blob: b8970e9e20f46d22411c6f4bf74e176f1679fc8d [file] [log] [blame]
/*
* Copyright 2018 Google Inc.
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#ifndef SkBitmapProcState_opts_DEFINED
#define SkBitmapProcState_opts_DEFINED
#include "include/private/SkVx.h"
#include "src/core/SkBitmapProcState.h"
#include "src/core/SkMSAN.h"
// SkBitmapProcState optimized Shader, Sample, or Matrix procs.
//
// Only S32_alpha_D32_filter_DX exploits instructions beyond
// our common baseline SSE2/NEON instruction sets, so that's
// all that lives here.
//
// The rest are scattershot at the moment but I want to get them
// all migrated to be normal code inside SkBitmapProcState.cpp.
#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
#include <immintrin.h>
#elif defined(SK_ARM_HAS_NEON)
#include <arm_neon.h>
#endif
namespace SK_OPTS_NS {
// This same basic packing scheme is used throughout the file.
template <typename U32, typename Out>
static void decode_packed_coordinates_and_weight(U32 packed, Out* v0, Out* v1, Out* w) {
*v0 = (packed >> 18); // Integer coordinate x0 or y0.
*v1 = (packed & 0x3fff); // Integer coordinate x1 or y1.
*w = (packed >> 14) & 0xf; // Lerp weight for v1; weight for v0 is 16-w.
}
#if 1 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2
/*not static*/ inline
void S32_alpha_D32_filter_DX(const SkBitmapProcState& s,
const uint32_t* xy, int count, uint32_t* colors) {
SkASSERT(count > 0 && colors != nullptr);
SkASSERT(s.fBilerp);
SkASSERT(kN32_SkColorType == s.fPixmap.colorType());
SkASSERT(s.fAlphaScale <= 256);
// In a _DX variant only X varies; all samples share y0/y1 coordinates and wy weight.
int y0, y1, wy;
decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
const uint32_t* row0 = s.fPixmap.addr32(0,y0);
const uint32_t* row1 = s.fPixmap.addr32(0,y1);
auto bilerp = [&](skvx::Vec<8,uint32_t> packed_x_coordinates) -> skvx::Vec<8,uint32_t> {
// Decode up to 8 output pixels' x-coordinates and weights.
skvx::Vec<8,uint32_t> x0,x1,wx;
decode_packed_coordinates_and_weight(packed_x_coordinates, &x0, &x1, &wx);
// Splat wx to each color channel.
wx = (wx << 0)
| (wx << 8)
| (wx << 16)
| (wx << 24);
auto gather = [](const uint32_t* ptr, skvx::Vec<8,uint32_t> ix) {
#if 1
// Drop into AVX2 intrinsics for vpgatherdd.
return skvx::bit_pun<skvx::Vec<8,uint32_t>>(
_mm256_i32gather_epi32((const int*)ptr, skvx::bit_pun<__m256i>(ix), 4));
#else
// Portable version... sometimes I don't trust vpgatherdd.
return skvx::Vec<8,uint32_t>{
ptr[ix[0]], ptr[ix[1]], ptr[ix[2]], ptr[ix[3]],
ptr[ix[4]], ptr[ix[5]], ptr[ix[6]], ptr[ix[7]],
};
#endif
};
// Gather the 32 32-bit pixels that we'll bilerp into our 8 output pixels.
skvx::Vec<8,uint32_t> tl = gather(row0, x0), tr = gather(row0, x1),
bl = gather(row1, x0), br = gather(row1, x1);
#if 1
// We'll use _mm256_maddubs_epi16() to lerp much like in the SSSE3 code.
auto lerp_x = [&](skvx::Vec<8,uint32_t> L, skvx::Vec<8,uint32_t> R) {
__m256i l = skvx::bit_pun<__m256i>(L),
r = skvx::bit_pun<__m256i>(R),
wr = skvx::bit_pun<__m256i>(wx),
wl = _mm256_sub_epi8(_mm256_set1_epi8(16), wr);
// Interlace l,r bytewise and line them up with their weights, then lerp.
