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> |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 19 | #include <xnnpack/params-init.h> |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 20 | #include <xnnpack/vcvt.h> |
| 21 | |
| 22 | |
| 23 | static void f16_f32_vcvt( |
| 24 | benchmark::State& state, |
| 25 | xnn_f16_f32_vcvt_ukernel_function cvt, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 26 | xnn_init_f16_f32_cvt_params_fn init_params = nullptr, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 27 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 28 | { |
| 29 | if (isa_check && !isa_check(state)) { |
| 30 | return; |
| 31 | } |
| 32 | |
| 33 | const size_t num_elements = state.range(0); |
| 34 | |
| 35 | std::random_device random_device; |
| 36 | auto rng = std::mt19937(random_device()); |
| 37 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
| 38 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 39 | |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 40 | 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] | 41 | std::vector<float, AlignedAllocator<float, 64>> y(num_elements); |
| 42 | std::generate(x.begin(), x.end(), std::ref(f16rng)); |
| 43 | std::fill(y.begin(), y.end(), std::nanf("")); |
| 44 | |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 45 | xnn_f16_f32_cvt_params params; |
| 46 | if (init_params != nullptr) { |
| 47 | init_params(¶ms); |
| 48 | } |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 49 | for (auto _ : state) { |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 50 | cvt(num_elements * sizeof(uint16_t), x.data(), y.data(), ¶ms); |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 51 | } |
| 52 | |
| 53 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 54 | if (cpu_frequency != 0) { |
| 55 | state.counters["cpufreq"] = cpu_frequency; |
| 56 | } |
| 57 | |
| 58 | const size_t elements_per_iteration = num_elements; |
| 59 | state.counters["elements"] = |
| 60 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 61 | |
| 62 | const size_t bytes_per_iteration = num_elements * (sizeof(uint16_t) + sizeof(float)); |
| 63 | state.counters["bytes"] = |
| 64 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 65 | } |
| 66 | |
| 67 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 68 | BENCHMARK_CAPTURE(f16_f32_vcvt, neonfp16_x8, |
| 69 | xnn_f16_f32_vcvt_ukernel__neonfp16_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 70 | nullptr /* init params */, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 71 | benchmark::utils::CheckNEONFP16) |
| 72 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 73 | ->UseRealTime(); |
| 74 | BENCHMARK_CAPTURE(f16_f32_vcvt, neonfp16_x16, |
| 75 | xnn_f16_f32_vcvt_ukernel__neonfp16_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 76 | nullptr /* init params */, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 77 | benchmark::utils::CheckNEONFP16) |
| 78 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 79 | ->UseRealTime(); |
| 80 | |
| 81 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x8, |
| 82 | xnn_f16_f32_vcvt_ukernel__neon_int16_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 83 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 84 | benchmark::utils::CheckNEON) |
| 85 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 86 | ->UseRealTime(); |
| 87 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x16, |
| 88 | xnn_f16_f32_vcvt_ukernel__neon_int16_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 89 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 90 | benchmark::utils::CheckNEON) |
| 91 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 92 | ->UseRealTime(); |
| 93 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x24, |
| 94 | xnn_f16_f32_vcvt_ukernel__neon_int16_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 95 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 96 | benchmark::utils::CheckNEON) |
| 97 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 98 | ->UseRealTime(); |
| 99 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int16_x32, |
| 100 | xnn_f16_f32_vcvt_ukernel__neon_int16_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 101 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 102 | benchmark::utils::CheckNEON) |
| 103 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 104 | ->UseRealTime(); |
| 105 | |
| 106 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x8, |
| 107 | xnn_f16_f32_vcvt_ukernel__neon_int32_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 108 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 109 | benchmark::utils::CheckNEON) |
| 110 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 111 | ->UseRealTime(); |
| 112 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x16, |
| 113 | xnn_f16_f32_vcvt_ukernel__neon_int32_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 114 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 115 | benchmark::utils::CheckNEON) |
| 116 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 117 | ->UseRealTime(); |
| 118 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x24, |
| 119 | xnn_f16_f32_vcvt_ukernel__neon_int32_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 120 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 121 | benchmark::utils::CheckNEON) |
| 122 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 123 | ->UseRealTime(); |
| 124 | BENCHMARK_CAPTURE(f16_f32_vcvt, neon_int32_x32, |
| 125 | xnn_f16_f32_vcvt_ukernel__neon_int32_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 126 | xnn_init_f16_f32_cvt_neon_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 127 | benchmark::utils::CheckNEON) |
| 128 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 129 | ->UseRealTime(); |
| 130 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 131 | |
| 132 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 133 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx512skx_x16, |
| 134 | xnn_f16_f32_vcvt_ukernel__avx512skx_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 135 | nullptr /* init params */, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 136 | benchmark::utils::CheckAVX512SKX) |
| 137 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 138 | ->UseRealTime(); |
| 139 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx512skx_x32, |
| 140 | xnn_f16_f32_vcvt_ukernel__avx512skx_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 141 | nullptr /* init params */, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 142 | benchmark::utils::CheckAVX512SKX) |
