jne           4b69 <_sk_store_f16_avx+0x24f>
  DB  196,65,120,17,28,248                ; vmovups       %xmm11,(%r8,%rdi,8)
  DB  196,65,120,17,84,248,16             ; vmovups       %xmm10,0x10(%r8,%rdi,8)
  DB  196,65,120,17,76,248,32             ; vmovups       %xmm9,0x20(%r8,%rdi,8)
  DB  196,65,122,127,68,248,48            ; vmovdqu       %xmm8,0x30(%r8,%rdi,8)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,16,4,36                     ; vmovups       (%rsp),%ymm0
  DB  197,252,16,76,36,32                 ; vmovups       0x20(%rsp),%ymm1
  DB  197,252,16,100,36,64                ; vmovups       0x40(%rsp),%ymm4
  DB  197,252,16,108,36,96                ; vmovups       0x60(%rsp),%ymm5
  DB  197,252,16,180,36,128,0,0,0         ; vmovups       0x80(%rsp),%ymm6
  DB  197,252,16,188,36,160,0,0,0         ; vmovups       0xa0(%rsp),%ymm7
  DB  72,129,196,216,0,0,0                ; add           $0xd8,%rsp
  DB  255,224                             ; jmpq          *%rax
  DB  196,65,121,214,28,248               ; vmovq         %xmm11,(%r8,%rdi,8)
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,192                             ; je            4b35 <_sk_store_f16_avx+0x21b>
  DB  196,65,121,23,92,248,8              ; vmovhpd       %xmm11,0x8(%r8,%rdi,8)
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,179                             ; jb            4b35 <_sk_store_f16_avx+0x21b>
  DB  196,65,121,214,84,248,16            ; vmovq         %xmm10,0x10(%r8,%rdi,8)
  DB  116,170                             ; je            4b35 <_sk_store_f16_avx+0x21b>
  DB  196,65,121,23,84,248,24             ; vmovhpd       %xmm10,0x18(%r8,%rdi,8)
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,157                             ; jb            4b35 <_sk_store_f16_avx+0x21b>
  DB  196,65,121,214,76,248,32            ; vmovq         %xmm9,0x20(%r8,%rdi,8)
  DB  116,148                             ; je            4b35 <_sk_store_f16_avx+0x21b>
  DB  196,65,121,23,76,248,40             ; vmovhpd       %xmm9,0x28(%r8,%rdi,8)
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,135                             ; jb            4b35 <_sk_store_f16_avx+0x21b>
  DB  196,65,121,214,68,248,48            ; vmovq         %xmm8,0x30(%r8,%rdi,8)
  DB  233,123,255,255,255                 ; jmpq          4b35 <_sk_store_f16_avx+0x21b>

PUBLIC _sk_load_u16_be_avx
_sk_load_u16_be_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  72,141,4,189,0,0,0,0                ; lea           0x0(,%rdi,4),%rax
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,5,1,0,0                      ; jne           4cd5 <_sk_load_u16_be_avx+0x11b>
  DB  196,65,121,16,4,64                  ; vmovupd       (%r8,%rax,2),%xmm8
  DB  196,193,121,16,84,64,16             ; vmovupd       0x10(%r8,%rax,2),%xmm2
  DB  196,193,121,16,92,64,32             ; vmovupd       0x20(%r8,%rax,2),%xmm3
  DB  196,65,122,111,76,64,48             ; vmovdqu       0x30(%r8,%rax,2),%xmm9
  DB  197,185,97,194                      ; vpunpcklwd    %xmm2,%xmm8,%xmm0
  DB  197,185,105,210                     ; vpunpckhwd    %xmm2,%xmm8,%xmm2
  DB  196,193,97,97,201                   ; vpunpcklwd    %xmm9,%xmm3,%xmm1
  DB  196,193,97,105,217                  ; vpunpckhwd    %xmm9,%xmm3,%xmm3
  DB  197,121,97,210                      ; vpunpcklwd    %xmm2,%xmm0,%xmm10
  DB  197,121,105,194                     ; vpunpckhwd    %xmm2,%xmm0,%xmm8
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,113,105,203                     ; vpunpckhwd    %xmm3,%xmm1,%xmm9
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,224,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm12
  DB  197,169,108,194                     ; vpunpcklqdq   %xmm2,%xmm10,%xmm0
  DB  197,241,113,240,8                   ; vpsllw        $0x8,%xmm0,%xmm1
  DB  197,249,113,208,8                   ; vpsrlw        $0x8,%xmm0,%xmm0
  DB  197,241,235,192                     ; vpor          %xmm0,%xmm1,%xmm0
  DB  196,65,33,239,219                   ; vpxor         %xmm11,%xmm11,%xmm11
  DB  196,193,121,105,203                 ; vpunpckhwd    %xmm11,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  197,156,89,192                      ; vmulps        %ymm0,%ymm12,%ymm0
  DB  197,169,109,202                     ; vpunpckhqdq   %xmm2,%xmm10,%xmm1
  DB  197,233,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm2
  DB  197,241,113,209,8                   ; vpsrlw        $0x8,%xmm1,%xmm1
  DB  197,233,235,201                     ; vpor          %xmm1,%xmm2,%xmm1
  DB  196,193,113,105,211                 ; vpunpckhwd    %xmm11,%xmm1,%xmm2
  DB  196,226,121,51,201                  ; vpmovzxwd     %xmm1,%xmm1
  DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,156,89,201                      ; vmulps        %ymm1,%ymm12,%ymm1
  DB  196,193,57,108,209                  ; vpunpcklqdq   %xmm9,%xmm8,%xmm2
  DB  197,169,113,242,8                   ; vpsllw        $0x8,%xmm2,%xmm10
  DB  197,233,113,210,8                   ; vpsrlw        $0x8,%xmm2,%xmm2
  DB  197,169,235,210                     ; vpor          %xmm2,%xmm10,%xmm2
  DB  196,65,105,105,211                  ; vpunpckhwd    %xmm11,%xmm2,%xmm10
  DB  196,226,121,51,210                  ; vpmovzxwd     %xmm2,%xmm2
  DB  196,195,109,24,210,1                ; vinsertf128   $0x1,%xmm10,%ymm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,156,89,210                      ; vmulps        %ymm2,%ymm12,%ymm2
  DB  196,193,57,109,217                  ; vpunpckhqdq   %xmm9,%xmm8,%xmm3
  DB  197,185,113,243,8                   ; vpsllw        $0x8,%xmm3,%xmm8
  DB  197,225,113,211,8                   ; vpsrlw        $0x8,%xmm3,%xmm3
  DB  197,185,235,219                     ; vpor          %xmm3,%xmm8,%xmm3
  DB  196,65,97,105,195                   ; vpunpckhwd    %xmm11,%xmm3,%xmm8
  DB  196,226,121,51,219                  ; vpmovzxwd     %xmm3,%xmm3
  DB  196,195,101,24,216,1                ; vinsertf128   $0x1,%xmm8,%ymm3,%ymm3
  DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
  DB  197,156,89,219                      ; vmulps        %ymm3,%ymm12,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,65,123,16,4,64                  ; vmovsd        (%r8,%rax,2),%xmm8
  DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,85                              ; je            4d3b <_sk_load_u16_be_avx+0x181>
  DB  196,65,57,22,68,64,8                ; vmovhpd       0x8(%r8,%rax,2),%xmm8,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,72                              ; jb            4d3b <_sk_load_u16_be_avx+0x181>
  DB  196,193,123,16,84,64,16             ; vmovsd        0x10(%r8,%rax,2),%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  116,72                              ; je            4d48 <_sk_load_u16_be_avx+0x18e>
  DB  196,193,105,22,84,64,24             ; vmovhpd       0x18(%r8,%rax,2),%xmm2,%xmm2
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,59                              ; jb            4d48 <_sk_load_u16_be_avx+0x18e>
  DB  196,193,123,16,92,64,32             ; vmovsd        0x20(%r8,%rax,2),%xmm3
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  15,132,205,254,255,255              ; je            4beb <_sk_load_u16_be_avx+0x31>
  DB  196,193,97,22,92,64,40              ; vmovhpd       0x28(%r8,%rax,2),%xmm3,%xmm3
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  15,130,188,254,255,255              ; jb            4beb <_sk_load_u16_be_avx+0x31>
  DB  196,65,122,126,76,64,48             ; vmovq         0x30(%r8,%rax,2),%xmm9
  DB  233,176,254,255,255                 ; jmpq          4beb <_sk_load_u16_be_avx+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  197,233,87,210                      ; vxorpd        %xmm2,%xmm2,%xmm2
  DB  233,163,254,255,255                 ; jmpq          4beb <_sk_load_u16_be_avx+0x31>
  DB  197,225,87,219                      ; vxorpd        %xmm3,%xmm3,%xmm3
  DB  233,154,254,255,255                 ; jmpq          4beb <_sk_load_u16_be_avx+0x31>

PUBLIC _sk_load_rgb_u16_be_avx
_sk_load_rgb_u16_be_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  72,141,4,127                        ; lea           (%rdi,%rdi,2),%rax
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,133,8,1,0,0                      ; jne           4e6b <_sk_load_rgb_u16_be_avx+0x11a>
  DB  196,193,122,111,4,64                ; vmovdqu       (%r8,%rax,2),%xmm0
  DB  196,193,122,111,84,64,12            ; vmovdqu       0xc(%r8,%rax,2),%xmm2
  DB  196,193,122,111,76,64,24            ; vmovdqu       0x18(%r8,%rax,2),%xmm1
  DB  196,193,122,111,92,64,32            ; vmovdqu       0x20(%r8,%rax,2),%xmm3
  DB  197,225,115,219,4                   ; vpsrldq       $0x4,%xmm3,%xmm3
  DB  197,185,115,216,6                   ; vpsrldq       $0x6,%xmm0,%xmm8
  DB  197,177,115,218,6                   ; vpsrldq       $0x6,%xmm2,%xmm9
  DB  197,161,115,217,6                   ; vpsrldq       $0x6,%xmm1,%xmm11
  DB  197,169,115,219,6                   ; vpsrldq       $0x6,%xmm3,%xmm10
  DB  197,249,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm0
  DB  196,193,57,97,209                   ; vpunpcklwd    %xmm9,%xmm8,%xmm2
  DB  197,241,97,203                      ; vpunpcklwd    %xmm3,%xmm1,%xmm1
  DB  196,193,33,97,218                   ; vpunpcklwd    %xmm10,%xmm11,%xmm3
  DB  197,121,97,194                      ; vpunpcklwd    %xmm2,%xmm0,%xmm8
  DB  197,121,105,202                     ; vpunpckhwd    %xmm2,%xmm0,%xmm9
  DB  197,241,97,211                      ; vpunpcklwd    %xmm3,%xmm1,%xmm2
  DB  197,113,105,219                     ; vpunpckhwd    %xmm3,%xmm1,%xmm11
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  197,249,110,192                     ; vmovd         %eax,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,208,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm10
  DB  197,185,108,194                     ; vpunpcklqdq   %xmm2,%xmm8,%xmm0
  DB  197,241,113,240,8                   ; vpsllw        $0x8,%xmm0,%xmm1
  DB  197,249,113,208,8                   ; vpsrlw        $0x8,%xmm0,%xmm0
  DB  197,241,235,192                     ; vpor          %xmm0,%xmm1,%xmm0
  DB  196,65,25,239,228                   ; vpxor         %xmm12,%xmm12,%xmm12
  DB  196,193,121,105,204                 ; vpunpckhwd    %xmm12,%xmm0,%xmm1
  DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
  DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
  DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  197,185,109,202                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm1
  DB  197,233,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm2
  DB  197,241,113,209,8                   ; vpsrlw        $0x8,%xmm1,%xmm1
  DB  197,233,235,201                     ; vpor          %xmm1,%xmm2,%xmm1
  DB  196,193,113,105,212                 ; vpunpckhwd    %xmm12,%xmm1,%xmm2
  DB  196,226,121,51,201                  ; vpmovzxwd     %xmm1,%xmm1
  DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
  DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  196,193,49,108,211                  ; vpunpcklqdq   %xmm11,%xmm9,%xmm2
  DB  197,225,113,242,8                   ; vpsllw        $0x8,%xmm2,%xmm3
  DB  197,233,113,210,8                   ; vpsrlw        $0x8,%xmm2,%xmm2
  DB  197,225,235,210                     ; vpor          %xmm2,%xmm3,%xmm2
  DB  196,193,105,105,220                 ; vpunpckhwd    %xmm12,%xmm2,%xmm3
  DB  196,226,121,51,210                  ; vpmovzxwd     %xmm2,%xmm2
  DB  196,227,109,24,211,1                ; vinsertf128   $0x1,%xmm3,%ymm2,%ymm2
  DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
  DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,193,121,110,4,64                ; vmovd         (%r8,%rax,2),%xmm0
  DB  196,193,121,196,68,64,4,2           ; vpinsrw       $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  117,5                               ; jne           4e84 <_sk_load_rgb_u16_be_avx+0x133>
  DB  233,19,255,255,255                  ; jmpq          4d97 <_sk_load_rgb_u16_be_avx+0x46>
  DB  196,193,121,110,76,64,6             ; vmovd         0x6(%r8,%rax,2),%xmm1
  DB  196,65,113,196,68,64,10,2           ; vpinsrw       $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,26                              ; jb            4eb3 <_sk_load_rgb_u16_be_avx+0x162>
  DB  196,193,121,110,76,64,12            ; vmovd         0xc(%r8,%rax,2),%xmm1
  DB  196,193,113,196,84,64,16,2          ; vpinsrw       $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  117,10                              ; jne           4eb8 <_sk_load_rgb_u16_be_avx+0x167>
  DB  233,228,254,255,255                 ; jmpq          4d97 <_sk_load_rgb_u16_be_avx+0x46>
  DB  233,223,254,255,255                 ; jmpq          4d97 <_sk_load_rgb_u16_be_avx+0x46>
  DB  196,193,121,110,76,64,18            ; vmovd         0x12(%r8,%rax,2),%xmm1
  DB  196,65,113,196,76,64,22,2           ; vpinsrw       $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,26                              ; jb            4ee7 <_sk_load_rgb_u16_be_avx+0x196>
  DB  196,193,121,110,76,64,24            ; vmovd         0x18(%r8,%rax,2),%xmm1
  DB  196,193,113,196,76,64,28,2          ; vpinsrw       $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  117,10                              ; jne           4eec <_sk_load_rgb_u16_be_avx+0x19b>
  DB  233,176,254,255,255                 ; jmpq          4d97 <_sk_load_rgb_u16_be_avx+0x46>
  DB  233,171,254,255,255                 ; jmpq          4d97 <_sk_load_rgb_u16_be_avx+0x46>
  DB  196,193,121,110,92,64,30            ; vmovd         0x1e(%r8,%rax,2),%xmm3
  DB  196,65,97,196,92,64,34,2            ; vpinsrw       $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,20                              ; jb            4f15 <_sk_load_rgb_u16_be_avx+0x1c4>
  DB  196,193,121,110,92,64,36            ; vmovd         0x24(%r8,%rax,2),%xmm3
  DB  196,193,97,196,92,64,40,2           ; vpinsrw       $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
  DB  233,130,254,255,255                 ; jmpq          4d97 <_sk_load_rgb_u16_be_avx+0x46>
  DB  233,125,254,255,255                 ; jmpq          4d97 <_sk_load_rgb_u16_be_avx+0x46>

PUBLIC _sk_store_u16_be_avx
_sk_store_u16_be_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  184,0,255,127,71                    ; mov           $0x477fff00,%eax
  DB  197,121,110,192                     ; vmovd         %eax,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
  DB  196,67,125,25,202,1                 ; vextractf128  $0x1,%ymm9,%xmm10
  DB  196,66,49,43,202                    ; vpackusdw     %xmm10,%xmm9,%xmm9
  DB  196,193,41,113,241,8                ; vpsllw        $0x8,%xmm9,%xmm10
  DB  196,193,49,113,209,8                ; vpsrlw        $0x8,%xmm9,%xmm9
  DB  196,65,41,235,201                   ; vpor          %xmm9,%xmm10,%xmm9
  DB  197,60,89,209                       ; vmulps        %ymm1,%ymm8,%ymm10
  DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
  DB  196,67,125,25,211,1                 ; vextractf128  $0x1,%ymm10,%xmm11
  DB  196,66,41,43,211                    ; vpackusdw     %xmm11,%xmm10,%xmm10
  DB  196,193,33,113,242,8                ; vpsllw        $0x8,%xmm10,%xmm11
  DB  196,193,41,113,210,8                ; vpsrlw        $0x8,%xmm10,%xmm10
  DB  196,65,33,235,210                   ; vpor          %xmm10,%xmm11,%xmm10
  DB  197,60,89,218                       ; vmulps        %ymm2,%ymm8,%ymm11
  DB  196,65,125,91,219                   ; vcvtps2dq     %ymm11,%ymm11
  DB  196,67,125,25,220,1                 ; vextractf128  $0x1,%ymm11,%xmm12
  DB  196,66,33,43,220                    ; vpackusdw     %xmm12,%xmm11,%xmm11
  DB  196,193,25,113,243,8                ; vpsllw        $0x8,%xmm11,%xmm12
  DB  196,193,33,113,211,8                ; vpsrlw        $0x8,%xmm11,%xmm11
  DB  196,65,25,235,219                   ; vpor          %xmm11,%xmm12,%xmm11
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
  DB  196,67,125,25,196,1                 ; vextractf128  $0x1,%ymm8,%xmm12
  DB  196,66,57,43,196                    ; vpackusdw     %xmm12,%xmm8,%xmm8
  DB  196,193,25,113,240,8                ; vpsllw        $0x8,%xmm8,%xmm12
  DB  196,193,57,113,208,8                ; vpsrlw        $0x8,%xmm8,%xmm8
  DB  196,65,25,235,192                   ; vpor          %xmm8,%xmm12,%xmm8
  DB  196,65,49,97,226                    ; vpunpcklwd    %xmm10,%xmm9,%xmm12
  DB  196,65,49,105,234                   ; vpunpckhwd    %xmm10,%xmm9,%xmm13
  DB  196,65,33,97,200                    ; vpunpcklwd    %xmm8,%xmm11,%xmm9
  DB  196,65,33,105,192                   ; vpunpckhwd    %xmm8,%xmm11,%xmm8
  DB  196,65,25,98,217                    ; vpunpckldq    %xmm9,%xmm12,%xmm11
  DB  196,65,25,106,209                   ; vpunpckhdq    %xmm9,%xmm12,%xmm10
  DB  196,65,17,98,200                    ; vpunpckldq    %xmm8,%xmm13,%xmm9
  DB  196,65,17,106,192                   ; vpunpckhdq    %xmm8,%xmm13,%xmm8
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,31                              ; jne           501c <_sk_store_u16_be_avx+0x102>
  DB  196,1,120,17,28,72                  ; vmovups       %xmm11,(%r8,%r9,2)
  DB  196,1,120,17,84,72,16               ; vmovups       %xmm10,0x10(%r8,%r9,2)
  DB  196,1,120,17,76,72,32               ; vmovups       %xmm9,0x20(%r8,%r9,2)
  DB  196,1,122,127,68,72,48              ; vmovdqu       %xmm8,0x30(%r8,%r9,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,1,121,214,28,72                 ; vmovq         %xmm11,(%r8,%r9,2)
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,240                             ; je            5018 <_sk_store_u16_be_avx+0xfe>
  DB  196,1,121,23,92,72,8                ; vmovhpd       %xmm11,0x8(%r8,%r9,2)
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,227                             ; jb            5018 <_sk_store_u16_be_avx+0xfe>
  DB  196,1,121,214,84,72,16              ; vmovq         %xmm10,0x10(%r8,%r9,2)
  DB  116,218                             ; je            5018 <_sk_store_u16_be_avx+0xfe>
  DB  196,1,121,23,84,72,24               ; vmovhpd       %xmm10,0x18(%r8,%r9,2)
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,205                             ; jb            5018 <_sk_store_u16_be_avx+0xfe>
  DB  196,1,121,214,76,72,32              ; vmovq         %xmm9,0x20(%r8,%r9,2)
  DB  116,196                             ; je            5018 <_sk_store_u16_be_avx+0xfe>
  DB  196,1,121,23,76,72,40               ; vmovhpd       %xmm9,0x28(%r8,%r9,2)
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,183                             ; jb            5018 <_sk_store_u16_be_avx+0xfe>
  DB  196,1,121,214,68,72,48              ; vmovq         %xmm8,0x30(%r8,%r9,2)
  DB  235,174                             ; jmp           5018 <_sk_store_u16_be_avx+0xfe>

PUBLIC _sk_load_f32_avx
_sk_load_f32_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  119,110                             ; ja            50e0 <_sk_load_f32_avx+0x76>
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  76,141,12,189,0,0,0,0               ; lea           0x0(,%rdi,4),%r9
  DB  76,141,21,132,0,0,0                 ; lea           0x84(%rip),%r10        # 5108 <_sk_load_f32_avx+0x9e>
  DB  73,99,4,138                         ; movslq        (%r10,%rcx,4),%rax
  DB  76,1,208                            ; add           %r10,%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,3,125,24,68,136,112,1           ; vinsertf128   $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8
  DB  196,131,125,24,92,136,96,1          ; vinsertf128   $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3
  DB  196,131,125,24,76,136,80,1          ; vinsertf128   $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1
  DB  196,131,125,24,84,136,64,1          ; vinsertf128   $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2
  DB  196,129,121,16,68,136,48            ; vmovupd       0x30(%r8,%r9,4),%xmm0
  DB  196,195,125,13,192,12               ; vblendpd      $0xc,%ymm8,%ymm0,%ymm0
  DB  196,1,121,16,68,136,32              ; vmovupd       0x20(%r8,%r9,4),%xmm8
  DB  196,99,61,13,203,12                 ; vblendpd      $0xc,%ymm3,%ymm8,%ymm9
  DB  196,129,121,16,92,136,16            ; vmovupd       0x10(%r8,%r9,4),%xmm3
  DB  196,99,101,13,209,12                ; vblendpd      $0xc,%ymm1,%ymm3,%ymm10
  DB  196,129,121,16,12,136               ; vmovupd       (%r8,%r9,4),%xmm1
  DB  196,227,117,13,202,12               ; vblendpd      $0xc,%ymm2,%ymm1,%ymm1
  DB  196,193,116,20,210                  ; vunpcklps     %ymm10,%ymm1,%ymm2
  DB  196,193,116,21,218                  ; vunpckhps     %ymm10,%ymm1,%ymm3
  DB  197,180,20,200                      ; vunpcklps     %ymm0,%ymm9,%ymm1
  DB  197,52,21,192                       ; vunpckhps     %ymm0,%ymm9,%ymm8
  DB  197,237,20,193                      ; vunpcklpd     %ymm1,%ymm2,%ymm0
  DB  197,237,21,201                      ; vunpckhpd     %ymm1,%ymm2,%ymm1
  DB  196,193,101,20,208                  ; vunpcklpd     %ymm8,%ymm3,%ymm2
  DB  196,193,101,21,216                  ; vunpckhpd     %ymm8,%ymm3,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  133,255                             ; test          %edi,%edi
  DB  255                                 ; (bad)
  DB  255,204                             ; dec           %esp
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  191,255,255,255,178                 ; mov           $0xb2ffffff,%edi
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,165,255,255,255,157             ; jmpq          *-0x62000001(%rbp)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255,149,255,255,255,141             ; callq         *-0x72000001(%rbp)
  DB  255                                 ; (bad)
  DB  255                                 ; (bad)
  DB  255                                 ; .byte         0xff

PUBLIC _sk_store_f32_avx
_sk_store_f32_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  72,141,4,189,0,0,0,0                ; lea           0x0(,%rdi,4),%rax
  DB  197,124,20,193                      ; vunpcklps     %ymm1,%ymm0,%ymm8
  DB  197,124,21,217                      ; vunpckhps     %ymm1,%ymm0,%ymm11
  DB  197,108,20,203                      ; vunpcklps     %ymm3,%ymm2,%ymm9
  DB  197,108,21,227                      ; vunpckhps     %ymm3,%ymm2,%ymm12
  DB  196,65,61,20,209                    ; vunpcklpd     %ymm9,%ymm8,%ymm10
  DB  196,65,61,21,201                    ; vunpckhpd     %ymm9,%ymm8,%ymm9
  DB  196,65,37,20,196                    ; vunpcklpd     %ymm12,%ymm11,%ymm8
  DB  196,65,37,21,220                    ; vunpckhpd     %ymm12,%ymm11,%ymm11
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  117,55                              ; jne           5195 <_sk_store_f32_avx+0x6d>
  DB  196,67,45,24,225,1                  ; vinsertf128   $0x1,%xmm9,%ymm10,%ymm12
  DB  196,67,61,24,235,1                  ; vinsertf128   $0x1,%xmm11,%ymm8,%ymm13
  DB  196,67,45,6,201,49                  ; vperm2f128    $0x31,%ymm9,%ymm10,%ymm9
  DB  196,67,61,6,195,49                  ; vperm2f128    $0x31,%ymm11,%ymm8,%ymm8
  DB  196,65,125,17,36,128                ; vmovupd       %ymm12,(%r8,%rax,4)
  DB  196,65,125,17,108,128,32            ; vmovupd       %ymm13,0x20(%r8,%rax,4)
  DB  196,65,125,17,76,128,64             ; vmovupd       %ymm9,0x40(%r8,%rax,4)
  DB  196,65,125,17,68,128,96             ; vmovupd       %ymm8,0x60(%r8,%rax,4)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax
  DB  196,65,121,17,20,128                ; vmovupd       %xmm10,(%r8,%rax,4)
  DB  72,131,249,1                        ; cmp           $0x1,%rcx
  DB  116,240                             ; je            5191 <_sk_store_f32_avx+0x69>
  DB  196,65,121,17,76,128,16             ; vmovupd       %xmm9,0x10(%r8,%rax,4)
  DB  72,131,249,3                        ; cmp           $0x3,%rcx
  DB  114,227                             ; jb            5191 <_sk_store_f32_avx+0x69>
  DB  196,65,121,17,68,128,32             ; vmovupd       %xmm8,0x20(%r8,%rax,4)
  DB  116,218                             ; je            5191 <_sk_store_f32_avx+0x69>
  DB  196,65,121,17,92,128,48             ; vmovupd       %xmm11,0x30(%r8,%rax,4)
  DB  72,131,249,5                        ; cmp           $0x5,%rcx
  DB  114,205                             ; jb            5191 <_sk_store_f32_avx+0x69>
  DB  196,67,125,25,84,128,64,1           ; vextractf128  $0x1,%ymm10,0x40(%r8,%rax,4)
  DB  116,195                             ; je            5191 <_sk_store_f32_avx+0x69>
  DB  196,67,125,25,76,128,80,1           ; vextractf128  $0x1,%ymm9,0x50(%r8,%rax,4)
  DB  72,131,249,7                        ; cmp           $0x7,%rcx
  DB  114,181                             ; jb            5191 <_sk_store_f32_avx+0x69>
  DB  196,67,125,25,68,128,96,1           ; vextractf128  $0x1,%ymm8,0x60(%r8,%rax,4)
  DB  235,171                             ; jmp           5191 <_sk_store_f32_avx+0x69>

PUBLIC _sk_clamp_x_avx
_sk_clamp_x_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,60,95,200                       ; vmaxps        %ymm0,%ymm8,%ymm9
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,99,125,25,192,1                 ; vextractf128  $0x1,%ymm8,%xmm0
  DB  196,65,41,118,210                   ; vpcmpeqd      %xmm10,%xmm10,%xmm10
  DB  196,193,121,254,194                 ; vpaddd        %xmm10,%xmm0,%xmm0
  DB  196,65,57,254,194                   ; vpaddd        %xmm10,%xmm8,%xmm8
  DB  196,227,61,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm8,%ymm0
  DB  197,180,93,192                      ; vminps        %ymm0,%ymm9,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_y_avx
_sk_clamp_y_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,60,95,201                       ; vmaxps        %ymm1,%ymm8,%ymm9
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,99,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm1
  DB  196,65,41,118,210                   ; vpcmpeqd      %xmm10,%xmm10,%xmm10
  DB  196,193,113,254,202                 ; vpaddd        %xmm10,%xmm1,%xmm1
  DB  196,65,57,254,194                   ; vpaddd        %xmm10,%xmm8,%xmm8
  DB  196,227,61,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm8,%ymm1
  DB  197,180,93,201                      ; vminps        %ymm1,%ymm9,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_x_avx
_sk_repeat_x_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,65,124,94,200                   ; vdivps        %ymm8,%ymm0,%ymm9
  DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
  DB  196,65,52,89,200                    ; vmulps        %ymm8,%ymm9,%ymm9
  DB  196,65,124,92,201                   ; vsubps        %ymm9,%ymm0,%ymm9
  DB  196,99,125,25,192,1                 ; vextractf128  $0x1,%ymm8,%xmm0
  DB  196,65,41,118,210                   ; vpcmpeqd      %xmm10,%xmm10,%xmm10
  DB  196,193,121,254,194                 ; vpaddd        %xmm10,%xmm0,%xmm0
  DB  196,65,57,254,194                   ; vpaddd        %xmm10,%xmm8,%xmm8
  DB  196,227,61,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm8,%ymm0
  DB  197,180,93,192                      ; vminps        %ymm0,%ymm9,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_y_avx
_sk_repeat_y_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,65,116,94,200                   ; vdivps        %ymm8,%ymm1,%ymm9
  DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
  DB  196,65,52,89,200                    ; vmulps        %ymm8,%ymm9,%ymm9
  DB  196,65,116,92,201                   ; vsubps        %ymm9,%ymm1,%ymm9
  DB  196,99,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm1
  DB  196,65,41,118,210                   ; vpcmpeqd      %xmm10,%xmm10,%xmm10
  DB  196,193,113,254,202                 ; vpaddd        %xmm10,%xmm1,%xmm1
  DB  196,65,57,254,194                   ; vpaddd        %xmm10,%xmm8,%xmm8
  DB  196,227,61,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm8,%ymm1
  DB  197,180,93,201                      ; vminps        %ymm1,%ymm9,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_x_avx
_sk_mirror_x_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,121,110,0                       ; vmovd         (%rax),%xmm8
  DB  196,65,121,112,200,0                ; vpshufd       $0x0,%xmm8,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  196,65,124,92,209                   ; vsubps        %ymm9,%ymm0,%ymm10
  DB  196,193,58,88,192                   ; vaddss        %xmm8,%xmm8,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,44,94,192                       ; vdivps        %ymm0,%ymm10,%ymm8
  DB  196,67,125,8,192,1                  ; vroundps      $0x1,%ymm8,%ymm8
  DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
  DB  197,172,92,192                      ; vsubps        %ymm0,%ymm10,%ymm0
  DB  196,193,124,92,193                  ; vsubps        %ymm9,%ymm0,%ymm0
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,60,92,192                       ; vsubps        %ymm0,%ymm8,%ymm8
  DB  197,60,84,192                       ; vandps        %ymm0,%ymm8,%ymm8
  DB  196,99,125,25,200,1                 ; vextractf128  $0x1,%ymm9,%xmm0
  DB  196,65,41,118,210                   ; vpcmpeqd      %xmm10,%xmm10,%xmm10
  DB  196,193,121,254,194                 ; vpaddd        %xmm10,%xmm0,%xmm0
  DB  196,65,49,254,202                   ; vpaddd        %xmm10,%xmm9,%xmm9
  DB  196,227,53,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm9,%ymm0
  DB  197,188,93,192                      ; vminps        %ymm0,%ymm8,%ymm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_y_avx
_sk_mirror_y_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,121,110,0                       ; vmovd         (%rax),%xmm8
  DB  196,65,121,112,200,0                ; vpshufd       $0x0,%xmm8,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  196,65,116,92,209                   ; vsubps        %ymm9,%ymm1,%ymm10
  DB  196,193,58,88,200                   ; vaddss        %xmm8,%xmm8,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,44,94,193                       ; vdivps        %ymm1,%ymm10,%ymm8
  DB  196,67,125,8,192,1                  ; vroundps      $0x1,%ymm8,%ymm8
  DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
  DB  197,172,92,201                      ; vsubps        %ymm1,%ymm10,%ymm1
  DB  196,193,116,92,201                  ; vsubps        %ymm9,%ymm1,%ymm1
  DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
  DB  197,60,92,193                       ; vsubps        %ymm1,%ymm8,%ymm8
  DB  197,60,84,193                       ; vandps        %ymm1,%ymm8,%ymm8
  DB  196,99,125,25,201,1                 ; vextractf128  $0x1,%ymm9,%xmm1
  DB  196,65,41,118,210                   ; vpcmpeqd      %xmm10,%xmm10,%xmm10
  DB  196,193,113,254,202                 ; vpaddd        %xmm10,%xmm1,%xmm1
  DB  196,65,49,254,202                   ; vpaddd        %xmm10,%xmm9,%xmm9
  DB  196,227,53,24,201,1                 ; vinsertf128   $0x1,%xmm1,%ymm9,%ymm1
  DB  197,188,93,201                      ; vminps        %ymm1,%ymm8,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_luminance_to_alpha_avx
_sk_luminance_to_alpha_avx LABEL PROC
  DB  184,208,179,89,62                   ; mov           $0x3e59b3d0,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,228,89,192                      ; vmulps        %ymm0,%ymm3,%ymm0
  DB  184,89,23,55,63                     ; mov           $0x3f371759,%eax
  DB  197,249,110,216                     ; vmovd         %eax,%xmm3
  DB  196,227,121,4,219,0                 ; vpermilps     $0x0,%xmm3,%xmm3
  DB  196,227,101,24,219,1                ; vinsertf128   $0x1,%xmm3,%ymm3,%ymm3
  DB  197,228,89,201                      ; vmulps        %ymm1,%ymm3,%ymm1
  DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
  DB  184,152,221,147,61                  ; mov           $0x3d93dd98,%eax
  DB  197,249,110,200                     ; vmovd         %eax,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
  DB  197,252,88,217                      ; vaddps        %ymm1,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
  DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
  DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_2x3_avx
_sk_matrix_2x3_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,98,125,24,72,8                  ; vbroadcastss  0x8(%rax),%ymm9
  DB  196,98,125,24,80,16                 ; vbroadcastss  0x10(%rax),%ymm10
  DB  197,52,89,201                       ; vmulps        %ymm1,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  197,60,89,192                       ; vmulps        %ymm0,%ymm8,%ymm8
  DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
  DB  196,98,125,24,72,4                  ; vbroadcastss  0x4(%rax),%ymm9
  DB  196,98,125,24,80,12                 ; vbroadcastss  0xc(%rax),%ymm10
  DB  196,98,125,24,88,20                 ; vbroadcastss  0x14(%rax),%ymm11
  DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
  DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
  DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
  DB  197,252,88,201                      ; vaddps        %ymm1,%ymm0,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_3x4_avx
_sk_matrix_3x4_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  196,98,125,24,88,36                 ; vbroadcastss  0x24(%rax),%ymm11
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  197,52,89,201                       ; vmulps        %ymm1,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  197,60,89,192                       ; vmulps        %ymm0,%ymm8,%ymm8
  DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
  DB  196,98,125,24,72,4                  ; vbroadcastss  0x4(%rax),%ymm9
  DB  196,98,125,24,80,16                 ; vbroadcastss  0x10(%rax),%ymm10
  DB  196,98,125,24,88,28                 ; vbroadcastss  0x1c(%rax),%ymm11
  DB  196,98,125,24,96,40                 ; vbroadcastss  0x28(%rax),%ymm12
  DB  197,36,89,218                       ; vmulps        %ymm2,%ymm11,%ymm11
  DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
  DB  197,44,89,209                       ; vmulps        %ymm1,%ymm10,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  197,52,89,200                       ; vmulps        %ymm0,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  196,98,125,24,80,8                  ; vbroadcastss  0x8(%rax),%ymm10
  DB  196,98,125,24,88,20                 ; vbroadcastss  0x14(%rax),%ymm11
  DB  196,98,125,24,96,32                 ; vbroadcastss  0x20(%rax),%ymm12
  DB  196,98,125,24,104,44                ; vbroadcastss  0x2c(%rax),%ymm13
  DB  197,156,89,210                      ; vmulps        %ymm2,%ymm12,%ymm2
  DB  196,193,108,88,213                  ; vaddps        %ymm13,%ymm2,%ymm2
  DB  197,164,89,201                      ; vmulps        %ymm1,%ymm11,%ymm1
  DB  197,244,88,202                      ; vaddps        %ymm2,%ymm1,%ymm1
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  197,252,88,209                      ; vaddps        %ymm1,%ymm0,%ymm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  197,124,41,201                      ; vmovaps       %ymm9,%ymm1
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_4x5_avx
_sk_matrix_4x5_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,98,125,24,72,16                 ; vbroadcastss  0x10(%rax),%ymm9
  DB  196,98,125,24,80,32                 ; vbroadcastss  0x20(%rax),%ymm10
  DB  196,98,125,24,88,48                 ; vbroadcastss  0x30(%rax),%ymm11
  DB  196,98,125,24,96,64                 ; vbroadcastss  0x40(%rax),%ymm12
  DB  197,36,89,219                       ; vmulps        %ymm3,%ymm11,%ymm11
  DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
  DB  197,44,89,210                       ; vmulps        %ymm2,%ymm10,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  197,52,89,201                       ; vmulps        %ymm1,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  197,60,89,192                       ; vmulps        %ymm0,%ymm8,%ymm8
  DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
  DB  196,98,125,24,72,4                  ; vbroadcastss  0x4(%rax),%ymm9
  DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
  DB  196,98,125,24,88,36                 ; vbroadcastss  0x24(%rax),%ymm11
  DB  196,98,125,24,96,52                 ; vbroadcastss  0x34(%rax),%ymm12
  DB  196,98,125,24,104,68                ; vbroadcastss  0x44(%rax),%ymm13
  DB  197,28,89,227                       ; vmulps        %ymm3,%ymm12,%ymm12
  DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
  DB  197,36,89,218                       ; vmulps        %ymm2,%ymm11,%ymm11
  DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
  DB  197,44,89,209                       ; vmulps        %ymm1,%ymm10,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  197,52,89,200                       ; vmulps        %ymm0,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  196,98,125,24,80,8                  ; vbroadcastss  0x8(%rax),%ymm10
  DB  196,98,125,24,88,24                 ; vbroadcastss  0x18(%rax),%ymm11
  DB  196,98,125,24,96,40                 ; vbroadcastss  0x28(%rax),%ymm12
  DB  196,98,125,24,104,56                ; vbroadcastss  0x38(%rax),%ymm13
  DB  196,98,125,24,112,72                ; vbroadcastss  0x48(%rax),%ymm14
  DB  197,20,89,235                       ; vmulps        %ymm3,%ymm13,%ymm13
  DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
  DB  197,28,89,226                       ; vmulps        %ymm2,%ymm12,%ymm12
  DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
  DB  197,36,89,217                       ; vmulps        %ymm1,%ymm11,%ymm11
  DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
  DB  197,44,89,208                       ; vmulps        %ymm0,%ymm10,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  196,98,125,24,88,12                 ; vbroadcastss  0xc(%rax),%ymm11
  DB  196,98,125,24,96,28                 ; vbroadcastss  0x1c(%rax),%ymm12
  DB  196,98,125,24,104,44                ; vbroadcastss  0x2c(%rax),%ymm13
  DB  196,98,125,24,112,60                ; vbroadcastss  0x3c(%rax),%ymm14
  DB  196,98,125,24,120,76                ; vbroadcastss  0x4c(%rax),%ymm15
  DB  197,140,89,219                      ; vmulps        %ymm3,%ymm14,%ymm3
  DB  196,193,100,88,223                  ; vaddps        %ymm15,%ymm3,%ymm3
  DB  197,148,89,210                      ; vmulps        %ymm2,%ymm13,%ymm2
  DB  197,236,88,211                      ; vaddps        %ymm3,%ymm2,%ymm2
  DB  197,156,89,201                      ; vmulps        %ymm1,%ymm12,%ymm1
  DB  197,244,88,202                      ; vaddps        %ymm2,%ymm1,%ymm1
  DB  197,164,89,192                      ; vmulps        %ymm0,%ymm11,%ymm0
  DB  197,252,88,217                      ; vaddps        %ymm1,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  197,124,41,201                      ; vmovaps       %ymm9,%ymm1
  DB  197,124,41,210                      ; vmovaps       %ymm10,%ymm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_perspective_avx
_sk_matrix_perspective_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
  DB  196,98,125,24,72,4                  ; vbroadcastss  0x4(%rax),%ymm9
  DB  196,98,125,24,80,8                  ; vbroadcastss  0x8(%rax),%ymm10
  DB  197,52,89,201                       ; vmulps        %ymm1,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  197,60,89,192                       ; vmulps        %ymm0,%ymm8,%ymm8
  DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
  DB  196,98,125,24,80,16                 ; vbroadcastss  0x10(%rax),%ymm10
  DB  196,98,125,24,88,20                 ; vbroadcastss  0x14(%rax),%ymm11
  DB  197,44,89,209                       ; vmulps        %ymm1,%ymm10,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  197,52,89,200                       ; vmulps        %ymm0,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  196,98,125,24,80,24                 ; vbroadcastss  0x18(%rax),%ymm10
  DB  196,98,125,24,88,28                 ; vbroadcastss  0x1c(%rax),%ymm11
  DB  196,98,125,24,96,32                 ; vbroadcastss  0x20(%rax),%ymm12
  DB  197,164,89,201                      ; vmulps        %ymm1,%ymm11,%ymm1
  DB  196,193,116,88,204                  ; vaddps        %ymm12,%ymm1,%ymm1
  DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
  DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
  DB  197,252,83,200                      ; vrcpps        %ymm0,%ymm1
  DB  197,188,89,193                      ; vmulps        %ymm1,%ymm8,%ymm0
  DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_avx
_sk_linear_gradient_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,98,125,24,64,16                 ; vbroadcastss  0x10(%rax),%ymm8
  DB  196,226,125,24,72,20                ; vbroadcastss  0x14(%rax),%ymm1
  DB  196,226,125,24,80,24                ; vbroadcastss  0x18(%rax),%ymm2
  DB  196,226,125,24,88,28                ; vbroadcastss  0x1c(%rax),%ymm3
  DB  76,139,0                            ; mov           (%rax),%r8
  DB  77,133,192                          ; test          %r8,%r8
  DB  15,132,146,0,0,0                    ; je            5749 <_sk_linear_gradient_avx+0xb8>
  DB  72,139,64,8                         ; mov           0x8(%rax),%rax
  DB  72,131,192,32                       ; add           $0x20,%rax
  DB  196,65,28,87,228                    ; vxorps        %ymm12,%ymm12,%ymm12
  DB  196,65,52,87,201                    ; vxorps        %ymm9,%ymm9,%ymm9
  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
  DB  196,65,36,87,219                    ; vxorps        %ymm11,%ymm11,%ymm11
  DB  196,98,125,24,104,224               ; vbroadcastss  -0x20(%rax),%ymm13
  DB  196,65,124,194,237,1                ; vcmpltps      %ymm13,%ymm0,%ymm13
  DB  196,98,125,24,112,228               ; vbroadcastss  -0x1c(%rax),%ymm14
  DB  196,67,13,74,228,208                ; vblendvps     %ymm13,%ymm12,%ymm14,%ymm12
  DB  196,98,125,24,112,232               ; vbroadcastss  -0x18(%rax),%ymm14
  DB  196,67,13,74,219,208                ; vblendvps     %ymm13,%ymm11,%ymm14,%ymm11
  DB  196,98,125,24,112,236               ; vbroadcastss  -0x14(%rax),%ymm14
  DB  196,67,13,74,210,208                ; vblendvps     %ymm13,%ymm10,%ymm14,%ymm10
  DB  196,98,125,24,112,240               ; vbroadcastss  -0x10(%rax),%ymm14
  DB  196,67,13,74,201,208                ; vblendvps     %ymm13,%ymm9,%ymm14,%ymm9
  DB  196,98,125,24,112,244               ; vbroadcastss  -0xc(%rax),%ymm14
  DB  196,67,13,74,192,208                ; vblendvps     %ymm13,%ymm8,%ymm14,%ymm8
  DB  196,98,125,24,112,248               ; vbroadcastss  -0x8(%rax),%ymm14
  DB  196,227,13,74,201,208               ; vblendvps     %ymm13,%ymm1,%ymm14,%ymm1
  DB  196,98,125,24,112,252               ; vbroadcastss  -0x4(%rax),%ymm14
  DB  196,227,13,74,210,208               ; vblendvps     %ymm13,%ymm2,%ymm14,%ymm2
  DB  196,98,125,24,48                    ; vbroadcastss  (%rax),%ymm14
  DB  196,227,13,74,219,208               ; vblendvps     %ymm13,%ymm3,%ymm14,%ymm3
  DB  72,131,192,36                       ; add           $0x24,%rax
  DB  73,255,200                          ; dec           %r8
  DB  117,140                             ; jne           56d3 <_sk_linear_gradient_avx+0x42>
  DB  235,20                              ; jmp           575d <_sk_linear_gradient_avx+0xcc>
  DB  196,65,36,87,219                    ; vxorps        %ymm11,%ymm11,%ymm11
  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
  DB  196,65,52,87,201                    ; vxorps        %ymm9,%ymm9,%ymm9
  DB  196,65,28,87,228                    ; vxorps        %ymm12,%ymm12,%ymm12
  DB  197,28,89,224                       ; vmulps        %ymm0,%ymm12,%ymm12
  DB  196,65,60,88,196                    ; vaddps        %ymm12,%ymm8,%ymm8
  DB  197,36,89,216                       ; vmulps        %ymm0,%ymm11,%ymm11
  DB  197,164,88,201                      ; vaddps        %ymm1,%ymm11,%ymm1
  DB  197,44,89,208                       ; vmulps        %ymm0,%ymm10,%ymm10
  DB  197,172,88,210                      ; vaddps        %ymm2,%ymm10,%ymm2
  DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
  DB  197,252,88,219                      ; vaddps        %ymm3,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_2stops_avx
_sk_linear_gradient_2stops_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  196,226,125,24,8                    ; vbroadcastss  (%rax),%ymm1
  DB  196,226,125,24,80,16                ; vbroadcastss  0x10(%rax),%ymm2
  DB  197,244,89,200                      ; vmulps        %ymm0,%ymm1,%ymm1
  DB  197,116,88,194                      ; vaddps        %ymm2,%ymm1,%ymm8
  DB  196,226,125,24,72,4                 ; vbroadcastss  0x4(%rax),%ymm1
  DB  196,226,125,24,80,20                ; vbroadcastss  0x14(%rax),%ymm2
  DB  197,244,89,200                      ; vmulps        %ymm0,%ymm1,%ymm1
  DB  197,244,88,202                      ; vaddps        %ymm2,%ymm1,%ymm1
  DB  196,226,125,24,80,8                 ; vbroadcastss  0x8(%rax),%ymm2
  DB  196,226,125,24,88,24                ; vbroadcastss  0x18(%rax),%ymm3
  DB  197,236,89,208                      ; vmulps        %ymm0,%ymm2,%ymm2
  DB  197,236,88,211                      ; vaddps        %ymm3,%ymm2,%ymm2
  DB  196,226,125,24,88,12                ; vbroadcastss  0xc(%rax),%ymm3
  DB  196,98,125,24,72,28                 ; vbroadcastss  0x1c(%rax),%ymm9
  DB  197,228,89,192                      ; vmulps        %ymm0,%ymm3,%ymm0
  DB  196,193,124,88,217                  ; vaddps        %ymm9,%ymm0,%ymm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,41,192                      ; vmovaps       %ymm8,%ymm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_save_xy_avx
_sk_save_xy_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,88,200                       ; vaddps        %ymm0,%ymm8,%ymm9
  DB  196,67,125,8,209,1                  ; vroundps      $0x1,%ymm9,%ymm10
  DB  196,65,52,92,202                    ; vsubps        %ymm10,%ymm9,%ymm9
  DB  197,60,88,193                       ; vaddps        %ymm1,%ymm8,%ymm8
  DB  196,67,125,8,208,1                  ; vroundps      $0x1,%ymm8,%ymm10
  DB  196,65,60,92,194                    ; vsubps        %ymm10,%ymm8,%ymm8
  DB  197,252,17,0                        ; vmovups       %ymm0,(%rax)
  DB  197,252,17,72,32                    ; vmovups       %ymm1,0x20(%rax)
  DB  197,124,17,72,64                    ; vmovups       %ymm9,0x40(%rax)
  DB  197,124,17,64,96                    ; vmovups       %ymm8,0x60(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_accumulate_avx
_sk_accumulate_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  197,124,16,128,128,0,0,0            ; vmovups       0x80(%rax),%ymm8
  DB  197,60,89,128,160,0,0,0             ; vmulps        0xa0(%rax),%ymm8,%ymm8
  DB  197,60,89,200                       ; vmulps        %ymm0,%ymm8,%ymm9
  DB  197,180,88,228                      ; vaddps        %ymm4,%ymm9,%ymm4
  DB  197,60,89,201                       ; vmulps        %ymm1,%ymm8,%ymm9
  DB  197,180,88,237                      ; vaddps        %ymm5,%ymm9,%ymm5
  DB  197,60,89,202                       ; vmulps        %ymm2,%ymm8,%ymm9
  DB  197,180,88,246                      ; vaddps        %ymm6,%ymm9,%ymm6
  DB  197,60,89,195                       ; vmulps        %ymm3,%ymm8,%ymm8
  DB  197,188,88,255                      ; vaddps        %ymm7,%ymm8,%ymm7
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_nx_avx
_sk_bilinear_nx_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_px_avx
_sk_bilinear_px_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,124,16,64,64                    ; vmovups       0x40(%rax),%ymm8
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_ny_avx
_sk_bilinear_ny_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_py_avx
_sk_bilinear_py_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,124,16,64,96                    ; vmovups       0x60(%rax),%ymm8
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3x_avx
_sk_bicubic_n3x_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,191                  ; mov           $0xbfc00000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  196,65,44,89,192                    ; vmulps        %ymm8,%ymm10,%ymm8
  DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
  DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1x_avx
_sk_bicubic_n1x_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,64,64                     ; vsubps        0x40(%rax),%ymm8,%ymm8
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,200                  ; vmovd         %r8d,%xmm9
  DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,52,89,200                    ; vmulps        %ymm8,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
  DB  196,65,44,88,201                    ; vaddps        %ymm9,%ymm10,%ymm9
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
  DB  196,65,44,88,192                    ; vaddps        %ymm8,%ymm10,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1x_avx
_sk_bicubic_p1x_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,99,125,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm8
  DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
  DB  197,124,16,72,64                    ; vmovups       0x40(%rax),%ymm9
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
  DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
  DB  196,65,44,88,192                    ; vaddps        %ymm8,%ymm10,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3x_avx
_sk_bicubic_p3x_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,193,121,110,192                 ; vmovd         %r8d,%xmm0
  DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
  DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
  DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
  DB  197,124,16,64,64                    ; vmovups       0x40(%rax),%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
  DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
  DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
  DB  197,124,17,128,128,0,0,0            ; vmovups       %ymm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3y_avx
_sk_bicubic_n3y_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,191                  ; mov           $0xbfc00000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  196,65,44,89,192                    ; vmulps        %ymm8,%ymm10,%ymm8
  DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
  DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1y_avx
_sk_bicubic_n1y_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,191                    ; mov           $0xbf000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  65,184,0,0,128,63                   ; mov           $0x3f800000,%r8d
  DB  196,65,121,110,192                  ; vmovd         %r8d,%xmm8
  DB  196,67,121,4,192,0                  ; vpermilps     $0x0,%xmm8,%xmm8
  DB  196,67,61,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm8,%ymm8
  DB  197,60,92,64,96                     ; vsubps        0x60(%rax),%ymm8,%ymm8
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,200                  ; vmovd         %r8d,%xmm9
  DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
  DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,52,89,200                    ; vmulps        %ymm8,%ymm9,%ymm9
  DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
  DB  196,65,44,88,201                    ; vaddps        %ymm9,%ymm10,%ymm9
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
  DB  196,65,44,88,192                    ; vaddps        %ymm8,%ymm10,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1y_avx
_sk_bicubic_p1y_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,0,63                     ; mov           $0x3f000000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,99,117,24,193,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm8
  DB  197,188,88,72,32                    ; vaddps        0x20(%rax),%ymm8,%ymm1
  DB  197,124,16,72,96                    ; vmovups       0x60(%rax),%ymm9
  DB  65,184,85,85,149,191                ; mov           $0xbf955555,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
  DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
  DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
  DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
  DB  65,184,57,142,99,61                 ; mov           $0x3d638e39,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
  DB  196,65,44,88,192                    ; vaddps        %ymm8,%ymm10,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3y_avx
_sk_bicubic_p3y_avx LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,184,0,0,192,63                   ; mov           $0x3fc00000,%r8d
  DB  196,193,121,110,200                 ; vmovd         %r8d,%xmm1
  DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
  DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
  DB  197,244,88,72,32                    ; vaddps        0x20(%rax),%ymm1,%ymm1
  DB  197,124,16,64,96                    ; vmovups       0x60(%rax),%ymm8
  DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
  DB  65,184,114,28,199,62                ; mov           $0x3ec71c72,%r8d
  DB  196,65,121,110,208                  ; vmovd         %r8d,%xmm10
  DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
  DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
  DB  65,184,171,170,170,190              ; mov           $0xbeaaaaab,%r8d
  DB  196,65,121,110,216                  ; vmovd         %r8d,%xmm11
  DB  196,67,121,4,219,0                  ; vpermilps     $0x0,%xmm11,%xmm11
  DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
  DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
  DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
  DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
  DB  197,124,17,128,160,0,0,0            ; vmovups       %ymm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_callback_avx
_sk_callback_avx LABEL PROC
  DB  65,86                               ; push          %r14
  DB  83                                  ; push          %rbx
  DB  72,129,236,40,1,0,0                 ; sub           $0x128,%rsp
  DB  197,252,17,188,36,0,1,0,0           ; vmovups       %ymm7,0x100(%rsp)
  DB  197,252,17,180,36,224,0,0,0         ; vmovups       %ymm6,0xe0(%rsp)
  DB  197,252,17,172,36,192,0,0,0         ; vmovups       %ymm5,0xc0(%rsp)
  DB  197,252,17,164,36,160,0,0,0         ; vmovups       %ymm4,0xa0(%rsp)
  DB  197,252,17,156,36,128,0,0,0         ; vmovups       %ymm3,0x80(%rsp)
  DB  197,252,17,84,36,96                 ; vmovups       %ymm2,0x60(%rsp)
  DB  197,252,17,76,36,64                 ; vmovups       %ymm1,0x40(%rsp)
  DB  197,252,17,68,36,32                 ; vmovups       %ymm0,0x20(%rsp)
  DB  72,137,203                          ; mov           %rcx,%rbx
  DB  73,137,214                          ; mov           %rdx,%r14
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,72,8                         ; mov           0x8(%rax),%rcx
  DB  72,133,219                          ; test          %rbx,%rbx
  DB  186,8,0,0,0                         ; mov           $0x8,%edx
  DB  15,69,211                           ; cmovne        %ebx,%edx
  DB  197,248,119                         ; vzeroupper
  DB  255,16                              ; callq         *(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  72,137,217                          ; mov           %rbx,%rcx
  DB  197,252,16,68,36,32                 ; vmovups       0x20(%rsp),%ymm0
  DB  197,252,16,76,36,64                 ; vmovups       0x40(%rsp),%ymm1
  DB  197,252,16,84,36,96                 ; vmovups       0x60(%rsp),%ymm2
  DB  197,252,16,156,36,128,0,0,0         ; vmovups       0x80(%rsp),%ymm3
  DB  197,252,16,164,36,160,0,0,0         ; vmovups       0xa0(%rsp),%ymm4
  DB  197,252,16,172,36,192,0,0,0         ; vmovups       0xc0(%rsp),%ymm5
  DB  197,252,16,180,36,224,0,0,0         ; vmovups       0xe0(%rsp),%ymm6
  DB  197,252,16,188,36,0,1,0,0           ; vmovups       0x100(%rsp),%ymm7
  DB  72,129,196,40,1,0,0                 ; add           $0x128,%rsp
  DB  91                                  ; pop           %rbx
  DB  65,94                               ; pop           %r14
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_start_pipeline_sse41
_sk_start_pipeline_sse41 LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  86                                  ; push          %rsi
  DB  87                                  ; push          %rdi
  DB  83                                  ; push          %rbx
  DB  72,129,236,160,0,0,0                ; sub           $0xa0,%rsp
  DB  68,15,41,188,36,144,0,0,0           ; movaps        %xmm15,0x90(%rsp)
  DB  68,15,41,180,36,128,0,0,0           ; movaps        %xmm14,0x80(%rsp)
  DB  68,15,41,108,36,112                 ; movaps        %xmm13,0x70(%rsp)
  DB  68,15,41,100,36,96                  ; movaps        %xmm12,0x60(%rsp)
  DB  68,15,41,92,36,80                   ; movaps        %xmm11,0x50(%rsp)
  DB  68,15,41,84,36,64                   ; movaps        %xmm10,0x40(%rsp)
  DB  68,15,41,76,36,48                   ; movaps        %xmm9,0x30(%rsp)
  DB  68,15,41,68,36,32                   ; movaps        %xmm8,0x20(%rsp)
  DB  15,41,124,36,16                     ; movaps        %xmm7,0x10(%rsp)
  DB  15,41,52,36                         ; movaps        %xmm6,(%rsp)
  DB  77,137,207                          ; mov           %r9,%r15
  DB  77,137,198                          ; mov           %r8,%r14
  DB  72,137,203                          ; mov           %rcx,%rbx
  DB  72,137,214                          ; mov           %rdx,%rsi
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,196                          ; mov           %rax,%r12
  DB  73,137,245                          ; mov           %rsi,%r13
  DB  72,141,67,4                         ; lea           0x4(%rbx),%rax
  DB  76,57,248                           ; cmp           %r15,%rax
  DB  118,5                               ; jbe           73 <_sk_start_pipeline_sse41+0x73>
  DB  72,137,216                          ; mov           %rbx,%rax
  DB  235,52                              ; jmp           a7 <_sk_start_pipeline_sse41+0xa7>
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  15,87,228                           ; xorps         %xmm4,%xmm4
  DB  15,87,237                           ; xorps         %xmm5,%xmm5
  DB  15,87,246                           ; xorps         %xmm6,%xmm6
  DB  15,87,255                           ; xorps         %xmm7,%xmm7
  DB  72,137,223                          ; mov           %rbx,%rdi
  DB  76,137,238                          ; mov           %r13,%rsi
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  65,255,212                          ; callq         *%r12
  DB  72,141,67,4                         ; lea           0x4(%rbx),%rax
  DB  72,131,195,8                        ; add           $0x8,%rbx
  DB  76,57,251                           ; cmp           %r15,%rbx
  DB  72,137,195                          ; mov           %rax,%rbx
  DB  118,204                             ; jbe           73 <_sk_start_pipeline_sse41+0x73>
  DB  15,40,52,36                         ; movaps        (%rsp),%xmm6
  DB  15,40,124,36,16                     ; movaps        0x10(%rsp),%xmm7
  DB  68,15,40,68,36,32                   ; movaps        0x20(%rsp),%xmm8
  DB  68,15,40,76,36,48                   ; movaps        0x30(%rsp),%xmm9
  DB  68,15,40,84,36,64                   ; movaps        0x40(%rsp),%xmm10
  DB  68,15,40,92,36,80                   ; movaps        0x50(%rsp),%xmm11
  DB  68,15,40,100,36,96                  ; movaps        0x60(%rsp),%xmm12
  DB  68,15,40,108,36,112                 ; movaps        0x70(%rsp),%xmm13
  DB  68,15,40,180,36,128,0,0,0           ; movaps        0x80(%rsp),%xmm14
  DB  68,15,40,188,36,144,0,0,0           ; movaps        0x90(%rsp),%xmm15
  DB  72,129,196,160,0,0,0                ; add           $0xa0,%rsp
  DB  91                                  ; pop           %rbx
  DB  95                                  ; pop           %rdi
  DB  94                                  ; pop           %rsi
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  195                                 ; retq

PUBLIC _sk_just_return_sse41
_sk_just_return_sse41 LABEL PROC
  DB  195                                 ; retq

PUBLIC _sk_seed_shader_sse41
_sk_seed_shader_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  102,15,110,199                      ; movd          %edi,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,88,202                           ; addps         %xmm2,%xmm1
  DB  15,16,2                             ; movups        (%rdx),%xmm0
  DB  15,88,193                           ; addps         %xmm1,%xmm0
  DB  102,15,110,8                        ; movd          (%rax),%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  15,88,202                           ; addps         %xmm2,%xmm1
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  15,87,228                           ; xorps         %xmm4,%xmm4
  DB  15,87,237                           ; xorps         %xmm5,%xmm5
  DB  15,87,246                           ; xorps         %xmm6,%xmm6
  DB  15,87,255                           ; xorps         %xmm7,%xmm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_constant_color_sse41
_sk_constant_color_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,15,16,80,8                      ; movss         0x8(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  243,15,16,88,12                     ; movss         0xc(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clear_sse41
_sk_clear_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcatop_sse41
_sk_srcatop_sse41 LABEL PROC
  DB  15,89,199                           ; mulps         %xmm7,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  65,15,88,193                        ; addps         %xmm9,%xmm0
  DB  15,89,207                           ; mulps         %xmm7,%xmm1
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  65,15,88,201                        ; addps         %xmm9,%xmm1
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstatop_sse41
_sk_dstatop_sse41 LABEL PROC
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,89,197                        ; mulps         %xmm5,%xmm8
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,89,198                        ; mulps         %xmm6,%xmm8
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  65,15,88,208                        ; addps         %xmm8,%xmm2
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcin_sse41
_sk_srcin_sse41 LABEL PROC
  DB  15,89,199                           ; mulps         %xmm7,%xmm0
  DB  15,89,207                           ; mulps         %xmm7,%xmm1
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstin_sse41
_sk_dstin_sse41 LABEL PROC
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  15,40,211                           ; movaps        %xmm3,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcout_sse41
_sk_srcout_sse41 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,199                        ; subps         %xmm7,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstout_sse41
_sk_dstout_sse41 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,216                        ; movaps        %xmm8,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcover_sse41
_sk_srcover_sse41 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  65,15,88,193                        ; addps         %xmm9,%xmm0
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  65,15,88,201                        ; addps         %xmm9,%xmm1
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstover_sse41
_sk_dstover_sse41 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,199                        ; subps         %xmm7,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_modulate_sse41
_sk_modulate_sse41 LABEL PROC
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_multiply_sse41
_sk_multiply_sse41 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,220                        ; mulps         %xmm4,%xmm11
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  65,15,88,195                        ; addps         %xmm11,%xmm0
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,221                        ; mulps         %xmm5,%xmm11
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  65,15,88,203                        ; addps         %xmm11,%xmm1
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,222                        ; mulps         %xmm6,%xmm11
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  65,15,88,211                        ; addps         %xmm11,%xmm2
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_plus__sse41
_sk_plus__sse41 LABEL PROC
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_screen_sse41
_sk_screen_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  68,15,88,196                        ; addps         %xmm4,%xmm8
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  68,15,92,192                        ; subps         %xmm0,%xmm8
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,88,205                        ; addps         %xmm5,%xmm9
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  68,15,92,201                        ; subps         %xmm1,%xmm9
  DB  68,15,40,210                        ; movaps        %xmm2,%xmm10
  DB  68,15,88,214                        ; addps         %xmm6,%xmm10
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  68,15,92,210                        ; subps         %xmm2,%xmm10
  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
  DB  68,15,88,223                        ; addps         %xmm7,%xmm11
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  68,15,92,219                        ; subps         %xmm3,%xmm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_xor__sse41
_sk_xor__sse41 LABEL PROC
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,92,216                        ; subps         %xmm8,%xmm3
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,89,212                        ; mulps         %xmm4,%xmm10
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,89,213                        ; mulps         %xmm5,%xmm10
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,89,214                        ; mulps         %xmm6,%xmm10
  DB  65,15,88,210                        ; addps         %xmm10,%xmm2
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_darken_sse41
_sk_darken_sse41 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  68,15,89,207                        ; mulps         %xmm7,%xmm9
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  68,15,95,201                        ; maxps         %xmm1,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  69,15,95,193                        ; maxps         %xmm9,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,95,209                        ; maxps         %xmm9,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lighten_sse41
_sk_lighten_sse41 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  68,15,89,207                        ; mulps         %xmm7,%xmm9
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  68,15,93,201                        ; minps         %xmm1,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  69,15,93,193                        ; minps         %xmm9,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,93,209                        ; minps         %xmm9,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_difference_sse41
_sk_difference_sse41 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  68,15,89,207                        ; mulps         %xmm7,%xmm9
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  68,15,93,201                        ; minps         %xmm1,%xmm9
  DB  69,15,88,201                        ; addps         %xmm9,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  69,15,93,193                        ; minps         %xmm9,%xmm8
  DB  69,15,88,192                        ; addps         %xmm8,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,93,209                        ; minps         %xmm9,%xmm2
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_exclusion_sse41
_sk_exclusion_sse41 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  15,40,200                           ; movaps        %xmm0,%xmm1
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  15,92,193                           ; subps         %xmm1,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,197                        ; mulps         %xmm5,%xmm8
  DB  69,15,88,192                        ; addps         %xmm8,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colorburn_sse41
_sk_colorburn_sse41 LABEL PROC
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
  DB  68,15,92,223                        ; subps         %xmm7,%xmm11
  DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
  DB  69,15,89,225                        ; mulps         %xmm9,%xmm12
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  15,40,199                           ; movaps        %xmm7,%xmm0
  DB  15,92,196                           ; subps         %xmm4,%xmm0
  DB  15,89,195                           ; mulps         %xmm3,%xmm0
  DB  65,15,94,193                        ; divps         %xmm9,%xmm0
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,93,232                        ; minps         %xmm0,%xmm13
  DB  68,15,40,247                        ; movaps        %xmm7,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  65,15,194,192,0                     ; cmpeqps       %xmm8,%xmm0
  DB  68,15,92,211                        ; subps         %xmm3,%xmm10
  DB  68,15,89,243                        ; mulps         %xmm3,%xmm14
  DB  69,15,88,244                        ; addps         %xmm12,%xmm14
  DB  102,69,15,56,20,241                 ; blendvps      %xmm0,%xmm9,%xmm14
  DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  68,15,88,228                        ; addps         %xmm4,%xmm12
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,194,199,0                        ; cmpeqps       %xmm7,%xmm0
  DB  69,15,88,206                        ; addps         %xmm14,%xmm9
  DB  102,69,15,56,20,204                 ; blendvps      %xmm0,%xmm12,%xmm9
  DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
  DB  68,15,89,225                        ; mulps         %xmm1,%xmm12
  DB  15,40,199                           ; movaps        %xmm7,%xmm0
  DB  15,92,197                           ; subps         %xmm5,%xmm0
  DB  15,89,195                           ; mulps         %xmm3,%xmm0
  DB  15,94,193                           ; divps         %xmm1,%xmm0
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,93,232                        ; minps         %xmm0,%xmm13
  DB  68,15,40,247                        ; movaps        %xmm7,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  15,40,193                           ; movaps        %xmm1,%xmm0
  DB  65,15,194,192,0                     ; cmpeqps       %xmm8,%xmm0
  DB  68,15,89,243                        ; mulps         %xmm3,%xmm14
  DB  69,15,88,244                        ; addps         %xmm12,%xmm14
  DB  102,68,15,56,20,241                 ; blendvps      %xmm0,%xmm1,%xmm14
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  68,15,88,229                        ; addps         %xmm5,%xmm12
  DB  65,15,88,206                        ; addps         %xmm14,%xmm1
  DB  15,40,197                           ; movaps        %xmm5,%xmm0
  DB  15,194,199,0                        ; cmpeqps       %xmm7,%xmm0
  DB  102,65,15,56,20,204                 ; blendvps      %xmm0,%xmm12,%xmm1
  DB  15,40,199                           ; movaps        %xmm7,%xmm0
  DB  15,92,198                           ; subps         %xmm6,%xmm0
  DB  15,89,195                           ; mulps         %xmm3,%xmm0
  DB  15,94,194                           ; divps         %xmm2,%xmm0
  DB  68,15,40,231                        ; movaps        %xmm7,%xmm12
  DB  68,15,93,224                        ; minps         %xmm0,%xmm12
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  68,15,194,194,0                     ; cmpeqps       %xmm2,%xmm8
  DB  68,15,89,235                        ; mulps         %xmm3,%xmm13
  DB  69,15,88,235                        ; addps         %xmm11,%xmm13
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  102,68,15,56,20,234                 ; blendvps      %xmm0,%xmm2,%xmm13
  DB  68,15,88,222                        ; addps         %xmm6,%xmm11
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  65,15,88,213                        ; addps         %xmm13,%xmm2
  DB  15,40,198                           ; movaps        %xmm6,%xmm0
  DB  15,194,199,0                        ; cmpeqps       %xmm7,%xmm0
  DB  102,65,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm2
  DB  68,15,89,215                        ; mulps         %xmm7,%xmm10
  DB  65,15,88,218                        ; addps         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colordodge_sse41
_sk_colordodge_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
  DB  68,15,92,223                        ; subps         %xmm7,%xmm11
  DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
  DB  69,15,89,224                        ; mulps         %xmm8,%xmm12
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  65,15,92,192                        ; subps         %xmm8,%xmm0
  DB  68,15,94,200                        ; divps         %xmm0,%xmm9
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,40,247                        ; movaps        %xmm7,%xmm14
  DB  69,15,93,241                        ; minps         %xmm9,%xmm14
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  15,194,195,0                        ; cmpeqps       %xmm3,%xmm0
  DB  68,15,89,243                        ; mulps         %xmm3,%xmm14
  DB  69,15,88,244                        ; addps         %xmm12,%xmm14
  DB  102,69,15,56,20,240                 ; blendvps      %xmm0,%xmm8,%xmm14
  DB  69,15,87,201                        ; xorps         %xmm9,%xmm9
  DB  68,15,92,211                        ; subps         %xmm3,%xmm10
  DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
  DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
  DB  68,15,88,228                        ; addps         %xmm4,%xmm12
  DB  69,15,88,198                        ; addps         %xmm14,%xmm8
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  65,15,194,193,0                     ; cmpeqps       %xmm9,%xmm0
  DB  102,69,15,56,20,196                 ; blendvps      %xmm0,%xmm12,%xmm8
  DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
  DB  68,15,89,229                        ; mulps         %xmm5,%xmm12
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  15,92,193                           ; subps         %xmm1,%xmm0
  DB  68,15,94,224                        ; divps         %xmm0,%xmm12
  DB  69,15,40,243                        ; movaps        %xmm11,%xmm14
  DB  68,15,89,241                        ; mulps         %xmm1,%xmm14
  DB  69,15,93,236                        ; minps         %xmm12,%xmm13
  DB  15,40,193                           ; movaps        %xmm1,%xmm0
  DB  15,194,195,0                        ; cmpeqps       %xmm3,%xmm0
  DB  68,15,89,235                        ; mulps         %xmm3,%xmm13
  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
  DB  102,68,15,56,20,233                 ; blendvps      %xmm0,%xmm1,%xmm13
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  68,15,88,245                        ; addps         %xmm5,%xmm14
  DB  65,15,88,205                        ; addps         %xmm13,%xmm1
  DB  15,40,197                           ; movaps        %xmm5,%xmm0
  DB  65,15,194,193,0                     ; cmpeqps       %xmm9,%xmm0
  DB  102,65,15,56,20,206                 ; blendvps      %xmm0,%xmm14,%xmm1
  DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
  DB  68,15,89,230                        ; mulps         %xmm6,%xmm12
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  15,92,194                           ; subps         %xmm2,%xmm0
  DB  68,15,94,224                        ; divps         %xmm0,%xmm12
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,93,236                        ; minps         %xmm12,%xmm13
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  15,194,195,0                        ; cmpeqps       %xmm3,%xmm0
  DB  68,15,89,235                        ; mulps         %xmm3,%xmm13
  DB  69,15,88,235                        ; addps         %xmm11,%xmm13
  DB  102,68,15,56,20,234                 ; blendvps      %xmm0,%xmm2,%xmm13
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  65,15,88,213                        ; addps         %xmm13,%xmm2
  DB  68,15,194,206,0                     ; cmpeqps       %xmm6,%xmm9
  DB  68,15,88,222                        ; addps         %xmm6,%xmm11
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  102,65,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm2
  DB  68,15,89,215                        ; mulps         %xmm7,%xmm10
  DB  65,15,88,218                        ; addps         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hardlight_sse41
_sk_hardlight_sse41 LABEL PROC
  DB  72,131,236,24                       ; sub           $0x18,%rsp
  DB  15,41,52,36                         ; movaps        %xmm6,(%rsp)
  DB  68,15,40,229                        ; movaps        %xmm5,%xmm12
  DB  15,40,244                           ; movaps        %xmm4,%xmm6
  DB  15,40,227                           ; movaps        %xmm3,%xmm4
  DB  15,40,234                           ; movaps        %xmm2,%xmm5
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,92,215                           ; subps         %xmm7,%xmm2
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  68,15,92,212                        ; subps         %xmm4,%xmm10
  DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
  DB  68,15,89,198                        ; mulps         %xmm6,%xmm8
  DB  68,15,88,192                        ; addps         %xmm0,%xmm8
  DB  68,15,40,252                        ; movaps        %xmm4,%xmm15
  DB  69,15,92,249                        ; subps         %xmm9,%xmm15
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,40,247                        ; movaps        %xmm7,%xmm14
  DB  15,40,199                           ; movaps        %xmm7,%xmm0
  DB  15,92,198                           ; subps         %xmm6,%xmm0
  DB  65,15,89,199                        ; mulps         %xmm15,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  68,15,40,251                        ; movaps        %xmm3,%xmm15
  DB  68,15,92,248                        ; subps         %xmm0,%xmm15
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,196,2                        ; cmpleps       %xmm4,%xmm0
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  69,15,88,201                        ; addps         %xmm9,%xmm9
  DB  102,69,15,56,20,249                 ; blendvps      %xmm0,%xmm9,%xmm15
  DB  68,15,40,218                        ; movaps        %xmm2,%xmm11
  DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,92,193                           ; subps         %xmm1,%xmm0
  DB  69,15,40,204                        ; movaps        %xmm12,%xmm9
  DB  69,15,92,233                        ; subps         %xmm9,%xmm13
  DB  68,15,89,232                        ; mulps         %xmm0,%xmm13
  DB  69,15,88,237                        ; addps         %xmm13,%xmm13
  DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
  DB  69,15,92,229                        ; subps         %xmm13,%xmm12
  DB  15,40,193                           ; movaps        %xmm1,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,196,2                        ; cmpleps       %xmm4,%xmm0
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  102,68,15,56,20,225                 ; blendvps      %xmm0,%xmm1,%xmm12
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
  DB  68,15,89,215                        ; mulps         %xmm7,%xmm10
  DB  69,15,88,199                        ; addps         %xmm15,%xmm8
  DB  65,15,89,205                        ; mulps         %xmm13,%xmm1
  DB  65,15,88,203                        ; addps         %xmm11,%xmm1
  DB  65,15,88,204                        ; addps         %xmm12,%xmm1
  DB  15,89,213                           ; mulps         %xmm5,%xmm2
  DB  68,15,40,28,36                      ; movaps        (%rsp),%xmm11
  DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
  DB  68,15,88,202                        ; addps         %xmm2,%xmm9
  DB  15,40,197                           ; movaps        %xmm5,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,196,2                        ; cmpleps       %xmm4,%xmm0
  DB  15,40,212                           ; movaps        %xmm4,%xmm2
  DB  15,92,213                           ; subps         %xmm5,%xmm2
  DB  65,15,89,235                        ; mulps         %xmm11,%xmm5
  DB  15,88,237                           ; addps         %xmm5,%xmm5
  DB  69,15,92,243                        ; subps         %xmm11,%xmm14
  DB  68,15,89,242                        ; mulps         %xmm2,%xmm14
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  65,15,92,222                        ; subps         %xmm14,%xmm3
  DB  102,15,56,20,221                    ; blendvps      %xmm0,%xmm5,%xmm3
  DB  68,15,88,203                        ; addps         %xmm3,%xmm9
  DB  65,15,88,226                        ; addps         %xmm10,%xmm4
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  15,40,220                           ; movaps        %xmm4,%xmm3
  DB  15,40,230                           ; movaps        %xmm6,%xmm4
  DB  65,15,40,237                        ; movaps        %xmm13,%xmm5
  DB  65,15,40,243                        ; movaps        %xmm11,%xmm6
  DB  72,131,196,24                       ; add           $0x18,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_overlay_sse41
_sk_overlay_sse41 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
  DB  68,15,92,223                        ; subps         %xmm7,%xmm11
  DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
  DB  65,15,89,198                        ; mulps         %xmm14,%xmm0
  DB  68,15,92,211                        ; subps         %xmm3,%xmm10
  DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
  DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
  DB  68,15,88,192                        ; addps         %xmm0,%xmm8
  DB  68,15,40,235                        ; movaps        %xmm3,%xmm13
  DB  69,15,92,238                        ; subps         %xmm14,%xmm13
  DB  68,15,89,244                        ; mulps         %xmm4,%xmm14
  DB  15,40,207                           ; movaps        %xmm7,%xmm1
  DB  15,92,204                           ; subps         %xmm4,%xmm1
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,199,2                        ; cmpleps       %xmm7,%xmm0
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
  DB  68,15,89,231                        ; mulps         %xmm7,%xmm12
  DB  65,15,89,205                        ; mulps         %xmm13,%xmm1
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
  DB  68,15,92,233                        ; subps         %xmm1,%xmm13
  DB  102,69,15,56,20,238                 ; blendvps      %xmm0,%xmm14,%xmm13
  DB  69,15,88,197                        ; addps         %xmm13,%xmm8
  DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  15,88,200                           ; addps         %xmm0,%xmm1
  DB  68,15,40,235                        ; movaps        %xmm3,%xmm13
  DB  69,15,92,233                        ; subps         %xmm9,%xmm13
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  68,15,40,247                        ; movaps        %xmm7,%xmm14
  DB  68,15,92,245                        ; subps         %xmm5,%xmm14
  DB  15,40,197                           ; movaps        %xmm5,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,199,2                        ; cmpleps       %xmm7,%xmm0
  DB  69,15,88,201                        ; addps         %xmm9,%xmm9
  DB  69,15,89,245                        ; mulps         %xmm13,%xmm14
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
  DB  69,15,92,238                        ; subps         %xmm14,%xmm13
  DB  102,69,15,56,20,233                 ; blendvps      %xmm0,%xmm9,%xmm13
  DB  65,15,88,205                        ; addps         %xmm13,%xmm1
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  69,15,88,203                        ; addps         %xmm11,%xmm9
  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
  DB  68,15,92,218                        ; subps         %xmm2,%xmm11
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,92,238                        ; subps         %xmm6,%xmm13
  DB  15,40,198                           ; movaps        %xmm6,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,199,2                        ; cmpleps       %xmm7,%xmm0
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
  DB  69,15,88,237                        ; addps         %xmm13,%xmm13
  DB  69,15,92,229                        ; subps         %xmm13,%xmm12
  DB  102,68,15,56,20,226                 ; blendvps      %xmm0,%xmm2,%xmm12
  DB  69,15,88,204                        ; addps         %xmm12,%xmm9
  DB  68,15,89,215                        ; mulps         %xmm7,%xmm10
  DB  65,15,88,218                        ; addps         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_softlight_sse41
_sk_softlight_sse41 LABEL PROC
  DB  72,131,236,56                       ; sub           $0x38,%rsp
  DB  15,41,116,36,16                     ; movaps        %xmm6,0x10(%rsp)
  DB  15,40,244                           ; movaps        %xmm4,%xmm6
  DB  15,41,84,36,32                      ; movaps        %xmm2,0x20(%rsp)
  DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  15,87,228                           ; xorps         %xmm4,%xmm4
  DB  15,194,231,1                        ; cmpltps       %xmm7,%xmm4
  DB  15,40,198                           ; movaps        %xmm6,%xmm0
  DB  15,94,199                           ; divps         %xmm7,%xmm0
  DB  15,84,196                           ; andps         %xmm4,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
  DB  68,15,92,208                        ; subps         %xmm0,%xmm10
  DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
  DB  68,15,40,248                        ; movaps        %xmm0,%xmm15
  DB  15,82,208                           ; rsqrtps       %xmm0,%xmm2
  DB  68,15,83,218                        ; rcpps         %xmm2,%xmm11
  DB  68,15,92,216                        ; subps         %xmm0,%xmm11
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  15,89,210                           ; mulps         %xmm2,%xmm2
  DB  15,88,208                           ; addps         %xmm0,%xmm2
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  69,15,88,201                        ; addps         %xmm9,%xmm9
  DB  68,15,92,241                        ; subps         %xmm1,%xmm14
  DB  68,15,89,242                        ; mulps         %xmm2,%xmm14
  DB  184,0,0,224,64                      ; mov           $0x40e00000,%eax
  DB  102,68,15,110,232                   ; movd          %eax,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,89,253                        ; mulps         %xmm13,%xmm15
  DB  69,15,88,254                        ; addps         %xmm14,%xmm15
  DB  15,40,198                           ; movaps        %xmm6,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,199,2                        ; cmpleps       %xmm7,%xmm0
  DB  102,69,15,56,20,223                 ; blendvps      %xmm0,%xmm15,%xmm11
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  15,92,195                           ; subps         %xmm3,%xmm0
  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
  DB  68,15,88,211                        ; addps         %xmm3,%xmm10
  DB  68,15,89,214                        ; mulps         %xmm6,%xmm10
  DB  15,40,211                           ; movaps        %xmm3,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,89,199                           ; mulps         %xmm7,%xmm0
  DB  68,15,89,216                        ; mulps         %xmm0,%xmm11
  DB  68,15,88,218                        ; addps         %xmm2,%xmm11
  DB  68,15,194,203,2                     ; cmpleps       %xmm3,%xmm9
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  102,69,15,56,20,218                 ; blendvps      %xmm0,%xmm10,%xmm11
  DB  68,15,40,213                        ; movaps        %xmm5,%xmm10
  DB  68,15,94,215                        ; divps         %xmm7,%xmm10
  DB  68,15,84,212                        ; andps         %xmm4,%xmm10
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  69,15,92,202                        ; subps         %xmm10,%xmm9
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  15,89,192                           ; mulps         %xmm0,%xmm0
  DB  15,88,194                           ; addps         %xmm2,%xmm0
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,92,209                           ; subps         %xmm1,%xmm2
  DB  15,89,208                           ; mulps         %xmm0,%xmm2
  DB  65,15,82,194                        ; rsqrtps       %xmm10,%xmm0
  DB  68,15,83,240                        ; rcpps         %xmm0,%xmm14
  DB  69,15,92,242                        ; subps         %xmm10,%xmm14
  DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
  DB  68,15,88,210                        ; addps         %xmm2,%xmm10
  DB  15,40,197                           ; movaps        %xmm5,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,199,2                        ; cmpleps       %xmm7,%xmm0
  DB  102,69,15,56,20,242                 ; blendvps      %xmm0,%xmm10,%xmm14
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  68,15,88,203                        ; addps         %xmm3,%xmm9
  DB  15,41,44,36                         ; movaps        %xmm5,(%rsp)
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,89,242                        ; mulps         %xmm2,%xmm14
  DB  15,40,211                           ; movaps        %xmm3,%xmm2
  DB  15,89,213                           ; mulps         %xmm5,%xmm2
  DB  68,15,88,242                        ; addps         %xmm2,%xmm14
  DB  68,15,40,249                        ; movaps        %xmm1,%xmm15
  DB  15,194,195,2                        ; cmpleps       %xmm3,%xmm0
  DB  102,69,15,56,20,241                 ; blendvps      %xmm0,%xmm9,%xmm14
  DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
  DB  15,40,108,36,16                     ; movaps        0x10(%rsp),%xmm5
  DB  15,40,197                           ; movaps        %xmm5,%xmm0
  DB  15,94,199                           ; divps         %xmm7,%xmm0
  DB  15,84,196                           ; andps         %xmm4,%xmm0
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  15,92,209                           ; subps         %xmm1,%xmm2
  DB  15,92,200                           ; subps         %xmm0,%xmm1
  DB  68,15,89,232                        ; mulps         %xmm0,%xmm13
  DB  15,82,224                           ; rsqrtps       %xmm0,%xmm4
  DB  68,15,83,204                        ; rcpps         %xmm4,%xmm9
  DB  68,15,92,200                        ; subps         %xmm0,%xmm9
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,40,224                           ; movaps        %xmm0,%xmm4
  DB  15,89,228                           ; mulps         %xmm4,%xmm4
  DB  15,88,224                           ; addps         %xmm0,%xmm4
  DB  15,89,226                           ; mulps         %xmm2,%xmm4
  DB  68,15,88,236                        ; addps         %xmm4,%xmm13
  DB  15,40,197                           ; movaps        %xmm5,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,194,199,2                        ; cmpleps       %xmm7,%xmm0
  DB  102,69,15,56,20,205                 ; blendvps      %xmm0,%xmm13,%xmm9
  DB  68,15,40,108,36,32                  ; movaps        0x20(%rsp),%xmm13
  DB  65,15,40,197                        ; movaps        %xmm13,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,202                           ; mulps         %xmm2,%xmm1
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  15,40,211                           ; movaps        %xmm3,%xmm2
  DB  15,89,213                           ; mulps         %xmm5,%xmm2
  DB  68,15,88,202                        ; addps         %xmm2,%xmm9
  DB  15,88,203                           ; addps         %xmm3,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  15,40,213                           ; movaps        %xmm5,%xmm2
  DB  15,194,195,2                        ; cmpleps       %xmm3,%xmm0
  DB  102,68,15,56,20,201                 ; blendvps      %xmm0,%xmm1,%xmm9
  DB  68,15,92,255                        ; subps         %xmm7,%xmm15
  DB  69,15,89,199                        ; mulps         %xmm15,%xmm8
  DB  69,15,89,231                        ; mulps         %xmm15,%xmm12
  DB  69,15,89,253                        ; mulps         %xmm13,%xmm15
  DB  68,15,92,211                        ; subps         %xmm3,%xmm10
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  15,89,198                           ; mulps         %xmm6,%xmm0
  DB  68,15,88,192                        ; addps         %xmm0,%xmm8
  DB  69,15,88,195                        ; addps         %xmm11,%xmm8
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  15,40,44,36                         ; movaps        (%rsp),%xmm5
  DB  15,89,197                           ; mulps         %xmm5,%xmm0
  DB  68,15,88,224                        ; addps         %xmm0,%xmm12
  DB  69,15,88,230                        ; addps         %xmm14,%xmm12
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  15,89,194                           ; mulps         %xmm2,%xmm0
  DB  65,15,88,199                        ; addps         %xmm15,%xmm0
  DB  68,15,88,200                        ; addps         %xmm0,%xmm9
  DB  68,15,89,215                        ; mulps         %xmm7,%xmm10
  DB  65,15,88,218                        ; addps         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,230                           ; movaps        %xmm6,%xmm4
  DB  15,40,242                           ; movaps        %xmm2,%xmm6
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,40,204                        ; movaps        %xmm12,%xmm1
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  72,131,196,56                       ; add           $0x38,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_0_sse41
_sk_clamp_0_sse41 LABEL PROC
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  65,15,95,192                        ; maxps         %xmm8,%xmm0
  DB  65,15,95,200                        ; maxps         %xmm8,%xmm1
  DB  65,15,95,208                        ; maxps         %xmm8,%xmm2
  DB  65,15,95,216                        ; maxps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_1_sse41
_sk_clamp_1_sse41 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_a_sse41
_sk_clamp_a_sse41 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
  DB  15,93,195                           ; minps         %xmm3,%xmm0
  DB  15,93,203                           ; minps         %xmm3,%xmm1
  DB  15,93,211                           ; minps         %xmm3,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_set_rgb_sse41
_sk_set_rgb_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,15,16,80,8                      ; movss         0x8(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_rb_sse41
_sk_swap_rb_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_sse41
_sk_swap_sse41 LABEL PROC
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,40,202                        ; movaps        %xmm2,%xmm9
  DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
  DB  68,15,40,216                        ; movaps        %xmm0,%xmm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,40,205                           ; movaps        %xmm5,%xmm1
  DB  15,40,214                           ; movaps        %xmm6,%xmm2
  DB  15,40,223                           ; movaps        %xmm7,%xmm3
  DB  65,15,40,227                        ; movaps        %xmm11,%xmm4
  DB  65,15,40,234                        ; movaps        %xmm10,%xmm5
  DB  65,15,40,241                        ; movaps        %xmm9,%xmm6
  DB  65,15,40,248                        ; movaps        %xmm8,%xmm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_src_dst_sse41
_sk_move_src_dst_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,224                           ; movaps        %xmm0,%xmm4
  DB  15,40,233                           ; movaps        %xmm1,%xmm5
  DB  15,40,242                           ; movaps        %xmm2,%xmm6
  DB  15,40,251                           ; movaps        %xmm3,%xmm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_dst_src_sse41
_sk_move_dst_src_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,40,205                           ; movaps        %xmm5,%xmm1
  DB  15,40,214                           ; movaps        %xmm6,%xmm2
  DB  15,40,223                           ; movaps        %xmm7,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_premul_sse41
_sk_premul_sse41 LABEL PROC
  DB  15,89,195                           ; mulps         %xmm3,%xmm0
  DB  15,89,203                           ; mulps         %xmm3,%xmm1
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_unpremul_sse41
_sk_unpremul_sse41 LABEL PROC
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,94,203                        ; divps         %xmm3,%xmm9
  DB  68,15,194,195,4                     ; cmpneqps      %xmm3,%xmm8
  DB  69,15,84,193                        ; andps         %xmm9,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_srgb_sse41
_sk_from_srgb_sse41 LABEL PROC
  DB  184,145,131,158,61                  ; mov           $0x3d9e8391,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
  DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
  DB  69,15,89,246                        ; mulps         %xmm14,%xmm14
  DB  184,154,153,153,62                  ; mov           $0x3e99999a,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  184,92,143,50,63                    ; mov           $0x3f328f5c,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  69,15,88,204                        ; addps         %xmm12,%xmm9
  DB  184,10,215,35,59                    ; mov           $0x3b23d70a,%eax
  DB  102,68,15,110,232                   ; movd          %eax,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,89,206                        ; mulps         %xmm14,%xmm9
  DB  69,15,88,205                        ; addps         %xmm13,%xmm9
  DB  184,174,71,97,61                    ; mov           $0x3d6147ae,%eax
  DB  102,68,15,110,240                   ; movd          %eax,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  65,15,194,198,1                     ; cmpltps       %xmm14,%xmm0
  DB  102,69,15,56,20,202                 ; blendvps      %xmm0,%xmm10,%xmm9
  DB  69,15,40,251                        ; movaps        %xmm11,%xmm15
  DB  68,15,89,249                        ; mulps         %xmm1,%xmm15
  DB  15,40,193                           ; movaps        %xmm1,%xmm0
  DB  15,89,192                           ; mulps         %xmm0,%xmm0
  DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  69,15,88,212                        ; addps         %xmm12,%xmm10
  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
  DB  69,15,88,213                        ; addps         %xmm13,%xmm10
  DB  65,15,194,206,1                     ; cmpltps       %xmm14,%xmm1
  DB  15,40,193                           ; movaps        %xmm1,%xmm0
  DB  102,69,15,56,20,215                 ; blendvps      %xmm0,%xmm15,%xmm10
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  15,89,192                           ; mulps         %xmm0,%xmm0
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  69,15,88,196                        ; addps         %xmm12,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  69,15,88,197                        ; addps         %xmm13,%xmm8
  DB  65,15,194,214,1                     ; cmpltps       %xmm14,%xmm2