__m256i lo = _mm256_maddubs_epi16(_mm256_unpacklo_epi8( l, r),
_mm256_unpacklo_epi8(wl,wr));
__m256i hi = _mm256_maddubs_epi16(_mm256_unpackhi_epi8( l, r),
_mm256_unpackhi_epi8(wl,wr));
// Those _mm256_unpack??_epi8() calls left us in a bit of an odd order:
//
// if l = a b c d | e f g h
// and r = A B C D | E F G H
//
// then lo = a A b B | e E f F (low half of each input)
// and hi = c C d D | g G h H (high half of each input)
//
// To get everything back in original order we need to transpose that.
__m256i abcd = _mm256_permute2x128_si256(lo, hi, 0x20),
efgh = _mm256_permute2x128_si256(lo, hi, 0x31);
return skvx::join(skvx::bit_pun<skvx::Vec<16,uint16_t>>(abcd),
skvx::bit_pun<skvx::Vec<16,uint16_t>>(efgh));
};
skvx::Vec<32, uint16_t> top = lerp_x(tl, tr),
bot = lerp_x(bl, br),
sum = 16*top + (bot-top)*wy;
#else
// Treat 32-bit pixels as 4 8-bit values, and expand to 16-bit for room to multiply.
auto to_16x4 = [](auto v) -> skvx::Vec<32, uint16_t> {
return skvx::cast<uint16_t>(skvx::bit_pun<skvx::Vec<32, uint8_t>>(v));
};
// Sum up weighted sample pixels. The naive, redundant math would be,
//
// sum = tl * (16-wy) * (16-wx)
// + bl * ( wy) * (16-wx)
// + tr * (16-wy) * ( wx)
// + br * ( wy) * ( wx)
//
// But we refactor to eliminate a bunch of those common factors.
auto lerp = [](auto lo, auto hi, auto w) {
return 16*lo + (hi-lo)*w;
};
skvx::Vec<32, uint16_t> sum = lerp(lerp(to_16x4(tl), to_16x4(bl), wy),
lerp(to_16x4(tr), to_16x4(br), wy), to_16x4(wx));
#endif
// Get back to [0,255] by dividing by maximum weight 16x16 = 256.
sum >>= 8;
// Scale by alpha if needed.
if(s.fAlphaScale < 256) {
sum *= s.fAlphaScale;
sum >>= 8;
}
// Pack back to 8-bit channels, undoing to_16x4().
return skvx::bit_pun<skvx::Vec<8,uint32_t>>(skvx::cast<uint8_t>(sum));
};
while (count >= 8) {
bilerp(skvx::Vec<8,uint32_t>::Load(xy)).store(colors);
xy += 8;
colors += 8;
count -= 8;
}
if (count > 0) {
__m256i active = skvx::bit_pun<__m256i>( count > skvx::Vec<8,int>{0,1,2,3, 4,5,6,7} ),
coords = _mm256_maskload_epi32((const int*)xy, active),
pixels;
bilerp(skvx::bit_pun<skvx::Vec<8,uint32_t>>(coords)).store(&pixels);
_mm256_maskstore_epi32((int*)colors, active, pixels);
sk_msan_mark_initialized(colors, colors+count,
"MSAN still doesn't understand AVX2 mask loads and stores.");
}
}
#elif 1 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
/*not static*/ inline
void S32_alpha_D32_filter_DX(const SkBitmapProcState& s,
const uint32_t* xy, int count, uint32_t* colors) {
SkASSERT(count > 0 && colors != nullptr);
SkASSERT(s.fBilerp);
SkASSERT(kN32_SkColorType == s.fPixmap.colorType());
SkASSERT(s.fAlphaScale <= 256);
// interpolate_in_x() is the crux of the SSSE3 implementation,
// interpolating in X for up to two output pixels (A and B) using _mm_maddubs_epi16().
auto interpolate_in_x = [](uint32_t A0, uint32_t A1,
uint32_t B0, uint32_t B1,
__m128i interlaced_x_weights) {
// _mm_maddubs_epi16() is a little idiosyncratic, but great as the core of a lerp.
