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. fix bugs

Browse Source
This commit is contained in:
homm 2016-07-03 23:46:33 +03:00 committed by Alexander Karpinsky
parent f64d6a19c1
commit f7f4e9e448
1 changed files with 10 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
src/libImaging/AlphaComposite.c
View File

@ -35,7 +35,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
__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_zero = _mm_setzero_si128();
__m128i mm_half = _mm_set1_epi32(128); __m128i mm_half = _mm_set1_epi16(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);
__m128i mm_get_hi = _mm_set_epi8( __m128i mm_get_hi = _mm_set_epi8(
@ -44,7 +44,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
__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_zero = _mm256_setzero_si256();
__m256i vmm_half = _mm256_set1_epi32(128); __m256i vmm_half = _mm256_set1_epi16(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,
-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);
@ -99,7 +99,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
__m256i mm_blend = _mm256_mullo_epi16(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_epi16(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_epi16(mm_src_a, vmm_max_alpha2); __m256i mm_coef1 = _mm256_mullo_epi32(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
@ -110,7 +110,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
_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, vmm_zero), 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_epi16(mm_out_lo, vmm_half);
mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo); mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo);
__m256i mm_out_hi = _mm256_add_epi16( __m256i mm_out_hi = _mm256_add_epi16(
@ -118,7 +118,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
_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, vmm_zero), 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_epi16(mm_out_hi, vmm_half);
mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi); mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi);
_mm256_storeu_si256((__m256i *) &out[x], _mm256_storeu_si256((__m256i *) &out[x],
@ -152,10 +152,11 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
// blend = dst->a * (255 - src->a) // blend = dst->a * (255 - src->a)
// [16] xx b3 xx b2 xx b1 xx b0 // [16] xx b3 xx b2 xx b1 xx b0
__m128i mm_blend = _mm_mullo_epi16(mm_dst_a, _mm_sub_epi32(mm_max_alpha, mm_src_a)); __m128i mm_blend = _mm_mullo_epi16(mm_dst_a, _mm_sub_epi32(mm_max_alpha, mm_src_a));
// outa = src->a * 255 + dst->a * 255 - src->a * dst->a // outa = src->a * 255 + blend
// [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); // coef1 = src->a * 255 * 255 / outa
__m128i mm_coef1 = _mm_mullo_epi32(mm_src_a, mm_max_alpha2);
// [8] xx xx xx c3 xx xx xx c2 xx xx xx c1 xx xx xx c0 // [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] xx xx xx c3 xx xx xx c2 xx xx xx c1 xx xx xx c0 // [8] xx xx xx c3 xx xx xx c2 xx xx xx c1 xx xx xx c0
@ -166,7 +167,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
_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_zero), 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_epi16(mm_out_lo, mm_half);
mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo); mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo);
__m128i mm_out_hi = _mm_add_epi16( __m128i mm_out_hi = _mm_add_epi16(
@ -174,7 +175,7 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) {
_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_zero), 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_epi16(mm_out_hi, mm_half);
mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi); mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi);
_mm_storeu_si128((__m128i *) &out[x], _mm_storeu_si128((__m128i *) &out[x],