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  102,69,15,56,20,195                 ; blendvps      %xmm0,%xmm11,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_srgb_sse41
_sk_to_srgb_sse41 LABEL PROC
  DB  72,131,236,24                       ; sub           $0x18,%rsp
  DB  15,41,60,36                         ; movaps        %xmm7,(%rsp)
  DB  15,40,254                           ; movaps        %xmm6,%xmm7
  DB  15,40,245                           ; movaps        %xmm5,%xmm6
  DB  15,40,236                           ; movaps        %xmm4,%xmm5
  DB  15,40,227                           ; movaps        %xmm3,%xmm4
  DB  15,40,218                           ; movaps        %xmm2,%xmm3
  DB  15,40,209                           ; movaps        %xmm1,%xmm2
  DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
  DB  69,15,83,200                        ; rcpps         %xmm8,%xmm9
  DB  69,15,82,248                        ; rsqrtps       %xmm8,%xmm15
  DB  184,41,92,71,65                     ; mov           $0x41475c29,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  184,194,135,210,62                  ; mov           $0x3ed287c2,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  184,206,111,48,63                   ; mov           $0x3f306fce,%eax
  DB  102,68,15,110,232                   ; movd          %eax,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  184,168,87,202,61                   ; mov           $0x3dca57a8,%eax
  DB  53,0,0,0,128                        ; xor           $0x80000000,%eax
  DB  102,68,15,110,240                   ; movd          %eax,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,89,205                        ; mulps         %xmm13,%xmm9
  DB  69,15,88,206                        ; addps         %xmm14,%xmm9
  DB  69,15,89,252                        ; mulps         %xmm12,%xmm15
  DB  69,15,88,249                        ; addps         %xmm9,%xmm15
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  69,15,93,207                        ; minps         %xmm15,%xmm9
  DB  184,4,231,140,59                    ; mov           $0x3b8ce704,%eax
  DB  102,68,15,110,248                   ; movd          %eax,%xmm15
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
  DB  102,69,15,56,20,202                 ; blendvps      %xmm0,%xmm10,%xmm9
  DB  68,15,82,210                        ; rsqrtps       %xmm2,%xmm10
  DB  65,15,83,194                        ; rcpps         %xmm10,%xmm0
  DB  69,15,82,210                        ; rsqrtps       %xmm10,%xmm10
  DB  65,15,89,197                        ; mulps         %xmm13,%xmm0
  DB  65,15,88,198                        ; addps         %xmm14,%xmm0
  DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
  DB  68,15,88,208                        ; addps         %xmm0,%xmm10
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  65,15,93,202                        ; minps         %xmm10,%xmm1
  DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
  DB  65,15,194,215,1                     ; cmpltps       %xmm15,%xmm2
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  102,65,15,56,20,202                 ; blendvps      %xmm0,%xmm10,%xmm1
  DB  15,82,195                           ; rsqrtps       %xmm3,%xmm0
  DB  15,83,208                           ; rcpps         %xmm0,%xmm2
  DB  65,15,89,213                        ; mulps         %xmm13,%xmm2
  DB  65,15,88,214                        ; addps         %xmm14,%xmm2
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  65,15,89,196                        ; mulps         %xmm12,%xmm0
  DB  15,88,194                           ; addps         %xmm2,%xmm0
  DB  68,15,93,192                        ; minps         %xmm0,%xmm8
  DB  68,15,89,219                        ; mulps         %xmm3,%xmm11
  DB  65,15,194,223,1                     ; cmpltps       %xmm15,%xmm3
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  102,69,15,56,20,195                 ; blendvps      %xmm0,%xmm11,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  15,40,220                           ; movaps        %xmm4,%xmm3
  DB  15,40,229                           ; movaps        %xmm5,%xmm4
  DB  15,40,238                           ; movaps        %xmm6,%xmm5
  DB  15,40,247                           ; movaps        %xmm7,%xmm6
  DB  15,40,60,36                         ; movaps        (%rsp),%xmm7
  DB  72,131,196,24                       ; add           $0x18,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_2dot2_sse41
_sk_from_2dot2_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  65,15,82,192                        ; rsqrtps       %xmm8,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  68,15,82,200                        ; rsqrtps       %xmm0,%xmm9
  DB  65,15,82,193                        ; rsqrtps       %xmm9,%xmm0
  DB  68,15,82,208                        ; rsqrtps       %xmm0,%xmm10
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  15,89,192                           ; mulps         %xmm0,%xmm0
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
  DB  68,15,82,193                        ; rsqrtps       %xmm1,%xmm8
  DB  69,15,82,192                        ; rsqrtps       %xmm8,%xmm8
  DB  69,15,82,192                        ; rsqrtps       %xmm8,%xmm8
  DB  69,15,82,200                        ; rsqrtps       %xmm8,%xmm9
  DB  69,15,82,193                        ; rsqrtps       %xmm9,%xmm8
  DB  69,15,82,216                        ; rsqrtps       %xmm8,%xmm11
  DB  15,89,201                           ; mulps         %xmm1,%xmm1
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  68,15,89,193                        ; mulps         %xmm1,%xmm8
  DB  69,15,89,195                        ; mulps         %xmm11,%xmm8
  DB  69,15,95,194                        ; maxps         %xmm10,%xmm8
  DB  15,82,202                           ; rsqrtps       %xmm2,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  68,15,82,217                        ; rsqrtps       %xmm1,%xmm11
  DB  65,15,82,203                        ; rsqrtps       %xmm11,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,89,210                           ; mulps         %xmm2,%xmm2
  DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
  DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
  DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  69,15,95,202                        ; maxps         %xmm10,%xmm9
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_2dot2_sse41
_sk_to_2dot2_sse41 LABEL PROC
  DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
  DB  65,15,82,192                        ; rsqrtps       %xmm8,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  68,15,82,200                        ; rsqrtps       %xmm0,%xmm9
  DB  69,15,83,192                        ; rcpps         %xmm8,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  65,15,83,193                        ; rcpps         %xmm9,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  65,15,95,192                        ; maxps         %xmm8,%xmm0
  DB  68,15,82,201                        ; rsqrtps       %xmm1,%xmm9
  DB  65,15,82,201                        ; rsqrtps       %xmm9,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  68,15,82,209                        ; rsqrtps       %xmm1,%xmm10
  DB  69,15,83,201                        ; rcpps         %xmm9,%xmm9
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  65,15,83,202                        ; rcpps         %xmm10,%xmm1
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  65,15,95,200                        ; maxps         %xmm8,%xmm1
  DB  68,15,82,202                        ; rsqrtps       %xmm2,%xmm9
  DB  65,15,82,209                        ; rsqrtps       %xmm9,%xmm2
  DB  15,82,210                           ; rsqrtps       %xmm2,%xmm2
  DB  15,82,210                           ; rsqrtps       %xmm2,%xmm2
  DB  15,82,210                           ; rsqrtps       %xmm2,%xmm2
  DB  68,15,82,210                        ; rsqrtps       %xmm2,%xmm10
  DB  69,15,83,201                        ; rcpps         %xmm9,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  65,15,83,210                        ; rcpps         %xmm10,%xmm2
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  65,15,95,208                        ; maxps         %xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_rgb_to_hsl_sse41
_sk_rgb_to_hsl_sse41 LABEL PROC
  DB  72,131,236,24                       ; sub           $0x18,%rsp
  DB  15,41,60,36                         ; movaps        %xmm7,(%rsp)
  DB  15,40,254                           ; movaps        %xmm6,%xmm7
  DB  15,40,245                           ; movaps        %xmm5,%xmm6
  DB  15,40,236                           ; movaps        %xmm4,%xmm5
  DB  15,40,227                           ; movaps        %xmm3,%xmm4
  DB  15,40,218                           ; movaps        %xmm2,%xmm3
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  65,184,171,170,42,62                ; mov           $0x3e2aaaab,%r8d
  DB  65,185,0,0,192,64                   ; mov           $0x40c00000,%r9d
  DB  184,0,0,0,64                        ; mov           $0x40000000,%eax
  DB  185,0,0,128,64                      ; mov           $0x40800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  68,15,40,224                        ; movaps        %xmm0,%xmm12
  DB  68,15,95,225                        ; maxps         %xmm1,%xmm12
  DB  68,15,95,227                        ; maxps         %xmm3,%xmm12
  DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
  DB  68,15,93,233                        ; minps         %xmm1,%xmm13
  DB  68,15,93,235                        ; minps         %xmm3,%xmm13
  DB  69,15,40,204                        ; movaps        %xmm12,%xmm9
  DB  68,15,194,200,0                     ; cmpeqps       %xmm0,%xmm9
  DB  68,15,40,241                        ; movaps        %xmm1,%xmm14
  DB  68,15,92,243                        ; subps         %xmm3,%xmm14
  DB  68,15,40,249                        ; movaps        %xmm1,%xmm15
  DB  68,15,194,251,1                     ; cmpltps       %xmm3,%xmm15
  DB  69,15,40,212                        ; movaps        %xmm12,%xmm10
  DB  68,15,194,209,0                     ; cmpeqps       %xmm1,%xmm10
  DB  15,92,216                           ; subps         %xmm0,%xmm3
  DB  15,92,193                           ; subps         %xmm1,%xmm0
  DB  65,15,40,212                        ; movaps        %xmm12,%xmm2
  DB  65,15,92,213                        ; subps         %xmm13,%xmm2
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,94,218                        ; divps         %xmm2,%xmm11
  DB  65,15,89,195                        ; mulps         %xmm11,%xmm0
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,88,192                        ; addps         %xmm0,%xmm8
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,88,217                           ; addps         %xmm1,%xmm3
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  102,68,15,56,20,195                 ; blendvps      %xmm0,%xmm3,%xmm8
  DB  69,15,89,243                        ; mulps         %xmm11,%xmm14
  DB  102,65,15,110,217                   ; movd          %r9d,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,84,223                        ; andps         %xmm15,%xmm3
  DB  65,15,88,222                        ; addps         %xmm14,%xmm3
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  102,68,15,56,20,195                 ; blendvps      %xmm0,%xmm3,%xmm8
  DB  65,15,40,220                        ; movaps        %xmm12,%xmm3
  DB  65,15,92,204                        ; subps         %xmm12,%xmm1
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,40,204                        ; movaps        %xmm12,%xmm9
  DB  69,15,89,202                        ; mulps         %xmm10,%xmm9
  DB  69,15,194,209,1                     ; cmpltps       %xmm9,%xmm10
  DB  65,15,92,205                        ; subps         %xmm13,%xmm1
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  102,68,15,56,20,225                 ; blendvps      %xmm0,%xmm1,%xmm12
  DB  65,15,194,221,4                     ; cmpneqps      %xmm13,%xmm3
  DB  102,65,15,110,192                   ; movd          %r8d,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  68,15,84,195                        ; andps         %xmm3,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  65,15,94,212                        ; divps         %xmm12,%xmm2
  DB  15,84,211                           ; andps         %xmm3,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  15,40,202                           ; movaps        %xmm2,%xmm1
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  15,40,220                           ; movaps        %xmm4,%xmm3
  DB  15,40,229                           ; movaps        %xmm5,%xmm4
  DB  15,40,238                           ; movaps        %xmm6,%xmm5
  DB  15,40,247                           ; movaps        %xmm7,%xmm6
  DB  15,40,60,36                         ; movaps        (%rsp),%xmm7
  DB  72,131,196,24                       ; add           $0x18,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hsl_to_rgb_sse41
_sk_hsl_to_rgb_sse41 LABEL PROC
  DB  72,129,236,152,0,0,0                ; sub           $0x98,%rsp
  DB  15,41,188,36,128,0,0,0              ; movaps        %xmm7,0x80(%rsp)
  DB  15,41,116,36,112                    ; movaps        %xmm6,0x70(%rsp)
  DB  15,41,108,36,96                     ; movaps        %xmm5,0x60(%rsp)
  DB  15,41,100,36,80                     ; movaps        %xmm4,0x50(%rsp)
  DB  15,41,92,36,64                      ; movaps        %xmm3,0x40(%rsp)
  DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,41,28,36                         ; movaps        %xmm3,(%rsp)
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  15,194,195,1                        ; cmpltps       %xmm3,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
  DB  15,88,225                           ; addps         %xmm1,%xmm4
  DB  15,89,226                           ; mulps         %xmm2,%xmm4
  DB  15,40,217                           ; movaps        %xmm1,%xmm3
  DB  15,40,249                           ; movaps        %xmm1,%xmm7
  DB  15,88,250                           ; addps         %xmm2,%xmm7
  DB  15,89,218                           ; mulps         %xmm2,%xmm3
  DB  15,40,234                           ; movaps        %xmm2,%xmm5
  DB  15,92,251                           ; subps         %xmm3,%xmm7
  DB  102,15,56,20,252                    ; blendvps      %xmm0,%xmm4,%xmm7
  DB  184,0,0,0,64                        ; mov           $0x40000000,%eax
  DB  185,171,170,170,62                  ; mov           $0x3eaaaaab,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,41,84,36,48                      ; movaps        %xmm2,0x30(%rsp)
  DB  65,15,88,210                        ; addps         %xmm10,%xmm2
  DB  185,0,0,0,0                         ; mov           $0x0,%ecx
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  15,194,194,1                        ; cmpltps       %xmm2,%xmm0
  DB  15,40,218                           ; movaps        %xmm2,%xmm3
  DB  65,15,92,216                        ; subps         %xmm8,%xmm3
  DB  68,15,40,226                        ; movaps        %xmm2,%xmm12
  DB  102,68,15,56,20,227                 ; blendvps      %xmm0,%xmm3,%xmm12
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  65,15,194,198,1                     ; cmpltps       %xmm14,%xmm0
  DB  68,15,41,116,36,32                  ; movaps        %xmm14,0x20(%rsp)
  DB  65,15,40,216                        ; movaps        %xmm8,%xmm3
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  102,68,15,56,20,227                 ; blendvps      %xmm0,%xmm3,%xmm12
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  15,40,229                           ; movaps        %xmm5,%xmm4
  DB  15,41,100,36,16                     ; movaps        %xmm4,0x10(%rsp)
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  184,171,170,42,62                   ; mov           $0x3e2aaaab,%eax
  DB  15,40,199                           ; movaps        %xmm7,%xmm0
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  185,0,0,192,64                      ; mov           $0x40c00000,%ecx
  DB  102,15,110,241                      ; movd          %ecx,%xmm6
  DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
  DB  15,89,240                           ; mulps         %xmm0,%xmm6
  DB  185,171,170,42,63                   ; mov           $0x3f2aaaab,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,40,235                           ; movaps        %xmm3,%xmm5
  DB  65,15,92,236                        ; subps         %xmm12,%xmm5
  DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
  DB  69,15,40,252                        ; movaps        %xmm12,%xmm15
  DB  68,15,194,227,1                     ; cmpltps       %xmm3,%xmm12
  DB  15,89,238                           ; mulps         %xmm6,%xmm5
  DB  65,15,88,233                        ; addps         %xmm9,%xmm5
  DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  102,68,15,56,20,221                 ; blendvps      %xmm0,%xmm5,%xmm11
  DB  68,15,194,60,36,1                   ; cmpltps       (%rsp),%xmm15
  DB  65,15,40,199                        ; movaps        %xmm15,%xmm0
  DB  102,68,15,56,20,223                 ; blendvps      %xmm0,%xmm7,%xmm11
  DB  102,15,110,232                      ; movd          %eax,%xmm5
  DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
  DB  68,15,194,237,1                     ; cmpltps       %xmm5,%xmm13
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
  DB  65,15,40,197                        ; movaps        %xmm13,%xmm0
  DB  102,68,15,56,20,218                 ; blendvps      %xmm0,%xmm2,%xmm11
  DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
  DB  68,15,194,225,0                     ; cmpeqps       %xmm1,%xmm12
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  102,68,15,56,20,220                 ; blendvps      %xmm0,%xmm4,%xmm11
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,194,194,1                     ; cmpltps       %xmm10,%xmm0
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  69,15,40,234                        ; movaps        %xmm10,%xmm13
  DB  102,68,15,56,20,233                 ; blendvps      %xmm0,%xmm1,%xmm13
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  65,15,194,198,1                     ; cmpltps       %xmm14,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  102,68,15,56,20,233                 ; blendvps      %xmm0,%xmm1,%xmm13
  DB  68,15,40,243                        ; movaps        %xmm3,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  65,15,40,229                        ; movaps        %xmm13,%xmm4
  DB  69,15,40,253                        ; movaps        %xmm13,%xmm15
  DB  68,15,194,235,1                     ; cmpltps       %xmm3,%xmm13
  DB  68,15,89,246                        ; mulps         %xmm6,%xmm14
  DB  69,15,88,241                        ; addps         %xmm9,%xmm14
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  65,15,40,197                        ; movaps        %xmm13,%xmm0
  DB  102,65,15,56,20,206                 ; blendvps      %xmm0,%xmm14,%xmm1
  DB  68,15,40,52,36                      ; movaps        (%rsp),%xmm14
  DB  69,15,194,254,1                     ; cmpltps       %xmm14,%xmm15
  DB  65,15,40,199                        ; movaps        %xmm15,%xmm0
  DB  102,15,56,20,207                    ; blendvps      %xmm0,%xmm7,%xmm1
  DB  15,194,229,1                        ; cmpltps       %xmm5,%xmm4
  DB  15,40,214                           ; movaps        %xmm6,%xmm2
  DB  65,15,89,210                        ; mulps         %xmm10,%xmm2
  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  102,15,56,20,202                    ; blendvps      %xmm0,%xmm2,%xmm1
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  68,15,40,124,36,16                  ; movaps        0x10(%rsp),%xmm15
  DB  102,65,15,56,20,207                 ; blendvps      %xmm0,%xmm15,%xmm1
  DB  68,15,92,84,36,48                   ; subps         0x30(%rsp),%xmm10
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  65,15,92,208                        ; subps         %xmm8,%xmm2
  DB  69,15,40,232                        ; movaps        %xmm8,%xmm13
  DB  69,15,194,194,1                     ; cmpltps       %xmm10,%xmm8
  DB  65,15,40,226                        ; movaps        %xmm10,%xmm4
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  102,15,56,20,226                    ; blendvps      %xmm0,%xmm2,%xmm4
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  15,194,68,36,32,1                   ; cmpltps       0x20(%rsp),%xmm0
  DB  69,15,88,234                        ; addps         %xmm10,%xmm13
  DB  102,65,15,56,20,229                 ; blendvps      %xmm0,%xmm13,%xmm4
  DB  68,15,89,214                        ; mulps         %xmm6,%xmm10
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,194,195,1                        ; cmpltps       %xmm3,%xmm0
  DB  15,92,220                           ; subps         %xmm4,%xmm3
  DB  15,89,222                           ; mulps         %xmm6,%xmm3
  DB  69,15,88,209                        ; addps         %xmm9,%xmm10
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  102,68,15,56,20,203                 ; blendvps      %xmm0,%xmm3,%xmm9
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  65,15,194,198,1                     ; cmpltps       %xmm14,%xmm0
  DB  102,68,15,56,20,207                 ; blendvps      %xmm0,%xmm7,%xmm9
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,194,197,1                        ; cmpltps       %xmm5,%xmm0
  DB  102,69,15,56,20,202                 ; blendvps      %xmm0,%xmm10,%xmm9
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  102,69,15,56,20,207                 ; blendvps      %xmm0,%xmm15,%xmm9
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  15,40,92,36,64                      ; movaps        0x40(%rsp),%xmm3
  DB  15,40,100,36,80                     ; movaps        0x50(%rsp),%xmm4
  DB  15,40,108,36,96                     ; movaps        0x60(%rsp),%xmm5
  DB  15,40,116,36,112                    ; movaps        0x70(%rsp),%xmm6
  DB  15,40,188,36,128,0,0,0              ; movaps        0x80(%rsp),%xmm7
  DB  72,129,196,152,0,0,0                ; add           $0x98,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_1_float_sse41
_sk_scale_1_float_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_u8_sse41
_sk_scale_u8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,68,15,56,49,4,56                ; pmovzxbd      (%rax,%rdi,1),%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_1_float_sse41
_sk_lerp_1_float_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,92,196                           ; subps         %xmm4,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,92,205                           ; subps         %xmm5,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,92,214                           ; subps         %xmm6,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  15,92,223                           ; subps         %xmm7,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_u8_sse41
_sk_lerp_u8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,68,15,56,49,4,56                ; pmovzxbd      (%rax,%rdi,1),%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  15,92,196                           ; subps         %xmm4,%xmm0
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,92,205                           ; subps         %xmm5,%xmm1
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,92,214                           ; subps         %xmm6,%xmm2
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  15,92,223                           ; subps         %xmm7,%xmm3
  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_565_sse41
_sk_lerp_565_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,68,15,56,51,4,120               ; pmovzxwd      (%rax,%rdi,2),%xmm8
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  68,15,91,203                        ; cvtdq2ps      %xmm3,%xmm9
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  68,15,91,203                        ; cvtdq2ps      %xmm3,%xmm9
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  68,15,91,195                        ; cvtdq2ps      %xmm3,%xmm8
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  15,92,196                           ; subps         %xmm4,%xmm0
  DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,92,205                           ; subps         %xmm5,%xmm1
  DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,92,214                           ; subps         %xmm6,%xmm2
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_tables_sse41
_sk_load_tables_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  243,68,15,111,4,185                 ; movdqu        (%rcx,%rdi,4),%xmm8
  DB  185,255,0,0,0                       ; mov           $0xff,%ecx
  DB  102,15,110,193                      ; movd          %ecx,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,65,15,111,200                   ; movdqa        %xmm8,%xmm1
  DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
  DB  102,15,219,200                      ; pand          %xmm0,%xmm1
  DB  102,65,15,111,208                   ; movdqa        %xmm8,%xmm2
  DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
  DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
  DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
  DB  65,137,201                          ; mov           %ecx,%r9d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,73,15,126,194                   ; movq          %xmm0,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  243,67,15,16,4,152                  ; movss         (%r8,%r11,4),%xmm0
  DB  102,67,15,58,33,4,144,16            ; insertps      $0x10,(%r8,%r10,4),%xmm0
  DB  102,67,15,58,33,4,136,32            ; insertps      $0x20,(%r8,%r9,4),%xmm0
  DB  102,65,15,58,33,4,136,48            ; insertps      $0x30,(%r8,%rcx,4),%xmm0
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  102,73,15,58,22,202,1               ; pextrq        $0x1,%xmm1,%r10
  DB  77,137,209                          ; mov           %r10,%r9
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
  DB  65,137,203                          ; mov           %ecx,%r11d
  DB  65,129,227,255,255,255,0            ; and           $0xffffff,%r11d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  65,129,226,255,255,255,0            ; and           $0xffffff,%r10d
  DB  243,67,15,16,12,152                 ; movss         (%r8,%r11,4),%xmm1
  DB  102,65,15,58,33,12,8,16             ; insertps      $0x10,(%r8,%rcx,1),%xmm1
  DB  243,67,15,16,28,144                 ; movss         (%r8,%r10,4),%xmm3
  DB  102,15,58,33,203,32                 ; insertps      $0x20,%xmm3,%xmm1
  DB  243,67,15,16,28,136                 ; movss         (%r8,%r9,4),%xmm3
  DB  102,15,58,33,203,48                 ; insertps      $0x30,%xmm3,%xmm1
  DB  76,139,72,24                        ; mov           0x18(%rax),%r9
  DB  102,72,15,58,22,209,1               ; pextrq        $0x1,%xmm2,%rcx
  DB  68,15,183,193                       ; movzwl        %cx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,72,15,126,208                   ; movq          %xmm2,%rax
  DB  68,15,183,208                       ; movzwl        %ax,%r10d
  DB  72,193,232,30                       ; shr           $0x1e,%rax
  DB  243,67,15,16,20,145                 ; movss         (%r9,%r10,4),%xmm2
  DB  102,65,15,58,33,20,1,16             ; insertps      $0x10,(%r9,%rax,1),%xmm2
  DB  243,67,15,16,28,129                 ; movss         (%r9,%r8,4),%xmm3
  DB  102,15,58,33,211,32                 ; insertps      $0x20,%xmm3,%xmm2
  DB  243,65,15,16,28,137                 ; movss         (%r9,%rcx,4),%xmm3
  DB  102,15,58,33,211,48                 ; insertps      $0x30,%xmm3,%xmm2
  DB  102,65,15,114,208,24                ; psrld         $0x18,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_tables_u16_be_sse41
_sk_load_tables_u16_be_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  243,15,111,4,249                    ; movdqu        (%rcx,%rdi,8),%xmm0
  DB  243,15,111,76,249,16                ; movdqu        0x10(%rcx,%rdi,8),%xmm1
  DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
  DB  102,15,97,217                       ; punpcklwd     %xmm1,%xmm3
  DB  102,15,105,193                      ; punpckhwd     %xmm1,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
  DB  102,15,105,216                      ; punpckhwd     %xmm0,%xmm3
  DB  102,15,56,51,193                    ; pmovzxwd      %xmm1,%xmm0
  DB  185,255,0,0,0                       ; mov           $0xff,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  102,68,15,112,194,0                 ; pshufd        $0x0,%xmm2,%xmm8
  DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
  DB  102,73,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%r9
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  73,193,233,30                       ; shr           $0x1e,%r9
  DB  243,67,15,16,4,144                  ; movss         (%r8,%r10,4),%xmm0
  DB  102,65,15,58,33,4,8,16              ; insertps      $0x10,(%r8,%rcx,1),%xmm0
  DB  102,67,15,58,33,4,152,32            ; insertps      $0x20,(%r8,%r11,4),%xmm0
  DB  102,67,15,58,33,4,8,48              ; insertps      $0x30,(%r8,%r9,1),%xmm0
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  102,15,115,217,8                    ; psrldq        $0x8,%xmm1
  DB  102,15,56,51,201                    ; pmovzxwd      %xmm1,%xmm1
  DB  102,65,15,219,200                   ; pand          %xmm8,%xmm1
  DB  102,73,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%r9
  DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  73,193,233,30                       ; shr           $0x1e,%r9
  DB  243,67,15,16,12,144                 ; movss         (%r8,%r10,4),%xmm1
  DB  102,65,15,58,33,12,8,16             ; insertps      $0x10,(%r8,%rcx,1),%xmm1
  DB  243,67,15,16,20,152                 ; movss         (%r8,%r11,4),%xmm2
  DB  102,15,58,33,202,32                 ; insertps      $0x20,%xmm2,%xmm1
  DB  243,67,15,16,20,8                   ; movss         (%r8,%r9,1),%xmm2
  DB  102,15,58,33,202,48                 ; insertps      $0x30,%xmm2,%xmm1
  DB  76,139,64,24                        ; mov           0x18(%rax),%r8
  DB  102,15,56,51,211                    ; pmovzxwd      %xmm3,%xmm2
  DB  102,65,15,219,208                   ; pand          %xmm8,%xmm2
  DB  102,72,15,58,22,209,1               ; pextrq        $0x1,%xmm2,%rcx
  DB  102,72,15,126,208                   ; movq          %xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,30                       ; shr           $0x1e,%rax
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  243,67,15,16,20,136                 ; movss         (%r8,%r9,4),%xmm2
  DB  102,65,15,58,33,20,0,16             ; insertps      $0x10,(%r8,%rax,1),%xmm2
  DB  102,67,15,58,33,20,144,32           ; insertps      $0x20,(%r8,%r10,4),%xmm2
  DB  102,65,15,58,33,20,8,48             ; insertps      $0x30,(%r8,%rcx,1),%xmm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  102,15,112,219,78                   ; pshufd        $0x4e,%xmm3,%xmm3
  DB  102,68,15,111,203                   ; movdqa        %xmm3,%xmm9
  DB  102,65,15,113,241,8                 ; psllw         $0x8,%xmm9
  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
  DB  102,65,15,235,217                   ; por           %xmm9,%xmm3
  DB  102,15,56,51,219                    ; pmovzxwd      %xmm3,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_tables_rgb_u16_be_sse41
_sk_load_tables_rgb_u16_be_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,141,12,127                       ; lea           (%rdi,%rdi,2),%r9
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  243,66,15,111,20,73                 ; movdqu        (%rcx,%r9,2),%xmm2
  DB  243,66,15,111,68,73,8               ; movdqu        0x8(%rcx,%r9,2),%xmm0
  DB  102,15,115,216,4                    ; psrldq        $0x4,%xmm0
  DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
  DB  102,15,115,217,6                    ; psrldq        $0x6,%xmm1
  DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
  DB  102,15,115,216,6                    ; psrldq        $0x6,%xmm0
  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
  DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  102,68,15,112,192,78                ; pshufd        $0x4e,%xmm0,%xmm8
  DB  102,15,105,209                      ; punpckhwd     %xmm1,%xmm2
  DB  102,15,56,51,192                    ; pmovzxwd      %xmm0,%xmm0
  DB  185,255,0,0,0                       ; mov           $0xff,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  102,68,15,112,201,0                 ; pshufd        $0x0,%xmm1,%xmm9
  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
  DB  102,73,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%r9
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  73,193,233,30                       ; shr           $0x1e,%r9
  DB  243,67,15,16,4,144                  ; movss         (%r8,%r10,4),%xmm0
  DB  102,65,15,58,33,4,8,16              ; insertps      $0x10,(%r8,%rcx,1),%xmm0
  DB  102,67,15,58,33,4,152,32            ; insertps      $0x20,(%r8,%r11,4),%xmm0
  DB  102,67,15,58,33,4,8,48              ; insertps      $0x30,(%r8,%r9,1),%xmm0
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  102,65,15,56,51,200                 ; pmovzxwd      %xmm8,%xmm1
  DB  102,65,15,219,201                   ; pand          %xmm9,%xmm1
  DB  102,73,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%r9
  DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  69,137,203                          ; mov           %r9d,%r11d
  DB  73,193,233,30                       ; shr           $0x1e,%r9
  DB  243,67,15,16,12,144                 ; movss         (%r8,%r10,4),%xmm1
  DB  102,65,15,58,33,12,8,16             ; insertps      $0x10,(%r8,%rcx,1),%xmm1
  DB  243,67,15,16,28,152                 ; movss         (%r8,%r11,4),%xmm3
  DB  102,15,58,33,203,32                 ; insertps      $0x20,%xmm3,%xmm1
  DB  243,67,15,16,28,8                   ; movss         (%r8,%r9,1),%xmm3
  DB  102,15,58,33,203,48                 ; insertps      $0x30,%xmm3,%xmm1
  DB  76,139,64,24                        ; mov           0x18(%rax),%r8
  DB  102,15,56,51,210                    ; pmovzxwd      %xmm2,%xmm2
  DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
  DB  102,72,15,58,22,209,1               ; pextrq        $0x1,%xmm2,%rcx
  DB  102,72,15,126,208                   ; movq          %xmm2,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,30                       ; shr           $0x1e,%rax
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  243,67,15,16,20,136                 ; movss         (%r8,%r9,4),%xmm2
  DB  102,65,15,58,33,20,0,16             ; insertps      $0x10,(%r8,%rax,1),%xmm2
  DB  102,67,15,58,33,20,144,32           ; insertps      $0x20,(%r8,%r10,4),%xmm2
  DB  102,65,15,58,33,20,8,48             ; insertps      $0x30,(%r8,%rcx,1),%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_byte_tables_sse41
_sk_byte_tables_sse41 LABEL PROC
  DB  65,86                               ; push          %r14
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,127,67                      ; mov           $0x437f0000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
  DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,73,15,126,193                   ; movq          %xmm0,%r9
  DB  69,137,202                          ; mov           %r9d,%r10d
  DB  77,137,203                          ; mov           %r9,%r11
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  76,139,48                           ; mov           (%rax),%r14
  DB  76,139,72,8                         ; mov           0x8(%rax),%r9
  DB  102,67,15,58,32,4,22,0              ; pinsrb        $0x0,(%r14,%r10,1),%xmm0
  DB  102,67,15,58,32,4,30,1              ; pinsrb        $0x1,(%r14,%r11,1),%xmm0
  DB  67,15,182,28,6                      ; movzbl        (%r14,%r8,1),%ebx
  DB  102,15,58,32,195,2                  ; pinsrb        $0x2,%ebx,%xmm0
  DB  65,15,182,12,14                     ; movzbl        (%r14,%rcx,1),%ecx
  DB  102,15,58,32,193,3                  ; pinsrb        $0x3,%ecx,%xmm0
  DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  185,129,128,128,59                  ; mov           $0x3b808081,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
  DB  102,72,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  102,67,15,58,32,12,17,0             ; pinsrb        $0x0,(%r9,%r10,1),%xmm1
  DB  102,65,15,58,32,12,25,1             ; pinsrb        $0x1,(%r9,%rbx,1),%xmm1
  DB  67,15,182,28,1                      ; movzbl        (%r9,%r8,1),%ebx
  DB  102,15,58,32,203,2                  ; pinsrb        $0x2,%ebx,%xmm1
  DB  65,15,182,12,9                      ; movzbl        (%r9,%rcx,1),%ecx
  DB  102,15,58,32,201,3                  ; pinsrb        $0x3,%ecx,%xmm1
  DB  102,15,56,49,201                    ; pmovzxbd      %xmm1,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  76,139,72,16                        ; mov           0x10(%rax),%r9
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,91,210                       ; cvtps2dq      %xmm2,%xmm2
  DB  102,72,15,58,22,211,1               ; pextrq        $0x1,%xmm2,%rbx
  DB  65,137,216                          ; mov           %ebx,%r8d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  102,72,15,126,209                   ; movq          %xmm2,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,58,32,20,17,0             ; pinsrb        $0x0,(%r9,%r10,1),%xmm2
  DB  102,65,15,58,32,20,9,1              ; pinsrb        $0x1,(%r9,%rcx,1),%xmm2
  DB  67,15,182,12,1                      ; movzbl        (%r9,%r8,1),%ecx
  DB  102,15,58,32,209,2                  ; pinsrb        $0x2,%ecx,%xmm2
  DB  65,15,182,12,25                     ; movzbl        (%r9,%rbx,1),%ecx
  DB  102,15,58,32,209,3                  ; pinsrb        $0x3,%ecx,%xmm2
  DB  102,15,56,49,210                    ; pmovzxbd      %xmm2,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,65,15,91,216                    ; cvtps2dq      %xmm8,%xmm3
  DB  102,72,15,58,22,217,1               ; pextrq        $0x1,%xmm3,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,72,15,126,219                   ; movq          %xmm3,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  102,66,15,58,32,28,8,0              ; pinsrb        $0x0,(%rax,%r9,1),%xmm3
  DB  102,15,58,32,28,24,1                ; pinsrb        $0x1,(%rax,%rbx,1),%xmm3
  DB  66,15,182,28,0                      ; movzbl        (%rax,%r8,1),%ebx
  DB  102,15,58,32,219,2                  ; pinsrb        $0x2,%ebx,%xmm3
  DB  15,182,4,8                          ; movzbl        (%rax,%rcx,1),%eax
  DB  102,15,58,32,216,3                  ; pinsrb        $0x3,%eax,%xmm3
  DB  102,15,56,49,219                    ; pmovzxbd      %xmm3,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,94                               ; pop           %r14
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_byte_tables_rgb_sse41
_sk_byte_tables_rgb_sse41 LABEL PROC
  DB  65,86                               ; push          %r14
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  139,72,24                           ; mov           0x18(%rax),%ecx
  DB  255,201                             ; dec           %ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
  DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,73,15,126,193                   ; movq          %xmm0,%r9
  DB  69,137,202                          ; mov           %r9d,%r10d
  DB  77,137,203                          ; mov           %r9,%r11
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  76,139,48                           ; mov           (%rax),%r14
  DB  76,139,72,8                         ; mov           0x8(%rax),%r9
  DB  102,67,15,58,32,4,22,0              ; pinsrb        $0x0,(%r14,%r10,1),%xmm0
  DB  102,67,15,58,32,4,30,1              ; pinsrb        $0x1,(%r14,%r11,1),%xmm0
  DB  67,15,182,28,6                      ; movzbl        (%r14,%r8,1),%ebx
  DB  102,15,58,32,195,2                  ; pinsrb        $0x2,%ebx,%xmm0
  DB  65,15,182,12,14                     ; movzbl        (%r14,%rcx,1),%ecx
  DB  102,15,58,32,193,3                  ; pinsrb        $0x3,%ecx,%xmm0
  DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  185,129,128,128,59                  ; mov           $0x3b808081,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
  DB  102,72,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  102,67,15,58,32,12,17,0             ; pinsrb        $0x0,(%r9,%r10,1),%xmm1
  DB  102,65,15,58,32,12,25,1             ; pinsrb        $0x1,(%r9,%rbx,1),%xmm1
  DB  67,15,182,28,1                      ; movzbl        (%r9,%r8,1),%ebx
  DB  102,15,58,32,203,2                  ; pinsrb        $0x2,%ebx,%xmm1
  DB  65,15,182,12,9                      ; movzbl        (%r9,%rcx,1),%ecx
  DB  102,15,58,32,201,3                  ; pinsrb        $0x3,%ecx,%xmm1
  DB  102,15,56,49,201                    ; pmovzxbd      %xmm1,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  72,139,64,16                        ; mov           0x10(%rax),%rax
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,91,210                       ; cvtps2dq      %xmm2,%xmm2
  DB  102,72,15,58,22,209,1               ; pextrq        $0x1,%xmm2,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,72,15,126,211                   ; movq          %xmm2,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  102,66,15,58,32,20,8,0              ; pinsrb        $0x0,(%rax,%r9,1),%xmm2
  DB  102,15,58,32,20,24,1                ; pinsrb        $0x1,(%rax,%rbx,1),%xmm2
  DB  66,15,182,28,0                      ; movzbl        (%rax,%r8,1),%ebx
  DB  102,15,58,32,211,2                  ; pinsrb        $0x2,%ebx,%xmm2
  DB  15,182,4,8                          ; movzbl        (%rax,%rcx,1),%eax
  DB  102,15,58,32,208,3                  ; pinsrb        $0x3,%eax,%xmm2
  DB  102,15,56,49,210                    ; pmovzxbd      %xmm2,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,94                               ; pop           %r14
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_r_sse41
_sk_table_r_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  102,65,15,91,192                    ; cvtps2dq      %xmm8,%xmm0
  DB  102,72,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,67,15,16,4,145                  ; movss         (%r9,%r10,4),%xmm0
  DB  102,65,15,58,33,4,137,16            ; insertps      $0x10,(%r9,%rcx,4),%xmm0
  DB  243,71,15,16,4,129                  ; movss         (%r9,%r8,4),%xmm8
  DB  102,65,15,58,33,192,32              ; insertps      $0x20,%xmm8,%xmm0
  DB  243,69,15,16,4,129                  ; movss         (%r9,%rax,4),%xmm8
  DB  102,65,15,58,33,192,48              ; insertps      $0x30,%xmm8,%xmm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_g_sse41
_sk_table_g_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,193                        ; mulps         %xmm1,%xmm8
  DB  102,65,15,91,200                    ; cvtps2dq      %xmm8,%xmm1
  DB  102,72,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,67,15,16,12,145                 ; movss         (%r9,%r10,4),%xmm1
  DB  102,65,15,58,33,12,137,16           ; insertps      $0x10,(%r9,%rcx,4),%xmm1
  DB  243,71,15,16,4,129                  ; movss         (%r9,%r8,4),%xmm8
  DB  102,65,15,58,33,200,32              ; insertps      $0x20,%xmm8,%xmm1
  DB  243,69,15,16,4,129                  ; movss         (%r9,%rax,4),%xmm8
  DB  102,65,15,58,33,200,48              ; insertps      $0x30,%xmm8,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_b_sse41
_sk_table_b_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  102,65,15,91,208                    ; cvtps2dq      %xmm8,%xmm2
  DB  102,72,15,58,22,208,1               ; pextrq        $0x1,%xmm2,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,209                   ; movq          %xmm2,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,67,15,16,20,145                 ; movss         (%r9,%r10,4),%xmm2
  DB  102,65,15,58,33,20,137,16           ; insertps      $0x10,(%r9,%rcx,4),%xmm2
  DB  243,71,15,16,4,129                  ; movss         (%r9,%r8,4),%xmm8
  DB  102,65,15,58,33,208,32              ; insertps      $0x20,%xmm8,%xmm2
  DB  243,69,15,16,4,129                  ; movss         (%r9,%rax,4),%xmm8
  DB  102,65,15,58,33,208,48              ; insertps      $0x30,%xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_a_sse41
_sk_table_a_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,65,15,91,216                    ; cvtps2dq      %xmm8,%xmm3
  DB  102,72,15,58,22,216,1               ; pextrq        $0x1,%xmm3,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,217                   ; movq          %xmm3,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,67,15,16,28,145                 ; movss         (%r9,%r10,4),%xmm3
  DB  102,65,15,58,33,28,137,16           ; insertps      $0x10,(%r9,%rcx,4),%xmm3
  DB  243,71,15,16,4,129                  ; movss         (%r9,%r8,4),%xmm8
  DB  102,65,15,58,33,216,32              ; insertps      $0x20,%xmm8,%xmm3
  DB  243,69,15,16,4,129                  ; movss         (%r9,%rax,4),%xmm8
  DB  102,65,15,58,33,216,48              ; insertps      $0x30,%xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_r_sse41
_sk_parametric_r_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,64,16                  ; movss         0x10(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,216                        ; mulps         %xmm0,%xmm11
  DB  65,15,194,192,2                     ; cmpleps       %xmm8,%xmm0
  DB  243,68,15,16,64,24                  ; movss         0x18(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,88,200                        ; addps         %xmm8,%xmm9
  DB  243,68,15,16,16                     ; movss         (%rax),%xmm10
  DB  243,68,15,16,64,8                   ; movss         0x8(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,88,216                        ; addps         %xmm8,%xmm11
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  102,69,15,112,228,0                 ; pshufd        $0x0,%xmm12,%xmm12
  DB  102,69,15,219,227                   ; pand          %xmm11,%xmm12
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  102,69,15,112,219,0                 ; pshufd        $0x0,%xmm11,%xmm11
  DB  102,69,15,235,220                   ; por           %xmm12,%xmm11
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,92,196                        ; subps         %xmm12,%xmm8
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,89,227                        ; mulps         %xmm11,%xmm12
  DB  69,15,92,196                        ; subps         %xmm12,%xmm8
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,88,235                        ; addps         %xmm11,%xmm13
  DB  69,15,94,229                        ; divps         %xmm13,%xmm12
  DB  69,15,92,196                        ; subps         %xmm12,%xmm8
  DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
  DB  102,69,15,58,8,216,1                ; roundps       $0x1,%xmm8,%xmm11
  DB  185,0,0,0,75                        ; mov           $0x4b000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,224                        ; addps         %xmm8,%xmm12
  DB  69,15,92,195                        ; subps         %xmm11,%xmm8
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
  DB  69,15,92,227                        ; subps         %xmm11,%xmm12
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,92,232                        ; subps         %xmm8,%xmm13
  DB  69,15,94,221                        ; divps         %xmm13,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
  DB  102,69,15,91,211                    ; cvtps2dq      %xmm11,%xmm10
  DB  243,68,15,16,64,20                  ; movss         0x14(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,88,194                        ; addps         %xmm10,%xmm8
  DB  102,69,15,56,20,193                 ; blendvps      %xmm0,%xmm9,%xmm8
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  68,15,93,192                        ; minps         %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_g_sse41
_sk_parametric_g_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
  DB  15,40,193                           ; movaps        %xmm1,%xmm0
  DB  65,15,194,194,2                     ; cmpleps       %xmm10,%xmm0
  DB  243,15,16,72,24                     ; movss         0x18(%rax),%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  68,15,88,201                        ; addps         %xmm1,%xmm9
  DB  243,68,15,16,16                     ; movss         (%rax),%xmm10
  DB  243,15,16,72,8                      ; movss         0x8(%rax),%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  68,15,88,217                        ; addps         %xmm1,%xmm11
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  102,68,15,112,225,0                 ; pshufd        $0x0,%xmm1,%xmm12
  DB  102,69,15,219,227                   ; pand          %xmm11,%xmm12
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  102,68,15,112,217,0                 ; pshufd        $0x0,%xmm1,%xmm11
  DB  102,69,15,235,220                   ; por           %xmm12,%xmm11
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  68,15,92,233                        ; subps         %xmm1,%xmm13
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
  DB  68,15,92,233                        ; subps         %xmm1,%xmm13
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  65,15,88,203                        ; addps         %xmm11,%xmm1
  DB  68,15,94,225                        ; divps         %xmm1,%xmm12
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
  DB  102,69,15,58,8,221,1                ; roundps       $0x1,%xmm13,%xmm11
  DB  185,0,0,0,75                        ; mov           $0x4b000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,92,235                        ; subps         %xmm11,%xmm13
