From e5e636892636d4de128a30c310cd630fa2147b86 Mon Sep 17 00:00:00 2001 From: Sebastian Frysztak Date: Fri, 11 Nov 2016 16:46:53 +0100 Subject: [PATCH] Remove SSSE3 version. --- Makefile | 1 - blur.h | 3 -- blur_simd.c | 146 ---------------------------------------------------- 3 files changed, 150 deletions(-) diff --git a/Makefile b/Makefile index da23bf3..4edcfe5 100644 --- a/Makefile +++ b/Makefile @@ -17,7 +17,6 @@ CFLAGS += -Wall CFLAGS += -O2 SIMD_CFLAGS += -mavx SIMD_CFLAGS += -mno-sse2avx -SIMD_CFLAGS += -mssse3 SIMD_CFLAGS += -funroll-loops CPPFLAGS += -D_GNU_SOURCE CPPFLAGS += -DXKBCOMPOSE=$(shell if test -e /usr/include/xkbcommon/xkbcommon-compose.h ; then echo 1 ; else echo 0 ; fi ) diff --git a/blur.h b/blur.h index 7469871..5c08411 100644 --- a/blur.h +++ b/blur.h @@ -15,8 +15,5 @@ void blur_impl_horizontal_pass_sse2(uint32_t *src, uint32_t *dst, float *kernel, void blur_impl_avx(uint32_t* src, uint32_t* dst, int width, int height, float sigma); void blur_impl_horizontal_pass_avx(uint32_t *src, uint32_t *dst, float *kernel, int width, int height); -void blur_impl_ssse3(uint32_t* src, uint32_t* dst, int width, int height, float sigma); -void blur_impl_horizontal_pass_ssse3(uint32_t *src, uint32_t *dst, int16_t *kernel, int width, int height); - #endif diff --git a/blur_simd.c b/blur_simd.c index 4bf4e38..91325a3 100644 --- a/blur_simd.c +++ b/blur_simd.c @@ -10,7 +10,6 @@ #include "blur.h" #include #include -#include #include #define ALIGN16 __attribute__((aligned(16))) @@ -21,11 +20,6 @@ // input pixels for given kernel size #define REGISTERS_CNT (KERNEL_SIZE + 4/2) / 4 -// scaling factor for kernel coefficients. -// higher values cause desaturation. -// used in SSSE3 implementation. -#define SCALE_FACTOR 14 - // AVX intrinsics missing in GCC #define _mm256_set_m128i(v0, v1) _mm256_insertf128_si256(_mm256_castsi128_si256(v1), (v0), 1) #define _mm256_setr_m128i(v0, v1) _mm256_set_m128i((v1), (v0)) @@ -215,143 +209,3 @@ void blur_impl_horizontal_pass_avx(uint32_t *src, uint32_t *dst, float *kernel, } } } - -void blur_impl_ssse3(uint32_t *src, uint32_t *dst, int width, int height, float sigma) { - // prepare kernel - float kernelf[KERNEL_SIZE]; - int16_t kernel[KERNEL_SIZE + 1]; - float coeff = 1.0 / sqrtf(2 * M_PI * sigma * sigma), sum = 0; - - for (int i = 0; i < KERNEL_SIZE; i++) { - float x = HALF_KERNEL - i; - kernelf[i] = coeff * expf(-x * x / (2.0 * sigma * sigma)); - sum += kernelf[i]; - } - - // normalize kernel - for (int i = 0; i < KERNEL_SIZE; i++) - kernelf[i] /= sum; - - // round to nearest integer and convert to int - for (int i = 0; i < KERNEL_SIZE; i++) - kernel[i] = (int16_t)lrintf(kernelf[i] * (1 << SCALE_FACTOR)); - kernel[KERNEL_SIZE] = 0; - - // horizontal pass includes image transposition: - // instead of writing pixel src[x] to dst[x], - // we write it to transposed location. - // (to be exact: dst[height * current_column + current_row]) - blur_impl_horizontal_pass_ssse3(src, dst, kernel, width, height); - blur_impl_horizontal_pass_ssse3(dst, src, kernel, height, width); -} - - -void blur_impl_horizontal_pass_ssse3(uint32_t *src, uint32_t *dst, int16_t *kernel, int width, int height) { - __m128i _kern[2]; - _kern[0] = _mm_loadu_si128((__m128i*)kernel); - _kern[1] = _mm_loadu_si128((__m128i*)(kernel + 8)); - __m128i rgbaIn[REGISTERS_CNT]; - - for (int row = 0; row < height; row++) { - for (int column = 0; column < width; column++, src++) { - uint32_t _rgbaIn[KERNEL_SIZE] ALIGN16; - // handle borders - int leftBorder = column < HALF_KERNEL; - int rightBorder = column > width - HALF_KERNEL; - if (leftBorder || rightBorder) { - int i = 0; - if (leftBorder) { - // for kernel size 7x7 and column == 0, we have: - // x