Neon: Intrinsics impl. of h2v1 & h2v2 fancy upsamp
The previous AArch32 GAS implementation of h2v1 fancy upsampling has
been removed, since the intrinsics implementation provides the same or
better performance. There was no previous GAS implementation of h2v2
fancy upsampling, and there was no previous AArch64 GAS implementation
of h2v1 fancy upsampling.
diff --git a/simd/CMakeLists.txt b/simd/CMakeLists.txt
index db1d966..5f56c13 100644
--- a/simd/CMakeLists.txt
+++ b/simd/CMakeLists.txt
@@ -265,8 +265,8 @@
file(REMOVE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S)
-set(SIMD_SOURCES arm/jcgray-neon.c arm/jcsample-neon.c arm/jfdctfst-neon.c
- arm/jquanti-neon.c)
+set(SIMD_SOURCES arm/jcgray-neon.c arm/jcsample-neon.c arm/jdsample-neon.c
+ arm/jfdctfst-neon.c arm/jquanti-neon.c)
if(NEON_INTRINSICS)
set(SIMD_SOURCES ${SIMD_SOURCES} arm/jccolor-neon.c)
endif()
diff --git a/simd/arm/aarch32/jsimd.c b/simd/arm/aarch32/jsimd.c
index d35672e..3c9a3f6 100644
--- a/simd/arm/aarch32/jsimd.c
+++ b/simd/arm/aarch32/jsimd.c
@@ -423,6 +423,17 @@
GLOBAL(int)
jsimd_can_h2v2_fancy_upsample(void)
{
+ init_simd();
+
+ /* The code is optimised for these values only */
+ if (BITS_IN_JSAMPLE != 8)
+ return 0;
+ if (sizeof(JDIMENSION) != 4)
+ return 0;
+
+ if (simd_support & JSIMD_NEON)
+ return 1;
+
return 0;
}
@@ -447,6 +458,9 @@
jsimd_h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
{
+ jsimd_h2v2_fancy_upsample_neon(cinfo->max_v_samp_factor,
+ compptr->downsampled_width, input_data,
+ output_data_ptr);
}
GLOBAL(void)
diff --git a/simd/arm/aarch32/jsimd_neon.S b/simd/arm/aarch32/jsimd_neon.S
index c45f63c..d2ab02d 100644
--- a/simd/arm/aarch32/jsimd_neon.S
+++ b/simd/arm/aarch32/jsimd_neon.S
@@ -1579,240 +1579,3 @@
.purgem do_store
#endif /* NEON_INTRINSICS */
-
-
-/*****************************************************************************/
-
-/*
- * GLOBAL(void)
- * jsimd_h2v1_fancy_upsample_neon(int max_v_samp_factor,
- * JDIMENSION downsampled_width,
- * JSAMPARRAY input_data,
- * JSAMPARRAY *output_data_ptr);
- *
- * Note: the use of unaligned writes is the main remaining bottleneck in
- * this code, which can be potentially solved to get up to tens
- * of percents performance improvement on Cortex-A8/Cortex-A9.
- */
-
-/*
- * Upsample 16 source pixels to 32 destination pixels. The new 16 source
- * pixels are loaded to q0. The previous 16 source pixels are in q1. The
- * shifted-by-one source pixels are constructed in q2 by using q0 and q1.
- * Register d28 is used for multiplication by 3. Register q15 is used
- * for adding +1 bias.
- */
-.macro upsample16 OUTPTR, INPTR
- vld1.8 {q0}, [\INPTR]!
- vmovl.u8 q8, d0
- vext.8 q2, q1, q0, #15
- vmovl.u8 q9, d1
- vaddw.u8 q10, q15, d4
- vaddw.u8 q11, q15, d5
- vmlal.u8 q8, d4, d28
- vmlal.u8 q9, d5, d28
- vmlal.u8 q10, d0, d28
- vmlal.u8 q11, d1, d28
- vmov q1, q0 /* backup source pixels to q1 */
- vrshrn.u16 d6, q8, #2
- vrshrn.u16 d7, q9, #2
- vshrn.u16 d8, q10, #2
- vshrn.u16 d9, q11, #2
- vst2.8 {d6, d7, d8, d9}, [\OUTPTR]!
