PolytreeDE
/
i3lock-color
mirror of https://github.com/Raymo111/i3lock-color.git
1
0
Fork 0
Code Releases Activity

Extend kernel size to 15x15.

This commit is contained in:
Sebastian Frysztak 2016-10-28 17:35:33 +02:00
parent a48ddb61db
commit afe41c5754
2 changed files with 33 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
blur.c
View File

@ -70,7 +70,7 @@ blur_image_surface (cairo_surface_t *surface, int radius)
dst_stride = cairo_image_surface_get_stride (tmp); dst_stride = cairo_image_surface_get_stride (tmp);
//blur_impl_naive(src, dst, width, height, src_stride, dst_stride, 10000); //blur_impl_naive(src, dst, width, height, src_stride, dst_stride, 10000);
blur_impl_sse2(src, dst, width, height, 2.5); blur_impl_sse2(src, dst, width, height, 4.5);
cairo_surface_destroy (tmp); cairo_surface_destroy (tmp);
cairo_surface_flush (surface); cairo_surface_flush (surface);

57
blur_simd.c
View File

@ -12,9 +12,13 @@
#include <xmmintrin.h> #include <xmmintrin.h>
#define ALIGN16 __attribute__((aligned(16))) #define ALIGN16 __attribute__((aligned(16)))
#define KERNEL_SIZE 7 #define KERNEL_SIZE 15
#define HALF_KERNEL KERNEL_SIZE / 2 #define HALF_KERNEL KERNEL_SIZE / 2
// number of xmm registers needed to store
// input pixels for given kernel size
#define REGISTERS_CNT (KERNEL_SIZE + 4/2) / 4
void blur_impl_sse2(uint32_t *src, uint32_t *dst, int width, int height, float sigma) { void blur_impl_sse2(uint32_t *src, uint32_t *dst, int width, int height, float sigma) {
// prepare kernel // prepare kernel
float kernel[KERNEL_SIZE]; float kernel[KERNEL_SIZE];
@ -46,34 +50,34 @@ void blur_impl_horizontal_pass_sse2(uint32_t *src, uint32_t *dst, float *kernel,
uint32_t lastPixel = *(src + width - 1); uint32_t lastPixel = *(src + width - 1);
for (int column = 0; column < width; column++, src++) { for (int column = 0; column < width; column++, src++) {
__m128i rgbaIn1, rgbaIn2; __m128i rgbaIn[REGISTERS_CNT];
// handle borders // handle borders
int leftBorder = column < HALF_KERNEL; int leftBorder = column < HALF_KERNEL;
int rightBorder = column + HALF_KERNEL >= width; int rightBorder = column + HALF_KERNEL >= width;
if (leftBorder || rightBorder) { if (leftBorder || rightBorder) {
uint32_t rgbaIn[KERNEL_SIZE] ALIGN16; uint32_t _rgbaIn[KERNEL_SIZE] ALIGN16;
int i = 0; int i = 0;
if (leftBorder) { if (leftBorder) {
// for kernel size 7x7 and column == 0, we have: // for kernel size 7x7 and column == 0, we have:
// x x x P0 P1 P2 P3 // x x x P0 P1 P2 P3
// first loop fills x's with P0, second one loads P{0..3} // first loop fills x's with P0, second one loads P{0..3}
for (; i < HALF_KERNEL - column; i++) for (; i < HALF_KERNEL - column; i++)
rgbaIn[i] = firstPixel; _rgbaIn[i] = firstPixel;
for (; i < KERNEL_SIZE; i++) for (; i < KERNEL_SIZE; i++)
rgbaIn[i] = *(src + i - HALF_KERNEL); _rgbaIn[i] = *(src + i - HALF_KERNEL);
} else { } else {
for (; width < column; i++) for (; width < column; i++)
rgbaIn[i] = *(src - i - HALF_KERNEL); _rgbaIn[i] = *(src - i - HALF_KERNEL);
for (; i < KERNEL_SIZE; i++) for (; i < KERNEL_SIZE; i++)
rgbaIn[i] = lastPixel; _rgbaIn[i] = lastPixel;
} }
rgbaIn1 = _mm_load_si128((__m128i *)(rgbaIn)); for (int k = 0; k < REGISTERS_CNT; k++)
rgbaIn2 = _mm_load_si128((__m128i *)(rgbaIn + 4)); rgbaIn[k] = _mm_load_si128((__m128i*)(_rgbaIn + 4*k));
} else { } else {
rgbaIn1 = _mm_loadu_si128((__m128i *)(src - 3)); for (int k = 0; k < REGISTERS_CNT; k++)
rgbaIn2 = _mm_loadu_si128((__m128i *)(src + 1)); rgbaIn[k] = _mm_loadu_si128((__m128i*)(src + 4*k - HALF_KERNEL));
} }
// unpack each pixel, convert to float, // unpack each pixel, convert to float,
@ -85,25 +89,28 @@ void blur_impl_horizontal_pass_sse2(uint32_t *src, uint32_t *dst, float *kernel,
__m128 acc = _mm_setzero_ps(); __m128 acc = _mm_setzero_ps();
int counter = 0; int counter = 0;
tmp = _mm_unpacklo_epi8(rgbaIn1, zero); for (int i = 0; i < 3; i++)
{
tmp = _mm_unpacklo_epi8(rgbaIn[i], zero);
rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
tmp = _mm_unpackhi_epi8(rgbaIn[i], zero);
rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
}
tmp = _mm_unpacklo_epi8(rgbaIn[3], zero);
rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero)); rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++]))); acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero)); rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++]))); acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
tmp = _mm_unpackhi_epi8(rgbaIn1, zero); tmp = _mm_unpackhi_epi8(rgbaIn[3], zero);
rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
tmp = _mm_unpacklo_epi8(rgbaIn2, zero);
rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
tmp = _mm_unpackhi_epi8(rgbaIn2, zero);
rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero)); rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++]))); acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));