Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <algorithm> |
| 7 | #include <cmath> |
| 8 | #include <functional> |
| 9 | #include <random> |
| 10 | #include <vector> |
| 11 | |
| 12 | #include <benchmark/benchmark.h> |
| 13 | #include "bench/utils.h" |
| 14 | #include <xnnpack/AlignedAllocator.h> |
| 15 | #include <xnnpack/common.h> |
| 16 | #include <xnnpack/params.h> |
| 17 | #include <xnnpack/raddstoreexpminusmax.h> |
| 18 | #include <xnnpack/rmax.h> |
| 19 | |
| 20 | |
| 21 | static void f32_raddstoreexpminusmax( |
| 22 | benchmark::State& state, |
| 23 | xnn_f32_rmax_ukernel_function rmax, |
| 24 | xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax, |
| 25 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 26 | { |
| 27 | if (isa_check && !isa_check(state)) { |
| 28 | return; |
| 29 | } |
| 30 | |
| 31 | const size_t n = state.range(0); |
| 32 | const size_t cache_line_size_max = 128; |
| 33 | const size_t packed_n = benchmark::utils::RoundUp(n, cache_line_size_max / sizeof(float)); |
| 34 | |
| 35 | std::random_device random_device; |
| 36 | auto rng = std::mt19937(random_device()); |
| 37 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-1000.0f, 1000.0f), rng); |
| 38 | |
| 39 | const size_t num_buffers = 1 + |
| 40 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), packed_n * sizeof(float)); |
| 41 | std::vector<float, AlignedAllocator<float, 64>> x(n); |
| 42 | std::vector<float, AlignedAllocator<float, 64>> y(packed_n * num_buffers); |
| 43 | |
| 44 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 45 | |
| 46 | benchmark::utils::DisableDenormals(); |
| 47 | |
| 48 | size_t buffer_index = 0; |
| 49 | for (auto _ : state) { |
| 50 | state.PauseTiming(); |
| 51 | float x_max = nanf(""); |
| 52 | rmax(n * sizeof(float), x.data(), &x_max); |
| 53 | if (++buffer_index == num_buffers) { |
| 54 | buffer_index = 0; |
| 55 | } |
| 56 | state.ResumeTiming(); |
| 57 | |
| 58 | float y_sum = nanf(""); |
| 59 | raddstoreexpminusmax(n * sizeof(float), x.data(), y.data() + buffer_index * packed_n, &y_sum, x_max); |
| 60 | } |
| 61 | |
| 62 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 63 | state.counters["elements"] = |
| 64 | benchmark::Counter(uint64_t(state.iterations()) * n, benchmark::Counter::kIsRate); |
| 65 | state.counters["bytes"] = |
| 66 | benchmark::Counter(uint64_t(state.iterations()) * 2 * sizeof(float) * n, benchmark::Counter::kIsRate); |
| 67 | } |
| 68 | |
| 69 | static void CharacteristicArguments(benchmark::internal::Benchmark* b) { |
| 70 | b->ArgName("N"); |
| 71 | for (int32_t n = 10000; n <= 100000000; n *= 10) { |
| 72 | b->Arg(n); |
| 73 | } |
| 74 | } |
| 75 | |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 76 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 77 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x4, |
| 78 | xnn_f32_rmax_ukernel__neon, |
| 79 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x4, |
| 80 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 81 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x8, |
| 82 | xnn_f32_rmax_ukernel__neon, |
| 83 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8, |
| 84 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 85 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x8_acc2, |
| 86 | xnn_f32_rmax_ukernel__neon, |
| 87 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2, |
| 88 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 89 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x12, |
| 90 | xnn_f32_rmax_ukernel__neon, |
| 91 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12, |
| 92 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 93 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x12_acc2, |
| 94 | xnn_f32_rmax_ukernel__neon, |
| 95 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2, |
| 96 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 97 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x12_acc3, |
| 98 | xnn_f32_rmax_ukernel__neon, |
| 99 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3, |
| 100 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 101 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x16, |
| 102 | xnn_f32_rmax_ukernel__neon, |
| 103 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16, |
| 104 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 105 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x16_acc2, |
| 106 | xnn_f32_rmax_ukernel__neon, |
| 107 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2, |
| 108 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 109 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x16_acc4, |
| 110 | xnn_f32_rmax_ukernel__neon, |
| 111 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4, |
| 112 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 113 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x20, |
| 114 | xnn_f32_rmax_ukernel__neon, |
| 115 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20, |
| 116 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 117 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x20_acc2, |
| 118 | xnn_f32_rmax_ukernel__neon, |
| 119 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2, |
| 120 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 121 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_p5_x20_acc5, |
| 122 | xnn_f32_rmax_ukernel__neon, |
| 123 | xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5, |
| 124 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 125 | |
| 126 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x4, |
| 127 | xnn_f32_rmax_ukernel__neon, |
| 128 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x4, |
| 129 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 130 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x8, |
| 131 | xnn_f32_rmax_ukernel__neon, |
| 132 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8, |
| 133 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 134 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x8_acc2, |
| 135 | xnn_f32_rmax_ukernel__neon, |
| 136 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2, |
| 137 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 138 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x12, |
| 139 | xnn_f32_rmax_ukernel__neon, |
| 140 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12, |
| 141 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 142 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x12_acc2, |
| 143 | xnn_f32_rmax_ukernel__neon, |
| 144 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2, |
| 145 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 146 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x12_acc3, |
| 147 | xnn_f32_rmax_ukernel__neon, |
| 148 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3, |