-.endm
-
-/*
- * Upsample 32 source pixels to 64 destination pixels. Compared to 'usample16'
- * macro, the roles of q0 and q1 registers are reversed for even and odd
- * groups of 16 pixels, that's why "vmov q1, q0" instructions are not needed.
- * Also this unrolling allows to reorder loads and stores to compensate
- * multiplication latency and reduce stalls.
- */
-.macro upsample32 OUTPTR, INPTR
- /* even 16 pixels group */
- vld1.8 {q0}, [\INPTR]!
- vmovl.u8 q8, d0
- vext.8 q2, q1, q0, #15
- vmovl.u8 q9, d1
- vaddw.u8 q10, q15, d4
- vaddw.u8 q11, q15, d5
- vmlal.u8 q8, d4, d28
- vmlal.u8 q9, d5, d28
- vmlal.u8 q10, d0, d28
- vmlal.u8 q11, d1, d28
- /* odd 16 pixels group */
- vld1.8 {q1}, [\INPTR]!
- vrshrn.u16 d6, q8, #2
- vrshrn.u16 d7, q9, #2
- vshrn.u16 d8, q10, #2
- vshrn.u16 d9, q11, #2
- vmovl.u8 q8, d2
- vext.8 q2, q0, q1, #15
- vmovl.u8 q9, d3
- vaddw.u8 q10, q15, d4
- vaddw.u8 q11, q15, d5
- vmlal.u8 q8, d4, d28
- vmlal.u8 q9, d5, d28
- vmlal.u8 q10, d2, d28
- vmlal.u8 q11, d3, d28
- vst2.8 {d6, d7, d8, d9}, [\OUTPTR]!
- vrshrn.u16 d6, q8, #2
- vrshrn.u16 d7, q9, #2
- vshrn.u16 d8, q10, #2
- vshrn.u16 d9, q11, #2
- vst2.8 {d6, d7, d8, d9}, [\OUTPTR]!
-.endm
-
-/*
- * Upsample a row of WIDTH pixels from INPTR to OUTPTR.
- */
-.macro upsample_row OUTPTR, INPTR, WIDTH, TMP1
- /* special case for the first and last pixels */
- sub \WIDTH, \WIDTH, #1
- add \OUTPTR, \OUTPTR, #1
- ldrb \TMP1, [\INPTR, \WIDTH]
- strb \TMP1, [\OUTPTR, \WIDTH, asl #1]
- ldrb \TMP1, [\INPTR], #1
- strb \TMP1, [\OUTPTR, #-1]
- vmov.8 d3[7], \TMP1
-
- subs \WIDTH, \WIDTH, #32
- blt 5f
-0: /* process 32 pixels per iteration */
- upsample32 \OUTPTR, \INPTR
- subs \WIDTH, \WIDTH, #32
- bge 0b
-5:
- adds \WIDTH, \WIDTH, #16
- blt 1f
-0: /* process 16 pixels if needed */
- upsample16 \OUTPTR, \INPTR
- subs \WIDTH, \WIDTH, #16
-1:
- adds \WIDTH, \WIDTH, #16
- beq 9f
-
- /* load the remaining 1-15 pixels */
- add \INPTR, \INPTR, \WIDTH
- tst \WIDTH, #1
- beq 2f
- sub \INPTR, \INPTR, #1
- vld1.8 {d0[0]}, [\INPTR]
-2:
- tst \WIDTH, #2
- beq 2f
- vext.8 d0, d0, d0, #6
- sub \INPTR, \INPTR, #1
- vld1.8 {d0[1]}, [\INPTR]
- sub \INPTR, \INPTR, #1
- vld1.8 {d0[0]}, [\INPTR]
-2:
- tst \WIDTH, #4
- beq 2f
- vrev64.32 d0, d0
- sub \INPTR, \INPTR, #1
- vld1.8 {d0[3]}, [\INPTR]
- sub \INPTR, \INPTR, #1
- vld1.8 {d0[2]}, [\INPTR]
- sub \INPTR, \INPTR, #1
- vld1.8 {d0[1]}, [\INPTR]
- sub \INPTR, \INPTR, #1
- vld1.8 {d0[0]}, [\INPTR]
-2:
- tst \WIDTH, #8
- beq 2f
- vmov d1, d0
- sub \INPTR, \INPTR, #8
- vld1.8 {d0}, [\INPTR]
-2: /* upsample the remaining pixels */
- vmovl.u8 q8, d0
- vext.8 q2, q1, q0, #15
- vmovl.u8 q9, d1
- vaddw.u8 q10, q15, d4
- vaddw.u8 q11, q15, d5
- vmlal.u8 q8, d4, d28
- vmlal.u8 q9, d5, d28
- vmlal.u8 q10, d0, d28
- vmlal.u8 q11, d1, d28
- vrshrn.u16 d10, q8, #2
- vrshrn.u16 d12, q9, #2
- vshrn.u16 d11, q10, #2
- vshrn.u16 d13, q11, #2
- vzip.8 d10, d11
- vzip.8 d12, d13
- /* store the remaining pixels */
- tst \WIDTH, #8
- beq 2f
- vst1.8 {d10, d11}, [\OUTPTR]!
