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
14 namespace {
15 enum 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 */
expand(int a)22 inline __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
35 template<BlurDirection srcDirection, BlurDirection dstDirection>
SkBoxBlur_SSE2(const SkPMColor * src,int srcStride,SkPMColor * dst,int kernelSize,int leftOffset,int rightOffset,int width,int height)36 void 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 // SSE2 has no PMULLUD, so we must do AG and RB separately.
59 __m128i tmp1 = _mm_mul_epu32(sum, scale);
60 __m128i tmp2 = _mm_mul_epu32(_mm_srli_si128(sum, 4),
61 _mm_srli_si128(scale, 4));
62 __m128i result = _mm_unpacklo_epi32(_mm_shuffle_epi32(tmp1, _MM_SHUFFLE(0,0,2,0)),
63 _mm_shuffle_epi32(tmp2, _MM_SHUFFLE(0,0,2,0)));
64
65 // sumA*scale+.5 sumB*scale+.5 sumG*scale+.5 sumB*scale+.5
66 result = _mm_add_epi32(result, half);
67
68 // 0 0 0 A 0 0 0 R 0 0 0 G 0 0 0 B
69 result = _mm_srli_epi32(result, 24);
70
71 // 0 0 0 0 0 0 0 0 0 A 0 R 0 G 0 B
72 result = _mm_packs_epi32(result, zero);
73
74 // 0 0 0 0 0 0 0 0 0 0 0 0 A R G B
75 result = _mm_packus_epi16(result, zero);
76 *dptr = _mm_cvtsi128_si32(result);
77 if (x >= leftOffset) {
78 SkColor l = *(sptr - leftOffset * srcStrideX);
79 sum = _mm_sub_epi32(sum, expand(l));
80 }
81 if (x + rightOffset + 1 < width) {
82 SkColor r = *(sptr + (rightOffset + 1) * srcStrideX);
83 sum = _mm_add_epi32(sum, expand(r));
84 }
85 sptr += srcStrideX;
86 if (srcDirection == kY) {
87 _mm_prefetch(reinterpret_cast<const char*>(sptr + (rightOffset + 1) * srcStrideX),
88 _MM_HINT_T0);
89 }
90 dptr += dstStrideX;
91 }
92 src += srcStrideY;
93 dst += dstStrideY;
94 }
95 }
96
97 } // namespace
98
SkBoxBlurGetPlatformProcs_SSE2(SkBoxBlurProc * boxBlurX,SkBoxBlurProc * boxBlurXY,SkBoxBlurProc * boxBlurYX)99 bool SkBoxBlurGetPlatformProcs_SSE2(SkBoxBlurProc* boxBlurX,
100 SkBoxBlurProc* boxBlurXY,
101 SkBoxBlurProc* boxBlurYX) {
102 *boxBlurX = SkBoxBlur_SSE2<kX, kX>;
103 *boxBlurXY = SkBoxBlur_SSE2<kX, kY>;
104 *boxBlurYX = SkBoxBlur_SSE2<kY, kX>;
105 return true;
106 }
107