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  65,15,89,205                        ; mulps         %xmm13,%xmm1
  DB  68,15,92,225                        ; subps         %xmm1,%xmm12
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,15,110,201                      ; movd          %ecx,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  65,15,92,205                        ; subps         %xmm13,%xmm1
  DB  68,15,94,217                        ; divps         %xmm1,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
  DB  102,69,15,91,211                    ; cvtps2dq      %xmm11,%xmm10
  DB  243,15,16,72,20                     ; movss         0x14(%rax),%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  102,65,15,56,20,201                 ; blendvps      %xmm0,%xmm9,%xmm1
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,95,200                           ; maxps         %xmm0,%xmm1
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,93,200                           ; minps         %xmm0,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_b_sse41
_sk_parametric_b_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  65,15,194,194,2                     ; cmpleps       %xmm10,%xmm0
  DB  243,15,16,80,24                     ; movss         0x18(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  68,15,88,202                        ; addps         %xmm2,%xmm9
  DB  243,68,15,16,16                     ; movss         (%rax),%xmm10
  DB  243,15,16,80,8                      ; movss         0x8(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  68,15,88,218                        ; addps         %xmm2,%xmm11
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  102,68,15,112,226,0                 ; pshufd        $0x0,%xmm2,%xmm12
  DB  102,69,15,219,227                   ; pand          %xmm11,%xmm12
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  102,68,15,112,218,0                 ; pshufd        $0x0,%xmm2,%xmm11
  DB  102,69,15,235,220                   ; por           %xmm12,%xmm11
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  68,15,92,234                        ; subps         %xmm2,%xmm13
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,211                        ; mulps         %xmm11,%xmm2
  DB  68,15,92,234                        ; subps         %xmm2,%xmm13
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,88,211                        ; addps         %xmm11,%xmm2
  DB  68,15,94,226                        ; divps         %xmm2,%xmm12
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
  DB  102,69,15,58,8,221,1                ; roundps       $0x1,%xmm13,%xmm11
  DB  185,0,0,0,75                        ; mov           $0x4b000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,92,235                        ; subps         %xmm11,%xmm13
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,213                        ; mulps         %xmm13,%xmm2
  DB  68,15,92,226                        ; subps         %xmm2,%xmm12
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,92,213                        ; subps         %xmm13,%xmm2
  DB  68,15,94,218                        ; divps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
  DB  102,69,15,91,211                    ; cvtps2dq      %xmm11,%xmm10
  DB  243,15,16,80,20                     ; movss         0x14(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,88,210                        ; addps         %xmm10,%xmm2
  DB  102,65,15,56,20,209                 ; blendvps      %xmm0,%xmm9,%xmm2
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,95,208                           ; maxps         %xmm0,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,93,208                           ; minps         %xmm0,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_a_sse41
_sk_parametric_a_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,219                        ; mulps         %xmm3,%xmm11
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  65,15,194,194,2                     ; cmpleps       %xmm10,%xmm0
  DB  243,15,16,88,24                     ; movss         0x18(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  68,15,88,203                        ; addps         %xmm3,%xmm9
  DB  243,68,15,16,16                     ; movss         (%rax),%xmm10
  DB  243,15,16,88,8                      ; movss         0x8(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  68,15,88,219                        ; addps         %xmm3,%xmm11
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  102,68,15,112,227,0                 ; pshufd        $0x0,%xmm3,%xmm12
  DB  102,69,15,219,227                   ; pand          %xmm11,%xmm12
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  102,68,15,112,219,0                 ; pshufd        $0x0,%xmm3,%xmm11
  DB  102,69,15,235,220                   ; por           %xmm12,%xmm11
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  68,15,92,235                        ; subps         %xmm3,%xmm13
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
  DB  68,15,92,235                        ; subps         %xmm3,%xmm13
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,88,219                        ; addps         %xmm11,%xmm3
  DB  68,15,94,227                        ; divps         %xmm3,%xmm12
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
  DB  102,69,15,58,8,221,1                ; roundps       $0x1,%xmm13,%xmm11
  DB  185,0,0,0,75                        ; mov           $0x4b000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,92,235                        ; subps         %xmm11,%xmm13
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,221                        ; mulps         %xmm13,%xmm3
  DB  68,15,92,227                        ; subps         %xmm3,%xmm12
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,92,221                        ; subps         %xmm13,%xmm3
  DB  68,15,94,219                        ; divps         %xmm3,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
  DB  102,69,15,91,211                    ; cvtps2dq      %xmm11,%xmm10
  DB  243,15,16,88,20                     ; movss         0x14(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,88,218                        ; addps         %xmm10,%xmm3
  DB  102,65,15,56,20,217                 ; blendvps      %xmm0,%xmm9,%xmm3
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,95,216                           ; maxps         %xmm0,%xmm3
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,93,216                           ; minps         %xmm0,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_a8_sse41
_sk_load_a8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,15,56,49,4,56                   ; pmovzxbd      (%rax,%rdi,1),%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_a8_sse41
_sk_gather_a8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,56,64,209                    ; pmulld        %xmm1,%xmm2
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
  DB  102,72,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,58,32,4,17,0              ; pinsrb        $0x0,(%r9,%r10,1),%xmm0
  DB  102,65,15,58,32,4,9,1               ; pinsrb        $0x1,(%r9,%rcx,1),%xmm0
  DB  67,15,182,12,1                      ; movzbl        (%r9,%r8,1),%ecx
  DB  102,15,58,32,193,2                  ; pinsrb        $0x2,%ecx,%xmm0
  DB  65,15,182,4,1                       ; movzbl        (%r9,%rax,1),%eax
  DB  102,15,58,32,192,3                  ; pinsrb        $0x3,%eax,%xmm0
  DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_a8_sse41
_sk_store_a8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,127,67                      ; mov           $0x437f0000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,69,15,56,43,192                 ; packusdw      %xmm8,%xmm8
  DB  102,69,15,103,192                   ; packuswb      %xmm8,%xmm8
  DB  102,68,15,126,4,56                  ; movd          %xmm8,(%rax,%rdi,1)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_g8_sse41
_sk_load_g8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,15,56,49,4,56                   ; pmovzxbd      (%rax,%rdi,1),%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,200                           ; movaps        %xmm0,%xmm1
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_g8_sse41
_sk_gather_g8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,56,64,209                    ; pmulld        %xmm1,%xmm2
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
  DB  102,72,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,58,32,4,17,0              ; pinsrb        $0x0,(%r9,%r10,1),%xmm0
  DB  102,65,15,58,32,4,9,1               ; pinsrb        $0x1,(%r9,%rcx,1),%xmm0
  DB  67,15,182,12,1                      ; movzbl        (%r9,%r8,1),%ecx
  DB  102,15,58,32,193,2                  ; pinsrb        $0x2,%ecx,%xmm0
  DB  65,15,182,4,1                       ; movzbl        (%r9,%rax,1),%eax
  DB  102,15,58,32,192,3                  ; pinsrb        $0x3,%eax,%xmm0
  DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,200                           ; movaps        %xmm0,%xmm1
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_i8_sse41
_sk_gather_i8_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,192                          ; mov           %rax,%r8
  DB  77,133,192                          ; test          %r8,%r8
  DB  116,5                               ; je            2757 <_sk_gather_i8_sse41+0xf>
  DB  76,137,192                          ; mov           %r8,%rax
  DB  235,2                               ; jmp           2759 <_sk_gather_i8_sse41+0x11>
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,56,64,209                    ; pmulld        %xmm1,%xmm2
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
  DB  102,72,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,203                          ; mov           %ecx,%r11d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,58,32,4,26,0              ; pinsrb        $0x0,(%r10,%r11,1),%xmm0
  DB  102,65,15,58,32,4,10,1              ; pinsrb        $0x1,(%r10,%rcx,1),%xmm0
  DB  102,67,15,58,32,4,10,2              ; pinsrb        $0x2,(%r10,%r9,1),%xmm0
  DB  102,65,15,58,32,4,2,3               ; pinsrb        $0x3,(%r10,%rax,1),%xmm0
  DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
  DB  102,73,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%r9
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  73,139,64,8                         ; mov           0x8(%r8),%rax
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  69,137,202                          ; mov           %r9d,%r10d
  DB  73,193,233,30                       ; shr           $0x1e,%r9
  DB  102,66,15,110,28,128                ; movd          (%rax,%r8,4),%xmm3
  DB  102,15,58,34,28,8,1                 ; pinsrd        $0x1,(%rax,%rcx,1),%xmm3
  DB  102,66,15,58,34,28,144,2            ; pinsrd        $0x2,(%rax,%r10,4),%xmm3
  DB  102,66,15,58,34,28,8,3              ; pinsrd        $0x3,(%rax,%r9,1),%xmm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
  DB  102,15,219,200                      ; pand          %xmm0,%xmm1
  DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
  DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_565_sse41
_sk_load_565_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,15,56,51,20,120                 ; pmovzxwd      (%rax,%rdi,2),%xmm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,219,194                      ; pand          %xmm2,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,15,219,202                      ; pand          %xmm2,%xmm1
  DB  15,91,217                           ; cvtdq2ps      %xmm1,%xmm3
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,203                           ; mulps         %xmm3,%xmm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,15,219,218                      ; pand          %xmm2,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_565_sse41
_sk_gather_565_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,56,64,209                    ; pmulld        %xmm1,%xmm2
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
  DB  102,72,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,196,4,81,0                ; pinsrw        $0x0,(%r9,%r10,2),%xmm0
  DB  102,65,15,196,4,73,1                ; pinsrw        $0x1,(%r9,%rcx,2),%xmm0
  DB  67,15,183,12,65                     ; movzwl        (%r9,%r8,2),%ecx
  DB  102,15,196,193,2                    ; pinsrw        $0x2,%ecx,%xmm0
  DB  65,15,183,4,65                      ; movzwl        (%r9,%rax,2),%eax
  DB  102,15,196,192,3                    ; pinsrw        $0x3,%eax,%xmm0
  DB  102,15,56,51,208                    ; pmovzxwd      %xmm0,%xmm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,219,194                      ; pand          %xmm2,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,15,219,202                      ; pand          %xmm2,%xmm1
  DB  15,91,217                           ; cvtdq2ps      %xmm1,%xmm3
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,203                           ; mulps         %xmm3,%xmm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,15,219,218                      ; pand          %xmm2,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_565_sse41
_sk_store_565_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,248,65                      ; mov           $0x41f80000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,11                ; pslld         $0xb,%xmm9
  DB  185,0,0,124,66                      ; mov           $0x427c0000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,65,15,114,242,5                 ; pslld         $0x5,%xmm10
  DB  102,69,15,235,209                   ; por           %xmm9,%xmm10
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,69,15,86,194                    ; orpd          %xmm10,%xmm8
  DB  102,69,15,56,43,192                 ; packusdw      %xmm8,%xmm8
  DB  102,68,15,214,4,120                 ; movq          %xmm8,(%rax,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_4444_sse41
_sk_load_4444_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,68,15,56,51,12,120              ; pmovzxwd      (%rax,%rdi,2),%xmm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,65,15,219,201                   ; pand          %xmm9,%xmm1
  DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,202                           ; mulps         %xmm2,%xmm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
  DB  68,15,91,194                        ; cvtdq2ps      %xmm2,%xmm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,217                   ; pand          %xmm9,%xmm3
  DB  68,15,91,195                        ; cvtdq2ps      %xmm3,%xmm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_4444_sse41
_sk_gather_4444_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,56,64,209                    ; pmulld        %xmm1,%xmm2
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
  DB  102,72,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,196,4,81,0                ; pinsrw        $0x0,(%r9,%r10,2),%xmm0
  DB  102,65,15,196,4,73,1                ; pinsrw        $0x1,(%r9,%rcx,2),%xmm0
  DB  67,15,183,12,65                     ; movzwl        (%r9,%r8,2),%ecx
  DB  102,15,196,193,2                    ; pinsrw        $0x2,%ecx,%xmm0
  DB  65,15,183,4,65                      ; movzwl        (%r9,%rax,2),%eax
  DB  102,15,196,192,3                    ; pinsrw        $0x3,%eax,%xmm0
  DB  102,68,15,56,51,200                 ; pmovzxwd      %xmm0,%xmm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,65,15,219,201                   ; pand          %xmm9,%xmm1
  DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,202                           ; mulps         %xmm2,%xmm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
  DB  68,15,91,194                        ; cvtdq2ps      %xmm2,%xmm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,217                   ; pand          %xmm9,%xmm3
  DB  68,15,91,195                        ; cvtdq2ps      %xmm3,%xmm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_4444_sse41
_sk_store_4444_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,112,65                      ; mov           $0x41700000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,12                ; pslld         $0xc,%xmm9
  DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,65,15,114,242,8                 ; pslld         $0x8,%xmm10
  DB  102,69,15,235,209                   ; por           %xmm9,%xmm10
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,4                 ; pslld         $0x4,%xmm9
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,69,15,86,193                    ; orpd          %xmm9,%xmm8
  DB  102,69,15,86,194                    ; orpd          %xmm10,%xmm8
  DB  102,69,15,56,43,192                 ; packusdw      %xmm8,%xmm8
  DB  102,68,15,214,4,120                 ; movq          %xmm8,(%rax,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_8888_sse41
_sk_load_8888_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,15,111,28,184                   ; movdqu        (%rax,%rdi,4),%xmm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
  DB  102,15,219,200                      ; pand          %xmm0,%xmm1
  DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
  DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_8888_sse41
_sk_gather_8888_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,56,64,209                    ; pmulld        %xmm1,%xmm2
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
  DB  102,72,15,126,192                   ; movq          %xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,110,28,129                ; movd          (%r9,%r8,4),%xmm3
  DB  102,65,15,58,34,28,129,1            ; pinsrd        $0x1,(%r9,%rax,4),%xmm3
  DB  102,67,15,58,34,28,145,2            ; pinsrd        $0x2,(%r9,%r10,4),%xmm3
  DB  102,65,15,58,34,28,137,3            ; pinsrd        $0x3,(%r9,%rcx,4),%xmm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
  DB  102,15,219,200                      ; pand          %xmm0,%xmm1
  DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
  DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_8888_sse41
_sk_store_8888_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,127,67                      ; mov           $0x437f0000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,65,15,114,242,8                 ; pslld         $0x8,%xmm10
  DB  102,69,15,235,209                   ; por           %xmm9,%xmm10
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,16                ; pslld         $0x10,%xmm9
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,114,240,24                ; pslld         $0x18,%xmm8
  DB  102,69,15,235,193                   ; por           %xmm9,%xmm8
  DB  102,69,15,235,194                   ; por           %xmm10,%xmm8
  DB  243,68,15,127,4,184                 ; movdqu        %xmm8,(%rax,%rdi,4)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_f16_sse41
_sk_load_f16_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,15,111,4,248                    ; movdqu        (%rax,%rdi,8),%xmm0
  DB  243,15,111,76,248,16                ; movdqu        0x10(%rax,%rdi,8),%xmm1
  DB  102,68,15,111,200                   ; movdqa        %xmm0,%xmm9
  DB  102,68,15,97,201                    ; punpcklwd     %xmm1,%xmm9
  DB  102,15,105,193                      ; punpckhwd     %xmm1,%xmm0
  DB  102,69,15,111,225                   ; movdqa        %xmm9,%xmm12
  DB  102,68,15,97,224                    ; punpcklwd     %xmm0,%xmm12
  DB  102,68,15,105,200                   ; punpckhwd     %xmm0,%xmm9
  DB  102,69,15,56,51,236                 ; pmovzxwd      %xmm12,%xmm13
  DB  184,0,128,0,0                       ; mov           $0x8000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,68,15,112,192,0                 ; pshufd        $0x0,%xmm0,%xmm8
  DB  102,65,15,111,213                   ; movdqa        %xmm13,%xmm2
  DB  102,65,15,219,208                   ; pand          %xmm8,%xmm2
  DB  184,0,124,0,0                       ; mov           $0x7c00,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,216,0                    ; pshufd        $0x0,%xmm0,%xmm3
  DB  102,65,15,111,197                   ; movdqa        %xmm13,%xmm0
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  102,68,15,239,234                   ; pxor          %xmm2,%xmm13
  DB  102,69,15,239,210                   ; pxor          %xmm10,%xmm10
  DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
  DB  102,65,15,114,245,13                ; pslld         $0xd,%xmm13
  DB  184,0,0,0,56                        ; mov           $0x38000000,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,68,15,112,217,0                 ; pshufd        $0x0,%xmm1,%xmm11
  DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
  DB  102,65,15,254,213                   ; paddd         %xmm13,%xmm2
  DB  102,65,15,118,194                   ; pcmpeqd       %xmm10,%xmm0
  DB  102,15,223,194                      ; pandn         %xmm2,%xmm0
  DB  102,65,15,115,220,8                 ; psrldq        $0x8,%xmm12
  DB  102,69,15,56,51,228                 ; pmovzxwd      %xmm12,%xmm12
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,65,15,219,208                   ; pand          %xmm8,%xmm2
  DB  102,65,15,111,204                   ; movdqa        %xmm12,%xmm1
  DB  102,15,219,203                      ; pand          %xmm3,%xmm1
  DB  102,68,15,239,226                   ; pxor          %xmm2,%xmm12
  DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
  DB  102,65,15,254,212                   ; paddd         %xmm12,%xmm2
  DB  102,65,15,118,202                   ; pcmpeqd       %xmm10,%xmm1
  DB  102,15,223,202                      ; pandn         %xmm2,%xmm1
  DB  102,69,15,56,51,225                 ; pmovzxwd      %xmm9,%xmm12
  DB  102,69,15,111,236                   ; movdqa        %xmm12,%xmm13
  DB  102,69,15,219,232                   ; pand          %xmm8,%xmm13
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,15,219,211                      ; pand          %xmm3,%xmm2
  DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
  DB  102,65,15,114,245,16                ; pslld         $0x10,%xmm13
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
  DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
  DB  102,65,15,118,210                   ; pcmpeqd       %xmm10,%xmm2
  DB  102,65,15,223,213                   ; pandn         %xmm13,%xmm2
  DB  102,65,15,115,217,8                 ; psrldq        $0x8,%xmm9
  DB  102,69,15,56,51,201                 ; pmovzxwd      %xmm9,%xmm9
  DB  102,69,15,219,193                   ; pand          %xmm9,%xmm8
  DB  102,65,15,219,217                   ; pand          %xmm9,%xmm3
  DB  102,69,15,239,200                   ; pxor          %xmm8,%xmm9
  DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
  DB  102,65,15,114,241,13                ; pslld         $0xd,%xmm9
  DB  102,69,15,254,195                   ; paddd         %xmm11,%xmm8
  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
  DB  102,65,15,118,218                   ; pcmpeqd       %xmm10,%xmm3
  DB  102,65,15,223,216                   ; pandn         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_f16_sse41
_sk_gather_f16_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,56,64,209                    ; pmulld        %xmm1,%xmm2
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
  DB  102,72,15,126,192                   ; movq          %xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,65,15,126,4,201                 ; movq          (%r9,%rcx,8),%xmm0
  DB  243,67,15,126,12,209                ; movq          (%r9,%r10,8),%xmm1
  DB  102,15,108,200                      ; punpcklqdq    %xmm0,%xmm1
  DB  243,65,15,126,4,193                 ; movq          (%r9,%rax,8),%xmm0
  DB  243,67,15,126,20,193                ; movq          (%r9,%r8,8),%xmm2
  DB  102,15,108,208                      ; punpcklqdq    %xmm0,%xmm2
  DB  102,68,15,111,202                   ; movdqa        %xmm2,%xmm9
  DB  102,68,15,97,201                    ; punpcklwd     %xmm1,%xmm9
  DB  102,15,105,209                      ; punpckhwd     %xmm1,%xmm2
  DB  102,69,15,111,225                   ; movdqa        %xmm9,%xmm12
  DB  102,68,15,97,226                    ; punpcklwd     %xmm2,%xmm12
  DB  102,68,15,105,202                   ; punpckhwd     %xmm2,%xmm9
  DB  102,69,15,56,51,236                 ; pmovzxwd      %xmm12,%xmm13
  DB  184,0,128,0,0                       ; mov           $0x8000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,68,15,112,192,0                 ; pshufd        $0x0,%xmm0,%xmm8
  DB  102,65,15,111,213                   ; movdqa        %xmm13,%xmm2
  DB  102,65,15,219,208                   ; pand          %xmm8,%xmm2
  DB  184,0,124,0,0                       ; mov           $0x7c00,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,216,0                    ; pshufd        $0x0,%xmm0,%xmm3
  DB  102,65,15,111,197                   ; movdqa        %xmm13,%xmm0
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  102,68,15,239,234                   ; pxor          %xmm2,%xmm13
  DB  102,69,15,239,210                   ; pxor          %xmm10,%xmm10
  DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
  DB  102,65,15,114,245,13                ; pslld         $0xd,%xmm13
  DB  184,0,0,0,56                        ; mov           $0x38000000,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,68,15,112,217,0                 ; pshufd        $0x0,%xmm1,%xmm11
  DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
  DB  102,65,15,254,213                   ; paddd         %xmm13,%xmm2
  DB  102,65,15,118,194                   ; pcmpeqd       %xmm10,%xmm0
  DB  102,15,223,194                      ; pandn         %xmm2,%xmm0
  DB  102,65,15,115,220,8                 ; psrldq        $0x8,%xmm12
  DB  102,69,15,56,51,228                 ; pmovzxwd      %xmm12,%xmm12
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,65,15,219,208                   ; pand          %xmm8,%xmm2
  DB  102,65,15,111,204                   ; movdqa        %xmm12,%xmm1
  DB  102,15,219,203                      ; pand          %xmm3,%xmm1
  DB  102,68,15,239,226                   ; pxor          %xmm2,%xmm12
  DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
  DB  102,65,15,254,212                   ; paddd         %xmm12,%xmm2
  DB  102,65,15,118,202                   ; pcmpeqd       %xmm10,%xmm1
  DB  102,15,223,202                      ; pandn         %xmm2,%xmm1
  DB  102,69,15,56,51,225                 ; pmovzxwd      %xmm9,%xmm12
  DB  102,69,15,111,236                   ; movdqa        %xmm12,%xmm13
  DB  102,69,15,219,232                   ; pand          %xmm8,%xmm13
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,15,219,211                      ; pand          %xmm3,%xmm2
  DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
  DB  102,65,15,114,245,16                ; pslld         $0x10,%xmm13
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
  DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
  DB  102,65,15,118,210                   ; pcmpeqd       %xmm10,%xmm2
  DB  102,65,15,223,213                   ; pandn         %xmm13,%xmm2
  DB  102,65,15,115,217,8                 ; psrldq        $0x8,%xmm9
  DB  102,69,15,56,51,201                 ; pmovzxwd      %xmm9,%xmm9
  DB  102,69,15,219,193                   ; pand          %xmm9,%xmm8
  DB  102,65,15,219,217                   ; pand          %xmm9,%xmm3
  DB  102,69,15,239,200                   ; pxor          %xmm8,%xmm9
  DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
  DB  102,65,15,114,241,13                ; pslld         $0xd,%xmm9
  DB  102,69,15,254,195                   ; paddd         %xmm11,%xmm8
  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
  DB  102,65,15,118,218                   ; pcmpeqd       %xmm10,%xmm3
  DB  102,65,15,223,216                   ; pandn         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_f16_sse41
_sk_store_f16_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,0,128                       ; mov           $0x80000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  102,69,15,112,200,0                 ; pshufd        $0x0,%xmm8,%xmm9
  DB  102,69,15,111,225                   ; movdqa        %xmm9,%xmm12
  DB  102,68,15,219,224                   ; pand          %xmm0,%xmm12
  DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
  DB  102,69,15,239,196                   ; pxor          %xmm12,%xmm8
  DB  185,0,0,128,56                      ; mov           $0x38800000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  102,65,15,114,212,16                ; psrld         $0x10,%xmm12
  DB  102,69,15,111,232                   ; movdqa        %xmm8,%xmm13
  DB  102,65,15,114,213,13                ; psrld         $0xd,%xmm13
  DB  185,0,192,1,0                       ; mov           $0x1c000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  102,69,15,112,219,0                 ; pshufd        $0x0,%xmm11,%xmm11
  DB  102,69,15,250,227                   ; psubd         %xmm11,%xmm12
  DB  102,69,15,254,229                   ; paddd         %xmm13,%xmm12
  DB  69,15,194,194,5                     ; cmpnltps      %xmm10,%xmm8
  DB  69,15,84,196                        ; andps         %xmm12,%xmm8
  DB  102,69,15,56,43,192                 ; packusdw      %xmm8,%xmm8
  DB  102,69,15,111,233                   ; movdqa        %xmm9,%xmm13
  DB  102,68,15,219,233                   ; pand          %xmm1,%xmm13
  DB  102,68,15,111,225                   ; movdqa        %xmm1,%xmm12
  DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
  DB  102,65,15,114,213,16                ; psrld         $0x10,%xmm13
  DB  102,69,15,111,244                   ; movdqa        %xmm12,%xmm14
  DB  102,65,15,114,214,13                ; psrld         $0xd,%xmm14
  DB  102,69,15,250,235                   ; psubd         %xmm11,%xmm13
  DB  102,69,15,254,238                   ; paddd         %xmm14,%xmm13
  DB  69,15,194,226,5                     ; cmpnltps      %xmm10,%xmm12
  DB  69,15,84,229                        ; andps         %xmm13,%xmm12
  DB  102,69,15,56,43,228                 ; packusdw      %xmm12,%xmm12
  DB  102,69,15,111,241                   ; movdqa        %xmm9,%xmm14
  DB  102,68,15,219,242                   ; pand          %xmm2,%xmm14
  DB  102,68,15,111,234                   ; movdqa        %xmm2,%xmm13
  DB  102,69,15,239,238                   ; pxor          %xmm14,%xmm13
  DB  102,65,15,114,214,16                ; psrld         $0x10,%xmm14
  DB  102,69,15,111,253                   ; movdqa        %xmm13,%xmm15
  DB  102,65,15,114,215,13                ; psrld         $0xd,%xmm15
  DB  102,69,15,250,243                   ; psubd         %xmm11,%xmm14
  DB  102,69,15,254,247                   ; paddd         %xmm15,%xmm14
  DB  69,15,194,234,5                     ; cmpnltps      %xmm10,%xmm13
  DB  69,15,84,238                        ; andps         %xmm14,%xmm13
  DB  102,69,15,56,43,237                 ; packusdw      %xmm13,%xmm13
  DB  102,68,15,219,203                   ; pand          %xmm3,%xmm9
  DB  102,68,15,111,243                   ; movdqa        %xmm3,%xmm14
  DB  102,69,15,239,241                   ; pxor          %xmm9,%xmm14
  DB  102,65,15,114,209,16                ; psrld         $0x10,%xmm9
  DB  102,69,15,111,254                   ; movdqa        %xmm14,%xmm15
  DB  102,65,15,114,215,13                ; psrld         $0xd,%xmm15
  DB  102,69,15,250,203                   ; psubd         %xmm11,%xmm9
  DB  102,69,15,254,207                   ; paddd         %xmm15,%xmm9
  DB  69,15,194,242,5                     ; cmpnltps      %xmm10,%xmm14
  DB  69,15,84,241                        ; andps         %xmm9,%xmm14
  DB  102,69,15,56,43,246                 ; packusdw      %xmm14,%xmm14
  DB  102,69,15,97,196                    ; punpcklwd     %xmm12,%xmm8
  DB  102,69,15,97,238                    ; punpcklwd     %xmm14,%xmm13
  DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
  DB  102,69,15,98,205                    ; punpckldq     %xmm13,%xmm9
  DB  243,68,15,127,12,248                ; movdqu        %xmm9,(%rax,%rdi,8)
  DB  102,69,15,106,197                   ; punpckhdq     %xmm13,%xmm8
  DB  243,68,15,127,68,248,16             ; movdqu        %xmm8,0x10(%rax,%rdi,8)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_u16_be_sse41
_sk_load_u16_be_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,15,111,4,248                    ; movdqu        (%rax,%rdi,8),%xmm0
  DB  243,15,111,76,248,16                ; movdqu        0x10(%rax,%rdi,8),%xmm1
  DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
  DB  102,15,97,209                       ; punpcklwd     %xmm1,%xmm2
  DB  102,15,105,193                      ; punpckhwd     %xmm1,%xmm0
  DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
  DB  102,15,105,208                      ; punpckhwd     %xmm0,%xmm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
  DB  102,15,113,240,8                    ; psllw         $0x8,%xmm0
  DB  102,15,112,217,78                   ; pshufd        $0x4e,%xmm1,%xmm3
  DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
  DB  102,15,235,200                      ; por           %xmm0,%xmm1
  DB  102,15,56,51,193                    ; pmovzxwd      %xmm1,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
  DB  102,15,235,217                      ; por           %xmm1,%xmm3
  DB  102,15,56,51,203                    ; pmovzxwd      %xmm3,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,68,15,111,202                   ; movdqa        %xmm2,%xmm9
  DB  102,65,15,113,241,8                 ; psllw         $0x8,%xmm9
  DB  102,15,112,218,78                   ; pshufd        $0x4e,%xmm2,%xmm3
  DB  102,15,113,210,8                    ; psrlw         $0x8,%xmm2
  DB  102,65,15,235,209                   ; por           %xmm9,%xmm2
  DB  102,15,56,51,210                    ; pmovzxwd      %xmm2,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,68,15,111,203                   ; movdqa        %xmm3,%xmm9
  DB  102,65,15,113,241,8                 ; psllw         $0x8,%xmm9
  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
  DB  102,65,15,235,217                   ; por           %xmm9,%xmm3
  DB  102,15,56,51,219                    ; pmovzxwd      %xmm3,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_rgb_u16_be_sse41
_sk_load_rgb_u16_be_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,141,12,127                       ; lea           (%rdi,%rdi,2),%rcx
  DB  243,15,111,20,72                    ; movdqu        (%rax,%rcx,2),%xmm2
  DB  243,15,111,68,72,8                  ; movdqu        0x8(%rax,%rcx,2),%xmm0
  DB  102,15,115,216,4                    ; psrldq        $0x4,%xmm0
  DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
  DB  102,15,115,217,6                    ; psrldq        $0x6,%xmm1
  DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
  DB  102,15,115,216,6                    ; psrldq        $0x6,%xmm0
  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
  DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  102,15,112,216,78                   ; pshufd        $0x4e,%xmm0,%xmm3
  DB  102,15,105,209                      ; punpckhwd     %xmm1,%xmm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
  DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
  DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
  DB  102,15,235,193                      ; por           %xmm1,%xmm0
  DB  102,15,56,51,192                    ; pmovzxwd      %xmm0,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
  DB  102,15,235,217                      ; por           %xmm1,%xmm3
  DB  102,15,56,51,203                    ; pmovzxwd      %xmm3,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
  DB  102,15,113,243,8                    ; psllw         $0x8,%xmm3
  DB  102,15,113,210,8                    ; psrlw         $0x8,%xmm2
  DB  102,15,235,211                      ; por           %xmm3,%xmm2
  DB  102,15,56,51,210                    ; pmovzxwd      %xmm2,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_u16_be_sse41
_sk_store_u16_be_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,255,127,71                    ; mov           $0x477fff00,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,69,15,56,43,192                 ; packusdw      %xmm8,%xmm8
  DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
  DB  102,65,15,113,242,8                 ; psllw         $0x8,%xmm10
  DB  102,65,15,113,208,8                 ; psrlw         $0x8,%xmm8
  DB  102,69,15,235,194                   ; por           %xmm10,%xmm8
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,69,15,56,43,210                 ; packusdw      %xmm10,%xmm10
  DB  102,69,15,111,218                   ; movdqa        %xmm10,%xmm11
  DB  102,65,15,113,243,8                 ; psllw         $0x8,%xmm11
  DB  102,65,15,113,210,8                 ; psrlw         $0x8,%xmm10
  DB  102,69,15,235,211                   ; por           %xmm11,%xmm10
  DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  102,69,15,91,219                    ; cvtps2dq      %xmm11,%xmm11
  DB  102,69,15,56,43,219                 ; packusdw      %xmm11,%xmm11
  DB  102,69,15,111,227                   ; movdqa        %xmm11,%xmm12
  DB  102,65,15,113,244,8                 ; psllw         $0x8,%xmm12
  DB  102,65,15,113,211,8                 ; psrlw         $0x8,%xmm11
  DB  102,69,15,235,220                   ; por           %xmm12,%xmm11
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,69,15,56,43,201                 ; packusdw      %xmm9,%xmm9
  DB  102,69,15,111,225                   ; movdqa        %xmm9,%xmm12
  DB  102,65,15,113,244,8                 ; psllw         $0x8,%xmm12
  DB  102,65,15,113,209,8                 ; psrlw         $0x8,%xmm9
  DB  102,69,15,235,204                   ; por           %xmm12,%xmm9
  DB  102,69,15,97,194                    ; punpcklwd     %xmm10,%xmm8
  DB  102,69,15,97,217                    ; punpcklwd     %xmm9,%xmm11
  DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
  DB  102,69,15,98,203                    ; punpckldq     %xmm11,%xmm9
  DB  243,68,15,127,12,248                ; movdqu        %xmm9,(%rax,%rdi,8)
  DB  102,69,15,106,195                   ; punpckhdq     %xmm11,%xmm8
  DB  243,68,15,127,68,248,16             ; movdqu        %xmm8,0x10(%rax,%rdi,8)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_f32_sse41
_sk_load_f32_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,137,249                          ; mov           %rdi,%rcx
  DB  72,193,225,4                        ; shl           $0x4,%rcx
  DB  68,15,16,4,8                        ; movups        (%rax,%rcx,1),%xmm8
  DB  15,16,68,8,16                       ; movups        0x10(%rax,%rcx,1),%xmm0
  DB  15,16,92,8,32                       ; movups        0x20(%rax,%rcx,1),%xmm3
  DB  68,15,16,76,8,48                    ; movups        0x30(%rax,%rcx,1),%xmm9
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  15,20,208                           ; unpcklps      %xmm0,%xmm2
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  65,15,20,201                        ; unpcklps      %xmm9,%xmm1
  DB  68,15,21,192                        ; unpckhps      %xmm0,%xmm8
  DB  65,15,21,217                        ; unpckhps      %xmm9,%xmm3
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  102,15,20,193                       ; unpcklpd      %xmm1,%xmm0
  DB  15,18,202                           ; movhlps       %xmm2,%xmm1
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  102,15,20,211                       ; unpcklpd      %xmm3,%xmm2
  DB  65,15,18,216                        ; movhlps       %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_f32_sse41
_sk_store_f32_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,137,249                          ; mov           %rdi,%rcx
  DB  72,193,225,4                        ; shl           $0x4,%rcx
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  68,15,20,201                        ; unpcklps      %xmm1,%xmm9
  DB  68,15,40,210                        ; movaps        %xmm2,%xmm10
  DB  68,15,40,218                        ; movaps        %xmm2,%xmm11
  DB  68,15,20,219                        ; unpcklps      %xmm3,%xmm11
  DB  68,15,21,193                        ; unpckhps      %xmm1,%xmm8
  DB  68,15,21,211                        ; unpckhps      %xmm3,%xmm10
  DB  69,15,40,225                        ; movaps        %xmm9,%xmm12
  DB  102,69,15,20,227                    ; unpcklpd      %xmm11,%xmm12
  DB  69,15,18,217                        ; movhlps       %xmm9,%xmm11
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  102,69,15,20,202                    ; unpcklpd      %xmm10,%xmm9
  DB  69,15,18,208                        ; movhlps       %xmm8,%xmm10
  DB  102,68,15,17,36,8                   ; movupd        %xmm12,(%rax,%rcx,1)
  DB  68,15,17,92,8,16                    ; movups        %xmm11,0x10(%rax,%rcx,1)
  DB  102,68,15,17,76,8,32                ; movupd        %xmm9,0x20(%rax,%rcx,1)
  DB  68,15,17,84,8,48                    ; movups        %xmm10,0x30(%rax,%rcx,1)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_x_sse41
_sk_clamp_x_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  102,15,118,192                      ; pcmpeqd       %xmm0,%xmm0
  DB  102,65,15,254,193                   ; paddd         %xmm9,%xmm0
  DB  68,15,93,192                        ; minps         %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_y_sse41
_sk_clamp_y_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,95,193                        ; maxps         %xmm1,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  102,15,118,201                      ; pcmpeqd       %xmm1,%xmm1
  DB  102,65,15,254,201                   ; paddd         %xmm9,%xmm1
  DB  68,15,93,193                        ; minps         %xmm1,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_x_sse41
_sk_repeat_x_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  69,15,94,200                        ; divps         %xmm8,%xmm9
  DB  102,69,15,58,8,201,1                ; roundps       $0x1,%xmm9,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
  DB  102,69,15,254,200                   ; paddd         %xmm8,%xmm9
  DB  65,15,93,193                        ; minps         %xmm9,%xmm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_y_sse41
_sk_repeat_y_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  69,15,94,200                        ; divps         %xmm8,%xmm9
  DB  102,69,15,58,8,201,1                ; roundps       $0x1,%xmm9,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  65,15,92,201                        ; subps         %xmm9,%xmm1
  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
  DB  102,69,15,254,200                   ; paddd         %xmm8,%xmm9
  DB  65,15,93,201                        ; minps         %xmm9,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_x_sse41
_sk_mirror_x_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  243,69,15,88,192                    ; addss         %xmm8,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
  DB  69,15,94,208                        ; divps         %xmm8,%xmm10
  DB  102,69,15,58,8,210,1                ; roundps       $0x1,%xmm10,%xmm10
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  65,15,92,194                        ; subps         %xmm10,%xmm0
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,92,192                        ; subps         %xmm0,%xmm8
  DB  65,15,84,192                        ; andps         %xmm8,%xmm0
  DB  102,69,15,118,192                   ; pcmpeqd       %xmm8,%xmm8
  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_y_sse41
_sk_mirror_y_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  65,15,92,201                        ; subps         %xmm9,%xmm1
  DB  243,69,15,88,192                    ; addss         %xmm8,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
  DB  69,15,94,208                        ; divps         %xmm8,%xmm10
  DB  102,69,15,58,8,210,1                ; roundps       $0x1,%xmm10,%xmm10
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  65,15,92,202                        ; subps         %xmm10,%xmm1
  DB  65,15,92,201                        ; subps         %xmm9,%xmm1
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,92,193                        ; subps         %xmm1,%xmm8
  DB  65,15,84,200                        ; andps         %xmm8,%xmm1
  DB  102,69,15,118,192                   ; pcmpeqd       %xmm8,%xmm8
  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_luminance_to_alpha_sse41
_sk_luminance_to_alpha_sse41 LABEL PROC
  DB  184,208,179,89,62                   ; mov           $0x3e59b3d0,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  184,89,23,55,63                     ; mov           $0x3f371759,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  15,88,195                           ; addps         %xmm3,%xmm0
  DB  184,152,221,147,61                  ; mov           $0x3d93dd98,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,218                           ; mulps         %xmm2,%xmm3
  DB  15,88,216                           ; addps         %xmm0,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_2x3_sse41
_sk_matrix_2x3_sse41 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,16                  ; movss         0x10(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,68,15,16,80,12                  ; movss         0xc(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,20                  ; movss         0x14(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_3x4_sse41
_sk_matrix_3x4_sse41 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,68,15,16,80,12                  ; movss         0xc(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,24                  ; movss         0x18(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,36                  ; movss         0x24(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,28                  ; movss         0x1c(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,40                  ; movss         0x28(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,20                  ; movss         0x14(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,32                  ; movss         0x20(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,44                 ; movss         0x2c(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  68,15,89,226                        ; mulps         %xmm2,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_4x5_sse41
_sk_matrix_4x5_sse41 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,32                  ; movss         0x20(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,48                  ; movss         0x30(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,64                 ; movss         0x40(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  68,15,89,227                        ; mulps         %xmm3,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,68,15,16,80,20                  ; movss         0x14(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,36                  ; movss         0x24(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,52                  ; movss         0x34(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,68                 ; movss         0x44(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  68,15,89,227                        ; mulps         %xmm3,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,24                  ; movss         0x18(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,40                  ; movss         0x28(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,56                 ; movss         0x38(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  243,68,15,16,112,72                 ; movss         0x48(%rax),%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  68,15,89,235                        ; mulps         %xmm3,%xmm13
  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
  DB  68,15,89,226                        ; mulps         %xmm2,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  243,68,15,16,88,12                  ; movss         0xc(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,28                  ; movss         0x1c(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,44                 ; movss         0x2c(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  243,68,15,16,112,60                 ; movss         0x3c(%rax),%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  243,68,15,16,120,76                 ; movss         0x4c(%rax),%xmm15
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  68,15,89,243                        ; mulps         %xmm3,%xmm14
  DB  69,15,88,247                        ; addps         %xmm15,%xmm14
  DB  68,15,89,234                        ; mulps         %xmm2,%xmm13
  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
  DB  69,15,89,225                        ; mulps         %xmm9,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_perspective_sse41
_sk_matrix_perspective_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,68,15,16,72,4                   ; movss         0x4(%rax),%xmm9
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,193                        ; addps         %xmm9,%xmm0
  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,20                  ; movss         0x14(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  243,68,15,16,80,24                  ; movss         0x18(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,28                  ; movss         0x1c(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,32                  ; movss         0x20(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,83,202                        ; rcpps         %xmm10,%xmm1
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_sse41
