1/* 2 * Copyright 2013 The Android Open Source Project 3 * 4 * Use of this source code is governed by a BSD-style license that can be 5 * found in the LICENSE file. 6 */ 7 8 9#include "SkBitmap.h" 10#include "SkColorPriv.h" 11#include "SkBlurImage_opts_SSE2.h" 12#include "SkRect.h" 13 14#include <emmintrin.h> 15 16namespace { 17 18enum BlurDirection { 19 kX, kY 20}; 21 22/** 23 * Helper function to spread the components of a 32-bit integer into the 24 * lower 8 bits of each 32-bit element of an SSE register. 25 */ 26 27inline __m128i expand(int a) { 28 const __m128i zero = _mm_setzero_si128(); 29 30 // 0 0 0 0 0 0 0 0 0 0 0 0 A R G B 31 __m128i result = _mm_cvtsi32_si128(a); 32 33 // 0 0 0 0 0 0 0 0 0 A 0 R 0 G 0 B 34 result = _mm_unpacklo_epi8(result, zero); 35 36 // 0 0 0 A 0 0 0 R 0 0 0 G 0 0 0 B 37 return _mm_unpacklo_epi16(result, zero); 38} 39 40template<BlurDirection srcDirection, BlurDirection dstDirection> 41void SkBoxBlur_SSE2(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize, 42 int leftOffset, int rightOffset, int width, int height) 43{ 44 const int rightBorder = SkMin32(rightOffset + 1, width); 45 const int srcStrideX = srcDirection == kX ? 1 : srcStride; 46 const int dstStrideX = dstDirection == kX ? 1 : height; 47 const int srcStrideY = srcDirection == kX ? srcStride : 1; 48 const int dstStrideY = dstDirection == kX ? width : 1; 49 const __m128i scale = _mm_set1_epi32((1 << 24) / kernelSize); 50 const __m128i half = _mm_set1_epi32(1 << 23); 51 const __m128i zero = _mm_setzero_si128(); 52 for (int y = 0; y < height; ++y) { 53 __m128i sum = zero; 54 const SkPMColor* p = src; 55 for (int i = 0; i < rightBorder; ++i) { 56 sum = _mm_add_epi32(sum, expand(*p)); 57 p += srcStrideX; 58 } 59 60 const SkPMColor* sptr = src; 61 SkColor* dptr = dst; 62 for (int x = 0; x < width; ++x) { 63#if 0 64 // In SSE4.1, this would be 65 __m128i result = _mm_mullo_epi32(sum, scale); 66#else 67 // But SSE2 has no PMULLUD, so we must do AG and RB separately. 68 __m128i tmp1 = _mm_mul_epu32(sum, scale); 69 __m128i tmp2 = _mm_mul_epu32(_mm_srli_si128(sum, 4), 70 _mm_srli_si128(scale, 4)); 71 __m128i result = _mm_unpacklo_epi32(_mm_shuffle_epi32(tmp1, _MM_SHUFFLE(0,0,2,0)), 72 _mm_shuffle_epi32(tmp2, _MM_SHUFFLE(0,0,2,0))); 73#endif 74 // sumA*scale+.5 sumB*scale+.5 sumG*scale+.5 sumB*scale+.5 75 result = _mm_add_epi32(result, half); 76 77 // 0 0 0 A 0 0 0 R 0 0 0 G 0 0 0 B 78 result = _mm_srli_epi32(result, 24); 79 80 // 0 0 0 0 0 0 0 0 0 A 0 R 0 G 0 B 81 result = _mm_packs_epi32(result, zero); 82 83 // 0 0 0 0 0 0 0 0 0 0 0 0 A R G B 84 result = _mm_packus_epi16(result, zero); 85 *dptr = _mm_cvtsi128_si32(result); 86 if (x >= leftOffset) { 87 SkColor l = *(sptr - leftOffset * srcStrideX); 88 sum = _mm_sub_epi32(sum, expand(l)); 89 } 90 if (x + rightOffset + 1 < width) { 91 SkColor r = *(sptr + (rightOffset + 1) * srcStrideX); 92 sum = _mm_add_epi32(sum, expand(r)); 93 } 94 sptr += srcStrideX; 95 if (srcDirection == kY) { 96 _mm_prefetch(reinterpret_cast<const char*>(sptr + (rightOffset + 1) * srcStrideX), 97 _MM_HINT_T0); 98 } 99 dptr += dstStrideX; 100 } 101 src += srcStrideY; 102 dst += dstStrideY; 103 } 104} 105 106} // namespace 107 108bool SkBoxBlurGetPlatformProcs_SSE2(SkBoxBlurProc* boxBlurX, 109 SkBoxBlurProc* boxBlurY, 110 SkBoxBlurProc* boxBlurXY, 111 SkBoxBlurProc* boxBlurYX) { 112 *boxBlurX = SkBoxBlur_SSE2<kX, kX>; 113 *boxBlurY = SkBoxBlur_SSE2<kY, kY>; 114 *boxBlurXY = SkBoxBlur_SSE2<kX, kY>; 115 *boxBlurYX = SkBoxBlur_SSE2<kY, kX>; 116 return true; 117} 118