| 149 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 150 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x16, |
| 151 | xnn_f32_rmax_ukernel__neon, |
| 152 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16, |
| 153 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 154 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x16_acc2, |
| 155 | xnn_f32_rmax_ukernel__neon, |
| 156 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2, |
| 157 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 158 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x16_acc4, |
| 159 | xnn_f32_rmax_ukernel__neon, |
| 160 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4, |
| 161 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 162 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x20, |
| 163 | xnn_f32_rmax_ukernel__neon, |
| 164 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20, |
| 165 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 166 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x20_acc2, |
| 167 | xnn_f32_rmax_ukernel__neon, |
| 168 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2, |
| 169 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 170 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neon_lut64_p2_x20_acc5, |
| 171 | xnn_f32_rmax_ukernel__neon, |
| 172 | xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5, |
| 173 | benchmark::utils::CheckNEON)->Apply(CharacteristicArguments)->UseRealTime(); |
| 174 | |
| 175 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x4, |
| 176 | xnn_f32_rmax_ukernel__neon, |
| 177 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x4, |
| 178 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 179 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x8, |
| 180 | xnn_f32_rmax_ukernel__neon, |
| 181 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8, |
| 182 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 183 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x8_acc2, |
| 184 | xnn_f32_rmax_ukernel__neon, |
| 185 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2, |
| 186 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 187 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x12, |
| 188 | xnn_f32_rmax_ukernel__neon, |
| 189 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12, |
| 190 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 191 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x12_acc2, |
| 192 | xnn_f32_rmax_ukernel__neon, |
| 193 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2, |
| 194 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 195 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x12_acc3, |
| 196 | xnn_f32_rmax_ukernel__neon, |
| 197 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3, |
| 198 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 199 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x16, |
| 200 | xnn_f32_rmax_ukernel__neon, |
| 201 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16, |
| 202 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 203 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x16_acc2, |
| 204 | xnn_f32_rmax_ukernel__neon, |
| 205 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2, |
| 206 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 207 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x16_acc4, |
| 208 | xnn_f32_rmax_ukernel__neon, |
| 209 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4, |
| 210 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 211 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x20, |
| 212 | xnn_f32_rmax_ukernel__neon, |
| 213 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20, |
| 214 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 215 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x20_acc2, |
| 216 | xnn_f32_rmax_ukernel__neon, |
| 217 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2, |
| 218 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 219 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_p5_x20_acc5, |
| 220 | xnn_f32_rmax_ukernel__neon, |
| 221 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5, |
| 222 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 223 | |
| 224 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x4, |
| 225 | xnn_f32_rmax_ukernel__neon, |
| 226 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x4, |
| 227 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 228 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x8, |
| 229 | xnn_f32_rmax_ukernel__neon, |
| 230 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8, |
| 231 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 232 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x8_acc2, |
| 233 | xnn_f32_rmax_ukernel__neon, |
| 234 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2, |
| 235 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 236 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x12, |
| 237 | xnn_f32_rmax_ukernel__neon, |
| 238 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12, |
| 239 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 240 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x12_acc2, |
| 241 | xnn_f32_rmax_ukernel__neon, |
| 242 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2, |
| 243 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 244 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x12_acc3, |
| 245 | xnn_f32_rmax_ukernel__neon, |
| 246 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3, |
| 247 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 248 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x16, |
| 249 | xnn_f32_rmax_ukernel__neon, |
| 250 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16, |
| 251 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 252 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x16_acc2, |
| 253 | xnn_f32_rmax_ukernel__neon, |
| 254 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2, |
| 255 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 256 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x16_acc4, |
| 257 | xnn_f32_rmax_ukernel__neon, |
| 258 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4, |
| 259 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 260 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x20, |
| 261 | xnn_f32_rmax_ukernel__neon, |
| 262 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20, |
| 263 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 264 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x20_acc2, |
| 265 | xnn_f32_rmax_ukernel__neon, |
| 266 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2, |
| 267 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 268 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, neonfma_lut64_p2_x20_acc5, |
