Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 1 | // Copyright 2021 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 | |
| 15 | #include <fp16/fp16.h> |
| 16 | #include <xnnpack/AlignedAllocator.h> |
| 17 | #include <xnnpack/common.h> |
| 18 | #include <xnnpack/params.h> |
| 19 | #include <xnnpack/vcvt.h> |
| 20 | |
| 21 | |
| 22 | static void f16_f32_vcvt( |
| 23 | benchmark::State& state, |
| 24 | xnn_f16_f32_vcvt_ukernel_function cvt, |
| 25 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 26 | { |
| 27 | if (isa_check && !isa_check(state)) { |
| 28 | return; |
| 29 | } |
| 30 | |
| 31 | const size_t num_elements = state.range(0); |
| 32 | |
| 33 | std::random_device random_device; |
| 34 | auto rng = std::mt19937(random_device()); |
| 35 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
| 36 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 37 | |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 38 | std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> x(num_elements + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 39 | std::vector<float, AlignedAllocator<float, 64>> y(num_elements); |
| 40 | std::generate(x.begin(), x.end(), std::ref(f16rng)); |
| 41 | std::fill(y.begin(), y.end(), std::nanf("")); |
| 42 | |
| 43 | for (auto _ : state) { |
| 44 | cvt(num_elements * sizeof(float), x.data(), y.data(), nullptr /* params */); |
| 45 | } |
| 46 | |
| 47 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 48 | if (cpu_frequency != 0) { |
| 49 | state.counters["cpufreq"] = cpu_frequency; |
| 50 | } |
| 51 | |
| 52 | const size_t elements_per_iteration = num_elements; |
| 53 | state.counters["elements"] = |
| 54 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 55 | |
| 56 | const size_t bytes_per_iteration = num_elements * (sizeof(uint16_t) + sizeof(float)); |
| 57 | state.counters["bytes"] = |
| 58 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 59 | } |
| 60 | |
| 61 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 62 | BENCHMARK_CAPTURE(f16_f32_vcvt, neonfp16_x8, |
| 63 | xnn_f16_f32_vcvt_ukernel__neonfp16_x8, |
| 64 | benchmark::utils::CheckNEONFP16) |
| 65 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 66 | ->UseRealTime(); |
| 67 | BENCHMARK_CAPTURE(f16_f32_vcvt, neonfp16_x16, |
| 68 | xnn_f16_f32_vcvt_ukernel__neonfp16_x16, |
| 69 | benchmark::utils::CheckNEONFP16) |
| 70 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 71 | ->UseRealTime(); |
| 72 | |
| 73 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x8, |
| 74 | xnn_f16_f32_vcvt_ukernel__neon_int16_x8, |
| 75 | benchmark::utils::CheckNEON) |
| 76 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 77 | ->UseRealTime(); |
| 78 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x16, |
| 79 | xnn_f16_f32_vcvt_ukernel__neon_int16_x16, |
| 80 | benchmark::utils::CheckNEON) |
| 81 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 82 | ->UseRealTime(); |
| 83 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x24, |
| 84 | xnn_f16_f32_vcvt_ukernel__neon_int16_x24, |
| 85 | benchmark::utils::CheckNEON) |
| 86 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 87 | ->UseRealTime(); |
| 88 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x32, |
| 89 | xnn_f16_f32_vcvt_ukernel__neon_int16_x32, |
| 90 | benchmark::utils::CheckNEON) |
| 91 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 92 | ->UseRealTime(); |
| 93 | |
| 94 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x8, |
| 95 | xnn_f16_f32_vcvt_ukernel__neon_int32_x8, |
| 96 | benchmark::utils::CheckNEON) |
| 97 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 98 | ->UseRealTime(); |
| 99 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x16, |
| 100 | xnn_f16_f32_vcvt_ukernel__neon_int32_x16, |
| 101 | benchmark::utils::CheckNEON) |
| 102 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 103 | ->UseRealTime(); |
| 104 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x24, |
| 105 | xnn_f16_f32_vcvt_ukernel__neon_int32_x24, |
| 106 | benchmark::utils::CheckNEON) |
| 107 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 108 | ->UseRealTime(); |
| 109 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x32, |
| 110 | xnn_f16_f32_vcvt_ukernel__neon_int32_x32, |
| 111 | benchmark::utils::CheckNEON) |
| 112 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 113 | ->UseRealTime(); |
| 114 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 115 | |
| 116 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 117 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx512skx_x16, |
| 118 | xnn_f16_f32_vcvt_ukernel__avx512skx_x16, |
| 119 | benchmark::utils::CheckAVX512SKX) |
| 120 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 121 | ->UseRealTime(); |
| 122 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx512skx_x32, |
| 123 | xnn_f16_f32_vcvt_ukernel__avx512skx_x32, |
| 124 | benchmark::utils::CheckAVX512SKX) |
| 125 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 126 | ->UseRealTime(); |
| 127 | |
| 128 | BENCHMARK_CAPTURE(f16_f32_vcvt, f16c_x8, |
| 129 | xnn_f16_f32_vcvt_ukernel__f16c_x8, |
| 130 | benchmark::utils::CheckF16C) |
| 131 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 132 | ->UseRealTime(); |
| 133 | BENCHMARK_CAPTURE(f16_f32_vcvt, f16c_x16, |
| 134 | xnn_f16_f32_vcvt_ukernel__f16c_x16, |
| 135 | benchmark::utils::CheckF16C) |
| 136 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 137 | ->UseRealTime(); |
| 138 | |
| 139 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x8, |
| 140 | xnn_f16_f32_vcvt_ukernel__avx_int16_x8, |
| 141 | benchmark::utils::CheckAVX) |
| 142 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 143 | ->UseRealTime(); |
| 144 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x16, |
| 145 | xnn_f16_f32_vcvt_ukernel__avx_int16_x16, |
| 146 | benchmark::utils::CheckAVX) |
| 147 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 148 | ->UseRealTime(); |
| 149 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x24, |
| 150 | xnn_f16_f32_vcvt_ukernel__avx_int16_x24, |
| 151 | benchmark::utils::CheckAVX) |
| 152 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 153 | ->UseRealTime(); |
| 154 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x32, |
| 155 | xnn_f16_f32_vcvt_ukernel__avx_int16_x32, |
| 156 | benchmark::utils::CheckAVX) |
| 157 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 158 | ->UseRealTime(); |
| 159 | |
| 160 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x8, |
| 161 | xnn_f16_f32_vcvt_ukernel__avx_int32_x8, |
| 162 | benchmark::utils::CheckAVX) |
| 163 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 164 | ->UseRealTime(); |
| 165 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x16, |
| 166 | xnn_f16_f32_vcvt_ukernel__avx_int32_x16, |
| 167 | benchmark::utils::CheckAVX) |
| 168 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 169 | ->UseRealTime(); |
| 170 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x24, |
| 171 | xnn_f16_f32_vcvt_ukernel__avx_int32_x24, |
| 172 | benchmark::utils::CheckAVX) |