//
// It takes two arguments interlaced byte-wise:
// - first arg: [ l,r, ... 7 more pairs of unsigned 8-bit values ...]
// - second arg: [ w,W, ... 7 more pairs of signed 8-bit values ...]
// and returns 8 signed 16-bit values: [ l*w + r*W, ... 7 more ... ].
//
// That's why we go to all this trouble to make interlaced_x_weights,
// and here we're about to interlace A0 with A1 and B0 with B1 to match.
//
// Our interlaced_x_weights are all in [0,16], and so we need not worry about
// the signedness of that input nor about the signedness of the output.
__m128i interlaced_A = _mm_unpacklo_epi8(_mm_cvtsi32_si128(A0), _mm_cvtsi32_si128(A1)),
interlaced_B = _mm_unpacklo_epi8(_mm_cvtsi32_si128(B0), _mm_cvtsi32_si128(B1));
return _mm_maddubs_epi16(_mm_unpacklo_epi64(interlaced_A, interlaced_B),
interlaced_x_weights);
};
// Interpolate {A0..A3} --> output pixel A, and {B0..B3} --> output pixel B.
// Returns two pixels, with each color channel in a 16-bit lane of the __m128i.
auto interpolate_in_x_and_y = [&](uint32_t A0, uint32_t A1,
uint32_t A2, uint32_t A3,
uint32_t B0, uint32_t B1,
uint32_t B2, uint32_t B3,
__m128i interlaced_x_weights,
int wy) {
// Interpolate each row in X, leaving 16-bit lanes scaled by interlaced_x_weights.
__m128i top = interpolate_in_x(A0,A1, B0,B1, interlaced_x_weights),
bot = interpolate_in_x(A2,A3, B2,B3, interlaced_x_weights);
// Interpolate in Y. As in the SSE2 code, we calculate top*(16-wy) + bot*wy
// as 16*top + (bot-top)*wy to save a multiply.
__m128i px = _mm_add_epi16(_mm_slli_epi16(top, 4),
_mm_mullo_epi16(_mm_sub_epi16(bot, top),
_mm_set1_epi16(wy)));
// Scale down by total max weight 16x16 = 256.
px = _mm_srli_epi16(px, 8);
// Scale by alpha if needed.
if (s.fAlphaScale < 256) {
px = _mm_srli_epi16(_mm_mullo_epi16(px, _mm_set1_epi16(s.fAlphaScale)), 8);
}
return px;
};
// We're in _DX mode here, so we're only varying in X.
// That means the first entry of xy is our constant pair of Y coordinates and weight in Y.
// All the other entries in xy will be pairs of X coordinates and the X weight.
int y0, y1, wy;
decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
auto row0 = (const uint32_t*)((const uint8_t*)s.fPixmap.addr() + y0 * s.fPixmap.rowBytes()),
row1 = (const uint32_t*)((const uint8_t*)s.fPixmap.addr() + y1 * s.fPixmap.rowBytes());
while (count >= 4) {
// We can really get going, loading 4 X-pairs at a time to produce 4 output pixels.
int x0[4],
x1[4];
__m128i wx;
// decode_packed_coordinates_and_weight(), 4x.
__m128i packed = _mm_loadu_si128((const __m128i*)xy);
_mm_storeu_si128((__m128i*)x0, _mm_srli_epi32(packed, 18));
_mm_storeu_si128((__m128i*)x1, _mm_and_si128 (packed, _mm_set1_epi32(0x3fff)));
wx = _mm_and_si128(_mm_srli_epi32(packed, 14), _mm_set1_epi32(0xf)); // [0,15]
// Splat each x weight 4x (for each color channel) as wr for pixels on the right at x1,
// and sixteen minus that as wl for pixels on the left at x0.
