1 /*
2 * Copyright 2014 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 "SkBitmap.h"
9 #include "SkBlurImage_opts_SSE4.h"
10 #include "SkColorPriv.h"
11 #include "SkRect.h"
12
13 /* With the exception of the compilers that don't support it, we always build the
14 * SSE4 functions and enable the caller to determine SSE4 support. However for
15 * compilers that do not support SSE4x we provide a stub implementation.
16 */
17 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
18
19 #include <smmintrin.h>
20
21 namespace {
22 enum BlurDirection {
23 kX, kY
24 };
25
26 /* Helper function to spread the components of a 32-bit integer into the
27 * lower 8 bits of each 32-bit element of an SSE register.
28 */
expand(int a)29 inline __m128i expand(int a) {
30 // ARGB -> 0000 0000 0000 ARGB
31 __m128i widened = _mm_cvtsi32_si128(a);
32 // SSE4.1 has xxxx xxxx xxxx ARGB -> 000A 000R 000G 000B as a one-stop-shop instruction.
33 // It can even work from memory, so a smart compiler probably merges in the _mm_cvtsi32_si128().
34 return _mm_cvtepu8_epi32(widened);
35 }
36
37 template<BlurDirection srcDirection, BlurDirection dstDirection>
SkBoxBlur_SSE4(const SkPMColor * src,int srcStride,SkPMColor * dst,int kernelSize,int leftOffset,int rightOffset,int width,int height)38 void SkBoxBlur_SSE4(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
39 int leftOffset, int rightOffset, int width, int height)
40 {
41 const int rightBorder = SkMin32(rightOffset + 1, width);
42 const int srcStrideX = srcDirection == kX ? 1 : srcStride;
43 const int dstStrideX = dstDirection == kX ? 1 : height;
44 const int srcStrideY = srcDirection == kX ? srcStride : 1;
45 const int dstStrideY = dstDirection == kX ? width : 1;
46 const __m128i scale = _mm_set1_epi32((1 << 24) / kernelSize);
47 const __m128i half = _mm_set1_epi32(1 << 23);
48 for (int y = 0; y < height; ++y) {
49 __m128i sum = _mm_setzero_si128();
50 const SkPMColor* p = src;
51 for (int i = 0; i < rightBorder; ++i) {
52 sum = _mm_add_epi32(sum, expand(*p));
53 p += srcStrideX;
54 }
55
56 const SkPMColor* sptr = src;
57 SkColor* dptr = dst;
58 for (int x = 0; x < width; ++x) {
59 // TODO(mtklein): We are working in 8.24 here. Drop to 8.8 when the kernel is narrow?
60
61 // Multiply each component by scale (i.e. divide by kernel size) and add half to round.
62 __m128i result = _mm_mullo_epi32(sum, scale);
63 result = _mm_add_epi32(result, half);
64
65 // Now pack the top byte of each 32-bit lane back down into one 32-bit color.
66 // Axxx Rxxx Gxxx Bxxx -> xxxx xxxx xxxx ARGB
67 const char _ = 0; // Don't care what ends up in these bytes. Happens to be byte 0.
68 result = _mm_shuffle_epi8(result, _mm_set_epi8(_,_,_,_, _,_,_,_, _,_,_,_, 15,11,7,3));
69
70 *dptr = _mm_cvtsi128_si32(result);
71
72 // TODO(mtklein): experiment with breaking this loop into 3 parts
73 if (x >= leftOffset) {
74 SkColor l = *(sptr - leftOffset * srcStrideX);
75 sum = _mm_sub_epi32(sum, expand(l));
76 }
77 if (x + rightOffset + 1 < width) {
78 SkColor r = *(sptr + (rightOffset + 1) * srcStrideX);
79 sum = _mm_add_epi32(sum, expand(r));
80 }
81 sptr += srcStrideX;
82 if (srcDirection == kY) {
83 // TODO(mtklein): experiment with moving this prefetch forward
84 _mm_prefetch(reinterpret_cast<const char*>(sptr + (rightOffset + 1) * srcStrideX),
85 _MM_HINT_T0);
86 }
87 dptr += dstStrideX;
88 }
89 src += srcStrideY;
90 dst += dstStrideY;
91 }
92 }
93
94 } // namespace
95
SkBoxBlurGetPlatformProcs_SSE4(SkBoxBlurProc * boxBlurX,SkBoxBlurProc * boxBlurXY,SkBoxBlurProc * boxBlurYX)96 bool SkBoxBlurGetPlatformProcs_SSE4(SkBoxBlurProc* boxBlurX,
97 SkBoxBlurProc* boxBlurXY,
98 SkBoxBlurProc* boxBlurYX) {
99 *boxBlurX = SkBoxBlur_SSE4<kX, kX>;
100 *boxBlurXY = SkBoxBlur_SSE4<kX, kY>;
101 *boxBlurYX = SkBoxBlur_SSE4<kY, kX>;
102 return true;
103 }
104
105 #else // SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
106
SkBoxBlurGetPlatformProcs_SSE4(SkBoxBlurProc * boxBlurX,SkBoxBlurProc * boxBlurXY,SkBoxBlurProc * boxBlurYX)107 bool SkBoxBlurGetPlatformProcs_SSE4(SkBoxBlurProc* boxBlurX,
108 SkBoxBlurProc* boxBlurXY,
109 SkBoxBlurProc* boxBlurYX) {
110 sk_throw();
111 return false;
112 }
113
114
115 #endif
116