python-pillow/Pillow
1
1
Fork 0
mirror of https://github.com/python-pillow/Pillow.git synced 2025-08-21 04:34:47 +03:00
Code Issues Packages Projects Releases Wiki Activity

SIMD AlphaComposite. speedup avx2 by using _mm256_mullo_epi16 instead of _mm256_mullo_epi32

Browse Source
This commit is contained in:
homm 2016-06-19 16:46:10 +03:00 committed by Alexander Karpinsky
parent 594b5d00cb
commit f64d6a19c1
1 changed files with 19 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
src/libImaging/AlphaComposite.c
View File

@ -34,6 +34,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
int xsize = imDst->xsize; int xsize = imDst->xsize;
__m128i mm_max_alpha = _mm_set1_epi32(255); __m128i mm_max_alpha = _mm_set1_epi32(255);
__m128i mm_max_alpha2 = _mm_set1_epi32(255 * 255); __m128i mm_max_alpha2 = _mm_set1_epi32(255 * 255);
__m128i mm_zero = _mm_setzero_si128();
__m128i mm_half = _mm_set1_epi32(128); __m128i mm_half = _mm_set1_epi32(128);
__m128i mm_get_lo = _mm_set_epi8( __m128i mm_get_lo = _mm_set_epi8(
-1,-1, 5,4, 5,4, 5,4, -1,-1, 1,0, 1,0, 1,0); -1,-1, 5,4, 5,4, 5,4, -1,-1, 1,0, 1,0, 1,0);
@ -42,6 +43,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
#if defined(__AVX2__) #if defined(__AVX2__)
__m256i vmm_max_alpha = _mm256_set1_epi32(255); __m256i vmm_max_alpha = _mm256_set1_epi32(255);
__m256i vmm_max_alpha2 = _mm256_set1_epi32(255 * 255); __m256i vmm_max_alpha2 = _mm256_set1_epi32(255 * 255);
__m256i vmm_zero = _mm256_setzero_si256();
__m256i vmm_half = _mm256_set1_epi32(128); __m256i vmm_half = _mm256_set1_epi32(128);
__m256i vmm_get_lo = _mm256_set_epi8( __m256i vmm_get_lo = _mm256_set_epi8(
-1,-1, 5,4, 5,4, 5,4, -1,-1, 1,0, 1,0, 1,0, -1,-1, 5,4, 5,4, 5,4, -1,-1, 1,0, 1,0, 1,0,
@ -83,21 +85,21 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
for (; x < xsize - 7; x += 8) { for (; x < xsize - 7; x += 8) {
__m256i mm_dst = _mm256_loadu_si256((__m256i *) &dst[x]); __m256i mm_dst = _mm256_loadu_si256((__m256i *) &dst[x]);
__m256i mm_dst_lo = _mm256_unpacklo_epi8(mm_dst, _mm256_setzero_si256()); __m256i mm_dst_lo = _mm256_unpacklo_epi8(mm_dst, vmm_zero);
__m256i mm_dst_hi = _mm256_unpackhi_epi8(mm_dst, _mm256_setzero_si256()); __m256i mm_dst_hi = _mm256_unpackhi_epi8(mm_dst, vmm_zero);
__m256i mm_src = _mm256_loadu_si256((__m256i *) &src[x]); __m256i mm_src = _mm256_loadu_si256((__m256i *) &src[x]);
__m256i mm_src_lo = _mm256_unpacklo_epi8(mm_src, _mm256_setzero_si256()); __m256i mm_src_lo = _mm256_unpacklo_epi8(mm_src, vmm_zero);
__m256i mm_src_hi = _mm256_unpackhi_epi8(mm_src, _mm256_setzero_si256()); __m256i mm_src_hi = _mm256_unpackhi_epi8(mm_src, vmm_zero);
__m256i mm_dst_a = _mm256_srli_epi32(mm_dst, 24); __m256i mm_dst_a = _mm256_srli_epi32(mm_dst, 24);
__m256i mm_src_a = _mm256_srli_epi32(mm_src, 24); __m256i mm_src_a = _mm256_srli_epi32(mm_src, 24);
// Compute coefficients // Compute coefficients
// blend = dst->a * (255 - src->a); 16 bits // blend = dst->a * (255 - src->a); 16 bits
