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 BoxBlur. sse4 ImagingBoxBlur implementation

Browse Source
This commit is contained in:
homm 2015-07-16 02:33:12 +03:00 committed by Alexander Karpinsky
parent 5a939e33f7
commit 9609a5849f
1 changed files with 53 additions and 41 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

88
src/libImaging/BoxBlur.c
View File

@ -1,60 +1,72 @@
#include "Imaging.h" #include "Imaging.h"
#include <emmintrin.h>
#include <mmintrin.h>
#include <smmintrin.h>
#define MAX(x, y) (((x) > (y)) ? (x) : (y)) #define MAX(x, y) (((x) > (y)) ? (x) : (y))
#define MIN(x, y) (((x) < (y)) ? (x) : (y)) #define MIN(x, y) (((x) < (y)) ? (x) : (y))
typedef UINT8 pixel[4];
void static inline ImagingLineBoxBlur32( void static inline
pixel *lineOut, ImagingLineBoxBlur32(UINT32 *lineOut, UINT32 *lineIn, int lastx, int radius, int edgeA,
pixel *lineIn, int edgeB, UINT32 ww, UINT32 fw)
int lastx, {
int radius,
int edgeA,
int edgeB,
UINT32 ww,
UINT32 fw) {
int x; int x;
UINT32 acc[4]; __m128i acc;
UINT32 bulk[4]; __m128i bulk;
__m128i wws = _mm_set1_epi32(ww);
__m128i fws = _mm_set1_epi32(fw);
__m128i b23 = _mm_set1_epi32(1 << 23);
__m128i shiftmask = _mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,3,7,11,15);
// #define MOVE_ACC(acc, subtract, add)
// acc[0] += lineIn[add][0] - lineIn[subtract][0];
#define MOVE_ACC(acc, subtract, add) \ #define MOVE_ACC(acc, subtract, add) \
acc[0] += lineIn[add][0] - lineIn[subtract][0]; \ __m128i tmp1 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128(lineIn[add])); \
acc[1] += lineIn[add][1] - lineIn[subtract][1]; \ __m128i tmp2 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128(lineIn[subtract])); \
acc[2] += lineIn[add][2] - lineIn[subtract][2]; \ acc = _mm_sub_epi32(_mm_add_epi32(acc, tmp1), tmp2);
acc[3] += lineIn[add][3] - lineIn[subtract][3];
// #define ADD_FAR(bulk, acc, left, right)
// bulk[0] = (acc[0] * ww) + (lineIn[left][0] + lineIn[right][0]) * fw;
#define ADD_FAR(bulk, acc, left, right) \ #define ADD_FAR(bulk, acc, left, right) \
bulk[0] = (acc[0] * ww) + (lineIn[left][0] + lineIn[right][0]) * fw; \ __m128i tmp3 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128(lineIn[left])); \
bulk[1] = (acc[1] * ww) + (lineIn[left][1] + lineIn[right][1]) * fw; \ __m128i tmp4 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128(lineIn[right])); \
bulk[2] = (acc[2] * ww) + (lineIn[left][2] + lineIn[right][2]) * fw; \ __m128i tmp5 = _mm_mullo_epi32(_mm_add_epi32(tmp3, tmp4), fws); \
bulk[3] = (acc[3] * ww) + (lineIn[left][3] + lineIn[right][3]) * fw; __m128i tmp6 = _mm_mullo_epi32(acc, wws); \
bulk = _mm_add_epi32(tmp5, tmp6);
// #define SAVE(x, bulk)
// lineOut[x][0] = (UINT8)((bulk[0] + (1 << 23)) >> 24);
#define SAVE(x, bulk) \ #define SAVE(x, bulk) \
lineOut[x][0] = (UINT8)((bulk[0] + (1 << 23)) >> 24); \ lineOut[x] = _mm_cvtsi128_si32( \
lineOut[x][1] = (UINT8)((bulk[1] + (1 << 23)) >> 24); \ _mm_shuffle_epi8(_mm_add_epi32(bulk, b23), shiftmask) \
lineOut[x][2] = (UINT8)((bulk[2] + (1 << 23)) >> 24); \ )
lineOut[x][3] = (UINT8)((bulk[3] + (1 << 23)) >> 24);
/* Compute acc for -1 pixel (outside of image): /* Compute acc for -1 pixel (outside of image):
From "-radius-1" to "-1" get first pixel, From "-radius-1" to "-1" get first pixel,
then from "0" to "radius-1". */ then from "0" to "radius-1". */
acc[0] = lineIn[0][0] * (radius + 1); // acc[0] = lineIn[0][0] * (radius + 1);
acc[1] = lineIn[0][1] * (radius + 1); acc = _mm_mullo_epi32(
acc[2] = lineIn[0][2] * (radius + 1); _mm_cvtepu8_epi32(_mm_cvtsi32_si128(lineIn[0])),
acc[3] = lineIn[0][3] * (radius + 1); _mm_set1_epi32(radius + 1)
);
/* As radius can be bigger than xsize, iterate to edgeA -1. */ /* As radius can be bigger than xsize, iterate to edgeA -1. */
for (x = 0; x < edgeA - 1; x++) { for (x = 0; x < edgeA - 1; x++) {
acc[0] += lineIn[x][0]; // acc[0] += lineIn[x][0];
acc[1] += lineIn[x][1]; acc = _mm_add_epi32(acc, _mm_cvtepu8_epi32(_mm_cvtsi32_si128(lineIn[x])));
acc[2] += lineIn[x][2];
acc[3] += lineIn[x][3];
} }
/* Then multiply remainder to last x. */ /* Then multiply remainder to last x. */
acc[0] += lineIn[lastx][0] * (radius - edgeA + 1); // acc[0] += lineIn[lastx][0] * (radius - edgeA + 1);
acc[1] += lineIn[lastx][1] * (radius - edgeA + 1); acc = _mm_add_epi32(acc, _mm_mullo_epi32(
acc[2] += lineIn[lastx][2] * (radius - edgeA + 1); _mm_cvtepu8_epi32(_mm_cvtsi32_si128(lineIn[x])),
acc[3] += lineIn[lastx][3] * (radius - edgeA + 1); _mm_set1_epi32(radius - edgeA + 1)
));
if (edgeA <= edgeB) { if (edgeA <= edgeB) {
/* Subtract pixel from left ("0"). /* Subtract pixel from left ("0").
@ -207,8 +219,8 @@ ImagingHorizontalBoxBlur(Imaging imOut, Imaging imIn, float floatRadius) {
} else { } else {
for (y = 0; y < imIn->ysize; y++) { for (y = 0; y < imIn->ysize; y++) {
ImagingLineBoxBlur32( ImagingLineBoxBlur32(
imIn == imOut ? (pixel *)lineOut : (pixel *)imOut->image32[y], imIn == imOut ? (UINT32 *) lineOut : (UINT32 *) imOut->image32[y],
(pixel *)imIn->image32[y], (UINT32 *) imIn->image32[y],
imIn->xsize - 1, imIn->xsize - 1,
radius, radius,
edgeA, edgeA,