__m128i wr = _mm_shuffle_epi8(wx, _mm_setr_epi8(0,0,0,0,4,4,4,4,8,8,8,8,12,12,12,12)),
wl = _mm_sub_epi8(_mm_set1_epi8(16), wr);
// We need to interlace wl and wr for _mm_maddubs_epi16().
__m128i interlaced_x_weights_AB = _mm_unpacklo_epi8(wl,wr),
interlaced_x_weights_CD = _mm_unpackhi_epi8(wl,wr);
enum { A,B,C,D };
// interpolate_in_x_and_y() can produce two output pixels (A and B) at a time
// from eight input pixels {A0..A3} and {B0..B3}, arranged in a 2x2 grid for each.
__m128i AB = interpolate_in_x_and_y(row0[x0[A]], row0[x1[A]],
row1[x0[A]], row1[x1[A]],
row0[x0[B]], row0[x1[B]],
row1[x0[B]], row1[x1[B]],
interlaced_x_weights_AB, wy);
// Once more with the other half of the x-weights for two more pixels C,D.
__m128i CD = interpolate_in_x_and_y(row0[x0[C]], row0[x1[C]],
row1[x0[C]], row1[x1[C]],
row0[x0[D]], row0[x1[D]],
row1[x0[D]], row1[x1[D]],
interlaced_x_weights_CD, wy);
// Scale by alpha, pack back together to 8-bit lanes, and write out four pixels!
_mm_storeu_si128((__m128i*)colors, _mm_packus_epi16(AB, CD));
xy += 4;
colors += 4;
count -= 4;
}
while (count --> 0) {
// This is exactly the same flow as the count >= 4 loop above, but writing one pixel.
int x0, x1, wx;
decode_packed_coordinates_and_weight(*xy++, &x0, &x1, &wx);
// As above, splat out wx four times as wr, and sixteen minus that as wl.
__m128i wr = _mm_set1_epi8(wx), // This splats it out 16 times, but that's fine.
wl = _mm_sub_epi8(_mm_set1_epi8(16), wr);
__m128i interlaced_x_weights = _mm_unpacklo_epi8(wl, wr);
__m128i A = interpolate_in_x_and_y(row0[x0], row0[x1],
row1[x0], row1[x1],
0, 0,
0, 0,
interlaced_x_weights, wy);
*colors++ = _mm_cvtsi128_si32(_mm_packus_epi16(A, _mm_setzero_si128()));
}
}
#elif 1 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
/*not static*/ inline
void S32_alpha_D32_filter_DX(const SkBitmapProcState& s,
const uint32_t* xy, int count, uint32_t* colors) {
SkASSERT(count > 0 && colors != nullptr);
SkASSERT(s.fBilerp);
SkASSERT(kN32_SkColorType == s.fPixmap.colorType());
SkASSERT(s.fAlphaScale <= 256);
int y0, y1, wy;
decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
auto row0 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y0 * s.fPixmap.rowBytes() ),
row1 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y1 * s.fPixmap.rowBytes() );
// We'll put one pixel in the low 4 16-bit lanes to line up with wy,
// and another in the upper 4 16-bit lanes to line up with 16 - wy.
const __m128i allY = _mm_unpacklo_epi64(_mm_set1_epi16( wy), // Bottom pixel goes here.
_mm_set1_epi16(16-wy)); // Top pixel goes here.
while (count --> 0) {
int x0, x1, wx;
decode_packed_coordinates_and_weight(*xy++, &x0, &x1, &wx);
// Load the 4 pixels we're interpolating, in this grid:
// | tl tr |
// | bl br |
const __m128i tl = _mm_cvtsi32_si128(row0[x0]), tr = _mm_cvtsi32_si128(row0[x1]),
bl = _mm_cvtsi32_si128(row1[x0]), br = _mm_cvtsi32_si128(row1[x1]);
// We want to calculate a sum of 4 pixels weighted in two directions:
//
// sum = tl * (16-wy) * (16-wx)
// + bl * ( wy) * (16-wx)
// + tr * (16-wy) * ( wx)
// + br * ( wy) * ( wx)
//
// (Notice top --> 16-wy, bottom --> wy, left --> 16-wx, right --> wx.)