_sk_linear_gradient_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,20                  ; movss         0x14(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,24                  ; movss         0x18(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,28                 ; movss         0x1c(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,132,4,1,0,0                      ; je            3b7a <_sk_linear_gradient_sse41+0x13e>
  DB  72,131,236,88                       ; sub           $0x58,%rsp
  DB  15,41,36,36                         ; movaps        %xmm4,(%rsp)
  DB  15,41,108,36,16                     ; movaps        %xmm5,0x10(%rsp)
  DB  15,41,116,36,32                     ; movaps        %xmm6,0x20(%rsp)
  DB  15,41,124,36,48                     ; movaps        %xmm7,0x30(%rsp)
  DB  72,139,64,8                         ; mov           0x8(%rax),%rax
  DB  72,131,192,32                       ; add           $0x20,%rax
  DB  69,15,87,201                        ; xorps         %xmm9,%xmm9
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,40,233                           ; movaps        %xmm1,%xmm5
  DB  15,40,242                           ; movaps        %xmm2,%xmm6
  DB  15,40,251                           ; movaps        %xmm3,%xmm7
  DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
  DB  69,15,40,243                        ; movaps        %xmm11,%xmm14
  DB  69,15,40,252                        ; movaps        %xmm12,%xmm15
  DB  68,15,41,108,36,64                  ; movaps        %xmm13,0x40(%rsp)
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  243,15,16,80,224                    ; movss         -0x20(%rax),%xmm2
  DB  243,68,15,16,72,228                 ; movss         -0x1c(%rax),%xmm9
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,40,224                           ; movaps        %xmm0,%xmm4
  DB  15,194,194,1                        ; cmpltps       %xmm2,%xmm0
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  102,68,15,56,20,201                 ; blendvps      %xmm0,%xmm1,%xmm9
  DB  243,15,16,72,232                    ; movss         -0x18(%rax),%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  102,15,56,20,205                    ; blendvps      %xmm0,%xmm5,%xmm1
  DB  243,15,16,80,236                    ; movss         -0x14(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  102,15,56,20,214                    ; blendvps      %xmm0,%xmm6,%xmm2
  DB  243,15,16,88,240                    ; movss         -0x10(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  102,15,56,20,223                    ; blendvps      %xmm0,%xmm7,%xmm3
  DB  243,68,15,16,80,244                 ; movss         -0xc(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  102,69,15,56,20,208                 ; blendvps      %xmm0,%xmm8,%xmm10
  DB  243,68,15,16,88,248                 ; movss         -0x8(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  102,69,15,56,20,222                 ; blendvps      %xmm0,%xmm14,%xmm11
  DB  243,68,15,16,96,252                 ; movss         -0x4(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  102,69,15,56,20,231                 ; blendvps      %xmm0,%xmm15,%xmm12
  DB  243,68,15,16,40                     ; movss         (%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  102,68,15,56,20,108,36,64           ; blendvps      %xmm0,0x40(%rsp),%xmm13
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  72,131,192,36                       ; add           $0x24,%rax
  DB  72,255,201                          ; dec           %rcx
  DB  15,133,65,255,255,255               ; jne           3aa2 <_sk_linear_gradient_sse41+0x66>
  DB  15,40,124,36,48                     ; movaps        0x30(%rsp),%xmm7
  DB  15,40,116,36,32                     ; movaps        0x20(%rsp),%xmm6
  DB  15,40,108,36,16                     ; movaps        0x10(%rsp),%xmm5
  DB  15,40,36,36                         ; movaps        (%rsp),%xmm4
  DB  72,131,196,88                       ; add           $0x58,%rsp
  DB  235,13                              ; jmp           3b87 <_sk_linear_gradient_sse41+0x14b>
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  69,15,87,201                        ; xorps         %xmm9,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  15,89,200                           ; mulps         %xmm0,%xmm1
  DB  65,15,88,203                        ; addps         %xmm11,%xmm1
  DB  15,89,208                           ; mulps         %xmm0,%xmm2
  DB  65,15,88,212                        ; addps         %xmm12,%xmm2
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  65,15,88,221                        ; addps         %xmm13,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_2stops_sse41
_sk_linear_gradient_2stops_sse41 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,15,16,80,16                     ; movss         0x10(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,88,194                           ; addps         %xmm2,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,15,16,80,20                     ; movss         0x14(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,88,202                           ; addps         %xmm2,%xmm1
  DB  243,15,16,80,8                      ; movss         0x8(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  243,15,16,88,24                     ; movss         0x18(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  15,88,211                           ; addps         %xmm3,%xmm2
  DB  243,15,16,88,12                     ; movss         0xc(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  243,68,15,16,72,28                  ; movss         0x1c(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_save_xy_sse41
_sk_save_xy_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,88,200                        ; addps         %xmm0,%xmm9
  DB  102,69,15,58,8,209,1                ; roundps       $0x1,%xmm9,%xmm10
  DB  69,15,92,202                        ; subps         %xmm10,%xmm9
  DB  68,15,88,193                        ; addps         %xmm1,%xmm8
  DB  102,69,15,58,8,208,1                ; roundps       $0x1,%xmm8,%xmm10
  DB  69,15,92,194                        ; subps         %xmm10,%xmm8
  DB  15,17,0                             ; movups        %xmm0,(%rax)
  DB  15,17,72,32                         ; movups        %xmm1,0x20(%rax)
  DB  68,15,17,72,64                      ; movups        %xmm9,0x40(%rax)
  DB  68,15,17,64,96                      ; movups        %xmm8,0x60(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_accumulate_sse41
_sk_accumulate_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
  DB  68,15,16,136,160,0,0,0              ; movups        0xa0(%rax),%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  65,15,88,224                        ; addps         %xmm8,%xmm4
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,193                        ; mulps         %xmm1,%xmm8
  DB  65,15,88,232                        ; addps         %xmm8,%xmm5
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  65,15,88,240                        ; addps         %xmm8,%xmm6
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  65,15,88,249                        ; addps         %xmm9,%xmm7
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_nx_sse41
_sk_bilinear_nx_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_px_sse41
_sk_bilinear_px_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_ny_sse41
_sk_bilinear_ny_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_py_sse41
_sk_bilinear_py_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3x_sse41
_sk_bicubic_n3x_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,191                     ; mov           $0xbfc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1x_sse41
_sk_bicubic_n1x_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1x_sse41
_sk_bicubic_p1x_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  68,15,17,144,128,0,0,0              ; movups        %xmm10,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3x_sse41
_sk_bicubic_p3x_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,194                        ; addps         %xmm10,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3y_sse41
_sk_bicubic_n3y_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,191                     ; mov           $0xbfc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1y_sse41
_sk_bicubic_n1y_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1y_sse41
_sk_bicubic_p1y_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  68,15,17,144,160,0,0,0              ; movups        %xmm10,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3y_sse41
_sk_bicubic_p3y_sse41 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,194                        ; addps         %xmm10,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_callback_sse41
_sk_callback_sse41 LABEL PROC
  DB  83                                  ; push          %rbx
  DB  72,131,236,32                       ; sub           $0x20,%rsp
  DB  68,15,40,197                        ; movaps        %xmm5,%xmm8
  DB  68,15,40,204                        ; movaps        %xmm4,%xmm9
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,40,218                        ; movaps        %xmm2,%xmm11
  DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
  DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
  DB  72,137,211                          ; mov           %rdx,%rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,72,8                         ; mov           0x8(%rax),%rcx
  DB  186,4,0,0,0                         ; mov           $0x4,%edx
  DB  255,16                              ; callq         *(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,137,218                          ; mov           %rbx,%rdx
  DB  65,15,40,197                        ; movaps        %xmm13,%xmm0
  DB  65,15,40,204                        ; movaps        %xmm12,%xmm1
  DB  65,15,40,211                        ; movaps        %xmm11,%xmm2
  DB  65,15,40,218                        ; movaps        %xmm10,%xmm3
  DB  65,15,40,225                        ; movaps        %xmm9,%xmm4
  DB  65,15,40,232                        ; movaps        %xmm8,%xmm5
  DB  72,131,196,32                       ; add           $0x20,%rsp
  DB  91                                  ; pop           %rbx
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_start_pipeline_sse2
_sk_start_pipeline_sse2 LABEL PROC
  DB  65,87                               ; push          %r15
  DB  65,86                               ; push          %r14
  DB  65,85                               ; push          %r13
  DB  65,84                               ; push          %r12
  DB  86                                  ; push          %rsi
  DB  87                                  ; push          %rdi
  DB  83                                  ; push          %rbx
  DB  72,129,236,160,0,0,0                ; sub           $0xa0,%rsp
  DB  68,15,41,188,36,144,0,0,0           ; movaps        %xmm15,0x90(%rsp)
  DB  68,15,41,180,36,128,0,0,0           ; movaps        %xmm14,0x80(%rsp)
  DB  68,15,41,108,36,112                 ; movaps        %xmm13,0x70(%rsp)
  DB  68,15,41,100,36,96                  ; movaps        %xmm12,0x60(%rsp)
  DB  68,15,41,92,36,80                   ; movaps        %xmm11,0x50(%rsp)
  DB  68,15,41,84,36,64                   ; movaps        %xmm10,0x40(%rsp)
  DB  68,15,41,76,36,48                   ; movaps        %xmm9,0x30(%rsp)
  DB  68,15,41,68,36,32                   ; movaps        %xmm8,0x20(%rsp)
  DB  15,41,124,36,16                     ; movaps        %xmm7,0x10(%rsp)
  DB  15,41,52,36                         ; movaps        %xmm6,(%rsp)
  DB  77,137,207                          ; mov           %r9,%r15
  DB  77,137,198                          ; mov           %r8,%r14
  DB  72,137,203                          ; mov           %rcx,%rbx
  DB  72,137,214                          ; mov           %rdx,%rsi
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,196                          ; mov           %rax,%r12
  DB  73,137,245                          ; mov           %rsi,%r13
  DB  72,141,67,4                         ; lea           0x4(%rbx),%rax
  DB  76,57,248                           ; cmp           %r15,%rax
  DB  118,5                               ; jbe           73 <_sk_start_pipeline_sse2+0x73>
  DB  72,137,216                          ; mov           %rbx,%rax
  DB  235,52                              ; jmp           a7 <_sk_start_pipeline_sse2+0xa7>
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  15,87,228                           ; xorps         %xmm4,%xmm4
  DB  15,87,237                           ; xorps         %xmm5,%xmm5
  DB  15,87,246                           ; xorps         %xmm6,%xmm6
  DB  15,87,255                           ; xorps         %xmm7,%xmm7
  DB  72,137,223                          ; mov           %rbx,%rdi
  DB  76,137,238                          ; mov           %r13,%rsi
  DB  76,137,242                          ; mov           %r14,%rdx
  DB  65,255,212                          ; callq         *%r12
  DB  72,141,67,4                         ; lea           0x4(%rbx),%rax
  DB  72,131,195,8                        ; add           $0x8,%rbx
  DB  76,57,251                           ; cmp           %r15,%rbx
  DB  72,137,195                          ; mov           %rax,%rbx
  DB  118,204                             ; jbe           73 <_sk_start_pipeline_sse2+0x73>
  DB  15,40,52,36                         ; movaps        (%rsp),%xmm6
  DB  15,40,124,36,16                     ; movaps        0x10(%rsp),%xmm7
  DB  68,15,40,68,36,32                   ; movaps        0x20(%rsp),%xmm8
  DB  68,15,40,76,36,48                   ; movaps        0x30(%rsp),%xmm9
  DB  68,15,40,84,36,64                   ; movaps        0x40(%rsp),%xmm10
  DB  68,15,40,92,36,80                   ; movaps        0x50(%rsp),%xmm11
  DB  68,15,40,100,36,96                  ; movaps        0x60(%rsp),%xmm12
  DB  68,15,40,108,36,112                 ; movaps        0x70(%rsp),%xmm13
  DB  68,15,40,180,36,128,0,0,0           ; movaps        0x80(%rsp),%xmm14
  DB  68,15,40,188,36,144,0,0,0           ; movaps        0x90(%rsp),%xmm15
  DB  72,129,196,160,0,0,0                ; add           $0xa0,%rsp
  DB  91                                  ; pop           %rbx
  DB  95                                  ; pop           %rdi
  DB  94                                  ; pop           %rsi
  DB  65,92                               ; pop           %r12
  DB  65,93                               ; pop           %r13
  DB  65,94                               ; pop           %r14
  DB  65,95                               ; pop           %r15
  DB  195                                 ; retq

PUBLIC _sk_just_return_sse2
_sk_just_return_sse2 LABEL PROC
  DB  195                                 ; retq

PUBLIC _sk_seed_shader_sse2
_sk_seed_shader_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  102,15,110,199                      ; movd          %edi,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,15,110,209                      ; movd          %ecx,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,88,202                           ; addps         %xmm2,%xmm1
  DB  15,16,2                             ; movups        (%rdx),%xmm0
  DB  15,88,193                           ; addps         %xmm1,%xmm0
  DB  102,15,110,8                        ; movd          (%rax),%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  15,88,202                           ; addps         %xmm2,%xmm1
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  15,87,228                           ; xorps         %xmm4,%xmm4
  DB  15,87,237                           ; xorps         %xmm5,%xmm5
  DB  15,87,246                           ; xorps         %xmm6,%xmm6
  DB  15,87,255                           ; xorps         %xmm7,%xmm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_constant_color_sse2
_sk_constant_color_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,15,16,80,8                      ; movss         0x8(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  243,15,16,88,12                     ; movss         0xc(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clear_sse2
_sk_clear_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcatop_sse2
_sk_srcatop_sse2 LABEL PROC
  DB  15,89,199                           ; mulps         %xmm7,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  65,15,88,193                        ; addps         %xmm9,%xmm0
  DB  15,89,207                           ; mulps         %xmm7,%xmm1
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  65,15,88,201                        ; addps         %xmm9,%xmm1
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstatop_sse2
_sk_dstatop_sse2 LABEL PROC
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,89,197                        ; mulps         %xmm5,%xmm8
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,89,198                        ; mulps         %xmm6,%xmm8
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  65,15,88,208                        ; addps         %xmm8,%xmm2
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcin_sse2
_sk_srcin_sse2 LABEL PROC
  DB  15,89,199                           ; mulps         %xmm7,%xmm0
  DB  15,89,207                           ; mulps         %xmm7,%xmm1
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstin_sse2
_sk_dstin_sse2 LABEL PROC
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  15,40,211                           ; movaps        %xmm3,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcout_sse2
_sk_srcout_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,199                        ; subps         %xmm7,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstout_sse2
_sk_dstout_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,216                        ; movaps        %xmm8,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_srcover_sse2
_sk_srcover_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  65,15,88,193                        ; addps         %xmm9,%xmm0
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  65,15,88,201                        ; addps         %xmm9,%xmm1
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_dstover_sse2
_sk_dstover_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,92,199                        ; subps         %xmm7,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_modulate_sse2
_sk_modulate_sse2 LABEL PROC
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_multiply_sse2
_sk_multiply_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,220                        ; mulps         %xmm4,%xmm11
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  65,15,88,195                        ; addps         %xmm11,%xmm0
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,221                        ; mulps         %xmm5,%xmm11
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  65,15,88,203                        ; addps         %xmm11,%xmm1
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,222                        ; mulps         %xmm6,%xmm11
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  65,15,88,211                        ; addps         %xmm11,%xmm2
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_plus__sse2
_sk_plus__sse2 LABEL PROC
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_screen_sse2
_sk_screen_sse2 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  68,15,88,196                        ; addps         %xmm4,%xmm8
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  68,15,92,192                        ; subps         %xmm0,%xmm8
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,88,205                        ; addps         %xmm5,%xmm9
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  68,15,92,201                        ; subps         %xmm1,%xmm9
  DB  68,15,40,210                        ; movaps        %xmm2,%xmm10
  DB  68,15,88,214                        ; addps         %xmm6,%xmm10
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  68,15,92,210                        ; subps         %xmm2,%xmm10
  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
  DB  68,15,88,223                        ; addps         %xmm7,%xmm11
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  68,15,92,219                        ; subps         %xmm3,%xmm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_xor__sse2
_sk_xor__sse2 LABEL PROC
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,92,216                        ; subps         %xmm8,%xmm3
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,89,212                        ; mulps         %xmm4,%xmm10
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,89,213                        ; mulps         %xmm5,%xmm10
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,89,214                        ; mulps         %xmm6,%xmm10
  DB  65,15,88,210                        ; addps         %xmm10,%xmm2
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  15,89,223                           ; mulps         %xmm7,%xmm3
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_darken_sse2
_sk_darken_sse2 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  68,15,89,207                        ; mulps         %xmm7,%xmm9
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  68,15,95,201                        ; maxps         %xmm1,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  69,15,95,193                        ; maxps         %xmm9,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,95,209                        ; maxps         %xmm9,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lighten_sse2
_sk_lighten_sse2 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  68,15,89,207                        ; mulps         %xmm7,%xmm9
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  68,15,93,201                        ; minps         %xmm1,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  69,15,93,193                        ; minps         %xmm9,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,93,209                        ; minps         %xmm9,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_difference_sse2
_sk_difference_sse2 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  68,15,89,207                        ; mulps         %xmm7,%xmm9
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  68,15,93,201                        ; minps         %xmm1,%xmm9
  DB  69,15,88,201                        ; addps         %xmm9,%xmm9
  DB  65,15,92,193                        ; subps         %xmm9,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  69,15,93,193                        ; minps         %xmm9,%xmm8
  DB  69,15,88,192                        ; addps         %xmm8,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
  DB  65,15,93,209                        ; minps         %xmm9,%xmm2
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_exclusion_sse2
_sk_exclusion_sse2 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  15,40,200                           ; movaps        %xmm0,%xmm1
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  15,92,193                           ; subps         %xmm1,%xmm0
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  68,15,89,197                        ; mulps         %xmm5,%xmm8
  DB  69,15,88,192                        ; addps         %xmm8,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  68,15,92,194                        ; subps         %xmm2,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,89,215                           ; mulps         %xmm7,%xmm2
  DB  15,88,218                           ; addps         %xmm2,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colorburn_sse2
_sk_colorburn_sse2 LABEL PROC
  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
  DB  68,15,40,224                        ; movaps        %xmm0,%xmm12
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
  DB  68,15,92,223                        ; subps         %xmm7,%xmm11
  DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
  DB  65,15,89,196                        ; mulps         %xmm12,%xmm0
  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
  DB  15,40,207                           ; movaps        %xmm7,%xmm1
  DB  15,92,204                           ; subps         %xmm4,%xmm1
  DB  15,89,203                           ; mulps         %xmm3,%xmm1
  DB  65,15,94,204                        ; divps         %xmm12,%xmm1
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,93,233                        ; minps         %xmm1,%xmm13
  DB  68,15,40,247                        ; movaps        %xmm7,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  65,15,40,204                        ; movaps        %xmm12,%xmm1
  DB  65,15,194,202,0                     ; cmpeqps       %xmm10,%xmm1
  DB  68,15,92,203                        ; subps         %xmm3,%xmm9
  DB  68,15,89,243                        ; mulps         %xmm3,%xmm14
  DB  68,15,88,240                        ; addps         %xmm0,%xmm14
  DB  68,15,84,225                        ; andps         %xmm1,%xmm12
  DB  65,15,85,206                        ; andnps        %xmm14,%xmm1
  DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
  DB  68,15,89,236                        ; mulps         %xmm4,%xmm13
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  65,15,86,204                        ; orps          %xmm12,%xmm1
  DB  68,15,40,228                        ; movaps        %xmm4,%xmm12
  DB  68,15,194,231,0                     ; cmpeqps       %xmm7,%xmm12
  DB  65,15,88,205                        ; addps         %xmm13,%xmm1
  DB  65,15,84,196                        ; andps         %xmm12,%xmm0
  DB  68,15,85,225                        ; andnps        %xmm1,%xmm12
  DB  65,15,86,196                        ; orps          %xmm12,%xmm0
  DB  65,15,40,203                        ; movaps        %xmm11,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  68,15,40,231                        ; movaps        %xmm7,%xmm12
  DB  68,15,92,229                        ; subps         %xmm5,%xmm12
  DB  68,15,89,227                        ; mulps         %xmm3,%xmm12
  DB  69,15,94,224                        ; divps         %xmm8,%xmm12
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  69,15,93,236                        ; minps         %xmm12,%xmm13
  DB  68,15,40,231                        ; movaps        %xmm7,%xmm12
  DB  69,15,92,229                        ; subps         %xmm13,%xmm12
  DB  69,15,40,232                        ; movaps        %xmm8,%xmm13
  DB  69,15,194,234,0                     ; cmpeqps       %xmm10,%xmm13
  DB  68,15,89,227                        ; mulps         %xmm3,%xmm12
  DB  68,15,88,225                        ; addps         %xmm1,%xmm12
  DB  69,15,84,197                        ; andps         %xmm13,%xmm8
  DB  69,15,85,236                        ; andnps        %xmm12,%xmm13
  DB  69,15,86,232                        ; orps          %xmm8,%xmm13
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,197                        ; mulps         %xmm5,%xmm8
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  69,15,88,232                        ; addps         %xmm8,%xmm13
  DB  68,15,40,197                        ; movaps        %xmm5,%xmm8
  DB  68,15,194,199,0                     ; cmpeqps       %xmm7,%xmm8
  DB  65,15,84,200                        ; andps         %xmm8,%xmm1
  DB  69,15,85,197                        ; andnps        %xmm13,%xmm8
  DB  65,15,86,200                        ; orps          %xmm8,%xmm1
  DB  68,15,40,199                        ; movaps        %xmm7,%xmm8
  DB  68,15,92,198                        ; subps         %xmm6,%xmm8
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  68,15,94,194                        ; divps         %xmm2,%xmm8
  DB  68,15,40,231                        ; movaps        %xmm7,%xmm12
  DB  69,15,93,224                        ; minps         %xmm8,%xmm12
  DB  68,15,40,199                        ; movaps        %xmm7,%xmm8
  DB  69,15,92,196                        ; subps         %xmm12,%xmm8
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  68,15,194,210,0                     ; cmpeqps       %xmm2,%xmm10
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  69,15,88,195                        ; addps         %xmm11,%xmm8
  DB  65,15,84,210                        ; andps         %xmm10,%xmm2
  DB  69,15,85,208                        ; andnps        %xmm8,%xmm10
  DB  69,15,40,195                        ; movaps        %xmm11,%xmm8
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  68,15,86,210                        ; orps          %xmm2,%xmm10
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  68,15,88,210                        ; addps         %xmm2,%xmm10
  DB  15,40,214                           ; movaps        %xmm6,%xmm2
  DB  15,194,215,0                        ; cmpeqps       %xmm7,%xmm2
  DB  68,15,84,194                        ; andps         %xmm2,%xmm8
  DB  65,15,85,210                        ; andnps        %xmm10,%xmm2
  DB  68,15,86,194                        ; orps          %xmm2,%xmm8
  DB  68,15,89,207                        ; mulps         %xmm7,%xmm9
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_colordodge_sse2
_sk_colordodge_sse2 LABEL PROC
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
  DB  68,15,92,223                        ; subps         %xmm7,%xmm11
  DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
  DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
  DB  69,15,92,225                        ; subps         %xmm9,%xmm12
  DB  69,15,94,196                        ; divps         %xmm12,%xmm8
  DB  68,15,40,231                        ; movaps        %xmm7,%xmm12
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  69,15,93,232                        ; minps         %xmm8,%xmm13
  DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
  DB  68,15,194,243,0                     ; cmpeqps       %xmm3,%xmm14
  DB  68,15,89,235                        ; mulps         %xmm3,%xmm13
  DB  68,15,88,232                        ; addps         %xmm0,%xmm13
  DB  69,15,84,206                        ; andps         %xmm14,%xmm9
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,92,211                        ; subps         %xmm3,%xmm10
  DB  69,15,86,241                        ; orps          %xmm9,%xmm14
  DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  69,15,88,241                        ; addps         %xmm9,%xmm14
  DB  68,15,40,204                        ; movaps        %xmm4,%xmm9
  DB  69,15,194,200,0                     ; cmpeqps       %xmm8,%xmm9
  DB  65,15,84,193                        ; andps         %xmm9,%xmm0
  DB  69,15,85,206                        ; andnps        %xmm14,%xmm9
  DB  65,15,86,193                        ; orps          %xmm9,%xmm0
  DB  68,15,40,235                        ; movaps        %xmm3,%xmm13
  DB  68,15,89,237                        ; mulps         %xmm5,%xmm13
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,92,201                        ; subps         %xmm1,%xmm9
  DB  69,15,94,233                        ; divps         %xmm9,%xmm13
  DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  69,15,93,229                        ; minps         %xmm13,%xmm12
  DB  68,15,40,233                        ; movaps        %xmm1,%xmm13
  DB  68,15,194,235,0                     ; cmpeqps       %xmm3,%xmm13
  DB  68,15,89,227                        ; mulps         %xmm3,%xmm12
  DB  69,15,88,225                        ; addps         %xmm9,%xmm12
  DB  65,15,84,205                        ; andps         %xmm13,%xmm1
  DB  69,15,85,236                        ; andnps        %xmm12,%xmm13
  DB  68,15,86,233                        ; orps          %xmm1,%xmm13
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  68,15,88,205                        ; addps         %xmm5,%xmm9
  DB  68,15,88,233                        ; addps         %xmm1,%xmm13
  DB  15,40,205                           ; movaps        %xmm5,%xmm1
  DB  65,15,194,200,0                     ; cmpeqps       %xmm8,%xmm1
  DB  68,15,84,201                        ; andps         %xmm1,%xmm9
  DB  65,15,85,205                        ; andnps        %xmm13,%xmm1
  DB  68,15,86,201                        ; orps          %xmm1,%xmm9
  DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
  DB  68,15,89,230                        ; mulps         %xmm6,%xmm12
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,92,202                           ; subps         %xmm2,%xmm1
  DB  68,15,94,225                        ; divps         %xmm1,%xmm12
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,93,236                        ; minps         %xmm12,%xmm13
  DB  15,40,202                           ; movaps        %xmm2,%xmm1
  DB  15,194,203,0                        ; cmpeqps       %xmm3,%xmm1
  DB  68,15,89,235                        ; mulps         %xmm3,%xmm13
  DB  69,15,88,235                        ; addps         %xmm11,%xmm13
  DB  15,84,209                           ; andps         %xmm1,%xmm2
  DB  65,15,85,205                        ; andnps        %xmm13,%xmm1
  DB  15,86,202                           ; orps          %xmm2,%xmm1
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  15,88,202                           ; addps         %xmm2,%xmm1
  DB  68,15,194,198,0                     ; cmpeqps       %xmm6,%xmm8
  DB  68,15,88,222                        ; addps         %xmm6,%xmm11
  DB  69,15,84,216                        ; andps         %xmm8,%xmm11
  DB  68,15,85,193                        ; andnps        %xmm1,%xmm8
  DB  69,15,86,195                        ; orps          %xmm11,%xmm8
  DB  68,15,89,215                        ; mulps         %xmm7,%xmm10
  DB  65,15,88,218                        ; addps         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hardlight_sse2
_sk_hardlight_sse2 LABEL PROC
  DB  72,131,236,24                       ; sub           $0x18,%rsp
  DB  15,41,52,36                         ; movaps        %xmm6,(%rsp)
  DB  15,40,245                           ; movaps        %xmm5,%xmm6
  DB  15,40,236                           ; movaps        %xmm4,%xmm5
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
  DB  68,15,92,215                        ; subps         %xmm7,%xmm10
  DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  68,15,92,219                        ; subps         %xmm3,%xmm11
  DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  69,15,88,200                        ; addps         %xmm8,%xmm9
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,92,192                        ; subps         %xmm0,%xmm8
  DB  15,40,227                           ; movaps        %xmm3,%xmm4
  DB  15,89,231                           ; mulps         %xmm7,%xmm4
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,40,247                        ; movaps        %xmm7,%xmm14
  DB  68,15,40,255                        ; movaps        %xmm7,%xmm15
  DB  68,15,92,253                        ; subps         %xmm5,%xmm15
  DB  69,15,89,248                        ; mulps         %xmm8,%xmm15
  DB  69,15,88,255                        ; addps         %xmm15,%xmm15
  DB  68,15,40,228                        ; movaps        %xmm4,%xmm12
  DB  69,15,92,231                        ; subps         %xmm15,%xmm12
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  69,15,88,192                        ; addps         %xmm8,%xmm8
  DB  68,15,194,195,2                     ; cmpleps       %xmm3,%xmm8
  DB  15,89,197                           ; mulps         %xmm5,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  65,15,84,192                        ; andps         %xmm8,%xmm0
  DB  69,15,85,196                        ; andnps        %xmm12,%xmm8
  DB  68,15,86,192                        ; orps          %xmm0,%xmm8
  DB  69,15,40,251                        ; movaps        %xmm11,%xmm15
  DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
  DB  68,15,89,223                        ; mulps         %xmm7,%xmm11
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  68,15,89,254                        ; mulps         %xmm6,%xmm15
  DB  68,15,88,248                        ; addps         %xmm0,%xmm15
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  15,92,193                           ; subps         %xmm1,%xmm0
  DB  68,15,92,238                        ; subps         %xmm6,%xmm13
  DB  68,15,89,232                        ; mulps         %xmm0,%xmm13
  DB  69,15,88,237                        ; addps         %xmm13,%xmm13
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  65,15,92,197                        ; subps         %xmm13,%xmm0
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  69,15,88,201                        ; addps         %xmm9,%xmm9
  DB  68,15,194,203,2                     ; cmpleps       %xmm3,%xmm9
  DB  15,89,206                           ; mulps         %xmm6,%xmm1
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  65,15,84,201                        ; andps         %xmm9,%xmm1
  DB  68,15,85,200                        ; andnps        %xmm0,%xmm9
  DB  68,15,86,201                        ; orps          %xmm1,%xmm9
  DB  69,15,88,207                        ; addps         %xmm15,%xmm9
  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
  DB  68,15,40,44,36                      ; movaps        (%rsp),%xmm13
  DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
  DB  69,15,88,226                        ; addps         %xmm10,%xmm12
  DB  68,15,40,210                        ; movaps        %xmm2,%xmm10
  DB  69,15,88,210                        ; addps         %xmm10,%xmm10
  DB  68,15,194,211,2                     ; cmpleps       %xmm3,%xmm10
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  15,92,194                           ; subps         %xmm2,%xmm0
  DB  65,15,89,213                        ; mulps         %xmm13,%xmm2
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  68,15,89,240                        ; mulps         %xmm0,%xmm14
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  65,15,92,230                        ; subps         %xmm14,%xmm4
  DB  65,15,84,210                        ; andps         %xmm10,%xmm2
  DB  68,15,85,212                        ; andnps        %xmm4,%xmm10
  DB  68,15,86,210                        ; orps          %xmm2,%xmm10
  DB  69,15,88,212                        ; addps         %xmm12,%xmm10
  DB  65,15,88,219                        ; addps         %xmm11,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,40,229                           ; movaps        %xmm5,%xmm4
  DB  15,40,238                           ; movaps        %xmm6,%xmm5
  DB  65,15,40,245                        ; movaps        %xmm13,%xmm6
  DB  72,131,196,24                       ; add           $0x18,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_overlay_sse2
_sk_overlay_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,92,207                        ; subps         %xmm7,%xmm9
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
  DB  68,15,92,195                        ; subps         %xmm3,%xmm8
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,220                        ; mulps         %xmm4,%xmm11
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
  DB  68,15,92,224                        ; subps         %xmm0,%xmm12
  DB  15,89,196                           ; mulps         %xmm4,%xmm0
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,92,236                        ; subps         %xmm4,%xmm13
  DB  68,15,40,244                        ; movaps        %xmm4,%xmm14
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  68,15,194,247,2                     ; cmpleps       %xmm7,%xmm14
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,89,215                        ; mulps         %xmm7,%xmm10
  DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
  DB  69,15,88,237                        ; addps         %xmm13,%xmm13
  DB  69,15,40,226                        ; movaps        %xmm10,%xmm12
  DB  69,15,92,229                        ; subps         %xmm13,%xmm12
  DB  65,15,84,198                        ; andps         %xmm14,%xmm0
  DB  69,15,85,244                        ; andnps        %xmm12,%xmm14
  DB  65,15,86,198                        ; orps          %xmm14,%xmm0
  DB  65,15,88,195                        ; addps         %xmm11,%xmm0
  DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
  DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
  DB  69,15,40,224                        ; movaps        %xmm8,%xmm12
  DB  68,15,89,229                        ; mulps         %xmm5,%xmm12
  DB  69,15,88,227                        ; addps         %xmm11,%xmm12
  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
  DB  68,15,92,217                        ; subps         %xmm1,%xmm11
  DB  15,89,205                           ; mulps         %xmm5,%xmm1
  DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
  DB  68,15,92,237                        ; subps         %xmm5,%xmm13
  DB  68,15,40,245                        ; movaps        %xmm5,%xmm14
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  68,15,194,247,2                     ; cmpleps       %xmm7,%xmm14
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
  DB  69,15,88,237                        ; addps         %xmm13,%xmm13
  DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
  DB  69,15,92,221                        ; subps         %xmm13,%xmm11
  DB  65,15,84,206                        ; andps         %xmm14,%xmm1
  DB  69,15,85,243                        ; andnps        %xmm11,%xmm14
  DB  65,15,86,206                        ; orps          %xmm14,%xmm1
  DB  65,15,88,204                        ; addps         %xmm12,%xmm1
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  68,15,89,222                        ; mulps         %xmm6,%xmm11
  DB  69,15,88,217                        ; addps         %xmm9,%xmm11
  DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
  DB  68,15,92,202                        ; subps         %xmm2,%xmm9
  DB  15,89,214                           ; mulps         %xmm6,%xmm2
  DB  68,15,40,231                        ; movaps        %xmm7,%xmm12
  DB  68,15,92,230                        ; subps         %xmm6,%xmm12
  DB  68,15,40,238                        ; movaps        %xmm6,%xmm13
  DB  69,15,88,237                        ; addps         %xmm13,%xmm13
  DB  68,15,194,239,2                     ; cmpleps       %xmm7,%xmm13
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  69,15,89,225                        ; mulps         %xmm9,%xmm12
  DB  69,15,88,228                        ; addps         %xmm12,%xmm12
  DB  69,15,92,212                        ; subps         %xmm12,%xmm10
  DB  65,15,84,213                        ; andps         %xmm13,%xmm2
  DB  69,15,85,234                        ; andnps        %xmm10,%xmm13
  DB  65,15,86,213                        ; orps          %xmm13,%xmm2
  DB  65,15,88,211                        ; addps         %xmm11,%xmm2
  DB  68,15,89,199                        ; mulps         %xmm7,%xmm8
  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_softlight_sse2
_sk_softlight_sse2 LABEL PROC
  DB  72,131,236,56                       ; sub           $0x38,%rsp
  DB  15,41,84,36,32                      ; movaps        %xmm2,0x20(%rsp)
  DB  15,40,209                           ; movaps        %xmm1,%xmm2
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
  DB  68,15,194,231,1                     ; cmpltps       %xmm7,%xmm12
  DB  68,15,40,212                        ; movaps        %xmm4,%xmm10
  DB  68,15,94,215                        ; divps         %xmm7,%xmm10
  DB  69,15,84,212                        ; andps         %xmm12,%xmm10
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
  DB  69,15,92,242                        ; subps         %xmm10,%xmm14
  DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
  DB  69,15,40,234                        ; movaps        %xmm10,%xmm13
  DB  65,15,82,194                        ; rsqrtps       %xmm10,%xmm0
  DB  68,15,83,248                        ; rcpps         %xmm0,%xmm15
  DB  69,15,92,250                        ; subps         %xmm10,%xmm15
  DB  69,15,88,210                        ; addps         %xmm10,%xmm10
  DB  69,15,88,210                        ; addps         %xmm10,%xmm10
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  15,89,192                           ; mulps         %xmm0,%xmm0
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  69,15,92,217                        ; subps         %xmm9,%xmm11
  DB  68,15,89,216                        ; mulps         %xmm0,%xmm11
  DB  184,0,0,224,64                      ; mov           $0x40e00000,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
  DB  69,15,88,235                        ; addps         %xmm11,%xmm13
  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
  DB  15,40,204                           ; movaps        %xmm4,%xmm1
  DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
  DB  15,88,228                           ; addps         %xmm4,%xmm4
  DB  15,88,228                           ; addps         %xmm4,%xmm4
  DB  15,194,231,2                        ; cmpleps       %xmm7,%xmm4
  DB  68,15,84,236                        ; andps         %xmm4,%xmm13
  DB  65,15,85,231                        ; andnps        %xmm15,%xmm4
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  15,88,192                           ; addps         %xmm0,%xmm0
  DB  65,15,86,229                        ; orps          %xmm13,%xmm4
  DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
  DB  68,15,92,235                        ; subps         %xmm3,%xmm13
  DB  69,15,89,245                        ; mulps         %xmm13,%xmm14
  DB  68,15,89,239                        ; mulps         %xmm7,%xmm13
  DB  65,15,89,229                        ; mulps         %xmm13,%xmm4
  DB  65,15,88,227                        ; addps         %xmm11,%xmm4
  DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
  DB  68,15,92,219                        ; subps         %xmm3,%xmm11
  DB  69,15,40,251                        ; movaps        %xmm11,%xmm15
  DB  15,41,76,36,16                      ; movaps        %xmm1,0x10(%rsp)
  DB  68,15,89,249                        ; mulps         %xmm1,%xmm15
  DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
  DB  68,15,92,239                        ; subps         %xmm7,%xmm13
  DB  69,15,89,197                        ; mulps         %xmm13,%xmm8
  DB  69,15,88,199                        ; addps         %xmm15,%xmm8
  DB  68,15,88,243                        ; addps         %xmm3,%xmm14
  DB  68,15,89,241                        ; mulps         %xmm1,%xmm14
  DB  15,194,195,2                        ; cmpleps       %xmm3,%xmm0
  DB  68,15,84,240                        ; andps         %xmm0,%xmm14
  DB  15,85,196                           ; andnps        %xmm4,%xmm0
  DB  65,15,86,198                        ; orps          %xmm14,%xmm0
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  15,41,44,36                         ; movaps        %xmm5,(%rsp)
  DB  68,15,40,197                        ; movaps        %xmm5,%xmm8
  DB  68,15,94,199                        ; divps         %xmm7,%xmm8
  DB  69,15,84,196                        ; andps         %xmm12,%xmm8
  DB  69,15,40,240                        ; movaps        %xmm8,%xmm14
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  69,15,88,246                        ; addps         %xmm14,%xmm14
  DB  65,15,40,230                        ; movaps        %xmm14,%xmm4
  DB  15,89,228                           ; mulps         %xmm4,%xmm4
  DB  65,15,88,230                        ; addps         %xmm14,%xmm4
  DB  69,15,40,248                        ; movaps        %xmm8,%xmm15
  DB  69,15,92,249                        ; subps         %xmm9,%xmm15
  DB  68,15,89,252                        ; mulps         %xmm4,%xmm15
  DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
  DB  69,15,92,240                        ; subps         %xmm8,%xmm14
  DB  65,15,82,224                        ; rsqrtps       %xmm8,%xmm4
  DB  15,83,228                           ; rcpps         %xmm4,%xmm4
  DB  65,15,92,224                        ; subps         %xmm8,%xmm4
  DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
  DB  69,15,88,199                        ; addps         %xmm15,%xmm8
  DB  68,15,40,253                        ; movaps        %xmm5,%xmm15
  DB  69,15,88,255                        ; addps         %xmm15,%xmm15
  DB  69,15,88,255                        ; addps         %xmm15,%xmm15
  DB  68,15,194,255,2                     ; cmpleps       %xmm7,%xmm15
  DB  69,15,84,199                        ; andps         %xmm15,%xmm8
  DB  68,15,85,252                        ; andnps        %xmm4,%xmm15
  DB  69,15,86,248                        ; orps          %xmm8,%xmm15
  DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
  DB  69,15,88,192                        ; addps         %xmm8,%xmm8
  DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
  DB  15,92,227                           ; subps         %xmm3,%xmm4
  DB  68,15,89,244                        ; mulps         %xmm4,%xmm14
  DB  15,89,231                           ; mulps         %xmm7,%xmm4
  DB  68,15,89,252                        ; mulps         %xmm4,%xmm15
  DB  15,40,227                           ; movaps        %xmm3,%xmm4
  DB  15,89,229                           ; mulps         %xmm5,%xmm4
  DB  68,15,88,252                        ; addps         %xmm4,%xmm15
  DB  65,15,40,227                        ; movaps        %xmm11,%xmm4
  DB  15,89,229                           ; mulps         %xmm5,%xmm4
  DB  65,15,89,213                        ; mulps         %xmm13,%xmm2
  DB  15,88,212                           ; addps         %xmm4,%xmm2