| 269 | xnn_f32_rmax_ukernel__neon, |
| 270 | xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5, |
| 271 | benchmark::utils::CheckNEONFMA)->Apply(CharacteristicArguments)->UseRealTime(); |
| 272 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 273 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 274 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 275 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x128, |
| 276 | xnn_f32_rmax_ukernel__avx, |
| 277 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128, |
| 278 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 279 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x128_acc2, |
| 280 | xnn_f32_rmax_ukernel__avx, |
| 281 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2, |
| 282 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 283 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x128_acc4, |
| 284 | xnn_f32_rmax_ukernel__avx, |
| 285 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4, |
| 286 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 287 | |
| 288 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x144, |
| 289 | xnn_f32_rmax_ukernel__avx, |
| 290 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144, |
| 291 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 292 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x144_acc3, |
| 293 | xnn_f32_rmax_ukernel__avx, |
| 294 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3, |
| 295 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 296 | |
| 297 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x160, |
| 298 | xnn_f32_rmax_ukernel__avx, |
| 299 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160, |
| 300 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 301 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x160_acc2, |
| 302 | xnn_f32_rmax_ukernel__avx, |
| 303 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2, |
| 304 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 305 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x160_acc5, |
| 306 | xnn_f32_rmax_ukernel__avx, |
| 307 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5, |
| 308 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 309 | |
| 310 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x192, |
| 311 | xnn_f32_rmax_ukernel__avx, |
| 312 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192, |
| 313 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 314 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x192_acc2, |
| 315 | xnn_f32_rmax_ukernel__avx, |
| 316 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2, |
| 317 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 318 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x192_acc3, |
| 319 | xnn_f32_rmax_ukernel__avx, |
| 320 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3, |
| 321 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 322 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx512f_p5_scalef_x192_acc6, |
| 323 | xnn_f32_rmax_ukernel__avx, |
| 324 | xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6, |
| 325 | benchmark::utils::CheckAVX512F)->Apply(CharacteristicArguments)->UseRealTime(); |
| 326 | |
| 327 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x64, |
| 328 | xnn_f32_rmax_ukernel__avx, |
| 329 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64, |
| 330 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 331 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x64_acc2, |
| 332 | xnn_f32_rmax_ukernel__avx, |
| 333 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2, |
| 334 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 335 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x64_acc4, |
| 336 | xnn_f32_rmax_ukernel__avx, |
| 337 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4, |
| 338 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 339 | |
| 340 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x72, |
| 341 | xnn_f32_rmax_ukernel__avx, |
| 342 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72, |
| 343 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 344 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x72_acc3, |
| 345 | xnn_f32_rmax_ukernel__avx, |
| 346 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72_acc3, |
| 347 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 348 | |
| 349 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x80, |
| 350 | xnn_f32_rmax_ukernel__avx, |
| 351 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80, |
| 352 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 353 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x80_acc2, |
| 354 | xnn_f32_rmax_ukernel__avx, |
| 355 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2, |
| 356 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 357 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x80_acc5, |
| 358 | xnn_f32_rmax_ukernel__avx, |
| 359 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5, |
| 360 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 361 | |
| 362 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x96, |
| 363 | xnn_f32_rmax_ukernel__avx, |
| 364 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96, |
| 365 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 366 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x96_acc2, |
| 367 | xnn_f32_rmax_ukernel__avx, |
| 368 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2, |
| 369 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 370 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x96_acc3, |
| 371 | xnn_f32_rmax_ukernel__avx, |
| 372 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3, |
| 373 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 374 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, avx2_p5_x96_acc6, |
| 375 | xnn_f32_rmax_ukernel__avx, |
| 376 | xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6, |
| 377 | benchmark::utils::CheckAVX2)->Apply(CharacteristicArguments)->UseRealTime(); |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 378 | |
| 379 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x4, |
| 380 | xnn_f32_rmax_ukernel__sse, |
| 381 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 382 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x8, |
| 383 | xnn_f32_rmax_ukernel__sse, |
| 384 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8)->Apply(CharacteristicArguments)->UseRealTime(); |
| 385 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x8_acc2, |
| 386 | xnn_f32_rmax_ukernel__sse, |
| 387 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 388 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x12, |
| 389 | xnn_f32_rmax_ukernel__sse, |
| 390 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12)->Apply(CharacteristicArguments)->UseRealTime(); |
