Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -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 | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 19 | #include <xnnpack/params-init.h> |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 20 | #include <xnnpack/vcvt.h> |
| 21 | |
| 22 | |
| 23 | static void f32_f16_vcvt( |
| 24 | benchmark::State& state, |
| 25 | xnn_f32_f16_vcvt_ukernel_function cvt, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 26 | xnn_init_f32_f16_cvt_params_fn init_params = nullptr, |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -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 | |
| 39 | std::vector<float, AlignedAllocator<float, 64>> x(num_elements + XNN_EXTRA_BYTES / sizeof(float)); |
| 40 | std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> y(num_elements); |
| 41 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 42 | std::fill(y.begin(), y.end(), UINT16_C(0x7E00)); |
| 43 | |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 44 | xnn_f32_f16_cvt_params params; |
| 45 | if (init_params != nullptr) { |
| 46 | init_params(¶ms); |
| 47 | } |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 48 | for (auto _ : state) { |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 49 | cvt(num_elements * sizeof(uint16_t), x.data(), y.data(), ¶ms); |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 50 | } |
| 51 | |
| 52 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 53 | if (cpu_frequency != 0) { |
| 54 | state.counters["cpufreq"] = cpu_frequency; |
| 55 | } |
| 56 | |
| 57 | const size_t elements_per_iteration = num_elements; |
| 58 | state.counters["elements"] = |
| 59 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 60 | |
| 61 | const size_t bytes_per_iteration = num_elements * (sizeof(uint16_t) + sizeof(float)); |
| 62 | state.counters["bytes"] = |
| 63 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 64 | } |
| 65 | |
| 66 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 67 | BENCHMARK_CAPTURE(f32_f16_vcvt, neonfp16_x8, |
| 68 | xnn_f32_f16_vcvt_ukernel__neonfp16_x8, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 69 | nullptr /* init params */, |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 70 | benchmark::utils::CheckNEONFP16) |
| 71 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 72 | ->UseRealTime(); |
| 73 | BENCHMARK_CAPTURE(f32_f16_vcvt, neonfp16_x16, |
| 74 | xnn_f32_f16_vcvt_ukernel__neonfp16_x16, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 75 | nullptr /* init params */, |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 76 | benchmark::utils::CheckNEONFP16) |
| 77 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 78 | ->UseRealTime(); |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 79 | |
| 80 | BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x8, |
| 81 | xnn_f32_f16_vcvt_ukernel__neon_x8, |
| 82 | xnn_init_f32_f16_cvt_neon_params, |
| 83 | benchmark::utils::CheckNEON) |
| 84 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 85 | ->UseRealTime(); |
| 86 | BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x16, |
| 87 | xnn_f32_f16_vcvt_ukernel__neon_x16, |
| 88 | xnn_init_f32_f16_cvt_neon_params, |
| 89 | benchmark::utils::CheckNEON) |
| 90 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 91 | ->UseRealTime(); |
| 92 | BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x24, |
| 93 | xnn_f32_f16_vcvt_ukernel__neon_x24, |
| 94 | xnn_init_f32_f16_cvt_neon_params, |
| 95 | benchmark::utils::CheckNEON) |
| 96 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 97 | ->UseRealTime(); |
| 98 | BENCHMARK_CAPTURE(f32_f16_vcvt, neon_x32, |
| 99 | xnn_f32_f16_vcvt_ukernel__neon_x32, |
| 100 | xnn_init_f32_f16_cvt_neon_params, |
| 101 | benchmark::utils::CheckNEON) |
| 102 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 103 | ->UseRealTime(); |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 104 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 105 | |
| 106 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 107 | BENCHMARK_CAPTURE(f32_f16_vcvt, avx512skx_x16, |
| 108 | xnn_f32_f16_vcvt_ukernel__avx512skx_x16, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 109 | nullptr /* init params */, |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 110 | benchmark::utils::CheckAVX512SKX) |
| 111 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 112 | ->UseRealTime(); |
| 113 | BENCHMARK_CAPTURE(f32_f16_vcvt, avx512skx_x32, |
| 114 | xnn_f32_f16_vcvt_ukernel__avx512skx_x32, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 115 | nullptr /* init params */, |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 116 | benchmark::utils::CheckAVX512SKX) |
| 117 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 118 | ->UseRealTime(); |
| 119 | |
| 120 | BENCHMARK_CAPTURE(f32_f16_vcvt, f16c_x8, |
| 121 | xnn_f32_f16_vcvt_ukernel__f16c_x8, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 122 | xnn_init_f32_f16_cvt_f16c_params, |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 123 | benchmark::utils::CheckF16C) |
| 124 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 125 | ->UseRealTime(); |
| 126 | BENCHMARK_CAPTURE(f32_f16_vcvt, f16c_x16, |
| 127 | xnn_f32_f16_vcvt_ukernel__f16c_x16, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 128 | xnn_init_f32_f16_cvt_f16c_params, |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 129 | benchmark::utils::CheckF16C) |
| 130 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 131 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 132 | |
| 133 | BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x8, |
| 134 | xnn_f32_f16_vcvt_ukernel__avx_x8, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 135 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 136 | benchmark::utils::CheckAVX) |
| 137 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 138 | ->UseRealTime(); |
| 139 | BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x16, |
| 140 | xnn_f32_f16_vcvt_ukernel__avx_x16, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 141 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 142 | benchmark::utils::CheckAVX) |
| 143 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 144 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 145 | BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x24, |
| 146 | xnn_f32_f16_vcvt_ukernel__avx_x24, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 147 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 148 | benchmark::utils::CheckAVX) |
| 149 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 150 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 151 | BENCHMARK_CAPTURE(f32_f16_vcvt, avx_x32, |
| 152 | xnn_f32_f16_vcvt_ukernel__avx_x32, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 153 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 154 | benchmark::utils::CheckAVX) |
| 155 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 156 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 157 | |
| 158 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x8, |
| 159 | xnn_f32_f16_vcvt_ukernel__sse41_x8, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 160 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 161 | benchmark::utils::CheckSSE41) |
| 162 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 163 | ->UseRealTime(); |
| 164 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x16, |
| 165 | xnn_f32_f16_vcvt_ukernel__sse41_x16, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 166 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 167 | benchmark::utils::CheckSSE41) |