  DB  68,15,88,243                        ; addps         %xmm3,%xmm14
  DB  68,15,89,245                        ; mulps         %xmm5,%xmm14
  DB  68,15,194,195,2                     ; cmpleps       %xmm3,%xmm8
  DB  69,15,84,240                        ; andps         %xmm8,%xmm14
  DB  69,15,85,199                        ; andnps        %xmm15,%xmm8
  DB  69,15,86,198                        ; orps          %xmm14,%xmm8
  DB  68,15,88,194                        ; addps         %xmm2,%xmm8
  DB  68,15,40,246                        ; movaps        %xmm6,%xmm14
  DB  65,15,40,206                        ; movaps        %xmm14,%xmm1
  DB  15,94,207                           ; divps         %xmm7,%xmm1
  DB  65,15,84,204                        ; andps         %xmm12,%xmm1
  DB  15,40,225                           ; movaps        %xmm1,%xmm4
  DB  65,15,92,225                        ; subps         %xmm9,%xmm4
  DB  68,15,92,201                        ; subps         %xmm1,%xmm9
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  15,82,241                           ; rsqrtps       %xmm1,%xmm6
  DB  15,83,246                           ; rcpps         %xmm6,%xmm6
  DB  15,92,241                           ; subps         %xmm1,%xmm6
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  15,40,233                           ; movaps        %xmm1,%xmm5
  DB  15,89,237                           ; mulps         %xmm5,%xmm5
  DB  15,88,233                           ; addps         %xmm1,%xmm5
  DB  15,89,236                           ; mulps         %xmm4,%xmm5
  DB  68,15,88,213                        ; addps         %xmm5,%xmm10
  DB  65,15,40,238                        ; movaps        %xmm14,%xmm5
  DB  15,40,205                           ; movaps        %xmm5,%xmm1
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  15,88,201                           ; addps         %xmm1,%xmm1
  DB  15,194,207,2                        ; cmpleps       %xmm7,%xmm1
  DB  68,15,84,209                        ; andps         %xmm1,%xmm10
  DB  15,85,206                           ; andnps        %xmm6,%xmm1
  DB  15,40,84,36,32                      ; movaps        0x20(%rsp),%xmm2
  DB  68,15,89,234                        ; mulps         %xmm2,%xmm13
  DB  15,88,210                           ; addps         %xmm2,%xmm2
  DB  65,15,86,202                        ; orps          %xmm10,%xmm1
  DB  15,40,226                           ; movaps        %xmm2,%xmm4
  DB  15,92,227                           ; subps         %xmm3,%xmm4
  DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
  DB  15,89,231                           ; mulps         %xmm7,%xmm4
  DB  15,89,204                           ; mulps         %xmm4,%xmm1
  DB  15,40,227                           ; movaps        %xmm3,%xmm4
  DB  15,89,229                           ; mulps         %xmm5,%xmm4
  DB  15,88,204                           ; addps         %xmm4,%xmm1
  DB  65,15,40,227                        ; movaps        %xmm11,%xmm4
  DB  15,89,229                           ; mulps         %xmm5,%xmm4
  DB  65,15,88,229                        ; addps         %xmm13,%xmm4
  DB  68,15,88,203                        ; addps         %xmm3,%xmm9
  DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
  DB  15,40,245                           ; movaps        %xmm5,%xmm6
  DB  15,194,211,2                        ; cmpleps       %xmm3,%xmm2
  DB  68,15,84,202                        ; andps         %xmm2,%xmm9
  DB  15,85,209                           ; andnps        %xmm1,%xmm2
  DB  65,15,86,209                        ; orps          %xmm9,%xmm2
  DB  15,88,212                           ; addps         %xmm4,%xmm2
  DB  68,15,89,223                        ; mulps         %xmm7,%xmm11
  DB  65,15,88,219                        ; addps         %xmm11,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,100,36,16                     ; movaps        0x10(%rsp),%xmm4
  DB  15,40,44,36                         ; movaps        (%rsp),%xmm5
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  72,131,196,56                       ; add           $0x38,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_0_sse2
_sk_clamp_0_sse2 LABEL PROC
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  65,15,95,192                        ; maxps         %xmm8,%xmm0
  DB  65,15,95,200                        ; maxps         %xmm8,%xmm1
  DB  65,15,95,208                        ; maxps         %xmm8,%xmm2
  DB  65,15,95,216                        ; maxps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_1_sse2
_sk_clamp_1_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_a_sse2
_sk_clamp_a_sse2 LABEL PROC
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
  DB  15,93,195                           ; minps         %xmm3,%xmm0
  DB  15,93,203                           ; minps         %xmm3,%xmm1
  DB  15,93,211                           ; minps         %xmm3,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_set_rgb_sse2
_sk_set_rgb_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,15,16,80,8                      ; movss         0x8(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_rb_sse2
_sk_swap_rb_sse2 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_swap_sse2
_sk_swap_sse2 LABEL PROC
  DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
  DB  68,15,40,202                        ; movaps        %xmm2,%xmm9
  DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
  DB  68,15,40,216                        ; movaps        %xmm0,%xmm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,40,205                           ; movaps        %xmm5,%xmm1
  DB  15,40,214                           ; movaps        %xmm6,%xmm2
  DB  15,40,223                           ; movaps        %xmm7,%xmm3
  DB  65,15,40,227                        ; movaps        %xmm11,%xmm4
  DB  65,15,40,234                        ; movaps        %xmm10,%xmm5
  DB  65,15,40,241                        ; movaps        %xmm9,%xmm6
  DB  65,15,40,248                        ; movaps        %xmm8,%xmm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_src_dst_sse2
_sk_move_src_dst_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,224                           ; movaps        %xmm0,%xmm4
  DB  15,40,233                           ; movaps        %xmm1,%xmm5
  DB  15,40,242                           ; movaps        %xmm2,%xmm6
  DB  15,40,251                           ; movaps        %xmm3,%xmm7
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_move_dst_src_sse2
_sk_move_dst_src_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,40,205                           ; movaps        %xmm5,%xmm1
  DB  15,40,214                           ; movaps        %xmm6,%xmm2
  DB  15,40,223                           ; movaps        %xmm7,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_premul_sse2
_sk_premul_sse2 LABEL PROC
  DB  15,89,195                           ; mulps         %xmm3,%xmm0
  DB  15,89,203                           ; mulps         %xmm3,%xmm1
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_unpremul_sse2
_sk_unpremul_sse2 LABEL PROC
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,94,203                        ; divps         %xmm3,%xmm9
  DB  68,15,194,195,4                     ; cmpneqps      %xmm3,%xmm8
  DB  69,15,84,193                        ; andps         %xmm9,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_srgb_sse2
_sk_from_srgb_sse2 LABEL PROC
  DB  184,145,131,158,61                  ; mov           $0x3d9e8391,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,232                        ; movaps        %xmm8,%xmm13
  DB  68,15,89,232                        ; mulps         %xmm0,%xmm13
  DB  68,15,40,224                        ; movaps        %xmm0,%xmm12
  DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
  DB  184,154,153,153,62                  ; mov           $0x3e99999a,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  184,92,143,50,63                    ; mov           $0x3f328f5c,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
  DB  68,15,89,240                        ; mulps         %xmm0,%xmm14
  DB  69,15,88,242                        ; addps         %xmm10,%xmm14
  DB  184,10,215,35,59                    ; mov           $0x3b23d70a,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,244                        ; mulps         %xmm12,%xmm14
  DB  69,15,88,243                        ; addps         %xmm11,%xmm14
  DB  184,174,71,97,61                    ; mov           $0x3d6147ae,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  65,15,194,196,1                     ; cmpltps       %xmm12,%xmm0
  DB  68,15,84,232                        ; andps         %xmm0,%xmm13
  DB  65,15,85,198                        ; andnps        %xmm14,%xmm0
  DB  65,15,86,197                        ; orps          %xmm13,%xmm0
  DB  69,15,40,232                        ; movaps        %xmm8,%xmm13
  DB  68,15,89,233                        ; mulps         %xmm1,%xmm13
  DB  68,15,40,241                        ; movaps        %xmm1,%xmm14
  DB  69,15,89,246                        ; mulps         %xmm14,%xmm14
  DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
  DB  68,15,89,249                        ; mulps         %xmm1,%xmm15
  DB  69,15,88,250                        ; addps         %xmm10,%xmm15
  DB  69,15,89,254                        ; mulps         %xmm14,%xmm15
  DB  69,15,88,251                        ; addps         %xmm11,%xmm15
  DB  65,15,194,204,1                     ; cmpltps       %xmm12,%xmm1
  DB  68,15,84,233                        ; andps         %xmm1,%xmm13
  DB  65,15,85,207                        ; andnps        %xmm15,%xmm1
  DB  65,15,86,205                        ; orps          %xmm13,%xmm1
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  68,15,40,234                        ; movaps        %xmm2,%xmm13
  DB  69,15,89,237                        ; mulps         %xmm13,%xmm13
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  69,15,89,205                        ; mulps         %xmm13,%xmm9
  DB  69,15,88,203                        ; addps         %xmm11,%xmm9
  DB  65,15,194,212,1                     ; cmpltps       %xmm12,%xmm2
  DB  68,15,84,194                        ; andps         %xmm2,%xmm8
  DB  65,15,85,209                        ; andnps        %xmm9,%xmm2
  DB  65,15,86,208                        ; orps          %xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_srgb_sse2
_sk_to_srgb_sse2 LABEL PROC
  DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
  DB  69,15,83,248                        ; rcpps         %xmm8,%xmm15
  DB  69,15,82,232                        ; rsqrtps       %xmm8,%xmm13
  DB  184,41,92,71,65                     ; mov           $0x41475c29,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,240                        ; movaps        %xmm8,%xmm14
  DB  68,15,89,240                        ; mulps         %xmm0,%xmm14
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  184,194,135,210,62                  ; mov           $0x3ed287c2,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  184,206,111,48,63                   ; mov           $0x3f306fce,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  184,168,87,202,61                   ; mov           $0x3dca57a8,%eax
  DB  53,0,0,0,128                        ; xor           $0x80000000,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,89,251                        ; mulps         %xmm11,%xmm15
  DB  69,15,88,252                        ; addps         %xmm12,%xmm15
  DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
  DB  69,15,88,239                        ; addps         %xmm15,%xmm13
  DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
  DB  69,15,93,253                        ; minps         %xmm13,%xmm15
  DB  184,4,231,140,59                    ; mov           $0x3b8ce704,%eax
  DB  102,68,15,110,232                   ; movd          %eax,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  65,15,194,197,1                     ; cmpltps       %xmm13,%xmm0
  DB  68,15,84,240                        ; andps         %xmm0,%xmm14
  DB  65,15,85,199                        ; andnps        %xmm15,%xmm0
  DB  65,15,86,198                        ; orps          %xmm14,%xmm0
  DB  68,15,82,241                        ; rsqrtps       %xmm1,%xmm14
  DB  69,15,83,254                        ; rcpps         %xmm14,%xmm15
  DB  69,15,82,246                        ; rsqrtps       %xmm14,%xmm14
  DB  69,15,89,251                        ; mulps         %xmm11,%xmm15
  DB  69,15,88,252                        ; addps         %xmm12,%xmm15
  DB  69,15,89,242                        ; mulps         %xmm10,%xmm14
  DB  69,15,88,247                        ; addps         %xmm15,%xmm14
  DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
  DB  69,15,93,254                        ; minps         %xmm14,%xmm15
  DB  69,15,40,240                        ; movaps        %xmm8,%xmm14
  DB  68,15,89,241                        ; mulps         %xmm1,%xmm14
  DB  65,15,194,205,1                     ; cmpltps       %xmm13,%xmm1
  DB  68,15,84,241                        ; andps         %xmm1,%xmm14
  DB  65,15,85,207                        ; andnps        %xmm15,%xmm1
  DB  65,15,86,206                        ; orps          %xmm14,%xmm1
  DB  68,15,82,242                        ; rsqrtps       %xmm2,%xmm14
  DB  69,15,83,254                        ; rcpps         %xmm14,%xmm15
  DB  69,15,89,251                        ; mulps         %xmm11,%xmm15
  DB  69,15,88,252                        ; addps         %xmm12,%xmm15
  DB  69,15,82,222                        ; rsqrtps       %xmm14,%xmm11
  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
  DB  69,15,88,223                        ; addps         %xmm15,%xmm11
  DB  69,15,93,203                        ; minps         %xmm11,%xmm9
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  65,15,194,213,1                     ; cmpltps       %xmm13,%xmm2
  DB  68,15,84,194                        ; andps         %xmm2,%xmm8
  DB  65,15,85,209                        ; andnps        %xmm9,%xmm2
  DB  65,15,86,208                        ; orps          %xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_from_2dot2_sse2
_sk_from_2dot2_sse2 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  65,15,82,192                        ; rsqrtps       %xmm8,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  68,15,82,200                        ; rsqrtps       %xmm0,%xmm9
  DB  65,15,82,193                        ; rsqrtps       %xmm9,%xmm0
  DB  68,15,82,208                        ; rsqrtps       %xmm0,%xmm10
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  15,89,192                           ; mulps         %xmm0,%xmm0
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
  DB  68,15,82,193                        ; rsqrtps       %xmm1,%xmm8
  DB  69,15,82,192                        ; rsqrtps       %xmm8,%xmm8
  DB  69,15,82,192                        ; rsqrtps       %xmm8,%xmm8
  DB  69,15,82,200                        ; rsqrtps       %xmm8,%xmm9
  DB  69,15,82,193                        ; rsqrtps       %xmm9,%xmm8
  DB  69,15,82,216                        ; rsqrtps       %xmm8,%xmm11
  DB  15,89,201                           ; mulps         %xmm1,%xmm1
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  68,15,89,193                        ; mulps         %xmm1,%xmm8
  DB  69,15,89,195                        ; mulps         %xmm11,%xmm8
  DB  69,15,95,194                        ; maxps         %xmm10,%xmm8
  DB  15,82,202                           ; rsqrtps       %xmm2,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  68,15,82,217                        ; rsqrtps       %xmm1,%xmm11
  DB  65,15,82,203                        ; rsqrtps       %xmm11,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,89,210                           ; mulps         %xmm2,%xmm2
  DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
  DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
  DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  69,15,95,202                        ; maxps         %xmm10,%xmm9
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_to_2dot2_sse2
_sk_to_2dot2_sse2 LABEL PROC
  DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
  DB  65,15,82,192                        ; rsqrtps       %xmm8,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  15,82,192                           ; rsqrtps       %xmm0,%xmm0
  DB  68,15,82,200                        ; rsqrtps       %xmm0,%xmm9
  DB  69,15,83,192                        ; rcpps         %xmm8,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  65,15,83,193                        ; rcpps         %xmm9,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  65,15,95,192                        ; maxps         %xmm8,%xmm0
  DB  68,15,82,201                        ; rsqrtps       %xmm1,%xmm9
  DB  65,15,82,201                        ; rsqrtps       %xmm9,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  15,82,201                           ; rsqrtps       %xmm1,%xmm1
  DB  68,15,82,209                        ; rsqrtps       %xmm1,%xmm10
  DB  69,15,83,201                        ; rcpps         %xmm9,%xmm9
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  65,15,83,202                        ; rcpps         %xmm10,%xmm1
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  65,15,95,200                        ; maxps         %xmm8,%xmm1
  DB  68,15,82,202                        ; rsqrtps       %xmm2,%xmm9
  DB  65,15,82,209                        ; rsqrtps       %xmm9,%xmm2
  DB  15,82,210                           ; rsqrtps       %xmm2,%xmm2
  DB  15,82,210                           ; rsqrtps       %xmm2,%xmm2
  DB  15,82,210                           ; rsqrtps       %xmm2,%xmm2
  DB  68,15,82,210                        ; rsqrtps       %xmm2,%xmm10
  DB  69,15,83,201                        ; rcpps         %xmm9,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  65,15,83,210                        ; rcpps         %xmm10,%xmm2
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  65,15,95,208                        ; maxps         %xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_rgb_to_hsl_sse2
_sk_rgb_to_hsl_sse2 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
  DB  69,15,95,209                        ; maxps         %xmm9,%xmm10
  DB  68,15,95,210                        ; maxps         %xmm2,%xmm10
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  69,15,93,217                        ; minps         %xmm9,%xmm11
  DB  68,15,93,218                        ; minps         %xmm2,%xmm11
  DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
  DB  65,15,92,203                        ; subps         %xmm11,%xmm1
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  68,15,94,225                        ; divps         %xmm1,%xmm12
  DB  65,184,171,170,42,62                ; mov           $0x3e2aaaab,%r8d
  DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
  DB  65,15,194,192,0                     ; cmpeqps       %xmm8,%xmm0
  DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
  DB  68,15,194,242,1                     ; cmpltps       %xmm2,%xmm14
  DB  185,0,0,192,64                      ; mov           $0x40c00000,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,84,238                        ; andps         %xmm14,%xmm13
  DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
  DB  68,15,92,242                        ; subps         %xmm2,%xmm14
  DB  69,15,89,244                        ; mulps         %xmm12,%xmm14
  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
  DB  69,15,40,242                        ; movaps        %xmm10,%xmm14
  DB  69,15,194,241,0                     ; cmpeqps       %xmm9,%xmm14
  DB  65,15,92,208                        ; subps         %xmm8,%xmm2
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  65,15,89,212                        ; mulps         %xmm12,%xmm2
  DB  185,0,0,0,64                        ; mov           $0x40000000,%ecx
  DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
  DB  184,0,0,128,64                      ; mov           $0x40800000,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,88,200                        ; addps         %xmm8,%xmm9
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,88,208                        ; addps         %xmm8,%xmm2
  DB  65,15,84,214                        ; andps         %xmm14,%xmm2
  DB  69,15,85,241                        ; andnps        %xmm9,%xmm14
  DB  68,15,86,242                        ; orps          %xmm2,%xmm14
  DB  68,15,84,232                        ; andps         %xmm0,%xmm13
  DB  65,15,85,198                        ; andnps        %xmm14,%xmm0
  DB  102,65,15,110,208                   ; movd          %r8d,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,86,197                        ; orps          %xmm13,%xmm0
  DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
  DB  69,15,194,203,4                     ; cmpneqps      %xmm11,%xmm9
  DB  65,15,84,193                        ; andps         %xmm9,%xmm0
  DB  15,89,194                           ; mulps         %xmm2,%xmm0
  DB  69,15,92,194                        ; subps         %xmm10,%xmm8
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,92,195                        ; subps         %xmm11,%xmm8
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  65,15,89,212                        ; mulps         %xmm12,%xmm2
  DB  68,15,194,226,1                     ; cmpltps       %xmm2,%xmm12
  DB  69,15,84,196                        ; andps         %xmm12,%xmm8
  DB  69,15,85,226                        ; andnps        %xmm10,%xmm12
  DB  69,15,86,224                        ; orps          %xmm8,%xmm12
  DB  65,15,94,204                        ; divps         %xmm12,%xmm1
  DB  65,15,84,201                        ; andps         %xmm9,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_hsl_to_rgb_sse2
_sk_hsl_to_rgb_sse2 LABEL PROC
  DB  72,131,236,104                      ; sub           $0x68,%rsp
  DB  15,41,124,36,80                     ; movaps        %xmm7,0x50(%rsp)
  DB  15,41,116,36,64                     ; movaps        %xmm6,0x40(%rsp)
  DB  15,41,108,36,48                     ; movaps        %xmm5,0x30(%rsp)
  DB  15,41,100,36,32                     ; movaps        %xmm4,0x20(%rsp)
  DB  15,41,92,36,16                      ; movaps        %xmm3,0x10(%rsp)
  DB  68,15,40,210                        ; movaps        %xmm2,%xmm10
  DB  15,40,209                           ; movaps        %xmm1,%xmm2
  DB  15,40,240                           ; movaps        %xmm0,%xmm6
  DB  184,0,0,0,63                        ; mov           $0x3f000000,%eax
  DB  102,68,15,110,240                   ; movd          %eax,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,40,202                        ; movaps        %xmm10,%xmm9
  DB  69,15,194,206,1                     ; cmpltps       %xmm14,%xmm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  15,88,194                           ; addps         %xmm2,%xmm0
  DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
  DB  15,40,218                           ; movaps        %xmm2,%xmm3
  DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
  DB  68,15,194,218,0                     ; cmpeqps       %xmm2,%xmm11
  DB  65,15,88,210                        ; addps         %xmm10,%xmm2
  DB  65,15,89,218                        ; mulps         %xmm10,%xmm3
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  65,15,84,193                        ; andps         %xmm9,%xmm0
  DB  68,15,85,202                        ; andnps        %xmm2,%xmm9
  DB  68,15,86,200                        ; orps          %xmm0,%xmm9
  DB  184,0,0,0,64                        ; mov           $0x40000000,%eax
  DB  185,171,170,170,62                  ; mov           $0x3eaaaaab,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,41,4,36                       ; movaps        %xmm8,(%rsp)
  DB  68,15,88,198                        ; addps         %xmm6,%xmm8
  DB  185,0,0,0,0                         ; mov           $0x0,%ecx
  DB  102,15,110,233                      ; movd          %ecx,%xmm5
  DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  65,15,194,192,1                     ; cmpltps       %xmm8,%xmm0
  DB  65,15,40,216                        ; movaps        %xmm8,%xmm3
  DB  65,15,92,220                        ; subps         %xmm12,%xmm3
  DB  15,84,216                           ; andps         %xmm0,%xmm3
  DB  65,15,85,192                        ; andnps        %xmm8,%xmm0
  DB  15,86,195                           ; orps          %xmm3,%xmm0
  DB  65,15,40,216                        ; movaps        %xmm8,%xmm3
  DB  15,194,221,1                        ; cmpltps       %xmm5,%xmm3
  DB  65,15,40,212                        ; movaps        %xmm12,%xmm2
  DB  65,15,88,208                        ; addps         %xmm8,%xmm2
  DB  15,84,211                           ; andps         %xmm3,%xmm2
  DB  15,85,216                           ; andnps        %xmm0,%xmm3
  DB  15,86,218                           ; orps          %xmm2,%xmm3
  DB  102,68,15,110,232                   ; movd          %eax,%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
  DB  69,15,92,233                        ; subps         %xmm9,%xmm13
  DB  184,171,170,42,62                   ; mov           $0x3e2aaaab,%eax
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  65,15,92,197                        ; subps         %xmm13,%xmm0
  DB  185,0,0,192,64                      ; mov           $0x40c00000,%ecx
  DB  102,68,15,110,249                   ; movd          %ecx,%xmm15
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  68,15,89,248                        ; mulps         %xmm0,%xmm15
  DB  185,171,170,42,63                   ; mov           $0x3f2aaaab,%ecx
  DB  102,15,110,225                      ; movd          %ecx,%xmm4
  DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
  DB  15,40,212                           ; movaps        %xmm4,%xmm2
  DB  15,92,211                           ; subps         %xmm3,%xmm2
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  15,40,195                           ; movaps        %xmm3,%xmm0
  DB  15,194,220,1                        ; cmpltps       %xmm4,%xmm3
  DB  65,15,89,215                        ; mulps         %xmm15,%xmm2
  DB  65,15,88,213                        ; addps         %xmm13,%xmm2
  DB  15,84,211                           ; andps         %xmm3,%xmm2
  DB  65,15,85,221                        ; andnps        %xmm13,%xmm3
  DB  15,86,218                           ; orps          %xmm2,%xmm3
  DB  65,15,194,198,1                     ; cmpltps       %xmm14,%xmm0
  DB  65,15,40,209                        ; movaps        %xmm9,%xmm2
  DB  15,84,208                           ; andps         %xmm0,%xmm2
  DB  15,85,195                           ; andnps        %xmm3,%xmm0
  DB  15,86,194                           ; orps          %xmm2,%xmm0
  DB  102,15,110,248                      ; movd          %eax,%xmm7
  DB  15,198,255,0                        ; shufps        $0x0,%xmm7,%xmm7
  DB  15,194,207,1                        ; cmpltps       %xmm7,%xmm1
  DB  69,15,89,199                        ; mulps         %xmm15,%xmm8
  DB  69,15,88,197                        ; addps         %xmm13,%xmm8
  DB  68,15,84,193                        ; andps         %xmm1,%xmm8
  DB  15,85,200                           ; andnps        %xmm0,%xmm1
  DB  65,15,86,200                        ; orps          %xmm8,%xmm1
  DB  69,15,40,195                        ; movaps        %xmm11,%xmm8
  DB  68,15,85,193                        ; andnps        %xmm1,%xmm8
  DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
  DB  15,194,198,1                        ; cmpltps       %xmm6,%xmm0
  DB  15,40,206                           ; movaps        %xmm6,%xmm1
  DB  65,15,92,204                        ; subps         %xmm12,%xmm1
  DB  15,84,200                           ; andps         %xmm0,%xmm1
  DB  15,85,198                           ; andnps        %xmm6,%xmm0
  DB  15,86,193                           ; orps          %xmm1,%xmm0
  DB  15,40,206                           ; movaps        %xmm6,%xmm1
  DB  15,194,205,1                        ; cmpltps       %xmm5,%xmm1
  DB  65,15,40,212                        ; movaps        %xmm12,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  15,84,209                           ; andps         %xmm1,%xmm2
  DB  15,85,200                           ; andnps        %xmm0,%xmm1
  DB  15,86,202                           ; orps          %xmm2,%xmm1
  DB  15,40,196                           ; movaps        %xmm4,%xmm0
  DB  15,92,193                           ; subps         %xmm1,%xmm0
  DB  15,40,217                           ; movaps        %xmm1,%xmm3
  DB  15,40,209                           ; movaps        %xmm1,%xmm2
  DB  15,194,204,1                        ; cmpltps       %xmm4,%xmm1
  DB  65,15,89,199                        ; mulps         %xmm15,%xmm0
  DB  65,15,88,197                        ; addps         %xmm13,%xmm0
  DB  15,84,193                           ; andps         %xmm1,%xmm0
  DB  65,15,85,205                        ; andnps        %xmm13,%xmm1
  DB  15,86,200                           ; orps          %xmm0,%xmm1
  DB  65,15,194,214,1                     ; cmpltps       %xmm14,%xmm2
  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
  DB  15,84,194                           ; andps         %xmm2,%xmm0
  DB  15,85,209                           ; andnps        %xmm1,%xmm2
  DB  15,86,208                           ; orps          %xmm0,%xmm2
  DB  15,194,223,1                        ; cmpltps       %xmm7,%xmm3
  DB  65,15,40,199                        ; movaps        %xmm15,%xmm0
  DB  15,89,198                           ; mulps         %xmm6,%xmm0
  DB  65,15,88,197                        ; addps         %xmm13,%xmm0
  DB  15,84,195                           ; andps         %xmm3,%xmm0
  DB  15,85,218                           ; andnps        %xmm2,%xmm3
  DB  15,86,216                           ; orps          %xmm0,%xmm3
  DB  65,15,40,203                        ; movaps        %xmm11,%xmm1
  DB  15,85,203                           ; andnps        %xmm3,%xmm1
  DB  15,92,52,36                         ; subps         (%rsp),%xmm6
  DB  15,40,198                           ; movaps        %xmm6,%xmm0
  DB  15,194,197,1                        ; cmpltps       %xmm5,%xmm0
  DB  15,40,214                           ; movaps        %xmm6,%xmm2
  DB  65,15,92,212                        ; subps         %xmm12,%xmm2
  DB  65,15,40,220                        ; movaps        %xmm12,%xmm3
  DB  68,15,194,230,1                     ; cmpltps       %xmm6,%xmm12
  DB  65,15,84,212                        ; andps         %xmm12,%xmm2
  DB  68,15,85,230                        ; andnps        %xmm6,%xmm12
  DB  68,15,86,226                        ; orps          %xmm2,%xmm12
  DB  15,88,222                           ; addps         %xmm6,%xmm3
  DB  15,84,216                           ; andps         %xmm0,%xmm3
  DB  65,15,85,196                        ; andnps        %xmm12,%xmm0
  DB  15,86,195                           ; orps          %xmm3,%xmm0
  DB  15,40,232                           ; movaps        %xmm0,%xmm5
  DB  15,194,239,1                        ; cmpltps       %xmm7,%xmm5
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  15,194,212,1                        ; cmpltps       %xmm4,%xmm2
  DB  15,92,224                           ; subps         %xmm0,%xmm4
  DB  65,15,194,198,1                     ; cmpltps       %xmm14,%xmm0
  DB  65,15,89,247                        ; mulps         %xmm15,%xmm6
  DB  65,15,89,231                        ; mulps         %xmm15,%xmm4
  DB  65,15,88,245                        ; addps         %xmm13,%xmm6
  DB  65,15,88,229                        ; addps         %xmm13,%xmm4
  DB  15,84,226                           ; andps         %xmm2,%xmm4
  DB  65,15,85,213                        ; andnps        %xmm13,%xmm2
  DB  15,86,212                           ; orps          %xmm4,%xmm2
  DB  68,15,84,200                        ; andps         %xmm0,%xmm9
  DB  15,85,194                           ; andnps        %xmm2,%xmm0
  DB  65,15,86,193                        ; orps          %xmm9,%xmm0
  DB  15,84,245                           ; andps         %xmm5,%xmm6
  DB  15,85,232                           ; andnps        %xmm0,%xmm5
  DB  15,86,238                           ; orps          %xmm6,%xmm5
  DB  69,15,84,211                        ; andps         %xmm11,%xmm10
  DB  68,15,85,221                        ; andnps        %xmm5,%xmm11
  DB  69,15,86,194                        ; orps          %xmm10,%xmm8
  DB  65,15,86,202                        ; orps          %xmm10,%xmm1
  DB  69,15,86,211                        ; orps          %xmm11,%xmm10
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  15,40,92,36,16                      ; movaps        0x10(%rsp),%xmm3
  DB  15,40,100,36,32                     ; movaps        0x20(%rsp),%xmm4
  DB  15,40,108,36,48                     ; movaps        0x30(%rsp),%xmm5
  DB  15,40,116,36,64                     ; movaps        0x40(%rsp),%xmm6
  DB  15,40,124,36,80                     ; movaps        0x50(%rsp),%xmm7
  DB  72,131,196,104                      ; add           $0x68,%rsp
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_1_float_sse2
_sk_scale_1_float_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_scale_u8_sse2
_sk_scale_u8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,68,15,110,4,56                  ; movd          (%rax,%rdi,1),%xmm8
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,69,15,96,193                    ; punpcklbw     %xmm9,%xmm8
  DB  102,69,15,97,193                    ; punpcklwd     %xmm9,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_1_float_sse2
_sk_lerp_1_float_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,92,196                           ; subps         %xmm4,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,92,205                           ; subps         %xmm5,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,92,214                           ; subps         %xmm6,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  15,92,223                           ; subps         %xmm7,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_u8_sse2
_sk_lerp_u8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,68,15,110,4,56                  ; movd          (%rax,%rdi,1),%xmm8
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,69,15,96,193                    ; punpcklbw     %xmm9,%xmm8
  DB  102,69,15,97,193                    ; punpcklwd     %xmm9,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  15,92,196                           ; subps         %xmm4,%xmm0
  DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,92,205                           ; subps         %xmm5,%xmm1
  DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,92,214                           ; subps         %xmm6,%xmm2
  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  15,92,223                           ; subps         %xmm7,%xmm3
  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
  DB  15,88,223                           ; addps         %xmm7,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_lerp_565_sse2
_sk_lerp_565_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,68,15,126,4,120                 ; movq          (%rax,%rdi,2),%xmm8
  DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
  DB  102,68,15,97,195                    ; punpcklwd     %xmm3,%xmm8
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  68,15,91,203                        ; cvtdq2ps      %xmm3,%xmm9
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  102,68,15,110,208                   ; movd          %eax,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  68,15,91,203                        ; cvtdq2ps      %xmm3,%xmm9
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  68,15,91,195                        ; cvtdq2ps      %xmm3,%xmm8
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  15,92,196                           ; subps         %xmm4,%xmm0
  DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
  DB  15,88,196                           ; addps         %xmm4,%xmm0
  DB  15,92,205                           ; subps         %xmm5,%xmm1
  DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
  DB  15,88,205                           ; addps         %xmm5,%xmm1
  DB  15,92,214                           ; subps         %xmm6,%xmm2
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  15,88,214                           ; addps         %xmm6,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_tables_sse2
_sk_load_tables_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  243,68,15,111,4,185                 ; movdqu        (%rcx,%rdi,4),%xmm8
  DB  185,255,0,0,0                       ; mov           $0xff,%ecx
  DB  102,15,110,193                      ; movd          %ecx,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
  DB  102,65,15,114,209,8                 ; psrld         $0x8,%xmm9
  DB  102,68,15,219,200                   ; pand          %xmm0,%xmm9
  DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
  DB  102,65,15,114,210,16                ; psrld         $0x10,%xmm10
  DB  102,68,15,219,208                   ; pand          %xmm0,%xmm10
  DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
  DB  102,15,112,216,78                   ; pshufd        $0x4e,%xmm0,%xmm3
  DB  102,72,15,126,217                   ; movq          %xmm3,%rcx
  DB  65,137,201                          ; mov           %ecx,%r9d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,73,15,126,194                   ; movq          %xmm0,%r10
  DB  69,137,211                          ; mov           %r10d,%r11d
  DB  73,193,234,32                       ; shr           $0x20,%r10
  DB  243,67,15,16,28,144                 ; movss         (%r8,%r10,4),%xmm3
  DB  243,65,15,16,4,136                  ; movss         (%r8,%rcx,4),%xmm0
  DB  15,20,216                           ; unpcklps      %xmm0,%xmm3
  DB  243,67,15,16,4,152                  ; movss         (%r8,%r11,4),%xmm0
  DB  243,67,15,16,12,136                 ; movss         (%r8,%r9,4),%xmm1
  DB  15,20,193                           ; unpcklps      %xmm1,%xmm0
  DB  15,20,195                           ; unpcklps      %xmm3,%xmm0
  DB  76,139,64,16                        ; mov           0x10(%rax),%r8
  DB  102,65,15,112,201,78                ; pshufd        $0x4e,%xmm9,%xmm1
  DB  102,73,15,126,202                   ; movq          %xmm1,%r10
  DB  77,137,209                          ; mov           %r10,%r9
  DB  73,193,233,32                       ; shr           $0x20,%r9
  DB  102,76,15,126,201                   ; movq          %xmm9,%rcx
  DB  65,137,203                          ; mov           %ecx,%r11d
  DB  65,129,227,255,255,255,0            ; and           $0xffffff,%r11d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  65,129,226,255,255,255,0            ; and           $0xffffff,%r10d
  DB  243,65,15,16,28,8                   ; movss         (%r8,%rcx,1),%xmm3
  DB  243,67,15,16,12,136                 ; movss         (%r8,%r9,4),%xmm1
  DB  15,20,217                           ; unpcklps      %xmm1,%xmm3
  DB  243,67,15,16,12,152                 ; movss         (%r8,%r11,4),%xmm1
  DB  243,67,15,16,20,144                 ; movss         (%r8,%r10,4),%xmm2
  DB  15,20,202                           ; unpcklps      %xmm2,%xmm1
  DB  15,20,203                           ; unpcklps      %xmm3,%xmm1
  DB  76,139,72,24                        ; mov           0x18(%rax),%r9
  DB  102,65,15,112,210,78                ; pshufd        $0x4e,%xmm10,%xmm2
  DB  102,72,15,126,209                   ; movq          %xmm2,%rcx
  DB  68,15,183,193                       ; movzwl        %cx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,76,15,126,208                   ; movq          %xmm10,%rax
  DB  68,15,183,208                       ; movzwl        %ax,%r10d
  DB  72,193,232,30                       ; shr           $0x1e,%rax
  DB  243,69,15,16,12,1                   ; movss         (%r9,%rax,1),%xmm9
  DB  243,65,15,16,20,137                 ; movss         (%r9,%rcx,4),%xmm2
  DB  68,15,20,202                        ; unpcklps      %xmm2,%xmm9
  DB  243,67,15,16,20,145                 ; movss         (%r9,%r10,4),%xmm2
  DB  243,67,15,16,28,129                 ; movss         (%r9,%r8,4),%xmm3
  DB  15,20,211                           ; unpcklps      %xmm3,%xmm2
  DB  65,15,20,209                        ; unpcklps      %xmm9,%xmm2
  DB  102,65,15,114,208,24                ; psrld         $0x18,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_tables_u16_be_sse2
_sk_load_tables_u16_be_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  243,15,111,4,249                    ; movdqu        (%rcx,%rdi,8),%xmm0
  DB  243,15,111,76,249,16                ; movdqu        0x10(%rcx,%rdi,8),%xmm1
  DB  102,68,15,111,208                   ; movdqa        %xmm0,%xmm10
  DB  102,68,15,97,209                    ; punpcklwd     %xmm1,%xmm10
  DB  102,15,105,193                      ; punpckhwd     %xmm1,%xmm0
  DB  102,65,15,111,202                   ; movdqa        %xmm10,%xmm1
  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
  DB  102,68,15,105,208                   ; punpckhwd     %xmm0,%xmm10
  DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
  DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
  DB  102,65,15,97,192                    ; punpcklwd     %xmm8,%xmm0
  DB  185,255,0,0,0                       ; mov           $0xff,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  102,68,15,112,203,0                 ; pshufd        $0x0,%xmm3,%xmm9
  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
  DB  102,15,112,216,78                   ; pshufd        $0x4e,%xmm0,%xmm3
  DB  102,72,15,126,217                   ; movq          %xmm3,%rcx
  DB  68,15,183,201                       ; movzwl        %cx,%r9d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,73,15,126,194                   ; movq          %xmm0,%r10
  DB  69,15,183,218                       ; movzwl        %r10w,%r11d
  DB  73,193,234,30                       ; shr           $0x1e,%r10
  DB  243,67,15,16,28,16                  ; movss         (%r8,%r10,1),%xmm3
  DB  243,65,15,16,4,136                  ; movss         (%r8,%rcx,4),%xmm0
  DB  15,20,216                           ; unpcklps      %xmm0,%xmm3
  DB  243,67,15,16,4,152                  ; movss         (%r8,%r11,4),%xmm0
  DB  243,67,15,16,20,136                 ; movss         (%r8,%r9,4),%xmm2
  DB  15,20,194                           ; unpcklps      %xmm2,%xmm0
  DB  15,20,195                           ; unpcklps      %xmm3,%xmm0
  DB  76,139,80,16                        ; mov           0x10(%rax),%r10
  DB  102,15,115,217,8                    ; psrldq        $0x8,%xmm1
  DB  102,65,15,97,200                    ; punpcklwd     %xmm8,%xmm1
  DB  102,65,15,219,201                   ; pand          %xmm9,%xmm1
  DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
  DB  102,73,15,126,208                   ; movq          %xmm2,%r8
  DB  69,15,183,200                       ; movzwl        %r8w,%r9d
  DB  73,193,232,32                       ; shr           $0x20,%r8
  DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
  DB  68,15,183,217                       ; movzwl        %cx,%r11d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  243,65,15,16,20,10                  ; movss         (%r10,%rcx,1),%xmm2
  DB  243,67,15,16,12,130                 ; movss         (%r10,%r8,4),%xmm1
  DB  15,20,209                           ; unpcklps      %xmm1,%xmm2
  DB  243,67,15,16,12,154                 ; movss         (%r10,%r11,4),%xmm1
  DB  243,67,15,16,28,138                 ; movss         (%r10,%r9,4),%xmm3
  DB  15,20,203                           ; unpcklps      %xmm3,%xmm1
  DB  15,20,202                           ; unpcklps      %xmm2,%xmm1
  DB  76,139,72,24                        ; mov           0x18(%rax),%r9
  DB  102,69,15,112,218,78                ; pshufd        $0x4e,%xmm10,%xmm11
  DB  102,69,15,97,208                    ; punpcklwd     %xmm8,%xmm10
  DB  102,69,15,219,209                   ; pand          %xmm9,%xmm10
  DB  102,65,15,112,210,78                ; pshufd        $0x4e,%xmm10,%xmm2
  DB  102,72,15,126,209                   ; movq          %xmm2,%rcx
  DB  68,15,183,193                       ; movzwl        %cx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,76,15,126,208                   ; movq          %xmm10,%rax
  DB  68,15,183,208                       ; movzwl        %ax,%r10d
  DB  72,193,232,30                       ; shr           $0x1e,%rax
  DB  243,69,15,16,12,1                   ; movss         (%r9,%rax,1),%xmm9
  DB  243,65,15,16,20,137                 ; movss         (%r9,%rcx,4),%xmm2
  DB  68,15,20,202                        ; unpcklps      %xmm2,%xmm9
  DB  243,67,15,16,20,145                 ; movss         (%r9,%r10,4),%xmm2
  DB  243,67,15,16,28,129                 ; movss         (%r9,%r8,4),%xmm3
  DB  15,20,211                           ; unpcklps      %xmm3,%xmm2
  DB  65,15,20,209                        ; unpcklps      %xmm9,%xmm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  102,68,15,110,200                   ; movd          %eax,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  102,65,15,111,219                   ; movdqa        %xmm11,%xmm3
  DB  102,15,113,243,8                    ; psllw         $0x8,%xmm3
  DB  102,65,15,113,211,8                 ; psrlw         $0x8,%xmm11
  DB  102,68,15,235,219                   ; por           %xmm3,%xmm11
  DB  102,69,15,97,216                    ; punpcklwd     %xmm8,%xmm11
  DB  65,15,91,219                        ; cvtdq2ps      %xmm11,%xmm3
  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_tables_rgb_u16_be_sse2
_sk_load_tables_rgb_u16_be_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,141,12,127                       ; lea           (%rdi,%rdi,2),%r9
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  76,139,64,8                         ; mov           0x8(%rax),%r8
  DB  243,70,15,111,28,73                 ; movdqu        (%rcx,%r9,2),%xmm11
  DB  243,66,15,111,68,73,8               ; movdqu        0x8(%rcx,%r9,2),%xmm0
  DB  102,15,115,216,4                    ; psrldq        $0x4,%xmm0
  DB  102,65,15,111,219                   ; movdqa        %xmm11,%xmm3
  DB  102,15,115,219,6                    ; psrldq        $0x6,%xmm3
  DB  102,68,15,97,216                    ; punpcklwd     %xmm0,%xmm11
  DB  102,15,115,216,6                    ; psrldq        $0x6,%xmm0
  DB  102,15,97,216                       ; punpcklwd     %xmm0,%xmm3
  DB  102,65,15,111,195                   ; movdqa        %xmm11,%xmm0
  DB  102,15,97,195                       ; punpcklwd     %xmm3,%xmm0
  DB  102,68,15,112,208,78                ; pshufd        $0x4e,%xmm0,%xmm10
  DB  102,68,15,105,219                   ; punpckhwd     %xmm3,%xmm11
  DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
  DB  102,65,15,97,192                    ; punpcklwd     %xmm8,%xmm0
  DB  185,255,0,0,0                       ; mov           $0xff,%ecx
  DB  102,15,110,217                      ; movd          %ecx,%xmm3
  DB  102,68,15,112,203,0                 ; pshufd        $0x0,%xmm3,%xmm9
  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
  DB  102,15,112,216,78                   ; pshufd        $0x4e,%xmm0,%xmm3
  DB  102,72,15,126,217                   ; movq          %xmm3,%rcx
  DB  68,15,183,201                       ; movzwl        %cx,%r9d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,73,15,126,194                   ; movq          %xmm0,%r10
  DB  69,15,183,218                       ; movzwl        %r10w,%r11d
  DB  73,193,234,30                       ; shr           $0x1e,%r10
  DB  243,67,15,16,28,16                  ; movss         (%r8,%r10,1),%xmm3
  DB  243,65,15,16,4,136                  ; movss         (%r8,%rcx,4),%xmm0
  DB  15,20,216                           ; unpcklps      %xmm0,%xmm3
  DB  243,67,15,16,4,152                  ; movss         (%r8,%r11,4),%xmm0
  DB  243,67,15,16,12,136                 ; movss         (%r8,%r9,4),%xmm1
  DB  15,20,193                           ; unpcklps      %xmm1,%xmm0
  DB  15,20,195                           ; unpcklps      %xmm3,%xmm0
  DB  76,139,80,16                        ; mov           0x10(%rax),%r10
  DB  102,69,15,97,208                    ; punpcklwd     %xmm8,%xmm10
  DB  102,69,15,219,209                   ; pand          %xmm9,%xmm10
  DB  102,65,15,112,202,78                ; pshufd        $0x4e,%xmm10,%xmm1
  DB  102,73,15,126,200                   ; movq          %xmm1,%r8
  DB  69,15,183,200                       ; movzwl        %r8w,%r9d
  DB  73,193,232,32                       ; shr           $0x20,%r8
  DB  102,76,15,126,209                   ; movq          %xmm10,%rcx
  DB  68,15,183,217                       ; movzwl        %cx,%r11d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  243,65,15,16,28,10                  ; movss         (%r10,%rcx,1),%xmm3
  DB  243,67,15,16,12,130                 ; movss         (%r10,%r8,4),%xmm1
  DB  15,20,217                           ; unpcklps      %xmm1,%xmm3
  DB  243,67,15,16,12,154                 ; movss         (%r10,%r11,4),%xmm1
  DB  243,67,15,16,20,138                 ; movss         (%r10,%r9,4),%xmm2
  DB  15,20,202                           ; unpcklps      %xmm2,%xmm1
  DB  15,20,203                           ; unpcklps      %xmm3,%xmm1
  DB  76,139,72,24                        ; mov           0x18(%rax),%r9
  DB  102,69,15,97,216                    ; punpcklwd     %xmm8,%xmm11
  DB  102,69,15,219,217                   ; pand          %xmm9,%xmm11
  DB  102,65,15,112,211,78                ; pshufd        $0x4e,%xmm11,%xmm2