| 173 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 174 | ->UseRealTime(); |
| 175 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x32, |
| 176 | xnn_f16_f32_vcvt_ukernel__avx_int32_x32, |
| 177 | benchmark::utils::CheckAVX) |
| 178 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 179 | ->UseRealTime(); |
| 180 | |
| 181 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x8, |
| 182 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x8, |
| 183 | benchmark::utils::CheckSSE41) |
| 184 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 185 | ->UseRealTime(); |
| 186 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x16, |
| 187 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x16, |
| 188 | benchmark::utils::CheckSSE41) |
| 189 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 190 | ->UseRealTime(); |
| 191 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x24, |
| 192 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x24, |
| 193 | benchmark::utils::CheckSSE41) |
| 194 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 195 | ->UseRealTime(); |
| 196 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x32, |
| 197 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x32, |
| 198 | benchmark::utils::CheckSSE41) |
| 199 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 200 | ->UseRealTime(); |
| 201 | |
| 202 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x8, |
| 203 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x8, |
| 204 | benchmark::utils::CheckSSE41) |
| 205 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 206 | ->UseRealTime(); |
| 207 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x16, |
| 208 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x16, |
| 209 | benchmark::utils::CheckSSE41) |
| 210 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 211 | ->UseRealTime(); |
| 212 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x24, |
| 213 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x24, |
| 214 | benchmark::utils::CheckSSE41) |
| 215 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 216 | ->UseRealTime(); |
| 217 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x32, |
| 218 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x32, |
| 219 | benchmark::utils::CheckSSE41) |
| 220 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 221 | ->UseRealTime(); |
| 222 | |
| 223 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x8, |
| 224 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x8) |
| 225 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 226 | ->UseRealTime(); |
| 227 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x16, |
| 228 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x16) |
| 229 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 230 | ->UseRealTime(); |
| 231 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x24, |
| 232 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x24) |
| 233 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 234 | ->UseRealTime(); |
| 235 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x32, |
| 236 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x32) |
| 237 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 238 | ->UseRealTime(); |
| 239 | |
| 240 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x8, |
| 241 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x8) |
| 242 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 243 | ->UseRealTime(); |
| 244 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x16, |
| 245 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x16) |
| 246 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 247 | ->UseRealTime(); |
| 248 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x24, |
| 249 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x24) |
| 250 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 251 | ->UseRealTime(); |
| 252 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x32, |
| 253 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x32) |
| 254 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 255 | ->UseRealTime(); |
| 256 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 257 | |
| 258 | #if XNN_ARCH_WASMSIMD |
| 259 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x8, |
| 260 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8) |
| 261 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 262 | ->UseRealTime(); |
| 263 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x16, |
| 264 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16) |
| 265 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 266 | ->UseRealTime(); |
| 267 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x24, |
| 268 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24) |
| 269 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 270 | ->UseRealTime(); |
| 271 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x32, |
| 272 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32) |
| 273 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 274 | ->UseRealTime(); |
| 275 | |
| 276 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x8, |
| 277 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8) |
| 278 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 279 | ->UseRealTime(); |
| 280 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x16, |
| 281 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16) |
| 282 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 283 | ->UseRealTime(); |
| 284 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x24, |
| 285 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24) |
| 286 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 287 | ->UseRealTime(); |
| 288 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x32, |
| 289 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32) |
| 290 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 291 | ->UseRealTime(); |
| 292 | #endif // XNN_ARCH_WASMSIMD |
| 293 | |
Marat Dukhan | e2c0001 | 2021-10-17 22:02:35 -0700 | [diff] [blame] | 294 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_float_x1, |
| 295 | xnn_f16_f32_vcvt_ukernel__scalar_float_x1) |
| 296 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 297 | ->UseRealTime(); |
| 298 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_float_x2, |
| 299 | xnn_f16_f32_vcvt_ukernel__scalar_float_x2) |
| 300 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 301 | ->UseRealTime(); |
| 302 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_float_x3, |
| 303 | xnn_f16_f32_vcvt_ukernel__scalar_float_x3) |
| 304 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 305 | ->UseRealTime(); |
| 306 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_float_x4, |
| 307 | xnn_f16_f32_vcvt_ukernel__scalar_float_x4) |
| 308 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 309 | ->UseRealTime(); |
| 310 | |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 311 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 312 | BENCHMARK_MAIN(); |
| 313 | #endif |