python-pillow/Pillow
1
1
Fork 0
mirror of https://github.com/python-pillow/Pillow.git synced 2025-08-20 20:24:45 +03:00
Code Issues Packages Projects Releases Wiki Activity

SIMD Filter. Correct offset for 3x3f_u8

Browse Source
This commit is contained in:
Alexander Karpinsky 2021-12-19 21:23:47 +03:00
parent 18783eeefa
commit fc915a4ed5
2 changed files with 10 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
src/libImaging/FilterSIMD_3x3f_u8.c
View File

@ -3,14 +3,12 @@ ImagingFilter3x3f_u8(Imaging imOut, Imaging im, const float* kernel,
float offset) float offset)
{ {
#define MM_KERNEL1x3_SUM1(ss, row, kernel) \ #define MM_KERNEL1x3_SUM1(ss, row, kernel) \
ss = _mm_set1_ps(offset); \ ss = _mm_mul_ps(pix0##row, kernel0##kernel); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix0##row, kernel0##kernel)); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix1##row, kernel1##kernel)); \ ss = _mm_add_ps(ss, _mm_mul_ps(pix1##row, kernel1##kernel)); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix2##row, kernel2##kernel)); ss = _mm_add_ps(ss, _mm_mul_ps(pix2##row, kernel2##kernel));
#define MM_KERNEL1x3_SUM1_2(ss, row, kernel) \ #define MM_KERNEL1x3_SUM1_2(ss, row, kernel) \
ss = _mm_set1_ps(offset); \ ss = _mm_mul_ps(pix3##row, kernel0##kernel); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix3##row, kernel0##kernel)); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix0##row, kernel1##kernel)); \ ss = _mm_add_ps(ss, _mm_mul_ps(pix0##row, kernel1##kernel)); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix1##row, kernel2##kernel)); ss = _mm_add_ps(ss, _mm_mul_ps(pix1##row, kernel2##kernel));
@ -26,6 +24,7 @@ ImagingFilter3x3f_u8(Imaging imOut, Imaging im, const float* kernel,
__m128 kernel01 = _mm_set_ps(kernel[2], kernel[1], kernel[0], 0); __m128 kernel01 = _mm_set_ps(kernel[2], kernel[1], kernel[0], 0);
__m128 kernel11 = _mm_set_ps(kernel[5], kernel[4], kernel[3], 0); __m128 kernel11 = _mm_set_ps(kernel[5], kernel[4], kernel[3], 0);
__m128 kernel21 = _mm_set_ps(kernel[8], kernel[7], kernel[6], 0); __m128 kernel21 = _mm_set_ps(kernel[8], kernel[7], kernel[6], 0);
__m128 mm_offset = _mm_set1_ps(offset);
memcpy(imOut->image8[0], im->image8[0], im->linesize); memcpy(imOut->image8[0], im->image8[0], im->linesize);
y = 1; y = 1;
@ -64,12 +63,14 @@ ImagingFilter3x3f_u8(Imaging imOut, Imaging im, const float* kernel,
ss4 = _mm_hadd_ps(ss4, ss5); ss4 = _mm_hadd_ps(ss4, ss5);
ss0 = _mm_hadd_ps(ss0, ss1); ss0 = _mm_hadd_ps(ss0, ss1);
ss0 = _mm_add_ps(ss0, mm_offset);
ssi0 = _mm_cvtps_epi32(ss0); ssi0 = _mm_cvtps_epi32(ss0);
ssi0 = _mm_packs_epi32(ssi0, ssi0); ssi0 = _mm_packs_epi32(ssi0, ssi0);
ssi0 = _mm_packus_epi16(ssi0, ssi0); ssi0 = _mm_packus_epi16(ssi0, ssi0);
*((UINT32*) &out0[x]) = _mm_cvtsi128_si32(ssi0); *((UINT32*) &out0[x]) = _mm_cvtsi128_si32(ssi0);
ss3 = _mm_hadd_ps(ss3, ss4); ss3 = _mm_hadd_ps(ss3, ss4);