- vmov q5, q6
-2:
- tst \WIDTH, #4
- beq 2f
- vst1.8 {d10}, [\OUTPTR]!
- vmov d10, d11
-2:
- tst \WIDTH, #2
- beq 2f
- vst1.8 {d10[0]}, [\OUTPTR]!
- vst1.8 {d10[1]}, [\OUTPTR]!
- vst1.8 {d10[2]}, [\OUTPTR]!
- vst1.8 {d10[3]}, [\OUTPTR]!
- vext.8 d10, d10, d10, #4
-2:
- tst \WIDTH, #1
- beq 2f
- vst1.8 {d10[0]}, [\OUTPTR]!
- vst1.8 {d10[1]}, [\OUTPTR]!
-2:
-9:
-.endm
-
-asm_function jsimd_h2v1_fancy_upsample_neon
-
- MAX_V_SAMP_FACTOR .req r0
- DOWNSAMPLED_WIDTH .req r1
- INPUT_DATA .req r2
- OUTPUT_DATA_PTR .req r3
- OUTPUT_DATA .req OUTPUT_DATA_PTR
-
- OUTPTR .req r4
- INPTR .req r5
- WIDTH .req ip
- TMP .req lr
-
- push {r4, r5, r6, lr}
- vpush {d8 - d15}
-
- ldr OUTPUT_DATA, [OUTPUT_DATA_PTR]
- cmp MAX_V_SAMP_FACTOR, #0
- ble 99f
-
- /* initialize constants */
- vmov.u8 d28, #3
- vmov.u16 q15, #1
-11:
- ldr INPTR, [INPUT_DATA], #4
- ldr OUTPTR, [OUTPUT_DATA], #4
- mov WIDTH, DOWNSAMPLED_WIDTH
- upsample_row OUTPTR, INPTR, WIDTH, TMP
- subs MAX_V_SAMP_FACTOR, MAX_V_SAMP_FACTOR, #1
- bgt 11b
-
-99:
- vpop {d8 - d15}
- pop {r4, r5, r6, pc}
-
- .unreq MAX_V_SAMP_FACTOR
- .unreq DOWNSAMPLED_WIDTH
- .unreq INPUT_DATA
- .unreq OUTPUT_DATA_PTR
- .unreq OUTPUT_DATA
-
- .unreq OUTPTR
- .unreq INPTR
- .unreq WIDTH
- .unreq TMP
-
-.purgem upsample16
-.purgem upsample32
-.purgem upsample_row
diff --git a/simd/arm/aarch64/jsimd.c b/simd/arm/aarch64/jsimd.c
index b588d8b..5040d50 100644
--- a/simd/arm/aarch64/jsimd.c
+++ b/simd/arm/aarch64/jsimd.c
@@ -491,12 +491,34 @@
GLOBAL(int)
jsimd_can_h2v2_fancy_upsample(void)
{
+ init_simd();
+
+ /* The code is optimised for these values only */
+ if (BITS_IN_JSAMPLE != 8)
+ return 0;
+ if (sizeof(JDIMENSION) != 4)
+ return 0;
+
+ if (simd_support & JSIMD_NEON)
+ return 1;
+
return 0;
}
GLOBAL(int)
jsimd_can_h2v1_fancy_upsample(void)
{
+ init_simd();
+
+ /* The code is optimised for these values only */
+ if (BITS_IN_JSAMPLE != 8)
+ return 0;
+ if (sizeof(JDIMENSION) != 4)
+ return 0;
+
+ if (simd_support & JSIMD_NEON)
+ return 1;
+
return 0;
}
@@ -504,12 +526,18 @@
jsimd_h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
{
+ jsimd_h2v2_fancy_upsample_neon(cinfo->max_v_samp_factor,
+ compptr->downsampled_width, input_data,
+ output_data_ptr);
}
GLOBAL(void)
jsimd_h2v1_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
{
+ jsimd_h2v1_fancy_upsample_neon(cinfo->max_v_samp_factor,
+ compptr->downsampled_width, input_data,
+ output_data_ptr);
}
GLOBAL(int)
diff --git a/simd/arm/jdsample-neon.c b/simd/arm/jdsample-neon.c
new file mode 100644
index 0000000..742ca58
--- /dev/null
+++ b/simd/arm/jdsample-neon.c
@@ -0,0 +1,374 @@
+/*
+ * jdsample-neon.c - upsampling (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+
+#include <arm_neon.h>
+
+