//
// We've already prepared allY as a vector containing [wy, 16-wy] as a way
// to apply those y-direction weights. So we'll start on the x-direction
// first, grouping into left and right halves, lined up with allY:
//
// L = [bl, tl]
// R = [br, tr]
//
// sum = horizontalSum( allY * (L*(16-wx) + R*wx) )
//
// Rewriting that one more step, we can replace a multiply with a shift:
//
// sum = horizontalSum( allY * (16*L + (R-L)*wx) )
//
// That's how we'll actually do this math.
__m128i L = _mm_unpacklo_epi8(_mm_unpacklo_epi32(bl, tl), _mm_setzero_si128()),
R = _mm_unpacklo_epi8(_mm_unpacklo_epi32(br, tr), _mm_setzero_si128());
__m128i inner = _mm_add_epi16(_mm_slli_epi16(L, 4),
_mm_mullo_epi16(_mm_sub_epi16(R,L), _mm_set1_epi16(wx)));
__m128i sum_in_x = _mm_mullo_epi16(inner, allY);
// sum = horizontalSum( ... )
__m128i sum = _mm_add_epi16(sum_in_x, _mm_srli_si128(sum_in_x, 8));
// Get back to [0,255] by dividing by maximum weight 16x16 = 256.
sum = _mm_srli_epi16(sum, 8);
if (s.fAlphaScale < 256) {
// Scale by alpha, which is in [0,256].
sum = _mm_mullo_epi16(sum, _mm_set1_epi16(s.fAlphaScale));
sum = _mm_srli_epi16(sum, 8);
}
// Pack back into 8-bit values and store.
*colors++ = _mm_cvtsi128_si32(_mm_packus_epi16(sum, _mm_setzero_si128()));
}
}
#else
// The NEON code only actually differs from the portable code in the
// filtering step after we've loaded all four pixels we want to bilerp.
#if defined(SK_ARM_HAS_NEON)
static void filter_and_scale_by_alpha(unsigned x, unsigned y,
SkPMColor a00, SkPMColor a01,
SkPMColor a10, SkPMColor a11,
SkPMColor *dst,
uint16_t scale) {
uint8x8_t vy, vconst16_8, v16_y, vres;
uint16x4_t vx, vconst16_16, v16_x, tmp, vscale;
uint32x2_t va0, va1;
uint16x8_t tmp1, tmp2;
vy = vdup_n_u8(y); // duplicate y into vy
vconst16_8 = vmov_n_u8(16); // set up constant in vconst16_8
v16_y = vsub_u8(vconst16_8, vy); // v16_y = 16-y
va0 = vdup_n_u32(a00); // duplicate a00
va1 = vdup_n_u32(a10); // duplicate a10
va0 = vset_lane_u32(a01, va0, 1); // set top to a01
va1 = vset_lane_u32(a11, va1, 1); // set top to a11
tmp1 = vmull_u8(vreinterpret_u8_u32(va0), v16_y); // tmp1 = [a01|a00] * (16-y)
tmp2 = vmull_u8(vreinterpret_u8_u32(va1), vy); // tmp2 = [a11|a10] * y
vx = vdup_n_u16(x); // duplicate x into vx
vconst16_16 = vmov_n_u16(16); // set up constant in vconst16_16
v16_x = vsub_u16(vconst16_16, vx); // v16_x = 16-x
tmp = vmul_u16(vget_high_u16(tmp1), vx); // tmp = a01 * x
tmp = vmla_u16(tmp, vget_high_u16(tmp2), vx); // tmp += a11 * x
tmp = vmla_u16(tmp, vget_low_u16(tmp1), v16_x); // tmp += a00 * (16-x)
tmp = vmla_u16(tmp, vget_low_u16(tmp2), v16_x); // tmp += a10 * (16-x)
if (scale < 256) {
vscale = vdup_n_u16(scale); // duplicate scale
tmp = vshr_n_u16(tmp, 8); // shift down result by 8
