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