__m256i mm_blend = _mm256_mullo_epi32(mm_dst_a, _mm256_sub_epi32(vmm_max_alpha, mm_src_a)); __m256i mm_blend = _mm256_mullo_epi16(mm_dst_a, _mm256_sub_epi32(vmm_max_alpha, mm_src_a));
// outa = src->a * 255 + dst->a * (255 - src->a); 16 bits // outa = src->a * 255 + dst->a * (255 - src->a); 16 bits
__m256i mm_outa = _mm256_add_epi32(_mm256_mullo_epi32(mm_src_a, vmm_max_alpha), mm_blend); __m256i mm_outa = _mm256_add_epi32(_mm256_mullo_epi16(mm_src_a, vmm_max_alpha), mm_blend);
__m256i mm_coef1 = _mm256_mullo_epi32(mm_src_a, vmm_max_alpha2); __m256i mm_coef1 = _mm256_mullo_epi16(mm_src_a, vmm_max_alpha2);
// 8 bits // 8 bits
mm_coef1 = _mm256_cvtps_epi32(_mm256_div_ps(_mm256_cvtepi32_ps(mm_coef1), _mm256_cvtepi32_ps(mm_outa))); mm_coef1 = _mm256_cvtps_epi32(_mm256_div_ps(_mm256_cvtepi32_ps(mm_coef1), _mm256_cvtepi32_ps(mm_outa)));
// 8 bits // 8 bits
@ -107,7 +109,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
_mm256_mullo_epi16(mm_src_lo, _mm256_shuffle_epi8(mm_coef1, vmm_get_lo)), _mm256_mullo_epi16(mm_src_lo, _mm256_shuffle_epi8(mm_coef1, vmm_get_lo)),
_mm256_mullo_epi16(mm_dst_lo, _mm256_shuffle_epi8(mm_coef2, vmm_get_lo))); _mm256_mullo_epi16(mm_dst_lo, _mm256_shuffle_epi8(mm_coef2, vmm_get_lo)));
mm_out_lo = _mm256_or_si256(mm_out_lo, _mm256_slli_epi64( mm_out_lo = _mm256_or_si256(mm_out_lo, _mm256_slli_epi64(
_mm256_unpacklo_epi32(mm_outa, _mm256_setzero_si256()), 48)); _mm256_unpacklo_epi32(mm_outa, vmm_zero), 48));
mm_out_lo = _mm256_add_epi32(mm_out_lo, vmm_half); mm_out_lo = _mm256_add_epi32(mm_out_lo, vmm_half);
mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo); mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo);
@ -115,7 +117,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
_mm256_mullo_epi16(mm_src_hi, _mm256_shuffle_epi8(mm_coef1, vmm_get_hi)), _mm256_mullo_epi16(mm_src_hi, _mm256_shuffle_epi8(mm_coef1, vmm_get_hi)),
_mm256_mullo_epi16(mm_dst_hi, _mm256_shuffle_epi8(mm_coef2, vmm_get_hi))); _mm256_mullo_epi16(mm_dst_hi, _mm256_shuffle_epi8(mm_coef2, vmm_get_hi)));
mm_out_hi = _mm256_or_si256(mm_out_hi, _mm256_slli_epi64( mm_out_hi = _mm256_or_si256(mm_out_hi, _mm256_slli_epi64(
_mm256_unpackhi_epi32(mm_outa, _mm256_setzero_si256()), 48)); _mm256_unpackhi_epi32(mm_outa, vmm_zero), 48));
mm_out_hi = _mm256_add_epi32(mm_out_hi, vmm_half); mm_out_hi = _mm256_add_epi32(mm_out_hi, vmm_half);
mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi); mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi);
@ -134,13 +136,13 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
__m128i mm_dst = _mm_loadu_si128((__m128i *) &dst[x]); __m128i mm_dst = _mm_loadu_si128((__m128i *) &dst[x]);
// [16] a1 b1 g1 r1 a0 b0 g0 r0 // [16] a1 b1 g1 r1 a0 b0 g0 r0
__m128i mm_dst_lo = _mm_unpacklo_epi8(mm_dst, _mm_setzero_si128()); __m128i mm_dst_lo = _mm_unpacklo_epi8(mm_dst, mm_zero);
// [16] a3 b3 g3 r3 a2 b2 g2 r2 // [16] a3 b3 g3 r3 a2 b2 g2 r2