| 168 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 169 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 170 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x24, |
| 171 | xnn_f32_f16_vcvt_ukernel__sse41_x24, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 172 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 173 | benchmark::utils::CheckSSE41) |
| 174 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 175 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 176 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse41_x32, |
| 177 | xnn_f32_f16_vcvt_ukernel__sse41_x32, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 178 | xnn_init_f32_f16_cvt_sse2_params, |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 179 | benchmark::utils::CheckSSE41) |
| 180 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 181 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 182 | |
| 183 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x8, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 184 | xnn_f32_f16_vcvt_ukernel__sse2_x8, |
| 185 | xnn_init_f32_f16_cvt_sse2_params) |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 186 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 187 | ->UseRealTime(); |
| 188 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x16, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 189 | xnn_f32_f16_vcvt_ukernel__sse2_x16, |
| 190 | xnn_init_f32_f16_cvt_sse2_params) |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 191 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 192 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 193 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x24, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 194 | xnn_f32_f16_vcvt_ukernel__sse2_x24, |
| 195 | xnn_init_f32_f16_cvt_sse2_params) |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 196 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 197 | ->UseRealTime(); |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 198 | BENCHMARK_CAPTURE(f32_f16_vcvt, sse2_x32, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 199 | xnn_f32_f16_vcvt_ukernel__sse2_x32, |
| 200 | xnn_init_f32_f16_cvt_sse2_params) |
Marat Dukhan | eb84423 | 2021-11-08 23:07:53 -0800 | [diff] [blame] | 201 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 202 | ->UseRealTime(); |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 203 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 204 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 205 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 206 | BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x8, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 207 | xnn_f32_f16_vcvt_ukernel__wasmsimd_x8, |
| 208 | xnn_init_f32_f16_cvt_wasmsimd_params) |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 209 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 210 | ->UseRealTime(); |
| 211 | BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x16, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 212 | xnn_f32_f16_vcvt_ukernel__wasmsimd_x16, |
| 213 | xnn_init_f32_f16_cvt_wasmsimd_params) |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 214 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 215 | ->UseRealTime(); |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 216 | BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x24, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 217 | xnn_f32_f16_vcvt_ukernel__wasmsimd_x24, |
| 218 | xnn_init_f32_f16_cvt_wasmsimd_params) |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 219 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 220 | ->UseRealTime(); |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 221 | BENCHMARK_CAPTURE(f32_f16_vcvt, wasmsimd_x32, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 222 | xnn_f32_f16_vcvt_ukernel__wasmsimd_x32, |
| 223 | xnn_init_f32_f16_cvt_wasmsimd_params) |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 224 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
Marat Dukhan | 5740f75 | 2021-12-02 15:19:10 -0800 | [diff] [blame] | 225 | ->UseRealTime(); |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 226 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 22e31c8 | 2021-11-09 00:00:28 -0800 | [diff] [blame] | 227 | |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 228 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x1, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 229 | xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x1, |
| 230 | xnn_init_f32_f16_cvt_scalar_bitcast_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 231 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 232 | ->UseRealTime(); |
| 233 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x2, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 234 | xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x2, |
| 235 | xnn_init_f32_f16_cvt_scalar_bitcast_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 236 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 237 | ->UseRealTime(); |
| 238 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x3, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 239 | xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x3, |
| 240 | xnn_init_f32_f16_cvt_scalar_bitcast_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 241 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 242 | ->UseRealTime(); |
| 243 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_bitcast_x4, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 244 | xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x4, |
| 245 | xnn_init_f32_f16_cvt_scalar_bitcast_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 246 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 247 | ->UseRealTime(); |
| 248 | |
| 249 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x1, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 250 | xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x1, |
| 251 | xnn_init_f32_f16_cvt_scalar_fabsf_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 252 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 253 | ->UseRealTime(); |
| 254 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x2, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 255 | xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x2, |
| 256 | xnn_init_f32_f16_cvt_scalar_fabsf_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 257 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 258 | ->UseRealTime(); |
| 259 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x3, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 260 | xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x3, |
| 261 | xnn_init_f32_f16_cvt_scalar_fabsf_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 262 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 263 | ->UseRealTime(); |
| 264 | BENCHMARK_CAPTURE(f32_f16_vcvt, scalar_fabsf_x4, |
Marat Dukhan | b7c1b71 | 2021-12-30 07:23:57 -0800 | [diff] [blame] | 265 | xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x4, |
| 266 | xnn_init_f32_f16_cvt_scalar_fabsf_params) |
Marat Dukhan | 1fe8995 | 2021-11-10 01:27:15 -0800 | [diff] [blame] | 267 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, float>) |
| 268 | ->UseRealTime(); |
| 269 | |
Marat Dukhan | d77f77d | 2021-10-24 15:39:59 -0700 | [diff] [blame] | 270 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 271 | BENCHMARK_MAIN(); |
| 272 | #endif |