+/* The diagram below shows a row of samples produced by h2v1 downsampling.
+ *
+ * s0 s1 s2
+ * +---------+---------+---------+
+ * | | | |
+ * | p0 p1 | p2 p3 | p4 p5 |
+ * | | | |
+ * +---------+---------+---------+
+ *
+ * Samples s0-s2 were created by averaging the original pixel component values
+ * centered at positions p0-p5 above. To approximate those original pixel
+ * component values, we proportionally blend the adjacent samples in each row.
+ *
+ * An upsampled pixel component value is computed by blending the sample
+ * containing the pixel center with the nearest neighboring sample, in the
+ * ratio 3:1. For example:
+ * p1(upsampled) = 3/4 * s0 + 1/4 * s1
+ * p2(upsampled) = 3/4 * s1 + 1/4 * s0
+ * When computing the first and last pixel component values in the row, there
+ * is no adjacent sample to blend, so:
+ * p0(upsampled) = s0
+ * p5(upsampled) = s2
+ */
+
+void jsimd_h2v1_fancy_upsample_neon(int max_v_samp_factor,
+ JDIMENSION downsampled_width,
+ JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr)
+{
+ JSAMPARRAY output_data = *output_data_ptr;
+ JSAMPROW inptr, outptr;
+ int inrow;
+ unsigned colctr;
+ /* Set up constants. */
+ const uint16x8_t one_u16 = vdupq_n_u16(1);
+ const uint8x8_t three_u8 = vdup_n_u8(3);
+
+ for (inrow = 0; inrow < max_v_samp_factor; inrow++) {
+ inptr = input_data[inrow];
+ outptr = output_data[inrow];
+ /* First pixel component value in this row of the original image */
+ *outptr = (JSAMPLE)GETJSAMPLE(*inptr);
+
+ /* 3/4 * containing sample + 1/4 * nearest neighboring sample
+ * For p1: containing sample = s0, nearest neighboring sample = s1
+ * For p2: containing sample = s1, nearest neighboring sample = s0
+ */
+ uint8x16_t s0 = vld1q_u8(inptr);
+ uint8x16_t s1 = vld1q_u8(inptr + 1);
+ /* Multiplication makes vectors twice as wide. '_l' and '_h' suffixes
+ * denote low half and high half respectively.
+ */
+ uint16x8_t s1_add_3s0_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);
+ uint16x8_t s1_add_3s0_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);
+ uint16x8_t s0_add_3s1_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);
+ uint16x8_t s0_add_3s1_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);
+ /* Add ordered dithering bias to odd pixel values. */
+ s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
+ s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
+
+ /* The offset is initially 1, because the first pixel component has already
+ * been stored. However, in subsequent iterations of the SIMD loop, this
+ * offset is (2 * colctr - 1) to stay within the bounds of the sample
+ * buffers without having to resort to a slow scalar tail case for the last
+ * (downsampled_width % 16) samples. See "Creation of 2-D sample arrays"
+ * in jmemmgr.c for more details.
