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 Convert. avx2 implementation

Browse Source
This commit is contained in:
homm 2016-07-02 16:11:10 +03:00 committed by Alexander Karpinsky
parent 6e91b2b41f
commit 55b71f2edb
1 changed files with 47 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
src/libImaging/Convert.c
View File

@ -37,6 +37,9 @@
#include <emmintrin.h> #include <emmintrin.h>
#include <mmintrin.h> #include <mmintrin.h>
#include <smmintrin.h> #include <smmintrin.h>
#if defined(__AVX2__)
#include <immintrin.h>
#endif
#define MAX(a, b) (a)>(b) ? (a) : (b) #define MAX(a, b) (a)>(b) ? (a) : (b)
@ -463,10 +466,47 @@ rgbA2rgba(UINT8* out, const UINT8* in, int xsize)
unsigned int tmp; unsigned int tmp;
unsigned char alpha; unsigned char alpha;
int x = 0; int x = 0;
#if defined(__AVX2__)
__m256i zero = _mm256_setzero_si256();
__m256i half = _mm256_set1_epi16(128);
__m256i maxalpha = _mm256_set_epi32(
0xff000000, 0xff000000, 0xff000000, 0xff000000,
0xff000000, 0xff000000, 0xff000000, 0xff000000);
__m256i factormask = _mm256_set_epi8(
15,15,15,15, 11,11,11,11, 7,7,7,7, 3,3,3,3,
15,15,15,15, 11,11,11,11, 7,7,7,7, 3,3,3,3);
__m256i factorsource, source, pix1, pix2, factors;
for (; x < xsize - 7; x += 8) {
source = _mm256_loadu_si256((__m256i *) &in[x * 4]);
factorsource = _mm256_shuffle_epi8(source, factormask);
factorsource = _mm256_or_si256(factorsource, maxalpha);
pix1 = _mm256_unpacklo_epi8(source, zero);
factors = _mm256_unpacklo_epi8(factorsource, zero);
pix1 = _mm256_add_epi16(_mm256_mullo_epi16(pix1, factors), half);
pix1 = _mm256_add_epi16(pix1, _mm256_srli_epi16(pix1, 8));
pix1 = _mm256_srli_epi16(pix1, 8);
pix2 = _mm256_unpackhi_epi8(source, zero);
factors = _mm256_unpackhi_epi8(factorsource, zero);
pix2 = _mm256_add_epi16(_mm256_mullo_epi16(pix2, factors), half);
pix2 = _mm256_add_epi16(pix2, _mm256_srli_epi16(pix2, 8));
pix2 = _mm256_srli_epi16(pix2, 8);
source = _mm256_packus_epi16(pix1, pix2);
_mm256_storeu_si256((__m256i *) &out[x * 4], source);
}
#else
__m128i zero = _mm_setzero_si128(); __m128i zero = _mm_setzero_si128();
__m128i half = _mm_set1_epi16(128); __m128i half = _mm_set1_epi16(128);
__m128i maxalpha = _mm_set1_epi32(0xff000000); __m128i maxalpha = _mm_set1_epi32(0xff000000);
__m128i factormask = _mm_set_epi8(15,15,15,15, 11,11,11,11, 7,7,7,7, 3,3,3,3); __m128i factormask = _mm_set_epi8(
15,15,15,15, 11,11,11,11, 7,7,7,7, 3,3,3,3);
__m128i factorsource, source, pix1, pix2, factors; __m128i factorsource, source, pix1, pix2, factors;
for (; x < xsize - 3; x += 4) { for (; x < xsize - 3; x += 4) {
@ -486,8 +526,12 @@ rgbA2rgba(UINT8* out, const UINT8* in, int xsize)
pix2 = _mm_add_epi16(pix2, _mm_srli_epi16(pix2, 8)); pix2 = _mm_add_epi16(pix2, _mm_srli_epi16(pix2, 8));
pix2 = _mm_srli_epi16(pix2, 8); pix2 = _mm_srli_epi16(pix2, 8);
_mm_storeu_si128((__m128i *) &out[x * 4], _mm_packus_epi16(pix1, pix2)); source = _mm_packus_epi16(pix1, pix2);
_mm_storeu_si128((__m128i *) &out[x * 4], source);
} }
#endif
for (; x < xsize; x++) { for (; x < xsize; x++) {
alpha = in[x * 4 + 3]; alpha = in[x * 4 + 3];
out[x * 4 + 0] = MULDIV255(in[x * 4 + 0], alpha, tmp); out[x * 4 + 0] = MULDIV255(in[x * 4 + 0], alpha, tmp);