herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2016 Google Inc. |
| 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 <tuple> |
| 9 | |
| 10 | #include "Benchmark.h" |
| 11 | #include "Resources.h" |
| 12 | #include "SkCpu.h" |
| 13 | #include "SkImage.h" |
| 14 | #include "SkImage_Base.h" |
| 15 | #include "SkNx.h" |
| 16 | #include "SkOpts.h" |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 17 | #include "SkPM4fPriv.h" |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 18 | #include "SkString.h" |
| 19 | |
| 20 | #define INNER_LOOPS 10 |
| 21 | |
mtklein | 0c90247 | 2016-07-20 18:10:07 -0700 | [diff] [blame] | 22 | static inline void brute_srcover_srgb_srgb_1(uint32_t* dst, uint32_t src) { |
| 23 | auto d = Sk4f_fromS32(*dst), |
| 24 | s = Sk4f_fromS32( src); |
| 25 | *dst = Sk4f_toS32(s + d * (1.0f - s[3])); |
| 26 | } |
| 27 | |
| 28 | static inline void srcover_srgb_srgb_1(uint32_t* dst, uint32_t src) { |
| 29 | if (src >= 0xFF000000) { |
| 30 | *dst = src; |
| 31 | return; |
| 32 | } |
| 33 | brute_srcover_srgb_srgb_1(dst, src); |
| 34 | } |
| 35 | |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 36 | static void brute_force_srcover_srgb_srgb( |
| 37 | uint32_t* dst, const uint32_t* const src, int ndst, const int nsrc) { |
| 38 | while (ndst > 0) { |
| 39 | int n = SkTMin(ndst, nsrc); |
| 40 | |
| 41 | for (int i = 0; i < n; i++) { |
mtklein | 0c90247 | 2016-07-20 18:10:07 -0700 | [diff] [blame] | 42 | brute_srcover_srgb_srgb_1(dst++, src[i]); |
| 43 | } |
| 44 | ndst -= n; |
| 45 | } |
| 46 | } |
| 47 | |
| 48 | static void trivial_srcover_srgb_srgb( |
| 49 | uint32_t* dst, const uint32_t* const src, int ndst, const int nsrc) { |
| 50 | while (ndst > 0) { |
| 51 | int n = SkTMin(ndst, nsrc); |
| 52 | |
| 53 | for (int i = 0; i < n; i++) { |
| 54 | srcover_srgb_srgb_1(dst++, src[i]); |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 55 | } |
| 56 | ndst -= n; |
| 57 | } |
| 58 | } |
| 59 | |
| 60 | static void best_non_simd_srcover_srgb_srgb( |
| 61 | uint32_t* dst, const uint32_t* const src, int ndst, const int nsrc) { |
| 62 | uint64_t* ddst = reinterpret_cast<uint64_t*>(dst); |
| 63 | |
| 64 | auto srcover_srgb_srgb_2 = [](uint32_t* dst, const uint32_t* src) { |
mtklein | 0c90247 | 2016-07-20 18:10:07 -0700 | [diff] [blame] | 65 | srcover_srgb_srgb_1(dst++, *src++); |
| 66 | srcover_srgb_srgb_1(dst, *src); |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 67 | }; |
| 68 | |
| 69 | while (ndst >0) { |
| 70 | int count = SkTMin(ndst, nsrc); |
| 71 | ndst -= count; |
| 72 | const uint64_t* dsrc = reinterpret_cast<const uint64_t*>(src); |
| 73 | const uint64_t* end = dsrc + (count >> 1); |
| 74 | do { |
| 75 | if ((~*dsrc & 0xFF000000FF000000) == 0) { |
| 76 | do { |
| 77 | *ddst++ = *dsrc++; |
| 78 | } while (dsrc < end && (~*dsrc & 0xFF000000FF000000) == 0); |
| 79 | } else if ((*dsrc & 0xFF000000FF000000) == 0) { |
| 80 | do { |
| 81 | dsrc++; |
| 82 | ddst++; |
| 83 | } while (dsrc < end && (*dsrc & 0xFF000000FF000000) == 0); |
| 84 | } else { |
| 85 | srcover_srgb_srgb_2(reinterpret_cast<uint32_t*>(ddst++), |
| 86 | reinterpret_cast<const uint32_t*>(dsrc++)); |
| 87 | } |
| 88 | } while (dsrc < end); |
| 89 | |
| 90 | if ((count & 1) != 0) { |
mtklein | 0c90247 | 2016-07-20 18:10:07 -0700 | [diff] [blame] | 91 | srcover_srgb_srgb_1(reinterpret_cast<uint32_t*>(ddst), |
| 92 | *reinterpret_cast<const uint32_t*>(dsrc)); |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 93 | } |
| 94 | } |
| 95 | } |
| 96 | |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 97 | class SrcOverVSkOptsBruteForce { |
| 98 | public: |
| 99 | static SkString Name() { return SkString{"VSkOptsBruteForce"}; } |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 100 | static void BlendN(uint32_t* dst, const uint32_t* src, int count) { |
| 101 | brute_force_srcover_srgb_srgb(dst, src, count, count); |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 102 | } |
| 103 | }; |
| 104 | |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 105 | class SrcOverVSkOptsTrivial { |
| 106 | public: |
| 107 | static SkString Name() { return SkString{"VSkOptsTrivial"}; } |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 108 | static void BlendN(uint32_t* dst, const uint32_t* src, int count) { |
| 109 | trivial_srcover_srgb_srgb(dst, src, count, count); |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 110 | } |
| 111 | }; |
| 112 | |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 113 | class SrcOverVSkOptsNonSimdCore { |