+ */
+ unsigned outptr_offset = 1;
+ uint8x16x2_t output_pixels;
+
+ /* We use software pipelining to maximise performance. The code indented
+ * an extra two spaces begins the next iteration of the loop.
+ */
+ for (colctr = 16; colctr < downsampled_width; colctr += 16) {
+
+ s0 = vld1q_u8(inptr + colctr - 1);
+ s1 = vld1q_u8(inptr + colctr);
+
+ /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+ output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
+ vrshrn_n_u16(s1_add_3s0_h, 2));
+ output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
+ vshrn_n_u16(s0_add_3s1_h, 2));
+
+ /* Multiplication makes vectors twice as wide. '_l' and '_h' suffixes
+ * denote low half and high half respectively.
+ */
+ s1_add_3s0_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);
+ s1_add_3s0_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);
+ s0_add_3s1_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);
+ s0_add_3s1_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);
+ /* Add ordered dithering bias to odd pixel values. */
+ s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
+ s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
+
+ /* Store pixel component values to memory. */
+ vst2q_u8(outptr + outptr_offset, output_pixels);
+ outptr_offset = 2 * colctr - 1;
+ }
+
+ /* Complete the last iteration of the loop. */
+
+ /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+ output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
+ vrshrn_n_u16(s1_add_3s0_h, 2));
+ output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
+ vshrn_n_u16(s0_add_3s1_h, 2));
+ /* Store pixel component values to memory. */
+ vst2q_u8(outptr + outptr_offset, output_pixels);
+
+ /* Last pixel component value in this row of the original image */
+ outptr[2 * downsampled_width - 1] =
+ GETJSAMPLE(inptr[downsampled_width - 1]);
+ }
+}
+
+
+/* The diagram below shows an array of samples produced by h2v2 downsampling.
+ *
+ * s0 s1 s2
+ * +---------+---------+---------+
+ * | p0 p1 | p2 p3 | p4 p5 |
+ * sA | | | |
+ * | p6 p7 | p8 p9 | p10 p11|
+ * +---------+---------+---------+
+ * | p12 p13| p14 p15| p16 p17|
+ * sB | | | |
+ * | p18 p19| p20 p21| p22 p23|
+ * +---------+---------+---------+
+ * | p24 p25| p26 p27| p28 p29|
+ * sC | | | |
+ * | p30 p31| p32 p33| p34 p35|
+ * +---------+---------+---------+
+ *
+ * Samples s0A-s2C were created by averaging the original pixel component
+ * values centered at positions p0-p35 above. To approximate one of those
+ * original pixel component values, we proportionally blend the sample
+ * containing the pixel center with the nearest neighboring samples in each
+ * row, column, and diagonal.
+ *
+ * An upsampled pixel component value is computed by first blending the sample
+ * containing the pixel center with the nearest neighboring samples in the
+ * same column, in the ratio 3:1, and then blending each column sum with the
+ * nearest neighboring column sum, in the ratio 3:1. For example:
+ * p14(upsampled) = 3/4 * (3/4 * s1B + 1/4 * s1A) +
+ * 1/4 * (3/4 * s0B + 1/4 * s0A)
+ * = 9/16 * s1B + 3/16 * s1A + 3/16 * s0B + 1/16 * s0A
+ * When computing the first and last pixel component values in the row, there
+ * is no horizontally adjacent sample to blend, so:
+ * p12(upsampled) = 3/4 * s0B + 1/4 * s0A
+ * p23(upsampled) = 3/4 * s2B + 1/4 * s2C
+ * When computing the first and last pixel component values in the column,
+ * there is no vertically adjacent sample to blend, so:
+ * p2(upsampled) = 3/4 * s1A + 1/4 * s0A
+ * p33(upsampled) = 3/4 * s1C + 1/4 * s2C
+ * When computing the corner pixel component values, there is no adjacent
+ * sample to blend, so:
+ * p0(upsampled) = s0A
+ * p35(upsampled) = s2C
+ */
+
+void jsimd_h2v2_fancy_upsample_neon(int max_v_samp_factor,
+ JDIMENSION downsampled_width,
+ JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr)
+{
+ JSAMPARRAY output_data = *output_data_ptr;
+ JSAMPROW inptr0, inptr1, inptr2, outptr0, outptr1;
+ int inrow, outrow;
+ unsigned colctr;
+ /* Set up constants. */
+ const uint16x8_t seven_u16 = vdupq_n_u16(7);
+ const uint8x8_t three_u8 = vdup_n_u8(3);
+ const uint16x8_t three_u16 = vdupq_n_u16(3);
+
+ inrow = outrow = 0;
+ while (outrow < max_v_samp_factor) {
+ inptr0 = input_data[inrow - 1];
+ inptr1 = input_data[inrow];
+ inptr2 = input_data[inrow + 1];
+ /* Suffixes 0 and 1 denote the upper and lower rows of output pixels,
+ * respectively.
