mtklein | 281b33f | 2016-07-12 15:01:26 -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 "Benchmark.h" |
| 9 | #include "SkRasterPipeline.h" |
| 10 | #include "SkSRGB.h" |
| 11 | |
| 12 | static const int N = 1023; |
| 13 | |
| 14 | static uint32_t dst[N], |
| 15 | src[N]; |
| 16 | static uint8_t mask[N]; |
| 17 | |
| 18 | // We'll build up a somewhat realistic useful pipeline: |
| 19 | // - load srgb src |
| 20 | // - scale src by 8-bit mask |
| 21 | // - load srgb dst |
| 22 | // - src = srcover(dst, src) |
| 23 | // - store src back as srgb |
| 24 | // Every stage except for srcover interacts with memory, and so will need _tail variants. |
| 25 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 26 | SK_RASTER_STAGE(load_s_srgb) { |
| 27 | auto ptr = (const uint32_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 28 | |
| 29 | r = Sk4f{ sk_linear_from_srgb[(ptr[0] >> 0) & 0xff], |
| 30 | sk_linear_from_srgb[(ptr[1] >> 0) & 0xff], |
| 31 | sk_linear_from_srgb[(ptr[2] >> 0) & 0xff], |
| 32 | sk_linear_from_srgb[(ptr[3] >> 0) & 0xff] }; |
| 33 | |
| 34 | g = Sk4f{ sk_linear_from_srgb[(ptr[0] >> 8) & 0xff], |
| 35 | sk_linear_from_srgb[(ptr[1] >> 8) & 0xff], |
| 36 | sk_linear_from_srgb[(ptr[2] >> 8) & 0xff], |
| 37 | sk_linear_from_srgb[(ptr[3] >> 8) & 0xff] }; |
| 38 | |
| 39 | b = Sk4f{ sk_linear_from_srgb[(ptr[0] >> 16) & 0xff], |
| 40 | sk_linear_from_srgb[(ptr[1] >> 16) & 0xff], |
| 41 | sk_linear_from_srgb[(ptr[2] >> 16) & 0xff], |
| 42 | sk_linear_from_srgb[(ptr[3] >> 16) & 0xff] }; |
| 43 | |
| 44 | a = SkNx_cast<float>((Sk4i::Load(ptr) >> 24) & 0xff) * (1/255.0f); |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 45 | } |
| 46 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 47 | SK_RASTER_STAGE(load_s_srgb_tail) { |
| 48 | auto ptr = (const uint32_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 49 | |
| 50 | r = Sk4f{ sk_linear_from_srgb[(*ptr >> 0) & 0xff], 0,0,0 }; |
| 51 | g = Sk4f{ sk_linear_from_srgb[(*ptr >> 8) & 0xff], 0,0,0 }; |
| 52 | b = Sk4f{ sk_linear_from_srgb[(*ptr >> 16) & 0xff], 0,0,0 }; |
| 53 | a = Sk4f{ (*ptr >> 24) * (1/255.0f), 0,0,0 }; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 54 | } |
| 55 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 56 | SK_RASTER_STAGE(load_d_srgb) { |
| 57 | auto ptr = (const uint32_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 58 | |
| 59 | dr = Sk4f{ sk_linear_from_srgb[(ptr[0] >> 0) & 0xff], |
| 60 | sk_linear_from_srgb[(ptr[1] >> 0) & 0xff], |
| 61 | sk_linear_from_srgb[(ptr[2] >> 0) & 0xff], |
| 62 | sk_linear_from_srgb[(ptr[3] >> 0) & 0xff] }; |
| 63 | |
| 64 | dg = Sk4f{ sk_linear_from_srgb[(ptr[0] >> 8) & 0xff], |
| 65 | sk_linear_from_srgb[(ptr[1] >> 8) & 0xff], |
| 66 | sk_linear_from_srgb[(ptr[2] >> 8) & 0xff], |
| 67 | sk_linear_from_srgb[(ptr[3] >> 8) & 0xff] }; |
| 68 | |
| 69 | db = Sk4f{ sk_linear_from_srgb[(ptr[0] >> 16) & 0xff], |
| 70 | sk_linear_from_srgb[(ptr[1] >> 16) & 0xff], |
| 71 | sk_linear_from_srgb[(ptr[2] >> 16) & 0xff], |
| 72 | sk_linear_from_srgb[(ptr[3] >> 16) & 0xff] }; |
| 73 | |
| 74 | da = SkNx_cast<float>((Sk4i::Load(ptr) >> 24) & 0xff) * (1/255.0f); |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 75 | } |
| 76 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 77 | SK_RASTER_STAGE(load_d_srgb_tail) { |
| 78 | auto ptr = (const uint32_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 79 | |
| 80 | dr = Sk4f{ sk_linear_from_srgb[(*ptr >> 0) & 0xff], 0,0,0 }; |
| 81 | dg = Sk4f{ sk_linear_from_srgb[(*ptr >> 8) & 0xff], 0,0,0 }; |
| 82 | db = Sk4f{ sk_linear_from_srgb[(*ptr >> 16) & 0xff], 0,0,0 }; |
| 83 | da = Sk4f{ (*ptr >> 24) * (1/255.0f), 0,0,0 }; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 84 | } |
| 85 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 86 | SK_RASTER_STAGE(scale_u8) { |
| 87 | auto ptr = (const uint8_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 88 | |
| 89 | auto c = SkNx_cast<float>(Sk4b::Load(ptr)) * (1/255.0f); |
| 90 | r *= c; |
| 91 | g *= c; |
| 92 | b *= c; |
| 93 | a *= c; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 94 | } |