| 143 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 144 | ->UseRealTime(); |
| 145 | |
| 146 | BENCHMARK_CAPTURE(f16_f32_vcvt, f16c_x8, |
| 147 | xnn_f16_f32_vcvt_ukernel__f16c_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 148 | nullptr /* init params */, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 149 | benchmark::utils::CheckF16C) |
| 150 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 151 | ->UseRealTime(); |
| 152 | BENCHMARK_CAPTURE(f16_f32_vcvt, f16c_x16, |
| 153 | xnn_f16_f32_vcvt_ukernel__f16c_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 154 | nullptr /* init params */, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 155 | benchmark::utils::CheckF16C) |
| 156 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 157 | ->UseRealTime(); |
| 158 | |
| 159 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x8, |
| 160 | xnn_f16_f32_vcvt_ukernel__avx_int16_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 161 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 162 | benchmark::utils::CheckAVX) |
| 163 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 164 | ->UseRealTime(); |
| 165 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x16, |
| 166 | xnn_f16_f32_vcvt_ukernel__avx_int16_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 167 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 168 | benchmark::utils::CheckAVX) |
| 169 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 170 | ->UseRealTime(); |
| 171 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x24, |
| 172 | xnn_f16_f32_vcvt_ukernel__avx_int16_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 173 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 174 | benchmark::utils::CheckAVX) |
| 175 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 176 | ->UseRealTime(); |
| 177 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int16_x32, |
| 178 | xnn_f16_f32_vcvt_ukernel__avx_int16_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 179 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 180 | benchmark::utils::CheckAVX) |
| 181 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 182 | ->UseRealTime(); |
| 183 | |
| 184 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x8, |
| 185 | xnn_f16_f32_vcvt_ukernel__avx_int32_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 186 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 187 | benchmark::utils::CheckAVX) |
| 188 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 189 | ->UseRealTime(); |
| 190 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x16, |
| 191 | xnn_f16_f32_vcvt_ukernel__avx_int32_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 192 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 193 | benchmark::utils::CheckAVX) |
| 194 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 195 | ->UseRealTime(); |
| 196 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x24, |
| 197 | xnn_f16_f32_vcvt_ukernel__avx_int32_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 198 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 199 | benchmark::utils::CheckAVX) |
| 200 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 201 | ->UseRealTime(); |
| 202 | BENCHMARK_CAPTURE(f16_f32_vcvt, avx_int32_x32, |
| 203 | xnn_f16_f32_vcvt_ukernel__avx_int32_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 204 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 205 | benchmark::utils::CheckAVX) |
| 206 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 207 | ->UseRealTime(); |
| 208 | |
| 209 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x8, |
| 210 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 211 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 212 | benchmark::utils::CheckSSE41) |
| 213 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 214 | ->UseRealTime(); |
| 215 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x16, |
| 216 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 217 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 218 | benchmark::utils::CheckSSE41) |
| 219 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 220 | ->UseRealTime(); |
| 221 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x24, |
| 222 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 223 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 224 | benchmark::utils::CheckSSE41) |
| 225 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 226 | ->UseRealTime(); |
| 227 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int16_x32, |
| 228 | xnn_f16_f32_vcvt_ukernel__sse41_int16_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 229 | xnn_init_f16_f32_cvt_sse_int16_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 230 | benchmark::utils::CheckSSE41) |
| 231 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 232 | ->UseRealTime(); |
| 233 | |
| 234 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x8, |
| 235 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 236 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 237 | benchmark::utils::CheckSSE41) |
| 238 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 239 | ->UseRealTime(); |
| 240 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x16, |
| 241 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 242 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 243 | benchmark::utils::CheckSSE41) |
| 244 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 245 | ->UseRealTime(); |
| 246 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x24, |
| 247 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 248 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 249 | benchmark::utils::CheckSSE41) |
| 250 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 251 | ->UseRealTime(); |
| 252 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse41_int32_x32, |
| 253 | xnn_f16_f32_vcvt_ukernel__sse41_int32_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 254 | xnn_init_f16_f32_cvt_sse_int32_params, |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 255 | benchmark::utils::CheckSSE41) |
| 256 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 257 | ->UseRealTime(); |
| 258 | |
| 259 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 260 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x8, |
| 261 | xnn_init_f16_f32_cvt_sse_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 262 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 263 | ->UseRealTime(); |
| 264 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 265 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x16, |
