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 "SkColorPriv.h"
michael@0: #include "SkMorphology_opts_SSE2.h"
michael@0: 
michael@0: #include <emmintrin.h>
michael@0: 
michael@0: /* SSE2 version of dilateX, dilateY, erodeX, erodeY.
michael@0:  * portable versions are in src/effects/SkMorphologyImageFilter.cpp.
michael@0:  */
michael@0: 
michael@0: enum MorphType {
michael@0:     kDilate, kErode
michael@0: };
michael@0: 
michael@0: enum MorphDirection {
michael@0:     kX, kY
michael@0: };
michael@0: 
michael@0: template<MorphType type, MorphDirection direction>
michael@0: static void SkMorph_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
michael@0:                          int width, int height, int srcStride, int dstStride)
michael@0: {
michael@0:     const int srcStrideX = direction == kX ? 1 : srcStride;
michael@0:     const int dstStrideX = direction == kX ? 1 : dstStride;
michael@0:     const int srcStrideY = direction == kX ? srcStride : 1;
michael@0:     const int dstStrideY = direction == kX ? dstStride : 1;
michael@0:     radius = SkMin32(radius, width - 1);
michael@0:     const SkPMColor* upperSrc = src + radius * srcStrideX;
michael@0:     for (int x = 0; x < width; ++x) {
michael@0:         const SkPMColor* lp = src;
michael@0:         const SkPMColor* up = upperSrc;
michael@0:         SkPMColor* dptr = dst;
michael@0:         for (int y = 0; y < height; ++y) {
michael@0:             __m128i max = type == kDilate ? _mm_setzero_si128() : _mm_set1_epi32(0xFFFFFFFF);
michael@0:             for (const SkPMColor* p = lp; p <= up; p += srcStrideX) {
michael@0:                 __m128i src_pixel = _mm_cvtsi32_si128(*p);
michael@0:                 max = type == kDilate ? _mm_max_epu8(src_pixel, max) : _mm_min_epu8(src_pixel, max);
michael@0:             }
michael@0:             *dptr = _mm_cvtsi128_si32(max);
michael@0:             dptr += dstStrideY;
michael@0:             lp += srcStrideY;
michael@0:             up += srcStrideY;
michael@0:         }
michael@0:         if (x >= radius) src += srcStrideX;
michael@0:         if (x + radius < width - 1) upperSrc += srcStrideX;
michael@0:         dst += dstStrideX;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void SkDilateX_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
michael@0:                     int width, int height, int srcStride, int dstStride)
michael@0: {
michael@0:     SkMorph_SSE2<kDilate, kX>(src, dst, radius, width, height, srcStride, dstStride);
michael@0: }
michael@0: 
michael@0: void SkErodeX_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
michael@0:                    int width, int height, int srcStride, int dstStride)
michael@0: {
michael@0:     SkMorph_SSE2<kErode, kX>(src, dst, radius, width, height, srcStride, dstStride);
michael@0: }
michael@0: 
michael@0: void SkDilateY_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
michael@0:                     int width, int height, int srcStride, int dstStride)
michael@0: {
michael@0:     SkMorph_SSE2<kDilate, kY>(src, dst, radius, width, height, srcStride, dstStride);
michael@0: }
michael@0: 
michael@0: void SkErodeY_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
michael@0:                    int width, int height, int srcStride, int dstStride)
michael@0: {
michael@0:     SkMorph_SSE2<kErode, kY>(src, dst, radius, width, height, srcStride, dstStride);
michael@0: }