michael@0: /*
michael@0:  * Copyright 2013 The Android Open Source Project
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: 
michael@0: #include "SkBitmap.h"
michael@0: #include "SkColorPriv.h"
michael@0: #include "SkBlurImage_opts_SSE2.h"
michael@0: #include "SkRect.h"
michael@0: 
michael@0: #include <emmintrin.h>
michael@0: 
michael@0: namespace {
michael@0: 
michael@0: enum BlurDirection {
michael@0:     kX, kY
michael@0: };
michael@0: 
michael@0: /**
michael@0:  * Helper function to spread the components of a 32-bit integer into the
michael@0:  * lower 8 bits of each 32-bit element of an SSE register.
michael@0:  */
michael@0: 
michael@0: inline __m128i expand(int a) {
michael@0:       const __m128i zero = _mm_setzero_si128();
michael@0: 
michael@0:       // 0 0 0 0   0 0 0 0   0 0 0 0   A R G B
michael@0:       __m128i result = _mm_cvtsi32_si128(a);
michael@0: 
michael@0:       // 0 0 0 0   0 0 0 0   0 A 0 R   0 G 0 B
michael@0:       result = _mm_unpacklo_epi8(result, zero);
michael@0: 
michael@0:       // 0 0 0 A   0 0 0 R   0 0 0 G   0 0 0 B
michael@0:       return _mm_unpacklo_epi16(result, zero);
michael@0: }
michael@0: 
michael@0: template<BlurDirection srcDirection, BlurDirection dstDirection>
michael@0: void SkBoxBlur_SSE2(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
michael@0:                     int leftOffset, int rightOffset, int width, int height)
michael@0: {
michael@0:     const int rightBorder = SkMin32(rightOffset + 1, width);
michael@0:     const int srcStrideX = srcDirection == kX ? 1 : srcStride;
michael@0:     const int dstStrideX = dstDirection == kX ? 1 : height;
michael@0:     const int srcStrideY = srcDirection == kX ? srcStride : 1;
michael@0:     const int dstStrideY = dstDirection == kX ? width : 1;
michael@0:     const __m128i scale = _mm_set1_epi32((1 << 24) / kernelSize);
michael@0:     const __m128i half = _mm_set1_epi32(1 << 23);
michael@0:     const __m128i zero = _mm_setzero_si128();
michael@0:     for (int y = 0; y < height; ++y) {
michael@0:         __m128i sum = zero;
michael@0:         const SkPMColor* p = src;
michael@0:         for (int i = 0; i < rightBorder; ++i) {
michael@0:             sum = _mm_add_epi32(sum, expand(*p));
michael@0:             p += srcStrideX;
michael@0:         }
michael@0: 
michael@0:         const SkPMColor* sptr = src;
michael@0:         SkColor* dptr = dst;
michael@0:         for (int x = 0; x < width; ++x) {
michael@0: #if 0
michael@0:             // In SSE4.1, this would be
michael@0:             __m128i result = _mm_mullo_epi32(sum, scale);
michael@0: #else
michael@0:             // But SSE2 has no PMULLUD, so we must do AG and RB separately.
michael@0:             __m128i tmp1 = _mm_mul_epu32(sum, scale);
michael@0:             __m128i tmp2 = _mm_mul_epu32(_mm_srli_si128(sum, 4),
michael@0:                                          _mm_srli_si128(scale, 4));
michael@0:             __m128i result = _mm_unpacklo_epi32(_mm_shuffle_epi32(tmp1, _MM_SHUFFLE(0,0,2,0)),
michael@0:                                                 _mm_shuffle_epi32(tmp2, _MM_SHUFFLE(0,0,2,0)));
michael@0: #endif
michael@0:             // sumA*scale+.5 sumB*scale+.5 sumG*scale+.5 sumB*scale+.5
michael@0:             result = _mm_add_epi32(result, half);
michael@0: 
michael@0:             // 0 0 0 A   0 0 0 R   0 0 0 G   0 0 0 B
michael@0:             result = _mm_srli_epi32(result, 24);
michael@0: 
michael@0:             // 0 0 0 0   0 0 0 0   0 A 0 R   0 G 0 B
michael@0:             result = _mm_packs_epi32(result, zero);
michael@0: 
michael@0:             // 0 0 0 0   0 0 0 0   0 0 0 0   A R G B
michael@0:             result = _mm_packus_epi16(result, zero);
michael@0:             *dptr = _mm_cvtsi128_si32(result);
michael@0:             if (x >= leftOffset) {
michael@0:                 SkColor l = *(sptr - leftOffset * srcStrideX);
michael@0:                 sum = _mm_sub_epi32(sum, expand(l));
michael@0:             }
michael@0:             if (x + rightOffset + 1 < width) {
michael@0:                 SkColor r = *(sptr + (rightOffset + 1) * srcStrideX);
michael@0:                 sum = _mm_add_epi32(sum, expand(r));
michael@0:             }
michael@0:             sptr += srcStrideX;
michael@0:             if (srcDirection == kY) {
michael@0:                 _mm_prefetch(reinterpret_cast<const char*>(sptr + (rightOffset + 1) * srcStrideX),
michael@0:                              _MM_HINT_T0);
michael@0:             }
michael@0:             dptr += dstStrideX;
michael@0:         }
michael@0:         src += srcStrideY;
michael@0:         dst += dstStrideY;
michael@0:     }
michael@0: }
michael@0: 
michael@0: } // namespace
michael@0: 
michael@0: bool SkBoxBlurGetPlatformProcs_SSE2(SkBoxBlurProc* boxBlurX,
michael@0:                                     SkBoxBlurProc* boxBlurY,
michael@0:                                     SkBoxBlurProc* boxBlurXY,
michael@0:                                     SkBoxBlurProc* boxBlurYX) {
michael@0:     *boxBlurX = SkBoxBlur_SSE2<kX, kX>;
michael@0:     *boxBlurY = SkBoxBlur_SSE2<kY, kY>;
michael@0:     *boxBlurXY = SkBoxBlur_SSE2<kX, kY>;
michael@0:     *boxBlurYX = SkBoxBlur_SSE2<kY, kX>;
michael@0:     return true;
michael@0: }