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 Filter. 3x3 implementation

Browse Source
This commit is contained in:
Alexander 2017-08-13 17:16:16 +03:00 committed by Alexander Karpinsky
parent a794af7b00
commit e9e2ed124a
1 changed files with 38 additions and 64 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

102
src/libImaging/Filter.c
View File

@ -26,8 +26,18 @@
#include "Imaging.h" #include "Imaging.h"
static inline UINT8 #include <emmintrin.h>
clip8(float in) { #include <mmintrin.h>
#include <smmintrin.h>
#if defined(__AVX2__)
#include <immintrin.h>
#endif
static inline UINT8 clip8(float in)
{
if (in <= 0.0) { if (in <= 0.0) {
return 0; return 0;
} }
@ -99,6 +109,14 @@ ImagingFilter3x3(Imaging imOut, Imaging im, const float *kernel, float offset) {
(_i2f((UINT8)in0[x - d]) * (kernel)[0] + _i2f((UINT8)in0[x]) * (kernel)[1] + \ (_i2f((UINT8)in0[x - d]) * (kernel)[0] + _i2f((UINT8)in0[x]) * (kernel)[1] + \
_i2f((UINT8)in0[x + d]) * (kernel)[2]) _i2f((UINT8)in0[x + d]) * (kernel)[2])
#define MM_KERNEL1x3(ss, in0, x, kernel, d) \
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps((kernel)[0]), \
_mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x-d])))); \
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps((kernel)[1]), \
_mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x+0])))); \
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps((kernel)[2]), \
_mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x+d]))));
int x = 0, y = 0; int x = 0, y = 0;
memcpy(imOut->image[0], im->image[0], im->linesize); memcpy(imOut->image[0], im->image[0], im->linesize);
@ -122,69 +140,25 @@ ImagingFilter3x3(Imaging imOut, Imaging im, const float *kernel, float offset) {
out[x] = in0[x]; out[x] = in0[x];
} }
} else { } else {
// Add one time for rounding for (y = 1; y < im->ysize-1; y++) {
offset += 0.5; UINT32* in_1 = (UINT32*) im->image[y-1];
for (y = 1; y < im->ysize - 1; y++) { UINT32* in0 = (UINT32*) im->image[y];
UINT8 *in_1 = (UINT8 *)im->image[y - 1]; UINT32* in1 = (UINT32*) im->image[y+1];
UINT8 *in0 = (UINT8 *)im->image[y]; UINT32* out = (UINT32*) imOut->image[y];
UINT8 *in1 = (UINT8 *)im->image[y + 1];
UINT8 *out = (UINT8 *)imOut->image[y];
memcpy(out, in0, sizeof(UINT32)); out[0] = ((UINT32*) in0)[0];
if (im->bands == 2) { for (x = 1; x < im->xsize-1; x++) {
for (x = 1; x < im->xsize - 1; x++) { __m128 ss = _mm_set1_ps(offset);
float ss0 = offset; __m128i ssi;
float ss3 = offset;
UINT32 v; MM_KERNEL1x3(ss, in1, x, &kernel[0], 1);
ss0 += KERNEL1x3(in1, x * 4 + 0, &kernel[0], 4); MM_KERNEL1x3(ss, in0, x, &kernel[3], 1);
ss3 += KERNEL1x3(in1, x * 4 + 3, &kernel[0], 4); MM_KERNEL1x3(ss, in_1, x, &kernel[6], 1);
ss0 += KERNEL1x3(in0, x * 4 + 0, &kernel[3], 4);
ss3 += KERNEL1x3(in0, x * 4 + 3, &kernel[3], 4); ssi = _mm_cvtps_epi32(ss);
ss0 += KERNEL1x3(in_1, x * 4 + 0, &kernel[6], 4); ssi = _mm_packs_epi32(ssi, ssi);
ss3 += KERNEL1x3(in_1, x * 4 + 3, &kernel[6], 4); ssi = _mm_packus_epi16(ssi, ssi);
v = MAKE_UINT32(clip8(ss0), 0, 0, clip8(ss3)); out[x] = _mm_cvtsi128_si32(ssi);
memcpy(out + x * sizeof(v), &v, sizeof(v));
}
} else if (im->bands == 3) {
for (x = 1; x < im->xsize - 1; x++) {
float ss0 = offset;
float ss1 = offset;
float ss2 = offset;
UINT32 v;
ss0 += KERNEL1x3(in1, x * 4 + 0, &kernel[0], 4);
ss1 += KERNEL1x3(in1, x * 4 + 1, &kernel[0], 4);
ss2 += KERNEL1x3(in1, x * 4 + 2, &kernel[0], 4);
ss0 += KERNEL1x3(in0, x * 4 + 0, &kernel[3], 4);
ss1 += KERNEL1x3(in0, x * 4 + 1, &kernel[3], 4);
ss2 += KERNEL1x3(in0, x * 4 + 2, &kernel[3], 4);
ss0 += KERNEL1x3(in_1, x * 4 + 0, &kernel[6], 4);
ss1 += KERNEL1x3(in_1, x * 4 + 1, &kernel[6], 4);
ss2 += KERNEL1x3(in_1, x * 4 + 2, &kernel[6], 4);
v = MAKE_UINT32(clip8(ss0), clip8(ss1), clip8(ss2), 0);
memcpy(out + x * sizeof(v), &v, sizeof(v));
}
} else if (im->bands == 4) {
for (x = 1; x < im->xsize - 1; x++) {
float ss0 = offset;
float ss1 = offset;
float ss2 = offset;
float ss3 = offset;
UINT32 v;
ss0 += KERNEL1x3(in1, x * 4 + 0, &kernel[0], 4);
ss1 += KERNEL1x3(in1, x * 4 + 1, &kernel[0], 4);
ss2 += KERNEL1x3(in1, x * 4 + 2, &kernel[0], 4);
ss3 += KERNEL1x3(in1, x * 4 + 3, &kernel[0], 4);
ss0 += KERNEL1x3(in0, x * 4 + 0, &kernel[3], 4);
ss1 += KERNEL1x3(in0, x * 4 + 1, &kernel[3], 4);
ss2 += KERNEL1x3(in0, x * 4 + 2, &kernel[3], 4);
ss3 += KERNEL1x3(in0, x * 4 + 3, &kernel[3], 4);
ss0 += KERNEL1x3(in_1, x * 4 + 0, &kernel[6], 4);
ss1 += KERNEL1x3(in_1, x * 4 + 1, &kernel[6], 4);
ss2 += KERNEL1x3(in_1, x * 4 + 2, &kernel[6], 4);
ss3 += KERNEL1x3(in_1, x * 4 + 3, &kernel[6], 4);
v = MAKE_UINT32(clip8(ss0), clip8(ss1), clip8(ss2), clip8(ss3));
memcpy(out + x * sizeof(v), &v, sizeof(v));
}
} }
memcpy(out + x * sizeof(UINT32), in0 + x * sizeof(UINT32), sizeof(UINT32)); memcpy(out + x * sizeof(UINT32), in0 + x * sizeof(UINT32), sizeof(UINT32));
} }