python-pillow/Pillow
1
1
Fork 0
mirror of https://github.com/python-pillow/Pillow.git synced 2025-08-20 12:14:46 +03:00
Code Issues Packages Projects Releases Wiki Activity

SIMD ColorLUT. remove left_mask and right_mask

Browse Source
This commit is contained in:
Alexander 2018-03-27 13:10:20 +03:00 committed by Alexander Karpinsky
parent 0520521560
commit b4a77d406c
1 changed files with 32 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
src/libImaging/ColorLUT.c
View File

@ -60,8 +60,6 @@ ImagingColorLUT3D_linear(
size1D*table_channels, table_channels); size1D*table_channels, table_channels);
__m128i shuffle3 = _mm_set_epi8(-1,-1, -1,-1, 11,10, 5,4, 9,8, 3,2, 7,6, 1,0); __m128i shuffle3 = _mm_set_epi8(-1,-1, -1,-1, 11,10, 5,4, 9,8, 3,2, 7,6, 1,0);
__m128i shuffle4 = _mm_set_epi8(15,14, 7,6, 13,12, 5,4, 11,10, 3,2, 9,8, 1,0); __m128i shuffle4 = _mm_set_epi8(15,14, 7,6, 13,12, 5,4, 11,10, 3,2, 9,8, 1,0);
__m128i left_mask = _mm_set1_epi32(0x0000ffff);
__m128i right_mask = _mm_set1_epi32(0xffff0000);
int x, y; int x, y;
ImagingSectionCookie cookie; ImagingSectionCookie cookie;
@ -114,33 +112,34 @@ ImagingColorLUT3D_linear(
if (table_channels == 3) { if (table_channels == 3) {
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + 0]), shuffle3); _mm_loadu_si128((__m128i *) &table[idx + 0]), shuffle3);
leftleft = _mm_and_si128(_mm_srai_epi32(_mm_madd_epi16( leftleft = _mm_srai_epi32(_mm_madd_epi16(
source, shift1D), SHIFT_BITS), left_mask); source, shift1D), SHIFT_BITS);
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + size1D*3]), shuffle3); _mm_loadu_si128((__m128i *) &table[idx + size1D*3]), shuffle3);
leftright = _mm_and_si128(_mm_slli_epi32(_mm_madd_epi16( leftright = _mm_slli_epi32(_mm_madd_epi16(
source, shift1D), 16 - SHIFT_BITS), right_mask); source, shift1D), 16 - SHIFT_BITS);
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + size1D_2D*3]), shuffle3); _mm_loadu_si128((__m128i *) &table[idx + size1D_2D*3]), shuffle3);
rightleft = _mm_and_si128(_mm_srai_epi32(_mm_madd_epi16( rightleft = _mm_srai_epi32(_mm_madd_epi16(
source, shift1D), SHIFT_BITS), left_mask); source, shift1D), SHIFT_BITS);
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + size1D_2D*3 + size1D*3]), shuffle3); _mm_loadu_si128((__m128i *) &table[idx + size1D_2D*3 + size1D*3]), shuffle3);
rightright = _mm_and_si128(_mm_slli_epi32(_mm_madd_epi16( rightright = _mm_slli_epi32(_mm_madd_epi16(
source, shift1D), 16 - SHIFT_BITS), right_mask); source, shift1D), 16 - SHIFT_BITS);
left = _mm_and_si128(_mm_srai_epi32(_mm_madd_epi16( left = _mm_srai_epi32(_mm_madd_epi16(
_mm_or_si128(leftleft, leftright), shift2D), _mm_blend_epi16(leftleft, leftright, 0xaa), shift2D),
SHIFT_BITS), left_mask); SHIFT_BITS);
right = _mm_and_si128(_mm_slli_epi32(_mm_madd_epi16( right = _mm_slli_epi32(_mm_madd_epi16(
_mm_or_si128(rightleft, rightright), shift2D), _mm_blend_epi16(rightleft, rightright, 0xaa), shift2D),
16 - SHIFT_BITS), right_mask); 16 - SHIFT_BITS);
result = _mm_madd_epi16(_mm_or_si128(left, right), shift3D); result = _mm_madd_epi16(
_mm_blend_epi16(left, right, 0xaa), shift3D);
result = _mm_srai_epi32(_mm_add_epi32( result = _mm_srai_epi32(_mm_add_epi32(
_mm_set1_epi32(PRECISION_ROUNDING<<SHIFT_BITS), result), _mm_set1_epi32(PRECISION_ROUNDING<<SHIFT_BITS), result),
@ -153,33 +152,34 @@ ImagingColorLUT3D_linear(
if (table_channels == 4) { if (table_channels == 4) {
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + 0]), shuffle4); _mm_loadu_si128((__m128i *) &table[idx + 0]), shuffle4);
leftleft = _mm_and_si128(_mm_srai_epi32(_mm_madd_epi16( leftleft = _mm_srai_epi32(_mm_madd_epi16(
source, shift1D), SHIFT_BITS), left_mask); source, shift1D), SHIFT_BITS);
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + size1D*4]), shuffle4); _mm_loadu_si128((__m128i *) &table[idx + size1D*4]), shuffle4);
leftright = _mm_and_si128(_mm_slli_epi32(_mm_madd_epi16( leftright = _mm_slli_epi32(_mm_madd_epi16(
source, shift1D), 16 - SHIFT_BITS), right_mask); source, shift1D), 16 - SHIFT_BITS);
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + size1D_2D*4]), shuffle4); _mm_loadu_si128((__m128i *) &table[idx + size1D_2D*4]), shuffle4);
rightleft = _mm_and_si128(_mm_srai_epi32(_mm_madd_epi16( rightleft = _mm_srai_epi32(_mm_madd_epi16(
source, shift1D), SHIFT_BITS), left_mask); source, shift1D), SHIFT_BITS);
source = _mm_shuffle_epi8( source = _mm_shuffle_epi8(
_mm_loadu_si128((__m128i *) &table[idx + size1D_2D*4 + size1D*4]), shuffle4); _mm_loadu_si128((__m128i *) &table[idx + size1D_2D*4 + size1D*4]), shuffle4);
rightright = _mm_and_si128(_mm_slli_epi32(_mm_madd_epi16( rightright = _mm_slli_epi32(_mm_madd_epi16(
source, shift1D), 16 - SHIFT_BITS), right_mask); source, shift1D), 16 - SHIFT_BITS);
left = _mm_and_si128(_mm_srai_epi32(_mm_madd_epi16( left = _mm_srai_epi32(_mm_madd_epi16(
_mm_or_si128(leftleft, leftright), shift2D), _mm_blend_epi16(leftleft, leftright, 0xaa), shift2D),
SHIFT_BITS), left_mask); SHIFT_BITS);
right = _mm_and_si128(_mm_slli_epi32(_mm_madd_epi16( right = _mm_slli_epi32(_mm_madd_epi16(
_mm_or_si128(rightleft, rightright), shift2D), _mm_blend_epi16(rightleft, rightright, 0xaa), shift2D),
16 - SHIFT_BITS), right_mask); 16 - SHIFT_BITS);
result = _mm_madd_epi16(_mm_or_si128(left, right), shift3D); result = _mm_madd_epi16(
_mm_blend_epi16(left, right, 0xaa), shift3D);
result = _mm_srai_epi32(_mm_add_epi32( result = _mm_srai_epi32(_mm_add_epi32(
_mm_set1_epi32(PRECISION_ROUNDING<<SHIFT_BITS), result), _mm_set1_epi32(PRECISION_ROUNDING<<SHIFT_BITS), result),