ss3 = _mm_add_ps(ss3, mm_offset);
ssi0 = _mm_cvtps_epi32(ss3); ssi0 = _mm_cvtps_epi32(ss3);
ssi0 = _mm_packs_epi32(ssi0, ssi0); ssi0 = _mm_packs_epi32(ssi0, ssi0);
ssi0 = _mm_packus_epi16(ssi0, ssi0); ssi0 = _mm_packus_epi16(ssi0, ssi0);
@ -89,6 +90,7 @@ ImagingFilter3x3f_u8(Imaging imOut, Imaging im, const float* kernel,
ss0 = _mm_hadd_ps(ss0, ss0); ss0 = _mm_hadd_ps(ss0, ss0);
ss0 = _mm_hadd_ps(ss0, ss0); ss0 = _mm_hadd_ps(ss0, ss0);
ss0 = _mm_add_ps(ss0, mm_offset);
ssi0 = _mm_cvtps_epi32(ss0); ssi0 = _mm_cvtps_epi32(ss0);
ssi0 = _mm_packs_epi32(ssi0, ssi0); ssi0 = _mm_packs_epi32(ssi0, ssi0);
ssi0 = _mm_packus_epi16(ssi0, ssi0); ssi0 = _mm_packus_epi16(ssi0, ssi0);
@ -96,6 +98,7 @@ ImagingFilter3x3f_u8(Imaging imOut, Imaging im, const float* kernel,
ss1 = _mm_hadd_ps(ss1, ss1); ss1 = _mm_hadd_ps(ss1, ss1);
ss1 = _mm_hadd_ps(ss1, ss1); ss1 = _mm_hadd_ps(ss1, ss1);
ss1 = _mm_add_ps(ss1, mm_offset);
ssi0 = _mm_cvtps_epi32(ss1); ssi0 = _mm_cvtps_epi32(ss1);
ssi0 = _mm_packs_epi32(ssi0, ssi0); ssi0 = _mm_packs_epi32(ssi0, ssi0);
ssi0 = _mm_packus_epi16(ssi0, ssi0); ssi0 = _mm_packus_epi16(ssi0, ssi0);
@ -122,6 +125,7 @@ ImagingFilter3x3f_u8(Imaging imOut, Imaging im, const float* kernel,
ss = _mm_hadd_ps(ss, ss); ss = _mm_hadd_ps(ss, ss);
ss = _mm_hadd_ps(ss, ss); ss = _mm_hadd_ps(ss, ss);
ss = _mm_add_ps(ss, mm_offset);
ssi0 = _mm_cvtps_epi32(ss); ssi0 = _mm_cvtps_epi32(ss);
ssi0 = _mm_packs_epi32(ssi0, ssi0); ssi0 = _mm_packs_epi32(ssi0, ssi0);
ssi0 = _mm_packus_epi16(ssi0, ssi0); ssi0 = _mm_packus_epi16(ssi0, ssi0);
@ -139,6 +143,7 @@ ImagingFilter3x3f_u8(Imaging imOut, Imaging im, const float* kernel,
ss = _mm_hadd_ps(ss, ss); ss = _mm_hadd_ps(ss, ss);
ss = _mm_hadd_ps(ss, ss); ss = _mm_hadd_ps(ss, ss);
ss = _mm_add_ps(ss, mm_offset);
ssi0 = _mm_cvtps_epi32(ss); ssi0 = _mm_cvtps_epi32(ss);
ssi0 = _mm_packs_epi32(ssi0, ssi0); ssi0 = _mm_packs_epi32(ssi0, ssi0);
ssi0 = _mm_packus_epi16(ssi0, ssi0); ssi0 = _mm_packus_epi16(ssi0, ssi0);

3
src/libImaging/FilterSIMD_5x5f_u8.c
View File

@ -10,8 +10,7 @@ ImagingFilter5x5f_u8(Imaging imOut, Imaging im, const float* kernel,
pix4##row = _mm_cvtepi32_ps(mm_cvtepu8_epi32(&in_2[x])); pix4##row = _mm_cvtepi32_ps(mm_cvtepu8_epi32(&in_2[x]));
#define MM_KERNEL1x5_SUM(ss, row, krow) \ #define MM_KERNEL1x5_SUM(ss, row, krow) \
ss = _mm_setzero_ps(); \ ss = _mm_mul_ps(pix0##row, kernel0##krow); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix0##row, kernel0##krow)); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix1##row, kernel1##krow)); \ ss = _mm_add_ps(ss, _mm_mul_ps(pix1##row, kernel1##krow)); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix2##row, kernel2##krow)); \ ss = _mm_add_ps(ss, _mm_mul_ps(pix2##row, kernel2##krow)); \
ss = _mm_add_ps(ss, _mm_mul_ps(pix3##row, kernel3##krow)); \ ss = _mm_add_ps(ss, _mm_mul_ps(pix3##row, kernel3##krow)); \