  DB  102,72,15,126,209                   ; movq          %xmm2,%rcx
  DB  68,15,183,193                       ; movzwl        %cx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,76,15,126,216                   ; movq          %xmm11,%rax
  DB  68,15,183,208                       ; movzwl        %ax,%r10d
  DB  72,193,232,30                       ; shr           $0x1e,%rax
  DB  243,69,15,16,4,1                    ; movss         (%r9,%rax,1),%xmm8
  DB  243,65,15,16,20,137                 ; movss         (%r9,%rcx,4),%xmm2
  DB  68,15,20,194                        ; unpcklps      %xmm2,%xmm8
  DB  243,67,15,16,20,145                 ; movss         (%r9,%r10,4),%xmm2
  DB  243,67,15,16,28,129                 ; movss         (%r9,%r8,4),%xmm3
  DB  15,20,211                           ; unpcklps      %xmm3,%xmm2
  DB  65,15,20,208                        ; unpcklps      %xmm8,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_byte_tables_sse2
_sk_byte_tables_sse2 LABEL PROC
  DB  65,86                               ; push          %r14
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,127,67                      ; mov           $0x437f0000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,15,112,192,78                   ; pshufd        $0x4e,%xmm0,%xmm0
  DB  102,73,15,126,193                   ; movq          %xmm0,%r9
  DB  69,137,202                          ; mov           %r9d,%r10d
  DB  77,137,203                          ; mov           %r9,%r11
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  76,139,48                           ; mov           (%rax),%r14
  DB  76,139,72,8                         ; mov           0x8(%rax),%r9
  DB  71,15,182,20,22                     ; movzbl        (%r14,%r10,1),%r10d
  DB  67,15,182,28,30                     ; movzbl        (%r14,%r11,1),%ebx
  DB  193,227,8                           ; shl           $0x8,%ebx
  DB  68,9,211                            ; or            %r10d,%ebx
  DB  71,15,182,4,6                       ; movzbl        (%r14,%r8,1),%r8d
  DB  65,15,182,12,14                     ; movzbl        (%r14,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,193                            ; or            %r8d,%ecx
  DB  102,15,196,193,0                    ; pinsrw        $0x0,%ecx,%xmm0
  DB  102,15,196,195,1                    ; pinsrw        $0x1,%ebx,%xmm0
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,65,15,96,193                    ; punpcklbw     %xmm9,%xmm0
  DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  185,129,128,128,59                  ; mov           $0x3b808081,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
  DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,15,112,201,78                   ; pshufd        $0x4e,%xmm1,%xmm1
  DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  71,15,182,20,17                     ; movzbl        (%r9,%r10,1),%r10d
  DB  65,15,182,28,25                     ; movzbl        (%r9,%rbx,1),%ebx
  DB  193,227,8                           ; shl           $0x8,%ebx
  DB  68,9,211                            ; or            %r10d,%ebx
  DB  71,15,182,4,1                       ; movzbl        (%r9,%r8,1),%r8d
  DB  65,15,182,12,9                      ; movzbl        (%r9,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,193                            ; or            %r8d,%ecx
  DB  102,15,196,201,0                    ; pinsrw        $0x0,%ecx,%xmm1
  DB  102,15,196,203,1                    ; pinsrw        $0x1,%ebx,%xmm1
  DB  102,65,15,96,201                    ; punpcklbw     %xmm9,%xmm1
  DB  102,65,15,97,201                    ; punpcklwd     %xmm9,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,202                        ; mulps         %xmm10,%xmm1
  DB  76,139,72,16                        ; mov           0x10(%rax),%r9
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,91,210                       ; cvtps2dq      %xmm2,%xmm2
  DB  102,72,15,126,211                   ; movq          %xmm2,%rbx
  DB  65,137,216                          ; mov           %ebx,%r8d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  102,15,112,210,78                   ; pshufd        $0x4e,%xmm2,%xmm2
  DB  102,72,15,126,209                   ; movq          %xmm2,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  71,15,182,20,17                     ; movzbl        (%r9,%r10,1),%r10d
  DB  65,15,182,12,9                      ; movzbl        (%r9,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,209                            ; or            %r10d,%ecx
  DB  71,15,182,4,1                       ; movzbl        (%r9,%r8,1),%r8d
  DB  65,15,182,28,25                     ; movzbl        (%r9,%rbx,1),%ebx
  DB  193,227,8                           ; shl           $0x8,%ebx
  DB  68,9,195                            ; or            %r8d,%ebx
  DB  102,15,196,211,0                    ; pinsrw        $0x0,%ebx,%xmm2
  DB  102,15,196,209,1                    ; pinsrw        $0x1,%ecx,%xmm2
  DB  102,65,15,96,209                    ; punpcklbw     %xmm9,%xmm2
  DB  102,65,15,97,209                    ; punpcklwd     %xmm9,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,210                        ; mulps         %xmm10,%xmm2
  DB  72,139,64,24                        ; mov           0x18(%rax),%rax
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,65,15,91,216                    ; cvtps2dq      %xmm8,%xmm3
  DB  102,72,15,126,217                   ; movq          %xmm3,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,15,112,219,78                   ; pshufd        $0x4e,%xmm3,%xmm3
  DB  102,72,15,126,219                   ; movq          %xmm3,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  70,15,182,12,8                      ; movzbl        (%rax,%r9,1),%r9d
  DB  15,182,28,24                        ; movzbl        (%rax,%rbx,1),%ebx
  DB  193,227,8                           ; shl           $0x8,%ebx
  DB  68,9,203                            ; or            %r9d,%ebx
  DB  70,15,182,4,0                       ; movzbl        (%rax,%r8,1),%r8d
  DB  15,182,4,8                          ; movzbl        (%rax,%rcx,1),%eax
  DB  193,224,8                           ; shl           $0x8,%eax
  DB  68,9,192                            ; or            %r8d,%eax
  DB  102,15,196,216,0                    ; pinsrw        $0x0,%eax,%xmm3
  DB  102,15,196,219,1                    ; pinsrw        $0x1,%ebx,%xmm3
  DB  102,65,15,96,217                    ; punpcklbw     %xmm9,%xmm3
  DB  102,65,15,97,217                    ; punpcklwd     %xmm9,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,218                        ; mulps         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,94                               ; pop           %r14
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_byte_tables_rgb_sse2
_sk_byte_tables_rgb_sse2 LABEL PROC
  DB  65,86                               ; push          %r14
  DB  83                                  ; push          %rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  139,72,24                           ; mov           0x18(%rax),%ecx
  DB  255,201                             ; dec           %ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,15,112,192,78                   ; pshufd        $0x4e,%xmm0,%xmm0
  DB  102,73,15,126,193                   ; movq          %xmm0,%r9
  DB  69,137,202                          ; mov           %r9d,%r10d
  DB  77,137,203                          ; mov           %r9,%r11
  DB  73,193,235,32                       ; shr           $0x20,%r11
  DB  76,139,48                           ; mov           (%rax),%r14
  DB  76,139,72,8                         ; mov           0x8(%rax),%r9
  DB  71,15,182,20,22                     ; movzbl        (%r14,%r10,1),%r10d
  DB  67,15,182,28,30                     ; movzbl        (%r14,%r11,1),%ebx
  DB  193,227,8                           ; shl           $0x8,%ebx
  DB  68,9,211                            ; or            %r10d,%ebx
  DB  71,15,182,4,6                       ; movzbl        (%r14,%r8,1),%r8d
  DB  65,15,182,12,14                     ; movzbl        (%r14,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,193                            ; or            %r8d,%ecx
  DB  102,15,196,193,0                    ; pinsrw        $0x0,%ecx,%xmm0
  DB  102,15,196,195,1                    ; pinsrw        $0x1,%ebx,%xmm0
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,65,15,96,193                    ; punpcklbw     %xmm9,%xmm0
  DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  185,129,128,128,59                  ; mov           $0x3b808081,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
  DB  102,72,15,126,201                   ; movq          %xmm1,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,15,112,201,78                   ; pshufd        $0x4e,%xmm1,%xmm1
  DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
  DB  65,137,218                          ; mov           %ebx,%r10d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  71,15,182,20,17                     ; movzbl        (%r9,%r10,1),%r10d
  DB  65,15,182,28,25                     ; movzbl        (%r9,%rbx,1),%ebx
  DB  193,227,8                           ; shl           $0x8,%ebx
  DB  68,9,211                            ; or            %r10d,%ebx
  DB  71,15,182,4,1                       ; movzbl        (%r9,%r8,1),%r8d
  DB  65,15,182,12,9                      ; movzbl        (%r9,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,193                            ; or            %r8d,%ecx
  DB  102,15,196,201,0                    ; pinsrw        $0x0,%ecx,%xmm1
  DB  102,15,196,203,1                    ; pinsrw        $0x1,%ebx,%xmm1
  DB  102,65,15,96,201                    ; punpcklbw     %xmm9,%xmm1
  DB  102,65,15,97,201                    ; punpcklwd     %xmm9,%xmm1
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,202                        ; mulps         %xmm10,%xmm1
  DB  72,139,64,16                        ; mov           0x10(%rax),%rax
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,91,210                       ; cvtps2dq      %xmm2,%xmm2
  DB  102,72,15,126,209                   ; movq          %xmm2,%rcx
  DB  65,137,200                          ; mov           %ecx,%r8d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,15,112,210,78                   ; pshufd        $0x4e,%xmm2,%xmm2
  DB  102,72,15,126,211                   ; movq          %xmm2,%rbx
  DB  65,137,217                          ; mov           %ebx,%r9d
  DB  72,193,235,32                       ; shr           $0x20,%rbx
  DB  70,15,182,12,8                      ; movzbl        (%rax,%r9,1),%r9d
  DB  15,182,28,24                        ; movzbl        (%rax,%rbx,1),%ebx
  DB  193,227,8                           ; shl           $0x8,%ebx
  DB  68,9,203                            ; or            %r9d,%ebx
  DB  70,15,182,4,0                       ; movzbl        (%rax,%r8,1),%r8d
  DB  15,182,4,8                          ; movzbl        (%rax,%rcx,1),%eax
  DB  193,224,8                           ; shl           $0x8,%eax
  DB  68,9,192                            ; or            %r8d,%eax
  DB  102,15,196,208,0                    ; pinsrw        $0x0,%eax,%xmm2
  DB  102,15,196,211,1                    ; pinsrw        $0x1,%ebx,%xmm2
  DB  102,65,15,96,209                    ; punpcklbw     %xmm9,%xmm2
  DB  102,65,15,97,209                    ; punpcklwd     %xmm9,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,210                        ; mulps         %xmm10,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  91                                  ; pop           %rbx
  DB  65,94                               ; pop           %r14
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_r_sse2
_sk_table_r_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,112,192,78                ; pshufd        $0x4e,%xmm8,%xmm0
  DB  102,72,15,126,192                   ; movq          %xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,76,15,126,193                   ; movq          %xmm8,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,69,15,16,4,137                  ; movss         (%r9,%rcx,4),%xmm8
  DB  243,65,15,16,4,129                  ; movss         (%r9,%rax,4),%xmm0
  DB  68,15,20,192                        ; unpcklps      %xmm0,%xmm8
  DB  243,67,15,16,4,145                  ; movss         (%r9,%r10,4),%xmm0
  DB  243,71,15,16,12,129                 ; movss         (%r9,%r8,4),%xmm9
  DB  65,15,20,193                        ; unpcklps      %xmm9,%xmm0
  DB  65,15,20,192                        ; unpcklps      %xmm8,%xmm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_g_sse2
_sk_table_g_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,193                        ; mulps         %xmm1,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,112,200,78                ; pshufd        $0x4e,%xmm8,%xmm1
  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,76,15,126,193                   ; movq          %xmm8,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,69,15,16,4,137                  ; movss         (%r9,%rcx,4),%xmm8
  DB  243,65,15,16,12,129                 ; movss         (%r9,%rax,4),%xmm1
  DB  68,15,20,193                        ; unpcklps      %xmm1,%xmm8
  DB  243,67,15,16,12,145                 ; movss         (%r9,%r10,4),%xmm1
  DB  243,71,15,16,12,129                 ; movss         (%r9,%r8,4),%xmm9
  DB  65,15,20,201                        ; unpcklps      %xmm9,%xmm1
  DB  65,15,20,200                        ; unpcklps      %xmm8,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_b_sse2
_sk_table_b_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,112,208,78                ; pshufd        $0x4e,%xmm8,%xmm2
  DB  102,72,15,126,208                   ; movq          %xmm2,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,76,15,126,193                   ; movq          %xmm8,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,69,15,16,4,137                  ; movss         (%r9,%rcx,4),%xmm8
  DB  243,65,15,16,20,129                 ; movss         (%r9,%rax,4),%xmm2
  DB  68,15,20,194                        ; unpcklps      %xmm2,%xmm8
  DB  243,67,15,16,20,145                 ; movss         (%r9,%r10,4),%xmm2
  DB  243,71,15,16,12,129                 ; movss         (%r9,%r8,4),%xmm9
  DB  65,15,20,209                        ; unpcklps      %xmm9,%xmm2
  DB  65,15,20,208                        ; unpcklps      %xmm8,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_table_a_sse2
_sk_table_a_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  139,64,8                            ; mov           0x8(%rax),%eax
  DB  255,200                             ; dec           %eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
  DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,112,216,78                ; pshufd        $0x4e,%xmm8,%xmm3
  DB  102,72,15,126,216                   ; movq          %xmm3,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,76,15,126,193                   ; movq          %xmm8,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,69,15,16,4,137                  ; movss         (%r9,%rcx,4),%xmm8
  DB  243,65,15,16,28,129                 ; movss         (%r9,%rax,4),%xmm3
  DB  68,15,20,195                        ; unpcklps      %xmm3,%xmm8
  DB  243,67,15,16,28,145                 ; movss         (%r9,%r10,4),%xmm3
  DB  243,71,15,16,12,129                 ; movss         (%r9,%r8,4),%xmm9
  DB  65,15,20,217                        ; unpcklps      %xmm9,%xmm3
  DB  65,15,20,216                        ; unpcklps      %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_r_sse2
_sk_parametric_r_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,72,16                  ; movss         0x10(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,216                        ; mulps         %xmm0,%xmm11
  DB  65,15,194,193,2                     ; cmpleps       %xmm9,%xmm0
  DB  243,68,15,16,72,24                  ; movss         0x18(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  102,69,15,112,236,0                 ; pshufd        $0x0,%xmm12,%xmm13
  DB  102,69,15,219,235                   ; pand          %xmm11,%xmm13
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  102,69,15,112,227,0                 ; pshufd        $0x0,%xmm11,%xmm12
  DB  102,69,15,235,229                   ; por           %xmm13,%xmm12
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,220                        ; mulps         %xmm12,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,249                   ; movd          %ecx,%xmm15
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,88,244                        ; addps         %xmm12,%xmm14
  DB  69,15,94,238                        ; divps         %xmm14,%xmm13
  DB  69,15,92,213                        ; subps         %xmm13,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  243,69,15,91,202                    ; cvttps2dq     %xmm10,%xmm9
  DB  69,15,91,225                        ; cvtdq2ps      %xmm9,%xmm12
  DB  69,15,40,234                        ; movaps        %xmm10,%xmm13
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  69,15,88,250                        ; addps         %xmm10,%xmm15
  DB  69,15,194,212,1                     ; cmpltps       %xmm12,%xmm10
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,84,211                        ; andps         %xmm11,%xmm10
  DB  69,15,87,201                        ; xorps         %xmm9,%xmm9
  DB  69,15,92,226                        ; subps         %xmm10,%xmm12
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  102,69,15,110,208                   ; movd          %r8d,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
  DB  69,15,92,252                        ; subps         %xmm12,%xmm15
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  69,15,94,230                        ; divps         %xmm14,%xmm12
  DB  69,15,88,231                        ; addps         %xmm15,%xmm12
  DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
  DB  102,69,15,91,212                    ; cvtps2dq      %xmm12,%xmm10
  DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,226                        ; addps         %xmm10,%xmm12
  DB  68,15,84,192                        ; andps         %xmm0,%xmm8
  DB  65,15,85,196                        ; andnps        %xmm12,%xmm0
  DB  65,15,86,192                        ; orps          %xmm8,%xmm0
  DB  65,15,95,193                        ; maxps         %xmm9,%xmm0
  DB  65,15,93,195                        ; minps         %xmm11,%xmm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_g_sse2
_sk_parametric_g_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,72,16                  ; movss         0x10(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,89,193                        ; mulps         %xmm1,%xmm8
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
  DB  65,15,194,201,2                     ; cmpleps       %xmm9,%xmm1
  DB  243,68,15,16,72,24                  ; movss         0x18(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  102,69,15,112,236,0                 ; pshufd        $0x0,%xmm12,%xmm13
  DB  102,69,15,219,235                   ; pand          %xmm11,%xmm13
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  102,69,15,112,227,0                 ; pshufd        $0x0,%xmm11,%xmm12
  DB  102,69,15,235,229                   ; por           %xmm13,%xmm12
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,220                        ; mulps         %xmm12,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,249                   ; movd          %ecx,%xmm15
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,88,244                        ; addps         %xmm12,%xmm14
  DB  69,15,94,238                        ; divps         %xmm14,%xmm13
  DB  69,15,92,213                        ; subps         %xmm13,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  243,69,15,91,202                    ; cvttps2dq     %xmm10,%xmm9
  DB  69,15,91,225                        ; cvtdq2ps      %xmm9,%xmm12
  DB  69,15,40,234                        ; movaps        %xmm10,%xmm13
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  69,15,88,250                        ; addps         %xmm10,%xmm15
  DB  69,15,194,212,1                     ; cmpltps       %xmm12,%xmm10
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,84,211                        ; andps         %xmm11,%xmm10
  DB  69,15,87,201                        ; xorps         %xmm9,%xmm9
  DB  69,15,92,226                        ; subps         %xmm10,%xmm12
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  102,69,15,110,208                   ; movd          %r8d,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
  DB  69,15,92,252                        ; subps         %xmm12,%xmm15
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  69,15,94,230                        ; divps         %xmm14,%xmm12
  DB  69,15,88,231                        ; addps         %xmm15,%xmm12
  DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
  DB  102,69,15,91,212                    ; cvtps2dq      %xmm12,%xmm10
  DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,226                        ; addps         %xmm10,%xmm12
  DB  68,15,84,193                        ; andps         %xmm1,%xmm8
  DB  65,15,85,204                        ; andnps        %xmm12,%xmm1
  DB  65,15,86,200                        ; orps          %xmm8,%xmm1
  DB  65,15,95,201                        ; maxps         %xmm9,%xmm1
  DB  65,15,93,203                        ; minps         %xmm11,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_b_sse2
_sk_parametric_b_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,72,16                  ; movss         0x10(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  65,15,194,209,2                     ; cmpleps       %xmm9,%xmm2
  DB  243,68,15,16,72,24                  ; movss         0x18(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  102,69,15,112,236,0                 ; pshufd        $0x0,%xmm12,%xmm13
  DB  102,69,15,219,235                   ; pand          %xmm11,%xmm13
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  102,69,15,112,227,0                 ; pshufd        $0x0,%xmm11,%xmm12
  DB  102,69,15,235,229                   ; por           %xmm13,%xmm12
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,220                        ; mulps         %xmm12,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,249                   ; movd          %ecx,%xmm15
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,88,244                        ; addps         %xmm12,%xmm14
  DB  69,15,94,238                        ; divps         %xmm14,%xmm13
  DB  69,15,92,213                        ; subps         %xmm13,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  243,69,15,91,202                    ; cvttps2dq     %xmm10,%xmm9
  DB  69,15,91,225                        ; cvtdq2ps      %xmm9,%xmm12
  DB  69,15,40,234                        ; movaps        %xmm10,%xmm13
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  69,15,88,250                        ; addps         %xmm10,%xmm15
  DB  69,15,194,212,1                     ; cmpltps       %xmm12,%xmm10
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,84,211                        ; andps         %xmm11,%xmm10
  DB  69,15,87,201                        ; xorps         %xmm9,%xmm9
  DB  69,15,92,226                        ; subps         %xmm10,%xmm12
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  102,69,15,110,208                   ; movd          %r8d,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
  DB  69,15,92,252                        ; subps         %xmm12,%xmm15
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  69,15,94,230                        ; divps         %xmm14,%xmm12
  DB  69,15,88,231                        ; addps         %xmm15,%xmm12
  DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
  DB  102,69,15,91,212                    ; cvtps2dq      %xmm12,%xmm10
  DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,226                        ; addps         %xmm10,%xmm12
  DB  68,15,84,194                        ; andps         %xmm2,%xmm8
  DB  65,15,85,212                        ; andnps        %xmm12,%xmm2
  DB  65,15,86,208                        ; orps          %xmm8,%xmm2
  DB  65,15,95,209                        ; maxps         %xmm9,%xmm2
  DB  65,15,93,211                        ; minps         %xmm11,%xmm2
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_parametric_a_sse2
_sk_parametric_a_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,72,16                  ; movss         0x10(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  243,68,15,16,88,4                   ; movss         0x4(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,219                        ; mulps         %xmm3,%xmm11
  DB  65,15,194,217,2                     ; cmpleps       %xmm9,%xmm3
  DB  243,68,15,16,72,24                  ; movss         0x18(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,218                        ; addps         %xmm10,%xmm11
  DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
  DB  185,0,0,0,52                        ; mov           $0x34000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
  DB  185,255,255,127,0                   ; mov           $0x7fffff,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  102,69,15,112,236,0                 ; pshufd        $0x0,%xmm12,%xmm13
  DB  102,69,15,219,235                   ; pand          %xmm11,%xmm13
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  102,69,15,112,227,0                 ; pshufd        $0x0,%xmm11,%xmm12
  DB  102,69,15,235,229                   ; por           %xmm13,%xmm12
  DB  185,119,115,248,66                  ; mov           $0x42f87377,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,117,191,191,63                  ; mov           $0x3fbfbf75,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,220                        ; mulps         %xmm12,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  185,163,233,220,63                  ; mov           $0x3fdce9a3,%ecx
  DB  102,68,15,110,233                   ; movd          %ecx,%xmm13
  DB  185,249,68,180,62                   ; mov           $0x3eb444f9,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  65,184,0,0,0,75                     ; mov           $0x4b000000,%r8d
  DB  185,81,140,242,66                   ; mov           $0x42f28c51,%ecx
  DB  102,68,15,110,249                   ; movd          %ecx,%xmm15
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,88,244                        ; addps         %xmm12,%xmm14
  DB  69,15,94,238                        ; divps         %xmm14,%xmm13
  DB  69,15,92,213                        ; subps         %xmm13,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  243,69,15,91,202                    ; cvttps2dq     %xmm10,%xmm9
  DB  69,15,91,225                        ; cvtdq2ps      %xmm9,%xmm12
  DB  69,15,40,234                        ; movaps        %xmm10,%xmm13
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  69,15,88,250                        ; addps         %xmm10,%xmm15
  DB  69,15,194,212,1                     ; cmpltps       %xmm12,%xmm10
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,84,211                        ; andps         %xmm11,%xmm10
  DB  69,15,87,201                        ; xorps         %xmm9,%xmm9
  DB  69,15,92,226                        ; subps         %xmm10,%xmm12
  DB  69,15,92,236                        ; subps         %xmm12,%xmm13
  DB  102,69,15,110,208                   ; movd          %r8d,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,141,188,190,63                  ; mov           $0x3fbebc8d,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
  DB  69,15,92,252                        ; subps         %xmm12,%xmm15
  DB  185,254,210,221,65                  ; mov           $0x41ddd2fe,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  185,248,245,154,64                  ; mov           $0x409af5f8,%ecx
  DB  102,68,15,110,241                   ; movd          %ecx,%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,92,245                        ; subps         %xmm13,%xmm14
  DB  69,15,94,230                        ; divps         %xmm14,%xmm12
  DB  69,15,88,231                        ; addps         %xmm15,%xmm12
  DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
  DB  102,69,15,91,212                    ; cvtps2dq      %xmm12,%xmm10
  DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,88,226                        ; addps         %xmm10,%xmm12
  DB  68,15,84,195                        ; andps         %xmm3,%xmm8
  DB  65,15,85,220                        ; andnps        %xmm12,%xmm3
  DB  65,15,86,216                        ; orps          %xmm8,%xmm3
  DB  65,15,95,217                        ; maxps         %xmm9,%xmm3
  DB  65,15,93,219                        ; minps         %xmm11,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_a8_sse2
_sk_load_a8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,15,110,4,56                     ; movd          (%rax,%rdi,1),%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_a8_sse2
_sk_gather_a8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,112,217,245                  ; pshufd        $0xf5,%xmm1,%xmm3
  DB  102,15,244,218                      ; pmuludq       %xmm2,%xmm3
  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
  DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
  DB  102,15,112,202,232                  ; pshufd        $0xe8,%xmm2,%xmm1
  DB  102,15,98,203                       ; punpckldq     %xmm3,%xmm1
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,193                      ; paddd         %xmm1,%xmm0
  DB  102,72,15,126,192                   ; movq          %xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,15,112,192,78                   ; pshufd        $0x4e,%xmm0,%xmm0
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  71,15,182,20,17                     ; movzbl        (%r9,%r10,1),%r10d
  DB  65,15,182,12,9                      ; movzbl        (%r9,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,209                            ; or            %r10d,%ecx
  DB  71,15,182,4,1                       ; movzbl        (%r9,%r8,1),%r8d
  DB  65,15,182,4,1                       ; movzbl        (%r9,%rax,1),%eax
  DB  193,224,8                           ; shl           $0x8,%eax
  DB  68,9,192                            ; or            %r8d,%eax
  DB  102,15,196,192,0                    ; pinsrw        $0x0,%eax,%xmm0
  DB  102,15,196,193,1                    ; pinsrw        $0x1,%ecx,%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_a8_sse2
_sk_store_a8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,127,67                      ; mov           $0x437f0000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
  DB  102,65,15,114,224,16                ; psrad         $0x10,%xmm8
  DB  102,69,15,107,192                   ; packssdw      %xmm8,%xmm8
  DB  102,69,15,103,192                   ; packuswb      %xmm8,%xmm8
  DB  102,68,15,126,4,56                  ; movd          %xmm8,(%rax,%rdi,1)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_g8_sse2
_sk_load_g8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  102,15,110,4,56                     ; movd          (%rax,%rdi,1),%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,200                           ; movaps        %xmm0,%xmm1
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_g8_sse2
_sk_gather_g8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,112,217,245                  ; pshufd        $0xf5,%xmm1,%xmm3
  DB  102,15,244,218                      ; pmuludq       %xmm2,%xmm3
  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
  DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
  DB  102,15,112,202,232                  ; pshufd        $0xe8,%xmm2,%xmm1
  DB  102,15,98,203                       ; punpckldq     %xmm3,%xmm1
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,193                      ; paddd         %xmm1,%xmm0
  DB  102,72,15,126,192                   ; movq          %xmm0,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,15,112,192,78                   ; pshufd        $0x4e,%xmm0,%xmm0
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  71,15,182,20,17                     ; movzbl        (%r9,%r10,1),%r10d
  DB  65,15,182,12,9                      ; movzbl        (%r9,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,209                            ; or            %r10d,%ecx
  DB  71,15,182,4,1                       ; movzbl        (%r9,%r8,1),%r8d
  DB  65,15,182,4,1                       ; movzbl        (%r9,%rax,1),%eax
  DB  193,224,8                           ; shl           $0x8,%eax
  DB  68,9,192                            ; or            %r8d,%eax
  DB  102,15,196,192,0                    ; pinsrw        $0x0,%eax,%xmm0
  DB  102,15,196,193,1                    ; pinsrw        $0x1,%ecx,%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,40,200                           ; movaps        %xmm0,%xmm1
  DB  15,40,208                           ; movaps        %xmm0,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_i8_sse2
_sk_gather_i8_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  73,137,192                          ; mov           %rax,%r8
  DB  77,133,192                          ; test          %r8,%r8
  DB  116,5                               ; je            295d <_sk_gather_i8_sse2+0xf>
  DB  76,137,192                          ; mov           %r8,%rax
  DB  235,2                               ; jmp           295f <_sk_gather_i8_sse2+0x11>
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,16                           ; mov           (%rax),%r10
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,112,217,245                  ; pshufd        $0xf5,%xmm1,%xmm3
  DB  102,15,244,218                      ; pmuludq       %xmm2,%xmm3
  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
  DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
  DB  102,15,112,202,232                  ; pshufd        $0xe8,%xmm2,%xmm1
  DB  102,15,98,203                       ; punpckldq     %xmm3,%xmm1
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,193                      ; paddd         %xmm1,%xmm0
  DB  102,72,15,126,192                   ; movq          %xmm0,%rax
  DB  65,137,193                          ; mov           %eax,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,15,112,192,78                   ; pshufd        $0x4e,%xmm0,%xmm0
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,203                          ; mov           %ecx,%r11d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  71,15,182,28,26                     ; movzbl        (%r10,%r11,1),%r11d
  DB  65,15,182,12,10                     ; movzbl        (%r10,%rcx,1),%ecx
  DB  193,225,8                           ; shl           $0x8,%ecx
  DB  68,9,217                            ; or            %r11d,%ecx
  DB  71,15,182,12,10                     ; movzbl        (%r10,%r9,1),%r9d
  DB  65,15,182,4,2                       ; movzbl        (%r10,%rax,1),%eax
  DB  193,224,8                           ; shl           $0x8,%eax
  DB  68,9,200                            ; or            %r9d,%eax
  DB  102,15,196,192,0                    ; pinsrw        $0x0,%eax,%xmm0
  DB  102,15,196,193,1                    ; pinsrw        $0x1,%ecx,%xmm0
  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
  DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
  DB  68,15,182,200                       ; movzbl        %al,%r9d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  77,139,64,8                         ; mov           0x8(%r8),%r8
  DB  68,15,182,209                       ; movzbl        %cl,%r10d
  DB  72,193,233,30                       ; shr           $0x1e,%rcx
  DB  102,65,15,110,4,8                   ; movd          (%r8,%rcx,1),%xmm0
  DB  102,65,15,110,12,128                ; movd          (%r8,%rax,4),%xmm1
  DB  102,15,98,193                       ; punpckldq     %xmm1,%xmm0
  DB  102,67,15,110,28,144                ; movd          (%r8,%r10,4),%xmm3
  DB  102,67,15,110,12,136                ; movd          (%r8,%r9,4),%xmm1
  DB  102,15,98,217                       ; punpckldq     %xmm1,%xmm3
  DB  102,15,98,216                       ; punpckldq     %xmm0,%xmm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
  DB  102,15,219,200                      ; pand          %xmm0,%xmm1
  DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
  DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_565_sse2
_sk_load_565_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,15,126,20,120                   ; movq          (%rax,%rdi,2),%xmm2
  DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
  DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,219,194                      ; pand          %xmm2,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,15,219,202                      ; pand          %xmm2,%xmm1
  DB  15,91,217                           ; cvtdq2ps      %xmm1,%xmm3
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,203                           ; mulps         %xmm3,%xmm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,15,219,218                      ; pand          %xmm2,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_565_sse2
_sk_gather_565_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,112,217,245                  ; pshufd        $0xf5,%xmm1,%xmm3
  DB  102,15,244,218                      ; pmuludq       %xmm2,%xmm3
  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
  DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
  DB  102,15,112,202,232                  ; pshufd        $0xe8,%xmm2,%xmm1
  DB  102,15,98,203                       ; punpckldq     %xmm3,%xmm1
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,193                      ; paddd         %xmm1,%xmm0
  DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,67,15,196,20,81,0               ; pinsrw        $0x0,(%r9,%r10,2),%xmm2
  DB  102,65,15,196,20,73,1               ; pinsrw        $0x1,(%r9,%rcx,2),%xmm2
  DB  67,15,183,12,65                     ; movzwl        (%r9,%r8,2),%ecx
  DB  102,15,196,209,2                    ; pinsrw        $0x2,%ecx,%xmm2
  DB  65,15,183,4,65                      ; movzwl        (%r9,%rax,2),%eax
  DB  102,15,196,208,3                    ; pinsrw        $0x3,%eax,%xmm2
  DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
  DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
  DB  184,0,248,0,0                       ; mov           $0xf800,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,219,194                      ; pand          %xmm2,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,8,33,132,55                     ; mov           $0x37842108,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,224,7,0,0                       ; mov           $0x7e0,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,15,219,202                      ; pand          %xmm2,%xmm1
  DB  15,91,217                           ; cvtdq2ps      %xmm1,%xmm3
  DB  184,33,8,2,58                       ; mov           $0x3a020821,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,203                           ; mulps         %xmm3,%xmm1
  DB  184,31,0,0,0                        ; mov           $0x1f,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,15,219,218                      ; pand          %xmm2,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  184,8,33,4,61                       ; mov           $0x3d042108,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  15,89,211                           ; mulps         %xmm3,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_565_sse2
_sk_store_565_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,248,65                      ; mov           $0x41f80000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,11                ; pslld         $0xb,%xmm9
  DB  185,0,0,124,66                      ; mov           $0x427c0000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,65,15,114,242,5                 ; pslld         $0x5,%xmm10
  DB  102,69,15,235,209                   ; por           %xmm9,%xmm10
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,69,15,86,194                    ; orpd          %xmm10,%xmm8
  DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
  DB  102,65,15,114,224,16                ; psrad         $0x10,%xmm8
  DB  102,69,15,107,192                   ; packssdw      %xmm8,%xmm8
  DB  102,68,15,214,4,120                 ; movq          %xmm8,(%rax,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_4444_sse2
_sk_load_4444_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,68,15,126,12,120                ; movq          (%rax,%rdi,2),%xmm9
  DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
  DB  102,68,15,97,200                    ; punpcklwd     %xmm0,%xmm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,65,15,219,201                   ; pand          %xmm9,%xmm1
  DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,202                           ; mulps         %xmm2,%xmm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
  DB  68,15,91,194                        ; cvtdq2ps      %xmm2,%xmm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,217                   ; pand          %xmm9,%xmm3
  DB  68,15,91,195                        ; cvtdq2ps      %xmm3,%xmm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_4444_sse2
_sk_gather_4444_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,112,217,245                  ; pshufd        $0xf5,%xmm1,%xmm3
  DB  102,15,244,218                      ; pmuludq       %xmm2,%xmm3
  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
  DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
  DB  102,15,112,202,232                  ; pshufd        $0xe8,%xmm2,%xmm1
  DB  102,15,98,203                       ; punpckldq     %xmm3,%xmm1
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,193                      ; paddd         %xmm1,%xmm0
  DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,71,15,196,12,81,0               ; pinsrw        $0x0,(%r9,%r10,2),%xmm9
  DB  102,69,15,196,12,73,1               ; pinsrw        $0x1,(%r9,%rcx,2),%xmm9
  DB  67,15,183,12,65                     ; movzwl        (%r9,%r8,2),%ecx
  DB  102,68,15,196,201,2                 ; pinsrw        $0x2,%ecx,%xmm9
  DB  65,15,183,4,65                      ; movzwl        (%r9,%rax,2),%eax
  DB  102,68,15,196,200,3                 ; pinsrw        $0x3,%eax,%xmm9
  DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
  DB  102,68,15,97,200                    ; punpcklwd     %xmm0,%xmm9
  DB  184,0,240,0,0                       ; mov           $0xf000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
  DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
  DB  184,137,136,136,55                  ; mov           $0x37888889,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  184,0,15,0,0                        ; mov           $0xf00,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
  DB  102,65,15,219,201                   ; pand          %xmm9,%xmm1
  DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
  DB  184,137,136,136,57                  ; mov           $0x39888889,%eax
  DB  102,15,110,200                      ; movd          %eax,%xmm1
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  15,89,202                           ; mulps         %xmm2,%xmm1
  DB  184,240,0,0,0                       ; mov           $0xf0,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
  DB  68,15,91,194                        ; cvtdq2ps      %xmm2,%xmm8
  DB  184,137,136,136,59                  ; mov           $0x3b888889,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  184,15,0,0,0                        ; mov           $0xf,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
  DB  102,65,15,219,217                   ; pand          %xmm9,%xmm3
  DB  68,15,91,195                        ; cvtdq2ps      %xmm3,%xmm8
  DB  184,137,136,136,61                  ; mov           $0x3d888889,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_4444_sse2
_sk_store_4444_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,112,65                      ; mov           $0x41700000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,12                ; pslld         $0xc,%xmm9
  DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,65,15,114,242,8                 ; pslld         $0x8,%xmm10
  DB  102,69,15,235,209                   ; por           %xmm9,%xmm10
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,4                 ; pslld         $0x4,%xmm9
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,69,15,86,193                    ; orpd          %xmm9,%xmm8
  DB  102,69,15,86,194                    ; orpd          %xmm10,%xmm8
  DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
  DB  102,65,15,114,224,16                ; psrad         $0x10,%xmm8
  DB  102,69,15,107,192                   ; packssdw      %xmm8,%xmm8
  DB  102,68,15,214,4,120                 ; movq          %xmm8,(%rax,%rdi,2)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_8888_sse2
_sk_load_8888_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,15,111,28,184                   ; movdqu        (%rax,%rdi,4),%xmm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
  DB  102,15,219,200                      ; pand          %xmm0,%xmm1
  DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
  DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_8888_sse2
_sk_gather_8888_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,112,217,245                  ; pshufd        $0xf5,%xmm1,%xmm3
  DB  102,15,244,218                      ; pmuludq       %xmm2,%xmm3
  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
  DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
  DB  102,15,112,202,232                  ; pshufd        $0xe8,%xmm2,%xmm1
  DB  102,15,98,203                       ; punpckldq     %xmm3,%xmm1
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,193                      ; paddd         %xmm1,%xmm0
  DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  102,65,15,110,4,137                 ; movd          (%r9,%rcx,4),%xmm0
  DB  102,65,15,110,12,129                ; movd          (%r9,%rax,4),%xmm1
  DB  102,15,98,193                       ; punpckldq     %xmm1,%xmm0
  DB  102,67,15,110,28,145                ; movd          (%r9,%r10,4),%xmm3
  DB  102,67,15,110,12,129                ; movd          (%r9,%r8,4),%xmm1
  DB  102,15,98,217                       ; punpckldq     %xmm1,%xmm3
  DB  102,15,98,216                       ; punpckldq     %xmm0,%xmm3
  DB  184,255,0,0,0                       ; mov           $0xff,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,192,0                    ; pshufd        $0x0,%xmm0,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
  DB  102,15,219,200                      ; pand          %xmm0,%xmm1
  DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
  DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  184,129,128,128,59                  ; mov           $0x3b808081,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_8888_sse2
_sk_store_8888_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,127,67                      ; mov           $0x437f0000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,65,15,114,242,8                 ; pslld         $0x8,%xmm10
  DB  102,69,15,235,209                   ; por           %xmm9,%xmm10
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,89,202                        ; mulps         %xmm2,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,16                ; pslld         $0x10,%xmm9
  DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,114,240,24                ; pslld         $0x18,%xmm8
  DB  102,69,15,235,193                   ; por           %xmm9,%xmm8
  DB  102,69,15,235,194                   ; por           %xmm10,%xmm8
  DB  243,68,15,127,4,184                 ; movdqu        %xmm8,(%rax,%rdi,4)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_f16_sse2
_sk_load_f16_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,15,111,4,248                    ; movdqu        (%rax,%rdi,8),%xmm0
  DB  243,15,111,76,248,16                ; movdqu        0x10(%rax,%rdi,8),%xmm1
  DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
  DB  102,68,15,97,193                    ; punpcklwd     %xmm1,%xmm8
  DB  102,15,105,193                      ; punpckhwd     %xmm1,%xmm0
  DB  102,69,15,111,224                   ; movdqa        %xmm8,%xmm12
  DB  102,68,15,97,224                    ; punpcklwd     %xmm0,%xmm12
  DB  102,68,15,105,192                   ; punpckhwd     %xmm0,%xmm8
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,69,15,111,236                   ; movdqa        %xmm12,%xmm13
  DB  102,69,15,97,233                    ; punpcklwd     %xmm9,%xmm13
  DB  184,0,128,0,0                       ; mov           $0x8000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,68,15,112,208,0                 ; pshufd        $0x0,%xmm0,%xmm10
  DB  102,65,15,111,205                   ; movdqa        %xmm13,%xmm1
  DB  102,65,15,219,202                   ; pand          %xmm10,%xmm1
  DB  184,0,124,0,0                       ; mov           $0x7c00,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,216,0                    ; pshufd        $0x0,%xmm0,%xmm3
  DB  102,65,15,111,197                   ; movdqa        %xmm13,%xmm0
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  102,68,15,239,233                   ; pxor          %xmm1,%xmm13
  DB  102,15,114,241,16                   ; pslld         $0x10,%xmm1
  DB  102,65,15,114,245,13                ; pslld         $0xd,%xmm13
  DB  184,0,0,0,56                        ; mov           $0x38000000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  102,68,15,112,218,0                 ; pshufd        $0x0,%xmm2,%xmm11
