17a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org/*
27a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org * Copyright 2013 The Android Open Source Project
37a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org *
47a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org * Use of this source code is governed by a BSD-style license that can be
57a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org * found in the LICENSE file.
67a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org */
77a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
88c4953c6f176469ad287c3270ab146e292b23badcommit-bot@chromium.org#include <emmintrin.h>
97a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org#include "SkColorPriv.h"
10f49a0955e26fe57bd6d4498cc71d3c90b6a800f6senorblanco@chromium.org#include "SkMorphology_opts_SSE2.h"
117a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
127a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org/* SSE2 version of dilateX, dilateY, erodeX, erodeY.
137a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org * portable versions are in src/effects/SkMorphologyImageFilter.cpp.
147a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org */
157a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
167a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgenum MorphType {
177a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    kDilate, kErode
187a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org};
197a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
207a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgenum MorphDirection {
217a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    kX, kY
227a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org};
237a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
247a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgtemplate<MorphType type, MorphDirection direction>
257a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgstatic void SkMorph_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
267a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org                         int width, int height, int srcStride, int dstStride)
277a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org{
287a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    const int srcStrideX = direction == kX ? 1 : srcStride;
297a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    const int dstStrideX = direction == kX ? 1 : dstStride;
307a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    const int srcStrideY = direction == kX ? srcStride : 1;
317a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    const int dstStrideY = direction == kX ? dstStride : 1;
327a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    radius = SkMin32(radius, width - 1);
337a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    const SkPMColor* upperSrc = src + radius * srcStrideX;
347a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    for (int x = 0; x < width; ++x) {
357a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org        const SkPMColor* lp = src;
367a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org        const SkPMColor* up = upperSrc;
377a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org        SkPMColor* dptr = dst;
387a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org        for (int y = 0; y < height; ++y) {
397a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org            __m128i max = type == kDilate ? _mm_setzero_si128() : _mm_set1_epi32(0xFFFFFFFF);
407a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org            for (const SkPMColor* p = lp; p <= up; p += srcStrideX) {
417a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org                __m128i src_pixel = _mm_cvtsi32_si128(*p);
427a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org                max = type == kDilate ? _mm_max_epu8(src_pixel, max) : _mm_min_epu8(src_pixel, max);
437a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org            }
447a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org            *dptr = _mm_cvtsi128_si32(max);
457a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org            dptr += dstStrideY;
467a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org            lp += srcStrideY;
477a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org            up += srcStrideY;
487a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org        }
498c4953c6f176469ad287c3270ab146e292b23badcommit-bot@chromium.org        if (x >= radius) {
508c4953c6f176469ad287c3270ab146e292b23badcommit-bot@chromium.org            src += srcStrideX;
518c4953c6f176469ad287c3270ab146e292b23badcommit-bot@chromium.org        }
528c4953c6f176469ad287c3270ab146e292b23badcommit-bot@chromium.org        if (x + radius < width - 1) {
538c4953c6f176469ad287c3270ab146e292b23badcommit-bot@chromium.org            upperSrc += srcStrideX;
548c4953c6f176469ad287c3270ab146e292b23badcommit-bot@chromium.org        }
557a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org        dst += dstStrideX;
567a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    }
577a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org}
587a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
597a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgvoid SkDilateX_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
607a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org                    int width, int height, int srcStride, int dstStride)
617a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org{
627a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    SkMorph_SSE2<kDilate, kX>(src, dst, radius, width, height, srcStride, dstStride);
637a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org}
647a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
657a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgvoid SkErodeX_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
667a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org                   int width, int height, int srcStride, int dstStride)
677a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org{
687a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    SkMorph_SSE2<kErode, kX>(src, dst, radius, width, height, srcStride, dstStride);
697a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org}
707a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
717a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgvoid SkDilateY_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
727a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org                    int width, int height, int srcStride, int dstStride)
737a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org{
747a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    SkMorph_SSE2<kDilate, kY>(src, dst, radius, width, height, srcStride, dstStride);
757a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org}
767a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org
777a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.orgvoid SkErodeY_SSE2(const SkPMColor* src, SkPMColor* dst, int radius,
787a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org                   int width, int height, int srcStride, int dstStride)
797a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org{
807a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org    SkMorph_SSE2<kErode, kY>(src, dst, radius, width, height, srcStride, dstStride);
817a47ad3bacb36ca7609490bd7cbd85d2c49a0042senorblanco@chromium.org}
82