x x P0 P1 P2 P3 - // first loop mirrors P{0..3} to fill x's, - // second one loads P{0..3} - for (; i < HALF_KERNEL - column; i++) - _rgbaIn[i] = *(src + (HALF_KERNEL - i)); - for (; i < KERNEL_SIZE; i++) - _rgbaIn[i] = *(src - (HALF_KERNEL - i)); - } else { - for (; i < width - column; i++) - _rgbaIn[i] = *(src + i); - for (int k = 0; i < KERNEL_SIZE; i++, k++) - _rgbaIn[i] = *(src - k); - } - - for (int k = 0; k < REGISTERS_CNT; k++) - rgbaIn[k] = _mm_load_si128((__m128i*)(_rgbaIn + 4*k)); - } else { - for (int k = 0; k < REGISTERS_CNT; k++) - rgbaIn[k] = _mm_loadu_si128((__m128i*)(src + 4*k - HALF_KERNEL)); - } - - // basis of this implementation is _mm_maddubs_epi16 (aka pmaddubsw). - // 'rgba' holds 16 unsigned bytes, so 4 pixels. - // 'kern' holds 16 signed bytes kernel values multiplied by (1 << SCALE_FACTOR). - // before multiplication takes place, vectors need to be prepared: - // 'rgba' is shuffled from R1B1G1A1...R4B4G4A4 to R1R2R3R4...A1A2A3A4 - // 'kern' is shuffled from w1w2w3w4...w13w14w15w16 to w1w2w3w4 repeated 4 times - // then we call _mm_maddubs_epi16 and we get: - // -------------------------------------------------------------------------------------- - // | R1*w1 + R2*w2 | R3*w3 + R4*w4 | G1*w1 + G2*w2 | G3*w3 + G4*w4 | repeat for B and A | - // -------------------------------------------------------------------------------------- - // each 'rectangle' is a 16-byte signed int. - // then we repeat the process for the rest of input pixels, - // call _mm_hadds_epi16 to add adjacent ints and shift right to scale by SCALE_FACTOR. - - __m128i rgba, rg, ba, kern; - __m128i zero = _mm_setzero_si128(); - __m128i acc_rg = _mm_setzero_si128(); - __m128i acc_ba = _mm_setzero_si128(); - - const __m128i rgba_shuf_mask = _mm_setr_epi8(0, 4, 8, 12, - 1, 5, 9, 13, - 2, 6, 10, 14, - 3, 7, 11, 15); - - const __m128i kern_shuf_mask = _mm_setr_epi8(0, 1, 2, 3, - 4, 5, 6, 7, - 0, 1, 2, 3, - 4, 5, 6, 7); - - rgba = _mm_shuffle_epi8(rgbaIn[0], rgba_shuf_mask); - rg = _mm_unpacklo_epi8(rgba, zero); - ba = _mm_unpackhi_epi8(rgba, zero); - kern = _mm_shuffle_epi8(_kern[0], kern_shuf_mask); - acc_rg = _mm_add_epi32(acc_rg, _mm_madd_epi16(rg, kern)); - acc_ba = _mm_add_epi32(acc_ba, _mm_madd_epi16(ba, kern)); - - rgba = _mm_shuffle_epi8(rgbaIn[1], rgba_shuf_mask); - rg = _mm_unpacklo_epi8(rgba, zero); - ba = _mm_unpackhi_epi8(rgba, zero); - kern = _mm_shuffle_epi8(_mm_srli_si128(_kern[0], 8), kern_shuf_mask); - acc_rg = _mm_add_epi32(acc_rg, _mm_madd_epi16(rg, kern)); - acc_ba = _mm_add_epi32(acc_ba, _mm_madd_epi16(ba, kern)); - - rgba = _mm_shuffle_epi8(rgbaIn[2], rgba_shuf_mask); - rg = _mm_unpacklo_epi8(rgba, zero); - ba = _mm_unpackhi_epi8(rgba, zero); - kern = _mm_shuffle_epi8(_kern[1], kern_shuf_mask); - acc_rg = _mm_add_epi32(acc_rg, _mm_madd_epi16(rg, kern)); - acc_ba = _mm_add_epi32(acc_ba, _mm_madd_epi16(ba, kern)); - - rgba = _mm_shuffle_epi8(rgbaIn[3], rgba_shuf_mask); - rg = _mm_unpacklo_epi8(rgba, zero); - ba = _mm_unpackhi_epi8(rgba, zero); - kern = _mm_shuffle_epi8(_mm_srli_si128(_kern[1], 8), kern_shuf_mask); - acc_rg = _mm_add_epi32(acc_rg, _mm_madd_epi16(rg, kern)); - acc_ba = _mm_add_epi32(acc_ba, _mm_madd_epi16(ba, kern)); - - rgba = _mm_hadd_epi32(acc_rg, acc_ba); - rgba = _mm_srai_epi32(rgba, SCALE_FACTOR); - - // Cairo sets alpha channel to 255 - // (or -1, depending how you look at it) - // this quickly overflows accumulator, - // and alpha is calculated completely wrong. - // I assume most people don't use semi-transparent - // lock screen images, so no one will mind if we - // 'correct it' by setting alpha to 255. - *(dst + height * column + row) = - _mm_cvtsi128_si32(_mm_shuffle_epi8(rgba, rgba_shuf_mask)); - } - } -}