__m128i mm_dst_hi = _mm_unpackhi_epi8(mm_dst, _mm_setzero_si128()); __m128i mm_dst_hi = _mm_unpackhi_epi8(mm_dst, mm_zero);
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
__m128i mm_src = _mm_loadu_si128((__m128i *) &src[x]); __m128i mm_src = _mm_loadu_si128((__m128i *) &src[x]);
__m128i mm_src_lo = _mm_unpacklo_epi8(mm_src, _mm_setzero_si128()); __m128i mm_src_lo = _mm_unpacklo_epi8(mm_src, mm_zero);
__m128i mm_src_hi = _mm_unpackhi_epi8(mm_src, _mm_setzero_si128()); __m128i mm_src_hi = _mm_unpackhi_epi8(mm_src, mm_zero);
// [32] a3 a2 a1 a0 // [32] a3 a2 a1 a0
__m128i mm_dst_a = _mm_srli_epi32(mm_dst, 24); __m128i mm_dst_a = _mm_srli_epi32(mm_dst, 24);
@ -154,16 +156,16 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
// [16] xx a3 xx a2 xx a1 xx a0 // [16] xx a3 xx a2 xx a1 xx a0
__m128i mm_outa = _mm_add_epi32(_mm_mullo_epi16(mm_src_a, mm_max_alpha), mm_blend); __m128i mm_outa = _mm_add_epi32(_mm_mullo_epi16(mm_src_a, mm_max_alpha), mm_blend);
__m128i mm_coef1 = _mm_mullo_epi16(mm_src_a, mm_max_alpha2); __m128i mm_coef1 = _mm_mullo_epi16(mm_src_a, mm_max_alpha2);
// 8 bits // [8] xx xx xx c3 xx xx xx c2 xx xx xx c1 xx xx xx c0
mm_coef1 = _mm_cvtps_epi32(_mm_div_ps(_mm_cvtepi32_ps(mm_coef1), _mm_cvtepi32_ps(mm_outa))); mm_coef1 = _mm_cvtps_epi32(_mm_div_ps(_mm_cvtepi32_ps(mm_coef1), _mm_cvtepi32_ps(mm_outa)));
// 8 bits // [8] xx xx xx c3 xx xx xx c2 xx xx xx c1 xx xx xx c0
__m128i mm_coef2 = _mm_sub_epi32(mm_max_alpha, mm_coef1); __m128i mm_coef2 = _mm_sub_epi32(mm_max_alpha, mm_coef1);
__m128i mm_out_lo = _mm_add_epi16( __m128i mm_out_lo = _mm_add_epi16(
_mm_mullo_epi16(mm_src_lo, _mm_shuffle_epi8(mm_coef1, mm_get_lo)), _mm_mullo_epi16(mm_src_lo, _mm_shuffle_epi8(mm_coef1, mm_get_lo)),
_mm_mullo_epi16(mm_dst_lo, _mm_shuffle_epi8(mm_coef2, mm_get_lo))); _mm_mullo_epi16(mm_dst_lo, _mm_shuffle_epi8(mm_coef2, mm_get_lo)));
mm_out_lo = _mm_or_si128(mm_out_lo, _mm_slli_epi64( mm_out_lo = _mm_or_si128(mm_out_lo, _mm_slli_epi64(
_mm_unpacklo_epi32(mm_outa, _mm_setzero_si128()), 48)); _mm_unpacklo_epi32(mm_outa, mm_zero), 48));
mm_out_lo = _mm_add_epi32(mm_out_lo, mm_half); mm_out_lo = _mm_add_epi32(mm_out_lo, mm_half);
mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo); mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo);
@ -171,7 +173,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
_mm_mullo_epi16(mm_src_hi, _mm_shuffle_epi8(mm_coef1, mm_get_hi)), _mm_mullo_epi16(mm_src_hi, _mm_shuffle_epi8(mm_coef1, mm_get_hi)),
_mm_mullo_epi16(mm_dst_hi, _mm_shuffle_epi8(mm_coef2, mm_get_hi))); _mm_mullo_epi16(mm_dst_hi, _mm_shuffle_epi8(mm_coef2, mm_get_hi)));
mm_out_hi = _mm_or_si128(mm_out_hi, _mm_slli_epi64( mm_out_hi = _mm_or_si128(mm_out_hi, _mm_slli_epi64(
_mm_unpackhi_epi32(mm_outa, _mm_setzero_si128()), 48)); _mm_unpackhi_epi32(mm_outa, mm_zero), 48));
mm_out_hi = _mm_add_epi32(mm_out_hi, mm_half); mm_out_hi = _mm_add_epi32(mm_out_hi, mm_half);
mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi); mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi);