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 Resample. optimize coefficients loading for horizontal pass

Browse Source 
wtf is xmax / 2

optimize coefficients loading for vertical pass
This commit is contained in:
Alexander 2017-02-25 11:26:24 +03:00 committed by Alexander Karpinsky
parent 8796b40ef6
commit 698e6a1698
2 changed files with 30 additions and 56 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
libImaging/ResampleSIMDHorizontalConv.c
View File

@ -17,22 +17,15 @@ ImagingResampleHorizontalConvolution8u4x(
__m256i sss0, sss1; __m256i sss0, sss1;
__m256i zero = _mm256_setzero_si256(); __m256i zero = _mm256_setzero_si256();
__m256i initial = _mm256_set1_epi32((1 << (coefs_precision -1))); __m256i initial = _mm256_set1_epi32(1 << (coefs_precision -1));
sss0 = initial; sss0 = initial;
sss1 = initial; sss1 = initial;
for (; x < xmax - 3; x += 4) { for (; x < xmax - 3; x += 4) {
__m256i pix, mmk0, mmk1, source, ksource; __m256i pix, mmk0, mmk1, source;
__m128i tmp = _mm_loadl_epi64((__m128i *) &k[x]);
ksource = _mm256_insertf128_si256(
_mm256_castsi128_si256(tmp), tmp, 1);
mmk0 = _mm256_shuffle_epi8(ksource, _mm256_set_epi8( mmk0 = _mm256_set1_epi32(*(INT32 *) &k[x]);
3,2, 1,0, 3,2, 1,0, 3,2, 1,0, 3,2, 1,0, mmk1 = _mm256_set1_epi32(*(INT32 *) &k[x + 2]);
3,2, 1,0, 3,2, 1,0, 3,2, 1,0, 3,2, 1,0));
mmk1 = _mm256_shuffle_epi8(ksource, _mm256_set_epi8(
7,6, 5,4, 7,6, 5,4, 7,6, 5,4, 7,6, 5,4,
7,6, 5,4, 7,6, 5,4, 7,6, 5,4, 7,6, 5,4));
source = _mm256_inserti128_si256(_mm256_castsi128_si256( source = _mm256_inserti128_si256(_mm256_castsi128_si256(
_mm_loadu_si128((__m128i *) &lineIn0[x + xmin])), _mm_loadu_si128((__m128i *) &lineIn0[x + xmin])),
@ -61,10 +54,8 @@ ImagingResampleHorizontalConvolution8u4x(
for (; x < xmax - 1; x += 2) { for (; x < xmax - 1; x += 2) {
__m256i pix, mmk; __m256i pix, mmk;
__m128i ksource = _mm_cvtsi32_si128(*(int *) &k[x]);
ksource = _mm_shuffle_epi8(ksource, _mm_set_epi8( mmk = _mm256_set1_epi32(*(INT32 *) &k[x]);
3,2, 1,0, 3,2, 1,0, 3,2, 1,0, 3,2, 1,0));
mmk = _mm256_inserti128_si256(_mm256_castsi128_si256(ksource), ksource, 1);
pix = _mm256_inserti128_si256(_mm256_castsi128_si256( pix = _mm256_inserti128_si256(_mm256_castsi128_si256(
_mm_loadl_epi64((__m128i *) &lineIn0[x + xmin])), _mm_loadl_epi64((__m128i *) &lineIn0[x + xmin])),
@ -115,7 +106,7 @@ ImagingResampleHorizontalConvolution8u4x(
#else #else