tmp = vmul_u16(tmp, vscale); // multiply result by scale
}
vres = vshrn_n_u16(vcombine_u16(tmp, vcreate_u16((uint64_t)0)), 8); // shift down result by 8
vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0); // store result
}
#else
static void filter_and_scale_by_alpha(unsigned x, unsigned y,
SkPMColor a00, SkPMColor a01,
SkPMColor a10, SkPMColor a11,
SkPMColor* dstColor,
unsigned alphaScale) {
SkASSERT((unsigned)x <= 0xF);
SkASSERT((unsigned)y <= 0xF);
SkASSERT(alphaScale <= 256);
int xy = x * y;
const uint32_t mask = 0xFF00FF;
int scale = 256 - 16*y - 16*x + xy;
uint32_t lo = (a00 & mask) * scale;
uint32_t hi = ((a00 >> 8) & mask) * scale;
scale = 16*x - xy;
lo += (a01 & mask) * scale;
hi += ((a01 >> 8) & mask) * scale;
scale = 16*y - xy;
lo += (a10 & mask) * scale;
hi += ((a10 >> 8) & mask) * scale;
lo += (a11 & mask) * xy;
hi += ((a11 >> 8) & mask) * xy;
if (alphaScale < 256) {
lo = ((lo >> 8) & mask) * alphaScale;
hi = ((hi >> 8) & mask) * alphaScale;
}
*dstColor = ((lo >> 8) & mask) | (hi & ~mask);
}
#endif
/*not static*/ inline
void S32_alpha_D32_filter_DX(const SkBitmapProcState& s,
const uint32_t* xy, int count, SkPMColor* colors) {
SkASSERT(count > 0 && colors != nullptr);
SkASSERT(s.fBilerp);
SkASSERT(4 == s.fPixmap.info().bytesPerPixel());
SkASSERT(s.fAlphaScale <= 256);
int y0, y1, wy;
decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
auto row0 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y0 * s.fPixmap.rowBytes() ),
row1 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y1 * s.fPixmap.rowBytes() );
while (count --> 0) {
int x0, x1, wx;
decode_packed_coordinates_and_weight(*xy++, &x0, &x1, &wx);
filter_and_scale_by_alpha(wx, wy,
row0[x0], row0[x1],
row1[x0], row1[x1],
colors++,
s.fAlphaScale);
}
}
#endif
#if defined(SK_ARM_HAS_NEON)
/*not static*/ inline
void S32_alpha_D32_filter_DXDY(const SkBitmapProcState& s,
const uint32_t* xy, int count, SkPMColor* colors) {
SkASSERT(count > 0 && colors != nullptr);
SkASSERT(s.fBilerp);
SkASSERT(4 == s.fPixmap.info().bytesPerPixel());
SkASSERT(s.fAlphaScale <= 256);
auto src = (const char*)s.fPixmap.addr();
size_t rb = s.fPixmap.rowBytes();
while (count --> 0) {
int y0, y1, wy,
x0, x1, wx;
decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
decode_packed_coordinates_and_weight(*xy++, &x0, &x1, &wx);
auto row0 = (const uint32_t*)(src + y0*rb),
row1 = (const uint32_t*)(src + y1*rb);
filter_and_scale_by_alpha(wx, wy,
row0[x0], row0[x1],
row1[x0], row1[x1],
colors++,
s.fAlphaScale);
}
}
#else
// It's not yet clear whether it's worthwhile specializing for SSE2/SSSE3/AVX2.
constexpr static void (*S32_alpha_D32_filter_DXDY)(const SkBitmapProcState&,
const uint32_t*, int, SkPMColor*) = nullptr;
#endif
} // namespace SK_OPTS_NS
#endif