| 391 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x12_acc2, |
| 392 | xnn_f32_rmax_ukernel__sse, |
| 393 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 394 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x12_acc3, |
| 395 | xnn_f32_rmax_ukernel__sse, |
| 396 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3)->Apply(CharacteristicArguments)->UseRealTime(); |
| 397 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x16, |
| 398 | xnn_f32_rmax_ukernel__sse, |
| 399 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16)->Apply(CharacteristicArguments)->UseRealTime(); |
| 400 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x16_acc2, |
| 401 | xnn_f32_rmax_ukernel__sse, |
| 402 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 403 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x16_acc4, |
| 404 | xnn_f32_rmax_ukernel__sse, |
| 405 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 406 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x20, |
| 407 | xnn_f32_rmax_ukernel__sse, |
| 408 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20)->Apply(CharacteristicArguments)->UseRealTime(); |
| 409 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x20_acc2, |
| 410 | xnn_f32_rmax_ukernel__sse, |
| 411 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 412 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, sse2_p5_x20_acc5, |
| 413 | xnn_f32_rmax_ukernel__sse, |
| 414 | xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5)->Apply(CharacteristicArguments)->UseRealTime(); |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 415 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 416 | |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 417 | #if !XNN_ARCH_WASM && !XNN_ARCH_ASMJS |
| 418 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x4, |
| 419 | xnn_f32_rmax_ukernel__psimd, |
| 420 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 421 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x8, |
| 422 | xnn_f32_rmax_ukernel__psimd, |
| 423 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8)->Apply(CharacteristicArguments)->UseRealTime(); |
| 424 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x8_acc2, |
| 425 | xnn_f32_rmax_ukernel__psimd, |
| 426 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 427 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x12, |
| 428 | xnn_f32_rmax_ukernel__psimd, |
| 429 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12)->Apply(CharacteristicArguments)->UseRealTime(); |
| 430 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x12_acc2, |
| 431 | xnn_f32_rmax_ukernel__psimd, |
| 432 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 433 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x12_acc3, |
| 434 | xnn_f32_rmax_ukernel__psimd, |
| 435 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3)->Apply(CharacteristicArguments)->UseRealTime(); |
| 436 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x16, |
| 437 | xnn_f32_rmax_ukernel__psimd, |
| 438 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16)->Apply(CharacteristicArguments)->UseRealTime(); |
| 439 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x16_acc2, |
| 440 | xnn_f32_rmax_ukernel__psimd, |
| 441 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 442 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x16_acc4, |
| 443 | xnn_f32_rmax_ukernel__psimd, |
| 444 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 445 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x20, |
| 446 | xnn_f32_rmax_ukernel__psimd, |
| 447 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20)->Apply(CharacteristicArguments)->UseRealTime(); |
| 448 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x20_acc2, |
| 449 | xnn_f32_rmax_ukernel__psimd, |
| 450 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 451 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, psimd_p5_x20_acc5, |
| 452 | xnn_f32_rmax_ukernel__psimd, |
| 453 | xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc5)->Apply(CharacteristicArguments)->UseRealTime(); |
| 454 | #endif // !XNN_ARCH_WASM && !XNN_ARCH_ASMJS |
| 455 | |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 456 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_lut64_p2_x1, |
| 457 | xnn_f32_rmax_ukernel__scalar, |
| 458 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x1)->Apply(CharacteristicArguments)->UseRealTime(); |
| 459 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_lut64_p2_x2, |
| 460 | xnn_f32_rmax_ukernel__scalar, |
| 461 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 462 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_lut64_p2_x2_acc2, |
| 463 | xnn_f32_rmax_ukernel__scalar, |
| 464 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x2_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 465 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_lut64_p2_x4, |
| 466 | xnn_f32_rmax_ukernel__scalar, |
| 467 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 468 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_lut64_p2_x4_acc2, |
| 469 | xnn_f32_rmax_ukernel__scalar, |
| 470 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 471 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_lut64_p2_x4_acc4, |
| 472 | xnn_f32_rmax_ukernel__scalar, |
| 473 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4_acc4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 474 | |
| 475 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_p5_x1, |
| 476 | xnn_f32_rmax_ukernel__scalar, |
| 477 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x1)->Apply(CharacteristicArguments)->UseRealTime(); |
| 478 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_p5_x2, |
| 479 | xnn_f32_rmax_ukernel__scalar, |
| 480 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 481 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_p5_x2_acc2, |
| 482 | xnn_f32_rmax_ukernel__scalar, |
| 483 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x2_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 484 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_p5_x4, |
| 485 | xnn_f32_rmax_ukernel__scalar, |
| 486 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 487 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_p5_x4_acc2, |
| 488 | xnn_f32_rmax_ukernel__scalar, |
| 489 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4_acc2)->Apply(CharacteristicArguments)->UseRealTime(); |
| 490 | BENCHMARK_CAPTURE(f32_raddstoreexpminusmax, scalar_p5_x4_acc4, |
| 491 | xnn_f32_rmax_ukernel__scalar, |
| 492 | xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4_acc4)->Apply(CharacteristicArguments)->UseRealTime(); |
| 493 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 494 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 495 | BENCHMARK_MAIN(); |
| 496 | #endif |