__m128i sss0, sss1, sss2, sss3; __m128i sss0, sss1, sss2, sss3;
__m128i initial = _mm_set1_epi32((1 << (coefs_precision -1))); __m128i initial = _mm_set1_epi32(1 << (coefs_precision -1));
sss0 = initial; sss0 = initial;
sss1 = initial; sss1 = initial;
sss2 = initial; sss2 = initial;
@ -128,13 +119,8 @@ ImagingResampleHorizontalConvolution8u4x(
__m128i mask_hi = _mm_set_epi8( __m128i mask_hi = _mm_set_epi8(
-1,15, -1,11, -1,14, -1,10, -1,13, -1,9, -1,12, -1,8); -1,15, -1,11, -1,14, -1,10, -1,13, -1,9, -1,12, -1,8);
// [16] xx xx xx xx k3 k2 k1 k0 mmk_lo = _mm_set1_epi32(*(INT32 *) &k[x]);
__m128i ksource = _mm_loadl_epi64((__m128i *) &k[x]); mmk_hi = _mm_set1_epi32(*(INT32 *) &k[x + 2]);
// [16] k1 k0 k1 k0 k1 k0 k1 k0
mmk_lo = _mm_shuffle_epi8(ksource, _mm_set_epi8(
3,2, 1,0, 3,2, 1,0, 3,2, 1,0, 3,2, 1,0));
mmk_hi = _mm_shuffle_epi8(ksource, _mm_set_epi8(
7,6, 5,4, 7,6, 5,4, 7,6, 5,4, 7,6, 5,4));
// [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
source = _mm_loadu_si128((__m128i *) &lineIn0[x + xmin]); source = _mm_loadu_si128((__m128i *) &lineIn0[x + xmin]);
@ -170,7 +156,7 @@ ImagingResampleHorizontalConvolution8u4x(
-1,7, -1,3, -1,6, -1,2, -1,5, -1,1, -1,4, -1,0); -1,7, -1,3, -1,6, -1,2, -1,5, -1,1, -1,4, -1,0);
// [16] k1 k0 k1 k0 k1 k0 k1 k0 // [16] k1 k0 k1 k0 k1 k0 k1 k0
mmk = _mm_set1_epi32(*(int *) &k[x]); mmk = _mm_set1_epi32(*(INT32 *) &k[x]);
// [8] x x x x x x x x a1 b1 g1 r1 a0 b0 g0 r0 // [8] x x x x x x x x a1 b1 g1 r1 a0 b0 g0 r0
pix = _mm_loadl_epi64((__m128i *) &lineIn0[x + xmin]); pix = _mm_loadl_epi64((__m128i *) &lineIn0[x + xmin]);
@ -359,14 +345,10 @@ ImagingResampleHorizontalConvolution8u(UINT32 *lineOut, UINT32 *lineIn,
#endif #endif
for (; x < xmax - 1; x += 2) { for (; x < xmax - 1; x += 2) {
__m128i pix, mmk; __m128i mmk = _mm_set1_epi32(*(INT32 *) &k[x]);
__m128i source = _mm_loadl_epi64((__m128i *) &lineIn[x + xmin]); __m128i source = _mm_loadl_epi64((__m128i *) &lineIn[x + xmin]);
__m128i ksource = _mm_cvtsi32_si128(*(int *) &k[x]); __m128i pix = _mm_shuffle_epi8(source, _mm_set_epi8(
pix = _mm_shuffle_epi8(source, _mm_set_epi8(
-1,7, -1,3, -1,6, -1,2, -1,5, -1,1, -1,4, -1,0)); -1,7, -1,3, -1,6, -1,2, -1,5, -1,1, -1,4, -1,0));
mmk = _mm_shuffle_epi8(ksource, _mm_set_epi8(
3,2, 1,0, 3,2, 1,0, 3,2, 1,0, 3,2, 1,0));
sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk)); sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk));
} }

44
libImaging/ResampleSIMDVerticalConv.c
View File

@ -7,11 +7,11 @@ ImagingResampleVerticalConvolution8u(UINT32 *lineOut, Imaging imIn,
int xx = 0; int xx = 0;
int xsize = imIn->xsize; int xsize = imIn->xsize;
__m128i initial = _mm_set1_epi32((1 << (coefs_precision -1)) + xmax / 2); __m128i initial = _mm_set1_epi32(1 << (coefs_precision -1));
#if defined(__AVX2__) #if defined(__AVX2__)
__m256i initial_256 = _mm256_set1_epi32((1 << (coefs_precision -1)) + xmax / 2); __m256i initial_256 = _mm256_set1_epi32(1 << (coefs_precision -1));
for (; xx < xsize - 7; xx += 8) { for (; xx < xsize - 7; xx += 8) {
__m256i sss0 = initial_256; __m256i sss0 = initial_256;
@ -21,12 +21,10 @@ ImagingResampleVerticalConvolution8u(UINT32 *lineOut, Imaging imIn,
x = 0; x = 0;
for (; x < xmax - 1; x += 2) { for (; x < xmax - 1; x += 2) {
__m256i source, source1, source2; __m256i source, source1, source2;
__m256i pix, mmk, mmk1; __m256i pix, mmk;
mmk = _mm256_set1_epi32(k[x]);
mmk1 = _mm256_set1_epi32(k[x + 1]); // Load two coefficients at once
mmk = _mm256_unpacklo_epi16( mmk = _mm256_set1_epi32(*(INT32 *) &k[x]);
_mm256_packs_epi32(mmk, mmk),
_mm256_packs_epi32(mmk1, mmk1));
source1 = _mm256_loadu_si256( // top line source1 = _mm256_loadu_si256( // top line
(__m256i *) &imIn->image32[x + xmin][xx]); (__m256i *) &imIn->image32[x + xmin][xx]);
@ -89,12 +87,10 @@ ImagingResampleVerticalConvolution8u(UINT32 *lineOut, Imaging imIn,
x = 0; x = 0;
for (; x < xmax - 1; x += 2) { for (; x < xmax - 1; x += 2) {
__m128i source, source1, source2; __m128i source, source1, source2;
__m128i pix, mmk, mmk1; __m128i pix, mmk;
mmk = _mm_set1_epi32(k[x]);
mmk1 = _mm_set1_epi32(k[x + 1]); // Load two coefficients at once
mmk = _mm_unpacklo_epi16( mmk = _mm_set1_epi32(*(INT32 *) &k[x]);
_mm_packs_epi32(mmk, mmk),
_mm_packs_epi32(mmk1, mmk1));
source1 = _mm_loadu_si128( // top line source1 = _mm_loadu_si128( // top line
(__m128i *) &imIn->image32[x + xmin][xx]); (__m128i *) &imIn->image32[x + xmin][xx]);
@ -191,12 +187,10 @@ ImagingResampleVerticalConvolution8u(UINT32 *lineOut, Imaging imIn,
x = 0; x = 0;
for (; x < xmax - 1; x += 2) { for (; x < xmax - 1; x += 2) {
__m128i source, source1, source2; __m128i source, source1, source2;
__m128i pix, mmk, mmk1; __m128i pix, mmk;
mmk = _mm_set1_epi32(k[x]);
mmk1 = _mm_set1_epi32(k[x + 1]); // Load two coefficients at once
mmk = _mm_unpacklo_epi16( mmk = _mm_set1_epi32(*(INT32 *) &k[x]);
_mm_packs_epi32(mmk, mmk),
_mm_packs_epi32(mmk1, mmk1));
source1 = _mm_loadl_epi64( // top line source1 = _mm_loadl_epi64( // top line
(__m128i *) &imIn->image32[x + xmin][xx]); (__m128i *) &imIn->image32[x + xmin][xx]);
@ -235,12 +229,10 @@ ImagingResampleVerticalConvolution8u(UINT32 *lineOut, Imaging imIn,
x = 0; x = 0;
for (; x < xmax - 1; x += 2) { for (; x < xmax - 1; x += 2) {
__m128i source, source1, source2; __m128i source, source1, source2;
__m128i pix, mmk, mmk1; __m128i pix, mmk;
mmk = _mm_set1_epi32(k[x]);
mmk1 = _mm_set1_epi32(k[x + 1]); // Load two coefficients at once
mmk = _mm_unpacklo_epi16( mmk = _mm_set1_epi32(*(INT32 *) &k[x]);
_mm_packs_epi32(mmk, mmk),
_mm_packs_epi32(mmk1, mmk1));
source1 = _mm_cvtsi32_si128( // top line source1 = _mm_cvtsi32_si128( // top line
*(int *) &imIn->image32[x + xmin][xx]); *(int *) &imIn->image32[x + xmin][xx]);