| 114 | public: |
| 115 | static SkString Name() { return SkString{"VSkOptsNonSimdCore"}; } |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 116 | static void BlendN(uint32_t* dst, const uint32_t* src, int count) { |
| 117 | best_non_simd_srcover_srgb_srgb(dst, src, count, count); |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 118 | } |
| 119 | }; |
| 120 | |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 121 | class SrcOverVSkOptsDefault { |
| 122 | public: |
| 123 | static SkString Name() { return SkString{"VSkOptsDefault"}; } |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 124 | static void BlendN(uint32_t* dst, const uint32_t* src, int count) { |
mtklein | 0358a6a | 2016-07-13 08:02:20 -0700 | [diff] [blame] | 125 | SkOpts::srcover_srgb_srgb(dst, src, count, count); |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 126 | } |
| 127 | }; |
| 128 | |
| 129 | /////////////////////////////////////////////////////////////////////////////////////////////////// |
| 130 | |
| 131 | template <typename Blender> |
| 132 | class LinearSrcOverBench : public Benchmark { |
| 133 | public: |
bsalomon | 809f258 | 2016-05-20 08:54:23 -0700 | [diff] [blame] | 134 | LinearSrcOverBench(const char* fileName) : fFileName(fileName) { |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 135 | fName = "LinearSrcOver_"; |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 136 | fName.append(fileName); |
| 137 | fName.append(Blender::Name()); |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 138 | } |
| 139 | |
| 140 | protected: |
mtklein | 0358a6a | 2016-07-13 08:02:20 -0700 | [diff] [blame] | 141 | bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; } |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 142 | const char* onGetName() override { return fName.c_str(); } |
bsalomon | 809f258 | 2016-05-20 08:54:23 -0700 | [diff] [blame] | 143 | |
| 144 | void onPreDraw(SkCanvas*) override { |
| 145 | if (!fPixmap.addr()) { |
| 146 | sk_sp<SkImage> image = GetResourceAsImage(fFileName.c_str()); |
| 147 | SkBitmap bm; |
| 148 | if (!as_IB(image)->getROPixels(&bm)) { |
| 149 | SkFAIL("Could not read resource"); |
| 150 | } |
| 151 | bm.peekPixels(&fPixmap); |
| 152 | fCount = fPixmap.rowBytesAsPixels(); |
| 153 | fDst.reset(fCount); |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 154 | sk_bzero(fDst.get(), fPixmap.rowBytes()); |
bsalomon | 809f258 | 2016-05-20 08:54:23 -0700 | [diff] [blame] | 155 | } |
| 156 | } |
| 157 | |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 158 | void onDraw(int loops, SkCanvas*) override { |
| 159 | SkASSERT(fPixmap.colorType() == kN32_SkColorType); |
| 160 | |
| 161 | const int width = fPixmap.rowBytesAsPixels(); |
| 162 | |
| 163 | for (int i = 0; i < loops * INNER_LOOPS; ++i) { |
| 164 | const uint32_t* src = fPixmap.addr32(); |
| 165 | for (int y = 0; y < fPixmap.height(); y++) { |
herb | 2edf0c6 | 2016-07-12 15:00:46 -0700 | [diff] [blame] | 166 | Blender::BlendN(fDst.get(), src, width); |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 167 | src += width; |
| 168 | } |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | void onPostDraw(SkCanvas*) override { |
| 173 | // Make sure the compiler does not optimize away the operation. |
| 174 | volatile uint32_t v = 0; |
| 175 | for (int i = 0; i < fCount; i++) { |
| 176 | v ^= fDst[i]; |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | private: |
| 181 | int fCount; |
| 182 | SkAutoTArray<uint32_t> fDst; |
bsalomon | 809f258 | 2016-05-20 08:54:23 -0700 | [diff] [blame] | 183 | SkString fFileName; |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 184 | SkString fName; |
| 185 | SkPixmap fPixmap; |
| 186 | |
| 187 | typedef Benchmark INHERITED; |
| 188 | }; |
| 189 | |
mtklein | 0358a6a | 2016-07-13 08:02:20 -0700 | [diff] [blame] | 190 | #define BENCHES(fileName) \ |
| 191 | DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsBruteForce>(fileName); ) \ |
| 192 | DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsTrivial>(fileName); ) \ |
| 193 | DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsNonSimdCore>(fileName); ) \ |
| 194 | DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsDefault>(fileName); ) |
herb | cc49e59 | 2016-05-17 09:57:34 -0700 | [diff] [blame] | 195 | |
| 196 | BENCHES("yellow_rose.png") |
| 197 | BENCHES("baby_tux.png") |
| 198 | BENCHES("plane.png") |
| 199 | BENCHES("mandrill_512.png") |
| 200 | BENCHES("iconstrip.png") |