+ */
+ outptr0 = output_data[outrow++];
+ outptr1 = output_data[outrow++];
+
+ /* First pixel component value in this row of the original image */
+ int s0colsum0 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr0);
+ *outptr0 = (JSAMPLE)((s0colsum0 * 4 + 8) >> 4);
+ int s0colsum1 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr2);
+ *outptr1 = (JSAMPLE)((s0colsum1 * 4 + 8) >> 4);
+
+ /* Step 1: Blend samples vertically in columns s0 and s1.
+ * Leave the divide by 4 until the end, when it can be done for both
+ * dimensions at once, right-shifting by 4.
+ */
+
+ /* Load and compute s0colsum0 and s0colsum1. */
+ uint8x16_t s0A = vld1q_u8(inptr0);
+ uint8x16_t s0B = vld1q_u8(inptr1);
+ uint8x16_t s0C = vld1q_u8(inptr2);
+ /* Multiplication makes vectors twice as wide. '_l' and '_h' suffixes
+ * denote low half and high half respectively.
+ */
+ uint16x8_t s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)),
+ vget_low_u8(s0B), three_u8);
+ uint16x8_t s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)),
+ vget_high_u8(s0B), three_u8);
+ uint16x8_t s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)),
+ vget_low_u8(s0B), three_u8);
+ uint16x8_t s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)),
+ vget_high_u8(s0B), three_u8);
+ /* Load and compute s1colsum0 and s1colsum1. */
+ uint8x16_t s1A = vld1q_u8(inptr0 + 1);
+ uint8x16_t s1B = vld1q_u8(inptr1 + 1);
+ uint8x16_t s1C = vld1q_u8(inptr2 + 1);
+ uint16x8_t s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)),
+ vget_low_u8(s1B), three_u8);
+ uint16x8_t s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)),
+ vget_high_u8(s1B), three_u8);
+ uint16x8_t s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)),
+ vget_low_u8(s1B), three_u8);
+ uint16x8_t s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)),
+ vget_high_u8(s1B), three_u8);
+
+ /* Step 2: Blend the already-blended columns. */
+
+ uint16x8_t output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
+ uint16x8_t output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
+ uint16x8_t output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
+ uint16x8_t output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
+ uint16x8_t output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
+ uint16x8_t output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
+ uint16x8_t output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
+ uint16x8_t output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
+ /* Add ordered dithering bias to odd pixel values. */
+ output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
+ output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
+ output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
+ output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
+ /* Right-shift by 4 (divide by 16), narrow to 8-bit, and combine. */
+ uint8x16x2_t output_pixels0 = { {
+ vcombine_u8(vshrn_n_u16(output0_p1_l, 4), vshrn_n_u16(output0_p1_h, 4)),
+ vcombine_u8(vrshrn_n_u16(output0_p2_l, 4), vrshrn_n_u16(output0_p2_h, 4))
+ } };
+ uint8x16x2_t output_pixels1 = { {
+ vcombine_u8(vshrn_n_u16(output1_p1_l, 4), vshrn_n_u16(output1_p1_h, 4)),
+ vcombine_u8(vrshrn_n_u16(output1_p2_l, 4), vrshrn_n_u16(output1_p2_h, 4))
+ } };
+
+ /* Store pixel component values to memory.
+ * The minimum size of the output buffer for each row is 64 bytes => no
+ * need to worry about buffer overflow here. See "Creation of 2-D sample
+ * arrays" in jmemmgr.c for more details.