| 95 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 96 | SK_RASTER_STAGE(scale_u8_tail) { |
| 97 | auto ptr = (const uint8_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 98 | |
| 99 | auto c = *ptr * (1/255.0f); |
| 100 | r *= c; |
| 101 | g *= c; |
| 102 | b *= c; |
| 103 | a *= c; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 104 | } |
| 105 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 106 | SK_RASTER_STAGE(srcover) { |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 107 | auto A = 1.0f - a; |
| 108 | r += dr * A; |
| 109 | g += dg * A; |
| 110 | b += db * A; |
| 111 | a += da * A; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 112 | } |
| 113 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 114 | SK_RASTER_STAGE(store_srgb) { |
| 115 | auto ptr = (uint32_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 116 | |
mtklein | 566ea9b | 2016-07-20 12:10:11 -0700 | [diff] [blame] | 117 | ( sk_linear_to_srgb(r) |
| 118 | | sk_linear_to_srgb(g) << 8 |
| 119 | | sk_linear_to_srgb(b) << 16 |
| 120 | | Sk4f_round(255.0f*a) << 24).store(ptr); |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 121 | } |
| 122 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 123 | SK_RASTER_STAGE(store_srgb_tail) { |
| 124 | auto ptr = (uint32_t*)ctx + x; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 125 | |
mtklein | 566ea9b | 2016-07-20 12:10:11 -0700 | [diff] [blame] | 126 | Sk4i rgba = sk_linear_to_srgb({r[0], g[0], b[0], 0}); |
| 127 | rgba = {rgba[0], rgba[1], rgba[2], (int)(255.0f * a[0] + 0.5f)}; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 128 | |
| 129 | SkNx_cast<uint8_t>(rgba).store(ptr); |
| 130 | } |
| 131 | |
| 132 | class SkRasterPipelineBench : public Benchmark { |
| 133 | public: |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 134 | SkRasterPipelineBench(bool fused) : fFused(fused) {} |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 135 | |
| 136 | bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; } |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 137 | const char* onGetName() override { return fFused ? "SkRasterPipelineBench_fused" |
| 138 | : "SkRasterPipelineBench_pipeline"; } |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 139 | |
| 140 | void onDraw(int loops, SkCanvas*) override { |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 141 | while (loops --> 0) { |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 142 | fFused ? this->runFused() : this->runPipeline(); |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 143 | } |
| 144 | } |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 145 | |
| 146 | void runFused() { |
| 147 | Sk4f r,g,b,a, dr,dg,db,da; |
| 148 | size_t x = 0, n = N; |
| 149 | while (n >= 4) { |
| 150 | load_s_srgb(src , x, r,g,b,a, dr,dg,db,da); |
| 151 | scale_u8 (mask , x, r,g,b,a, dr,dg,da,da); |
| 152 | load_d_srgb(dst , x, r,g,b,a, dr,dg,da,da); |
| 153 | srcover (nullptr, x, r,g,b,a, dr,dg,da,da); |
| 154 | store_srgb (dst , x, r,g,b,a, dr,dg,da,da); |
| 155 | |
| 156 | x += 4; |
| 157 | n -= 4; |
| 158 | } |
| 159 | while (n > 0) { |
| 160 | load_s_srgb_tail(src , x, r,g,b,a, dr,dg,db,da); |
| 161 | scale_u8_tail (mask , x, r,g,b,a, dr,dg,da,da); |
| 162 | load_d_srgb_tail(dst , x, r,g,b,a, dr,dg,da,da); |
| 163 | srcover (nullptr, x, r,g,b,a, dr,dg,da,da); |
| 164 | store_srgb_tail (dst , x, r,g,b,a, dr,dg,da,da); |
| 165 | |
| 166 | x += 1; |
| 167 | n -= 1; |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | void runPipeline() { |
| 172 | SkRasterPipeline p; |
| 173 | p.append<load_s_srgb, load_s_srgb_tail>( src); |
| 174 | p.append< scale_u8, scale_u8_tail>(mask); |
| 175 | p.append<load_d_srgb, load_d_srgb_tail>( dst); |
| 176 | p.append<srcover>(); |
| 177 | p.append< store_srgb, store_srgb_tail>( dst); |
| 178 | |
| 179 | p.run(N); |
| 180 | } |
| 181 | |
| 182 | bool fFused; |
mtklein | 281b33f | 2016-07-12 15:01:26 -0700 | [diff] [blame] | 183 | }; |
| 184 | |
mtklein | fe2042e | 2016-07-29 14:27:41 -0700 | [diff] [blame] | 185 | DEF_BENCH( return new SkRasterPipelineBench(true); ) |
| 186 | DEF_BENCH( return new SkRasterPipelineBench(false); ) |