  DB  102,65,15,254,203                   ; paddd         %xmm11,%xmm1
  DB  102,65,15,254,205                   ; paddd         %xmm13,%xmm1
  DB  102,65,15,118,193                   ; pcmpeqd       %xmm9,%xmm0
  DB  102,15,223,193                      ; pandn         %xmm1,%xmm0
  DB  102,65,15,115,220,8                 ; psrldq        $0x8,%xmm12
  DB  102,69,15,97,225                    ; punpcklwd     %xmm9,%xmm12
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,65,15,219,210                   ; pand          %xmm10,%xmm2
  DB  102,65,15,111,204                   ; movdqa        %xmm12,%xmm1
  DB  102,15,219,203                      ; pand          %xmm3,%xmm1
  DB  102,68,15,239,226                   ; pxor          %xmm2,%xmm12
  DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
  DB  102,65,15,254,212                   ; paddd         %xmm12,%xmm2
  DB  102,65,15,118,201                   ; pcmpeqd       %xmm9,%xmm1
  DB  102,15,223,202                      ; pandn         %xmm2,%xmm1
  DB  102,69,15,111,224                   ; movdqa        %xmm8,%xmm12
  DB  102,69,15,97,225                    ; punpcklwd     %xmm9,%xmm12
  DB  102,69,15,111,236                   ; movdqa        %xmm12,%xmm13
  DB  102,69,15,219,234                   ; pand          %xmm10,%xmm13
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,15,219,211                      ; pand          %xmm3,%xmm2
  DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
  DB  102,65,15,114,245,16                ; pslld         $0x10,%xmm13
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
  DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
  DB  102,65,15,118,209                   ; pcmpeqd       %xmm9,%xmm2
  DB  102,65,15,223,213                   ; pandn         %xmm13,%xmm2
  DB  102,65,15,115,216,8                 ; psrldq        $0x8,%xmm8
  DB  102,69,15,97,193                    ; punpcklwd     %xmm9,%xmm8
  DB  102,69,15,219,208                   ; pand          %xmm8,%xmm10
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  102,69,15,239,194                   ; pxor          %xmm10,%xmm8
  DB  102,65,15,114,242,16                ; pslld         $0x10,%xmm10
  DB  102,65,15,114,240,13                ; pslld         $0xd,%xmm8
  DB  102,69,15,254,211                   ; paddd         %xmm11,%xmm10
  DB  102,69,15,254,208                   ; paddd         %xmm8,%xmm10
  DB  102,65,15,118,217                   ; pcmpeqd       %xmm9,%xmm3
  DB  102,65,15,223,218                   ; pandn         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_gather_f16_sse2
_sk_gather_f16_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  76,139,8                            ; mov           (%rax),%r9
  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
  DB  102,15,110,80,16                    ; movd          0x10(%rax),%xmm2
  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
  DB  102,15,112,217,245                  ; pshufd        $0xf5,%xmm1,%xmm3
  DB  102,15,244,218                      ; pmuludq       %xmm2,%xmm3
  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
  DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
  DB  102,15,112,202,232                  ; pshufd        $0xe8,%xmm2,%xmm1
  DB  102,15,98,203                       ; punpckldq     %xmm3,%xmm1
  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
  DB  102,15,254,193                      ; paddd         %xmm1,%xmm0
  DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
  DB  65,137,192                          ; mov           %eax,%r8d
  DB  72,193,232,32                       ; shr           $0x20,%rax
  DB  102,72,15,126,193                   ; movq          %xmm0,%rcx
  DB  65,137,202                          ; mov           %ecx,%r10d
  DB  72,193,233,32                       ; shr           $0x20,%rcx
  DB  243,65,15,126,4,201                 ; movq          (%r9,%rcx,8),%xmm0
  DB  243,67,15,126,12,209                ; movq          (%r9,%r10,8),%xmm1
  DB  102,15,108,200                      ; punpcklqdq    %xmm0,%xmm1
  DB  243,65,15,126,4,193                 ; movq          (%r9,%rax,8),%xmm0
  DB  243,67,15,126,20,193                ; movq          (%r9,%r8,8),%xmm2
  DB  102,15,108,208                      ; punpcklqdq    %xmm0,%xmm2
  DB  102,68,15,111,193                   ; movdqa        %xmm1,%xmm8
  DB  102,68,15,97,194                    ; punpcklwd     %xmm2,%xmm8
  DB  102,15,105,202                      ; punpckhwd     %xmm2,%xmm1
  DB  102,69,15,111,224                   ; movdqa        %xmm8,%xmm12
  DB  102,68,15,97,225                    ; punpcklwd     %xmm1,%xmm12
  DB  102,68,15,105,193                   ; punpckhwd     %xmm1,%xmm8
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,69,15,111,236                   ; movdqa        %xmm12,%xmm13
  DB  102,69,15,97,233                    ; punpcklwd     %xmm9,%xmm13
  DB  184,0,128,0,0                       ; mov           $0x8000,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,68,15,112,208,0                 ; pshufd        $0x0,%xmm0,%xmm10
  DB  102,65,15,111,205                   ; movdqa        %xmm13,%xmm1
  DB  102,65,15,219,202                   ; pand          %xmm10,%xmm1
  DB  184,0,124,0,0                       ; mov           $0x7c00,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  102,15,112,216,0                    ; pshufd        $0x0,%xmm0,%xmm3
  DB  102,65,15,111,197                   ; movdqa        %xmm13,%xmm0
  DB  102,15,219,195                      ; pand          %xmm3,%xmm0
  DB  102,68,15,239,233                   ; pxor          %xmm1,%xmm13
  DB  102,15,114,241,16                   ; pslld         $0x10,%xmm1
  DB  102,65,15,114,245,13                ; pslld         $0xd,%xmm13
  DB  184,0,0,0,56                        ; mov           $0x38000000,%eax
  DB  102,15,110,208                      ; movd          %eax,%xmm2
  DB  102,68,15,112,218,0                 ; pshufd        $0x0,%xmm2,%xmm11
  DB  102,65,15,254,203                   ; paddd         %xmm11,%xmm1
  DB  102,65,15,254,205                   ; paddd         %xmm13,%xmm1
  DB  102,65,15,118,193                   ; pcmpeqd       %xmm9,%xmm0
  DB  102,15,223,193                      ; pandn         %xmm1,%xmm0
  DB  102,65,15,115,220,8                 ; psrldq        $0x8,%xmm12
  DB  102,69,15,97,225                    ; punpcklwd     %xmm9,%xmm12
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,65,15,219,210                   ; pand          %xmm10,%xmm2
  DB  102,65,15,111,204                   ; movdqa        %xmm12,%xmm1
  DB  102,15,219,203                      ; pand          %xmm3,%xmm1
  DB  102,68,15,239,226                   ; pxor          %xmm2,%xmm12
  DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
  DB  102,65,15,254,212                   ; paddd         %xmm12,%xmm2
  DB  102,65,15,118,201                   ; pcmpeqd       %xmm9,%xmm1
  DB  102,15,223,202                      ; pandn         %xmm2,%xmm1
  DB  102,69,15,111,224                   ; movdqa        %xmm8,%xmm12
  DB  102,69,15,97,225                    ; punpcklwd     %xmm9,%xmm12
  DB  102,69,15,111,236                   ; movdqa        %xmm12,%xmm13
  DB  102,69,15,219,234                   ; pand          %xmm10,%xmm13
  DB  102,65,15,111,212                   ; movdqa        %xmm12,%xmm2
  DB  102,15,219,211                      ; pand          %xmm3,%xmm2
  DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
  DB  102,65,15,114,245,16                ; pslld         $0x10,%xmm13
  DB  102,65,15,114,244,13                ; pslld         $0xd,%xmm12
  DB  102,69,15,254,235                   ; paddd         %xmm11,%xmm13
  DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
  DB  102,65,15,118,209                   ; pcmpeqd       %xmm9,%xmm2
  DB  102,65,15,223,213                   ; pandn         %xmm13,%xmm2
  DB  102,65,15,115,216,8                 ; psrldq        $0x8,%xmm8
  DB  102,69,15,97,193                    ; punpcklwd     %xmm9,%xmm8
  DB  102,69,15,219,208                   ; pand          %xmm8,%xmm10
  DB  102,65,15,219,216                   ; pand          %xmm8,%xmm3
  DB  102,69,15,239,194                   ; pxor          %xmm10,%xmm8
  DB  102,65,15,114,242,16                ; pslld         $0x10,%xmm10
  DB  102,65,15,114,240,13                ; pslld         $0xd,%xmm8
  DB  102,69,15,254,211                   ; paddd         %xmm11,%xmm10
  DB  102,69,15,254,208                   ; paddd         %xmm8,%xmm10
  DB  102,65,15,118,217                   ; pcmpeqd       %xmm9,%xmm3
  DB  102,65,15,223,218                   ; pandn         %xmm10,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_f16_sse2
_sk_store_f16_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,0,0,128                       ; mov           $0x80000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  102,69,15,112,200,0                 ; pshufd        $0x0,%xmm8,%xmm9
  DB  102,69,15,111,225                   ; movdqa        %xmm9,%xmm12
  DB  102,68,15,219,224                   ; pand          %xmm0,%xmm12
  DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
  DB  102,69,15,239,196                   ; pxor          %xmm12,%xmm8
  DB  185,0,0,128,56                      ; mov           $0x38800000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  102,65,15,114,212,16                ; psrld         $0x10,%xmm12
  DB  102,69,15,111,232                   ; movdqa        %xmm8,%xmm13
  DB  102,65,15,114,213,13                ; psrld         $0xd,%xmm13
  DB  185,0,192,1,0                       ; mov           $0x1c000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  102,69,15,112,219,0                 ; pshufd        $0x0,%xmm11,%xmm11
  DB  102,69,15,250,227                   ; psubd         %xmm11,%xmm12
  DB  102,69,15,254,229                   ; paddd         %xmm13,%xmm12
  DB  102,65,15,114,244,16                ; pslld         $0x10,%xmm12
  DB  102,65,15,114,228,16                ; psrad         $0x10,%xmm12
  DB  69,15,194,194,5                     ; cmpnltps      %xmm10,%xmm8
  DB  69,15,84,196                        ; andps         %xmm12,%xmm8
  DB  102,69,15,107,192                   ; packssdw      %xmm8,%xmm8
  DB  102,69,15,111,233                   ; movdqa        %xmm9,%xmm13
  DB  102,68,15,219,233                   ; pand          %xmm1,%xmm13
  DB  102,68,15,111,225                   ; movdqa        %xmm1,%xmm12
  DB  102,69,15,239,229                   ; pxor          %xmm13,%xmm12
  DB  102,65,15,114,213,16                ; psrld         $0x10,%xmm13
  DB  102,69,15,111,244                   ; movdqa        %xmm12,%xmm14
  DB  102,65,15,114,214,13                ; psrld         $0xd,%xmm14
  DB  102,69,15,250,235                   ; psubd         %xmm11,%xmm13
  DB  102,69,15,254,238                   ; paddd         %xmm14,%xmm13
  DB  102,65,15,114,245,16                ; pslld         $0x10,%xmm13
  DB  102,65,15,114,229,16                ; psrad         $0x10,%xmm13
  DB  69,15,194,226,5                     ; cmpnltps      %xmm10,%xmm12
  DB  69,15,84,229                        ; andps         %xmm13,%xmm12
  DB  102,69,15,107,228                   ; packssdw      %xmm12,%xmm12
  DB  102,69,15,111,241                   ; movdqa        %xmm9,%xmm14
  DB  102,68,15,219,242                   ; pand          %xmm2,%xmm14
  DB  102,68,15,111,234                   ; movdqa        %xmm2,%xmm13
  DB  102,69,15,239,238                   ; pxor          %xmm14,%xmm13
  DB  102,65,15,114,214,16                ; psrld         $0x10,%xmm14
  DB  102,69,15,111,253                   ; movdqa        %xmm13,%xmm15
  DB  102,65,15,114,215,13                ; psrld         $0xd,%xmm15
  DB  102,69,15,250,243                   ; psubd         %xmm11,%xmm14
  DB  102,69,15,254,247                   ; paddd         %xmm15,%xmm14
  DB  102,65,15,114,246,16                ; pslld         $0x10,%xmm14
  DB  102,65,15,114,230,16                ; psrad         $0x10,%xmm14
  DB  69,15,194,234,5                     ; cmpnltps      %xmm10,%xmm13
  DB  69,15,84,238                        ; andps         %xmm14,%xmm13
  DB  102,69,15,107,237                   ; packssdw      %xmm13,%xmm13
  DB  102,68,15,219,203                   ; pand          %xmm3,%xmm9
  DB  102,68,15,111,243                   ; movdqa        %xmm3,%xmm14
  DB  102,69,15,239,241                   ; pxor          %xmm9,%xmm14
  DB  102,65,15,114,209,16                ; psrld         $0x10,%xmm9
  DB  102,69,15,111,254                   ; movdqa        %xmm14,%xmm15
  DB  102,65,15,114,215,13                ; psrld         $0xd,%xmm15
  DB  102,69,15,250,203                   ; psubd         %xmm11,%xmm9
  DB  102,69,15,254,207                   ; paddd         %xmm15,%xmm9
  DB  102,65,15,114,241,16                ; pslld         $0x10,%xmm9
  DB  102,65,15,114,225,16                ; psrad         $0x10,%xmm9
  DB  69,15,194,242,5                     ; cmpnltps      %xmm10,%xmm14
  DB  69,15,84,241                        ; andps         %xmm9,%xmm14
  DB  102,69,15,107,246                   ; packssdw      %xmm14,%xmm14
  DB  102,69,15,97,196                    ; punpcklwd     %xmm12,%xmm8
  DB  102,69,15,97,238                    ; punpcklwd     %xmm14,%xmm13
  DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
  DB  102,69,15,98,205                    ; punpckldq     %xmm13,%xmm9
  DB  243,68,15,127,12,248                ; movdqu        %xmm9,(%rax,%rdi,8)
  DB  102,69,15,106,197                   ; punpckhdq     %xmm13,%xmm8
  DB  243,68,15,127,68,248,16             ; movdqu        %xmm8,0x10(%rax,%rdi,8)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_u16_be_sse2
_sk_load_u16_be_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  243,15,111,4,248                    ; movdqu        (%rax,%rdi,8),%xmm0
  DB  243,15,111,76,248,16                ; movdqu        0x10(%rax,%rdi,8),%xmm1
  DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
  DB  102,15,97,209                       ; punpcklwd     %xmm1,%xmm2
  DB  102,15,105,193                      ; punpckhwd     %xmm1,%xmm0
  DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
  DB  102,15,105,208                      ; punpckhwd     %xmm0,%xmm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
  DB  102,15,113,240,8                    ; psllw         $0x8,%xmm0
  DB  102,15,112,217,78                   ; pshufd        $0x4e,%xmm1,%xmm3
  DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
  DB  102,15,235,200                      ; por           %xmm0,%xmm1
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,65,15,97,201                    ; punpcklwd     %xmm9,%xmm1
  DB  15,91,193                           ; cvtdq2ps      %xmm1,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
  DB  102,15,235,217                      ; por           %xmm1,%xmm3
  DB  102,65,15,97,217                    ; punpcklwd     %xmm9,%xmm3
  DB  15,91,203                           ; cvtdq2ps      %xmm3,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,68,15,111,210                   ; movdqa        %xmm2,%xmm10
  DB  102,65,15,113,242,8                 ; psllw         $0x8,%xmm10
  DB  102,15,112,218,78                   ; pshufd        $0x4e,%xmm2,%xmm3
  DB  102,15,113,210,8                    ; psrlw         $0x8,%xmm2
  DB  102,65,15,235,210                   ; por           %xmm10,%xmm2
  DB  102,65,15,97,209                    ; punpcklwd     %xmm9,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  102,68,15,111,211                   ; movdqa        %xmm3,%xmm10
  DB  102,65,15,113,242,8                 ; psllw         $0x8,%xmm10
  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
  DB  102,65,15,235,218                   ; por           %xmm10,%xmm3
  DB  102,65,15,97,217                    ; punpcklwd     %xmm9,%xmm3
  DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_rgb_u16_be_sse2
_sk_load_rgb_u16_be_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,141,12,127                       ; lea           (%rdi,%rdi,2),%rcx
  DB  243,15,111,20,72                    ; movdqu        (%rax,%rcx,2),%xmm2
  DB  243,15,111,68,72,8                  ; movdqu        0x8(%rax,%rcx,2),%xmm0
  DB  102,15,115,216,4                    ; psrldq        $0x4,%xmm0
  DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
  DB  102,15,115,217,6                    ; psrldq        $0x6,%xmm1
  DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
  DB  102,15,115,216,6                    ; psrldq        $0x6,%xmm0
  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
  DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
  DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
  DB  102,15,112,216,78                   ; pshufd        $0x4e,%xmm0,%xmm3
  DB  102,15,105,209                      ; punpckhwd     %xmm1,%xmm2
  DB  184,128,0,128,55                    ; mov           $0x37800080,%eax
  DB  102,68,15,110,192                   ; movd          %eax,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
  DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
  DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
  DB  102,15,235,193                      ; por           %xmm1,%xmm0
  DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
  DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
  DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
  DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
  DB  102,15,235,217                      ; por           %xmm1,%xmm3
  DB  102,65,15,97,217                    ; punpcklwd     %xmm9,%xmm3
  DB  15,91,203                           ; cvtdq2ps      %xmm3,%xmm1
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
  DB  102,15,113,243,8                    ; psllw         $0x8,%xmm3
  DB  102,15,113,210,8                    ; psrlw         $0x8,%xmm2
  DB  102,15,235,211                      ; por           %xmm3,%xmm2
  DB  102,65,15,97,209                    ; punpcklwd     %xmm9,%xmm2
  DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_u16_be_sse2
_sk_store_u16_be_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  185,0,255,127,71                    ; mov           $0x477fff00,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
  DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
  DB  102,65,15,114,224,16                ; psrad         $0x10,%xmm8
  DB  102,69,15,107,192                   ; packssdw      %xmm8,%xmm8
  DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
  DB  102,65,15,113,242,8                 ; psllw         $0x8,%xmm10
  DB  102,65,15,113,208,8                 ; psrlw         $0x8,%xmm8
  DB  102,69,15,235,194                   ; por           %xmm10,%xmm8
  DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
  DB  102,65,15,114,242,16                ; pslld         $0x10,%xmm10
  DB  102,65,15,114,226,16                ; psrad         $0x10,%xmm10
  DB  102,69,15,107,210                   ; packssdw      %xmm10,%xmm10
  DB  102,69,15,111,218                   ; movdqa        %xmm10,%xmm11
  DB  102,65,15,113,243,8                 ; psllw         $0x8,%xmm11
  DB  102,65,15,113,210,8                 ; psrlw         $0x8,%xmm10
  DB  102,69,15,235,211                   ; por           %xmm11,%xmm10
  DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  102,69,15,91,219                    ; cvtps2dq      %xmm11,%xmm11
  DB  102,65,15,114,243,16                ; pslld         $0x10,%xmm11
  DB  102,65,15,114,227,16                ; psrad         $0x10,%xmm11
  DB  102,69,15,107,219                   ; packssdw      %xmm11,%xmm11
  DB  102,69,15,111,227                   ; movdqa        %xmm11,%xmm12
  DB  102,65,15,113,244,8                 ; psllw         $0x8,%xmm12
  DB  102,65,15,113,211,8                 ; psrlw         $0x8,%xmm11
  DB  102,69,15,235,220                   ; por           %xmm12,%xmm11
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
  DB  102,65,15,114,241,16                ; pslld         $0x10,%xmm9
  DB  102,65,15,114,225,16                ; psrad         $0x10,%xmm9
  DB  102,69,15,107,201                   ; packssdw      %xmm9,%xmm9
  DB  102,69,15,111,225                   ; movdqa        %xmm9,%xmm12
  DB  102,65,15,113,244,8                 ; psllw         $0x8,%xmm12
  DB  102,65,15,113,209,8                 ; psrlw         $0x8,%xmm9
  DB  102,69,15,235,204                   ; por           %xmm12,%xmm9
  DB  102,69,15,97,194                    ; punpcklwd     %xmm10,%xmm8
  DB  102,69,15,97,217                    ; punpcklwd     %xmm9,%xmm11
  DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
  DB  102,69,15,98,203                    ; punpckldq     %xmm11,%xmm9
  DB  243,68,15,127,12,248                ; movdqu        %xmm9,(%rax,%rdi,8)
  DB  102,69,15,106,195                   ; punpckhdq     %xmm11,%xmm8
  DB  243,68,15,127,68,248,16             ; movdqu        %xmm8,0x10(%rax,%rdi,8)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_load_f32_sse2
_sk_load_f32_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,137,249                          ; mov           %rdi,%rcx
  DB  72,193,225,4                        ; shl           $0x4,%rcx
  DB  68,15,16,4,8                        ; movups        (%rax,%rcx,1),%xmm8
  DB  15,16,68,8,16                       ; movups        0x10(%rax,%rcx,1),%xmm0
  DB  15,16,92,8,32                       ; movups        0x20(%rax,%rcx,1),%xmm3
  DB  68,15,16,76,8,48                    ; movups        0x30(%rax,%rcx,1),%xmm9
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  15,20,208                           ; unpcklps      %xmm0,%xmm2
  DB  15,40,203                           ; movaps        %xmm3,%xmm1
  DB  65,15,20,201                        ; unpcklps      %xmm9,%xmm1
  DB  68,15,21,192                        ; unpckhps      %xmm0,%xmm8
  DB  65,15,21,217                        ; unpckhps      %xmm9,%xmm3
  DB  15,40,194                           ; movaps        %xmm2,%xmm0
  DB  102,15,20,193                       ; unpcklpd      %xmm1,%xmm0
  DB  15,18,202                           ; movhlps       %xmm2,%xmm1
  DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
  DB  102,15,20,211                       ; unpcklpd      %xmm3,%xmm2
  DB  65,15,18,216                        ; movhlps       %xmm8,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_store_f32_sse2
_sk_store_f32_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,0                            ; mov           (%rax),%rax
  DB  72,137,249                          ; mov           %rdi,%rcx
  DB  72,193,225,4                        ; shl           $0x4,%rcx
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  68,15,20,201                        ; unpcklps      %xmm1,%xmm9
  DB  68,15,40,210                        ; movaps        %xmm2,%xmm10
  DB  68,15,40,218                        ; movaps        %xmm2,%xmm11
  DB  68,15,20,219                        ; unpcklps      %xmm3,%xmm11
  DB  68,15,21,193                        ; unpckhps      %xmm1,%xmm8
  DB  68,15,21,211                        ; unpckhps      %xmm3,%xmm10
  DB  69,15,40,225                        ; movaps        %xmm9,%xmm12
  DB  102,69,15,20,227                    ; unpcklpd      %xmm11,%xmm12
  DB  69,15,18,217                        ; movhlps       %xmm9,%xmm11
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  102,69,15,20,202                    ; unpcklpd      %xmm10,%xmm9
  DB  69,15,18,208                        ; movhlps       %xmm8,%xmm10
  DB  102,68,15,17,36,8                   ; movupd        %xmm12,(%rax,%rcx,1)
  DB  68,15,17,92,8,16                    ; movups        %xmm11,0x10(%rax,%rcx,1)
  DB  102,68,15,17,76,8,32                ; movupd        %xmm9,0x20(%rax,%rcx,1)
  DB  68,15,17,84,8,48                    ; movups        %xmm10,0x30(%rax,%rcx,1)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_x_sse2
_sk_clamp_x_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  102,15,118,192                      ; pcmpeqd       %xmm0,%xmm0
  DB  102,65,15,254,193                   ; paddd         %xmm9,%xmm0
  DB  68,15,93,192                        ; minps         %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_clamp_y_sse2
_sk_clamp_y_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,95,193                        ; maxps         %xmm1,%xmm8
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  102,15,118,201                      ; pcmpeqd       %xmm1,%xmm1
  DB  102,65,15,254,201                   ; paddd         %xmm9,%xmm1
  DB  68,15,93,193                        ; minps         %xmm1,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_x_sse2
_sk_repeat_x_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
  DB  69,15,94,200                        ; divps         %xmm8,%xmm9
  DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
  DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
  DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,84,217                        ; andps         %xmm9,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  65,15,92,194                        ; subps         %xmm10,%xmm0
  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
  DB  102,69,15,254,200                   ; paddd         %xmm8,%xmm9
  DB  65,15,93,193                        ; minps         %xmm9,%xmm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_repeat_y_sse2
_sk_repeat_y_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  69,15,94,200                        ; divps         %xmm8,%xmm9
  DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
  DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
  DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,216                   ; movd          %eax,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,84,217                        ; andps         %xmm9,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  65,15,92,202                        ; subps         %xmm10,%xmm1
  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
  DB  102,69,15,254,200                   ; paddd         %xmm8,%xmm9
  DB  65,15,93,201                        ; minps         %xmm9,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_x_sse2
_sk_mirror_x_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,92,192                        ; subps         %xmm8,%xmm0
  DB  243,69,15,88,201                    ; addss         %xmm9,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
  DB  69,15,94,209                        ; divps         %xmm9,%xmm10
  DB  243,69,15,91,218                    ; cvttps2dq     %xmm10,%xmm11
  DB  69,15,91,219                        ; cvtdq2ps      %xmm11,%xmm11
  DB  69,15,194,211,1                     ; cmpltps       %xmm11,%xmm10
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,84,226                        ; andps         %xmm10,%xmm12
  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
  DB  69,15,92,220                        ; subps         %xmm12,%xmm11
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  65,15,92,195                        ; subps         %xmm11,%xmm0
  DB  65,15,92,192                        ; subps         %xmm8,%xmm0
  DB  68,15,92,208                        ; subps         %xmm0,%xmm10
  DB  65,15,84,194                        ; andps         %xmm10,%xmm0
  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
  DB  102,69,15,254,200                   ; paddd         %xmm8,%xmm9
  DB  65,15,93,193                        ; minps         %xmm9,%xmm0
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_mirror_y_sse2
_sk_mirror_y_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,8                      ; movss         (%rax),%xmm9
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  243,69,15,88,201                    ; addss         %xmm9,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
  DB  69,15,94,209                        ; divps         %xmm9,%xmm10
  DB  243,69,15,91,218                    ; cvttps2dq     %xmm10,%xmm11
  DB  69,15,91,219                        ; cvtdq2ps      %xmm11,%xmm11
  DB  69,15,194,211,1                     ; cmpltps       %xmm11,%xmm10
  DB  184,0,0,128,63                      ; mov           $0x3f800000,%eax
  DB  102,68,15,110,224                   ; movd          %eax,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,84,226                        ; andps         %xmm10,%xmm12
  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
  DB  69,15,92,220                        ; subps         %xmm12,%xmm11
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  65,15,92,203                        ; subps         %xmm11,%xmm1
  DB  65,15,92,200                        ; subps         %xmm8,%xmm1
  DB  68,15,92,209                        ; subps         %xmm1,%xmm10
  DB  65,15,84,202                        ; andps         %xmm10,%xmm1
  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
  DB  102,69,15,254,200                   ; paddd         %xmm8,%xmm9
  DB  65,15,93,201                        ; minps         %xmm9,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_luminance_to_alpha_sse2
_sk_luminance_to_alpha_sse2 LABEL PROC
  DB  184,208,179,89,62                   ; mov           $0x3e59b3d0,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  184,89,23,55,63                     ; mov           $0x3f371759,%eax
  DB  102,15,110,192                      ; movd          %eax,%xmm0
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  15,88,195                           ; addps         %xmm3,%xmm0
  DB  184,152,221,147,61                  ; mov           $0x3d93dd98,%eax
  DB  102,15,110,216                      ; movd          %eax,%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  15,89,218                           ; mulps         %xmm2,%xmm3
  DB  15,88,216                           ; addps         %xmm0,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  15,87,192                           ; xorps         %xmm0,%xmm0
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_2x3_sse2
_sk_matrix_2x3_sse2 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,16                  ; movss         0x10(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,68,15,16,80,12                  ; movss         0xc(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,20                  ; movss         0x14(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_3x4_sse2
_sk_matrix_3x4_sse2 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,68,15,16,80,12                  ; movss         0xc(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,24                  ; movss         0x18(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,36                  ; movss         0x24(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,28                  ; movss         0x1c(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,40                  ; movss         0x28(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,20                  ; movss         0x14(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,32                  ; movss         0x20(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,44                 ; movss         0x2c(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  68,15,89,226                        ; mulps         %xmm2,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_4x5_sse2
_sk_matrix_4x5_sse2 LABEL PROC
  DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,32                  ; movss         0x20(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,48                  ; movss         0x30(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,64                 ; movss         0x40(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  68,15,89,227                        ; mulps         %xmm3,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,68,15,16,80,20                  ; movss         0x14(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,36                  ; movss         0x24(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,52                  ; movss         0x34(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,68                 ; movss         0x44(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  68,15,89,227                        ; mulps         %xmm3,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,24                  ; movss         0x18(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,40                  ; movss         0x28(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,56                 ; movss         0x38(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  243,68,15,16,112,72                 ; movss         0x48(%rax),%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  68,15,89,235                        ; mulps         %xmm3,%xmm13
  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
  DB  68,15,89,226                        ; mulps         %xmm2,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,89,217                        ; mulps         %xmm9,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  243,68,15,16,88,12                  ; movss         0xc(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,28                  ; movss         0x1c(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  243,68,15,16,104,44                 ; movss         0x2c(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  243,68,15,16,112,60                 ; movss         0x3c(%rax),%xmm14
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  243,68,15,16,120,76                 ; movss         0x4c(%rax),%xmm15
  DB  69,15,198,255,0                     ; shufps        $0x0,%xmm15,%xmm15
  DB  68,15,89,243                        ; mulps         %xmm3,%xmm14
  DB  69,15,88,247                        ; addps         %xmm15,%xmm14
  DB  68,15,89,234                        ; mulps         %xmm2,%xmm13
  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
  DB  69,15,89,225                        ; mulps         %xmm9,%xmm12
  DB  69,15,88,229                        ; addps         %xmm13,%xmm12
  DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
  DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_matrix_perspective_sse2
_sk_matrix_perspective_sse2 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,68,15,16,72,4                   ; movss         0x4(%rax),%xmm9
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  65,15,88,193                        ; addps         %xmm9,%xmm0
  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,20                  ; movss         0x14(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  243,68,15,16,80,24                  ; movss         0x18(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,28                  ; movss         0x1c(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,32                  ; movss         0x20(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
  DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  65,15,83,202                        ; rcpps         %xmm10,%xmm1
  DB  15,89,193                           ; mulps         %xmm1,%xmm0
  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,201                        ; movaps        %xmm9,%xmm1
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_sse2
_sk_linear_gradient_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,68,15,16,72,16                  ; movss         0x10(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  243,68,15,16,80,20                  ; movss         0x14(%rax),%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  243,68,15,16,88,24                  ; movss         0x18(%rax),%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  243,68,15,16,96,28                  ; movss         0x1c(%rax),%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  72,139,8                            ; mov           (%rax),%rcx
  DB  72,133,201                          ; test          %rcx,%rcx
  DB  15,132,15,1,0,0                     ; je            3f3e <_sk_linear_gradient_sse2+0x149>
  DB  72,139,64,8                         ; mov           0x8(%rax),%rax
  DB  72,131,192,32                       ; add           $0x20,%rax
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  243,68,15,16,112,224                ; movss         -0x20(%rax),%xmm14
  DB  243,68,15,16,104,228                ; movss         -0x1c(%rax),%xmm13
  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
  DB  69,15,40,252                        ; movaps        %xmm12,%xmm15
  DB  68,15,40,224                        ; movaps        %xmm0,%xmm12
  DB  69,15,194,230,1                     ; cmpltps       %xmm14,%xmm12
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,40,244                        ; movaps        %xmm12,%xmm14
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  69,15,84,196                        ; andps         %xmm12,%xmm8
  DB  69,15,86,198                        ; orps          %xmm14,%xmm8
  DB  243,68,15,16,104,232                ; movss         -0x18(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,40,244                        ; movaps        %xmm12,%xmm14
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  65,15,84,204                        ; andps         %xmm12,%xmm1
  DB  65,15,86,206                        ; orps          %xmm14,%xmm1
  DB  243,68,15,16,104,236                ; movss         -0x14(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,40,244                        ; movaps        %xmm12,%xmm14
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  65,15,84,212                        ; andps         %xmm12,%xmm2
  DB  65,15,86,214                        ; orps          %xmm14,%xmm2
  DB  243,68,15,16,104,240                ; movss         -0x10(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,40,244                        ; movaps        %xmm12,%xmm14
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  65,15,84,220                        ; andps         %xmm12,%xmm3
  DB  65,15,86,222                        ; orps          %xmm14,%xmm3
  DB  243,68,15,16,104,244                ; movss         -0xc(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,40,244                        ; movaps        %xmm12,%xmm14
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  69,15,84,204                        ; andps         %xmm12,%xmm9
  DB  69,15,86,206                        ; orps          %xmm14,%xmm9
  DB  243,68,15,16,104,248                ; movss         -0x8(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,40,244                        ; movaps        %xmm12,%xmm14
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  69,15,84,212                        ; andps         %xmm12,%xmm10
  DB  69,15,86,214                        ; orps          %xmm14,%xmm10
  DB  243,68,15,16,104,252                ; movss         -0x4(%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,40,244                        ; movaps        %xmm12,%xmm14
  DB  69,15,85,245                        ; andnps        %xmm13,%xmm14
  DB  69,15,84,220                        ; andps         %xmm12,%xmm11
  DB  69,15,86,222                        ; orps          %xmm14,%xmm11
  DB  243,68,15,16,40                     ; movss         (%rax),%xmm13
  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
  DB  69,15,84,252                        ; andps         %xmm12,%xmm15
  DB  69,15,85,229                        ; andnps        %xmm13,%xmm12
  DB  69,15,86,231                        ; orps          %xmm15,%xmm12
  DB  72,131,192,36                       ; add           $0x24,%rax
  DB  72,255,201                          ; dec           %rcx
  DB  15,133,8,255,255,255                ; jne           3e44 <_sk_linear_gradient_sse2+0x4f>
  DB  235,13                              ; jmp           3f4b <_sk_linear_gradient_sse2+0x156>
  DB  15,87,201                           ; xorps         %xmm1,%xmm1
  DB  15,87,210                           ; xorps         %xmm2,%xmm2
  DB  15,87,219                           ; xorps         %xmm3,%xmm3
  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  69,15,88,193                        ; addps         %xmm9,%xmm8
  DB  15,89,200                           ; mulps         %xmm0,%xmm1
  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
  DB  15,89,208                           ; mulps         %xmm0,%xmm2
  DB  65,15,88,211                        ; addps         %xmm11,%xmm2
  DB  15,89,216                           ; mulps         %xmm0,%xmm3
  DB  65,15,88,220                        ; addps         %xmm12,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_linear_gradient_2stops_sse2
_sk_linear_gradient_2stops_sse2 LABEL PROC
  DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  243,15,16,0                         ; movss         (%rax),%xmm0
  DB  243,15,16,72,4                      ; movss         0x4(%rax),%xmm1
  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
  DB  243,15,16,80,16                     ; movss         0x10(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
  DB  15,88,194                           ; addps         %xmm2,%xmm0
  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
  DB  243,15,16,80,20                     ; movss         0x14(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
  DB  15,88,202                           ; addps         %xmm2,%xmm1
  DB  243,15,16,80,8                      ; movss         0x8(%rax),%xmm2
  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
  DB  243,15,16,88,24                     ; movss         0x18(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
  DB  15,88,211                           ; addps         %xmm3,%xmm2
  DB  243,15,16,88,12                     ; movss         0xc(%rax),%xmm3
  DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
  DB  243,68,15,16,72,28                  ; movss         0x1c(%rax),%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  65,15,89,216                        ; mulps         %xmm8,%xmm3
  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_save_xy_sse2
_sk_save_xy_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
  DB  68,15,88,200                        ; addps         %xmm0,%xmm9
  DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
  DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
  DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
  DB  69,15,194,218,1                     ; cmpltps       %xmm10,%xmm11
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,225                   ; movd          %ecx,%xmm12
  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
  DB  69,15,84,220                        ; andps         %xmm12,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  69,15,92,202                        ; subps         %xmm10,%xmm9
  DB  68,15,88,193                        ; addps         %xmm1,%xmm8
  DB  243,69,15,91,208                    ; cvttps2dq     %xmm8,%xmm10
  DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
  DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
  DB  69,15,194,218,1                     ; cmpltps       %xmm10,%xmm11
  DB  69,15,84,220                        ; andps         %xmm12,%xmm11
  DB  69,15,92,211                        ; subps         %xmm11,%xmm10
  DB  69,15,92,194                        ; subps         %xmm10,%xmm8
  DB  15,17,0                             ; movups        %xmm0,(%rax)
  DB  15,17,72,32                         ; movups        %xmm1,0x20(%rax)
  DB  68,15,17,72,64                      ; movups        %xmm9,0x40(%rax)
  DB  68,15,17,64,96                      ; movups        %xmm8,0x60(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_accumulate_sse2
_sk_accumulate_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
  DB  68,15,16,136,160,0,0,0              ; movups        0xa0(%rax),%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
  DB  65,15,88,224                        ; addps         %xmm8,%xmm4
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,193                        ; mulps         %xmm1,%xmm8
  DB  65,15,88,232                        ; addps         %xmm8,%xmm5
  DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
  DB  68,15,89,194                        ; mulps         %xmm2,%xmm8
  DB  65,15,88,240                        ; addps         %xmm8,%xmm6
  DB  68,15,89,203                        ; mulps         %xmm3,%xmm9
  DB  65,15,88,249                        ; addps         %xmm9,%xmm7
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_nx_sse2
_sk_bilinear_nx_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_px_sse2
_sk_bilinear_px_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_ny_sse2
_sk_bilinear_ny_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bilinear_py_sse2
_sk_bilinear_py_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3x_sse2
_sk_bicubic_n3x_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,191                     ; mov           $0xbfc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1x_sse2
_sk_bicubic_n1x_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  68,15,17,136,128,0,0,0              ; movups        %xmm9,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1x_sse2
_sk_bicubic_p1x_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  68,15,17,144,128,0,0,0              ; movups        %xmm10,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3x_sse2
_sk_bicubic_p3x_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,0                             ; movups        (%rax),%xmm0
  DB  68,15,16,72,64                      ; movups        0x40(%rax),%xmm9
  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,194                        ; addps         %xmm10,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  68,15,17,128,128,0,0,0              ; movups        %xmm8,0x80(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n3y_sse2
_sk_bicubic_n3y_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,191                     ; mov           $0xbfc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_n1y_sse2
_sk_bicubic_n1y_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,191                       ; mov           $0xbf000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,0,0,128,63                      ; mov           $0x3f800000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,92,193                        ; subps         %xmm9,%xmm8
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
  DB  68,15,17,136,160,0,0,0              ; movups        %xmm9,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p1y_sse2
_sk_bicubic_p1y_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,0,63                        ; mov           $0x3f000000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,85,85,149,191                   ; mov           $0xbf955555,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,217                   ; movd          %ecx,%xmm11
  DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,211                        ; addps         %xmm11,%xmm10
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  185,57,142,99,61                    ; mov           $0x3d638e39,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
  DB  69,15,88,208                        ; addps         %xmm8,%xmm10
  DB  68,15,17,144,160,0,0,0              ; movups        %xmm10,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_bicubic_p3y_sse2
_sk_bicubic_p3y_sse2 LABEL PROC
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  185,0,0,192,63                      ; mov           $0x3fc00000,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  15,16,72,32                         ; movups        0x20(%rax),%xmm1
  DB  68,15,16,72,96                      ; movups        0x60(%rax),%xmm9
  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
  DB  185,114,28,199,62                   ; mov           $0x3ec71c72,%ecx
  DB  102,68,15,110,193                   ; movd          %ecx,%xmm8
  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
  DB  185,171,170,170,190                 ; mov           $0xbeaaaaab,%ecx
  DB  102,68,15,110,209                   ; movd          %ecx,%xmm10
  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
  DB  69,15,88,194                        ; addps         %xmm10,%xmm8
  DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
  DB  68,15,17,128,160,0,0,0              ; movups        %xmm8,0xa0(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  255,224                             ; jmpq          *%rax

PUBLIC _sk_callback_sse2
_sk_callback_sse2 LABEL PROC
  DB  83                                  ; push          %rbx
  DB  72,131,236,32                       ; sub           $0x20,%rsp
  DB  68,15,40,197                        ; movaps        %xmm5,%xmm8
  DB  68,15,40,204                        ; movaps        %xmm4,%xmm9
  DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
  DB  68,15,40,218                        ; movaps        %xmm2,%xmm11
  DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
  DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
  DB  72,137,211                          ; mov           %rdx,%rbx
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,139,72,8                         ; mov           0x8(%rax),%rcx
  DB  186,4,0,0,0                         ; mov           $0x4,%edx
  DB  255,16                              ; callq         *(%rax)
  DB  72,173                              ; lods          %ds:(%rsi),%rax
  DB  72,137,218                          ; mov           %rbx,%rdx
  DB  65,15,40,197                        ; movaps        %xmm13,%xmm0
  DB  65,15,40,204                        ; movaps        %xmm12,%xmm1
  DB  65,15,40,211                        ; movaps        %xmm11,%xmm2
  DB  65,15,40,218                        ; movaps        %xmm10,%xmm3
  DB  65,15,40,225                        ; movaps        %xmm9,%xmm4
  DB  65,15,40,232                        ; movaps        %xmm8,%xmm5
  DB  72,131,196,32                       ; add           $0x20,%rsp
  DB  91                                  ; pop           %rbx
  DB  255,224                             ; jmpq          *%rax
ENDIF
END