+ */
+ vst2q_u8(outptr0 + 1, output_pixels0);
+ vst2q_u8(outptr1 + 1, output_pixels1);
+
+ /* The first pixel of the image shifted our loads and stores by one byte.
+ * We have to re-align on a 32-byte boundary at some point before the end
+ * of the row (we do it now on the 32/33 pixel boundary) to stay within the
+ * bounds of the sample buffers without having to resort to a slow scalar
+ * tail case for the last (downsampled_width % 16) samples. See "Creation
+ * of 2-D sample arrays" in jmemmgr.c for more details.
+ */
+ for (colctr = 16; colctr < downsampled_width; colctr += 16) {
+ /* Step 1: Blend samples vertically in columns s0 and s1. */
+
+ /* Load and compute s0colsum0 and s0colsum1. */
+ s0A = vld1q_u8(inptr0 + colctr - 1);
+ s0B = vld1q_u8(inptr1 + colctr - 1);
+ s0C = vld1q_u8(inptr2 + colctr - 1);
+ s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)), vget_low_u8(s0B),
+ three_u8);
+ s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)), vget_high_u8(s0B),
+ three_u8);
+ s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)), vget_low_u8(s0B),
+ three_u8);
+ s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)), vget_high_u8(s0B),
+ three_u8);
+ /* Load and compute s1colsum0 and s1colsum1. */
+ s1A = vld1q_u8(inptr0 + colctr);
+ s1B = vld1q_u8(inptr1 + colctr);
+ s1C = vld1q_u8(inptr2 + colctr);
+ s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)), vget_low_u8(s1B),
+ three_u8);
+ s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)), vget_high_u8(s1B),
+ three_u8);
+ s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)), vget_low_u8(s1B),
+ three_u8);
+ s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)), vget_high_u8(s1B),
+ three_u8);
+
+ /* Step 2: Blend the already-blended columns. */
+
+ output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
+ output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
+ output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
+ output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
+ output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
+ output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
+ output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
+ output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
+ /* Add ordered dithering bias to odd pixel values. */
+ output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
+ output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
+ output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
+ output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
+ /* Right-shift by 4 (divide by 16), narrow to 8-bit, and combine. */
+ output_pixels0.val[0] = vcombine_u8(vshrn_n_u16(output0_p1_l, 4),
+ vshrn_n_u16(output0_p1_h, 4));
+ output_pixels0.val[1] = vcombine_u8(vrshrn_n_u16(output0_p2_l, 4),
+ vrshrn_n_u16(output0_p2_h, 4));
+ output_pixels1.val[0] = vcombine_u8(vshrn_n_u16(output1_p1_l, 4),
+ vshrn_n_u16(output1_p1_h, 4));
+ output_pixels1.val[1] = vcombine_u8(vrshrn_n_u16(output1_p2_l, 4),
+ vrshrn_n_u16(output1_p2_h, 4));
+ /* Store pixel component values to memory. */
+ vst2q_u8(outptr0 + 2 * colctr - 1, output_pixels0);
+ vst2q_u8(outptr1 + 2 * colctr - 1, output_pixels1);
+ }
+
+ /* Last pixel component value in this row of the original image */
+ int s1colsum0 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
+ GETJSAMPLE(inptr0[downsampled_width - 1]);
+ outptr0[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum0 * 4 + 7) >> 4);
+ int s1colsum1 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
+ GETJSAMPLE(inptr2[downsampled_width - 1]);
+ outptr1[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum1 * 4 + 7) >> 4);
+ inrow++;
+ }
+}
diff --git a/simd/jsimd.h b/simd/jsimd.h
index a2758b4..29920a0 100644
--- a/simd/jsimd.h
+++ b/simd/jsimd.h
@@ -661,6 +661,9 @@
EXTERN(void) jsimd_h2v1_fancy_upsample_neon
(int max_v_samp_factor, JDIMENSION downsampled_width, JSAMPARRAY input_data,
JSAMPARRAY *output_data_ptr);
+EXTERN(void) jsimd_h2v2_fancy_upsample_neon
+ (int max_v_samp_factor, JDIMENSION downsampled_width, JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr);
EXTERN(void) jsimd_h2v1_fancy_upsample_dspr2
(int max_v_samp_factor, JDIMENSION downsampled_width, JSAMPARRAY input_data,