| 266 | xnn_init_f16_f32_cvt_sse_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 267 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 268 | ->UseRealTime(); |
| 269 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 270 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x24, |
| 271 | xnn_init_f16_f32_cvt_sse_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 272 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 273 | ->UseRealTime(); |
| 274 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int16_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 275 | xnn_f16_f32_vcvt_ukernel__sse2_int16_x32, |
| 276 | xnn_init_f16_f32_cvt_sse_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 277 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 278 | ->UseRealTime(); |
| 279 | |
| 280 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 281 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x8, |
| 282 | xnn_init_f16_f32_cvt_sse_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 283 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 284 | ->UseRealTime(); |
| 285 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 286 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x16, |
| 287 | xnn_init_f16_f32_cvt_sse_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 288 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 289 | ->UseRealTime(); |
| 290 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 291 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x24, |
| 292 | xnn_init_f16_f32_cvt_sse_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 293 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 294 | ->UseRealTime(); |
| 295 | BENCHMARK_CAPTURE(f16_f32_vcvt, sse2_int32_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 296 | xnn_f16_f32_vcvt_ukernel__sse2_int32_x32, |
| 297 | xnn_init_f16_f32_cvt_sse_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 298 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 299 | ->UseRealTime(); |
| 300 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 301 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 302 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 303 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 304 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8, |
| 305 | xnn_init_f16_f32_cvt_wasmsimd_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 306 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 307 | ->UseRealTime(); |
| 308 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 309 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16, |
| 310 | xnn_init_f16_f32_cvt_wasmsimd_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 311 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 312 | ->UseRealTime(); |
| 313 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 314 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24, |
| 315 | xnn_init_f16_f32_cvt_wasmsimd_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 316 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 317 | ->UseRealTime(); |
| 318 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int16_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 319 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32, |
| 320 | xnn_init_f16_f32_cvt_wasmsimd_int16_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 321 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 322 | ->UseRealTime(); |
| 323 | |
| 324 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x8, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 325 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8, |
| 326 | xnn_init_f16_f32_cvt_wasmsimd_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 327 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 328 | ->UseRealTime(); |
| 329 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x16, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 330 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16, |
| 331 | xnn_init_f16_f32_cvt_wasmsimd_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 332 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 333 | ->UseRealTime(); |
| 334 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x24, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 335 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24, |
| 336 | xnn_init_f16_f32_cvt_wasmsimd_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 337 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 338 | ->UseRealTime(); |
| 339 | BENCHMARK_CAPTURE(f16_f32_vcvt, wasmsimd_int32_x32, |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 340 | xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32, |
| 341 | xnn_init_f16_f32_cvt_wasmsimd_int32_params) |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 342 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 343 | ->UseRealTime(); |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 344 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 345 | |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 346 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x1, |
| 347 | xnn_f16_f32_vcvt_ukernel__scalar_x1, |
| 348 | xnn_init_f16_f32_cvt_scalar_params) |
Marat Dukhan | e2c0001 | 2021-10-17 22:02:35 -0700 | [diff] [blame] | 349 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 350 | ->UseRealTime(); |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 351 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x2, |
| 352 | xnn_f16_f32_vcvt_ukernel__scalar_x2, |
| 353 | xnn_init_f16_f32_cvt_scalar_params) |
Marat Dukhan | e2c0001 | 2021-10-17 22:02:35 -0700 | [diff] [blame] | 354 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 355 | ->UseRealTime(); |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 356 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x3, |
| 357 | xnn_f16_f32_vcvt_ukernel__scalar_x3, |
| 358 | xnn_init_f16_f32_cvt_scalar_params) |
Marat Dukhan | e2c0001 | 2021-10-17 22:02:35 -0700 | [diff] [blame] | 359 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 360 | ->UseRealTime(); |
Marat Dukhan | 134f984 | 2021-12-29 19:57:31 -0800 | [diff] [blame] | 361 | BENCHMARK_CAPTURE(f16_f32_vcvt, scalar_x4, |
| 362 | xnn_f16_f32_vcvt_ukernel__scalar_x4, |
| 363 | xnn_init_f16_f32_cvt_scalar_params) |
Marat Dukhan | e2c0001 | 2021-10-17 22:02:35 -0700 | [diff] [blame] | 364 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 365 | ->UseRealTime(); |
| 366 | |
Marat Dukhan | 434352f | 2021-10-16 18:28:55 -0700 | [diff] [blame] | 367 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 368 | BENCHMARK_MAIN(); |
| 369 | #endif |