Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1 | // Copyright 2020 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 <xnnpack/AlignedAllocator.h> |
| 16 | #include <xnnpack/common.h> |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 17 | #include <xnnpack/params.h> |
| 18 | #include <xnnpack/params-init.h> |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 19 | #include <xnnpack/vunary.h> |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 20 | |
| 21 | |
| 22 | static void f32_vsqrt( |
| 23 | benchmark::State& state, |
| 24 | xnn_f32_vsqrt_ukernel_function vsqrt, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 25 | xnn_init_f32_sqrt_params_fn init_params = nullptr, |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 26 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 27 | { |
| 28 | if (isa_check && !isa_check(state)) { |
| 29 | return; |
| 30 | } |
| 31 | |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 32 | const size_t num_elements = state.range(0); |
| 33 | std::vector<float, AlignedAllocator<float, 64>> input(num_elements); |
| 34 | std::vector<float, AlignedAllocator<float, 64>> output(num_elements); |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 35 | |
| 36 | std::random_device random_device; |
| 37 | auto rng = std::mt19937(random_device()); |
| 38 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 10.0f), std::ref(rng)); |
| 39 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 40 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 41 | |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame] | 42 | union xnn_f32_sqrt_params params; |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 43 | if (init_params != nullptr) { |
| 44 | init_params(¶ms); |
| 45 | } |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 46 | for (auto _ : state) { |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 47 | vsqrt(num_elements * sizeof(float), input.data(), output.data(), ¶ms); |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 48 | } |
| 49 | |
| 50 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 51 | if (cpu_frequency != 0) { |
| 52 | state.counters["cpufreq"] = cpu_frequency; |
| 53 | } |
| 54 | |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 55 | const size_t elements_per_iteration = num_elements; |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 56 | state.counters["elements"] = |
| 57 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 58 | |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 59 | const size_t bytes_per_iteration = 2 * num_elements * sizeof(float); |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 60 | state.counters["bytes"] = |
| 61 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 62 | } |
| 63 | |
| 64 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 65 | BENCHMARK_CAPTURE(f32_vsqrt, neon_sqrt_x4, |
| 66 | xnn_f32_vsqrt_ukernel__neon_sqrt_x4) |
| 67 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 68 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 69 | BENCHMARK_CAPTURE(f32_vsqrt, neon_sqrt_x8, |
| 70 | xnn_f32_vsqrt_ukernel__neon_sqrt_x8) |
| 71 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 72 | ->UseRealTime(); |
| 73 | #endif // XNN_ARCH_ARM64 |
| 74 | |
| 75 | #if XNN_ARCH_ARM64 || XNN_ARCH_ARM64 |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 76 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x4, |
| 77 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 78 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 79 | benchmark::utils::CheckNEONFMA) |
| 80 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 81 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 82 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x8, |
| 83 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 84 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 85 | benchmark::utils::CheckNEONFMA) |
| 86 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 87 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 88 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x12, |
| 89 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 90 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 91 | benchmark::utils::CheckNEONFMA) |
| 92 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 93 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 94 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x16, |
| 95 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 96 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 97 | benchmark::utils::CheckNEONFMA) |
| 98 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 99 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 100 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x20, |
| 101 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 102 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 103 | benchmark::utils::CheckNEONFMA) |
| 104 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 105 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 106 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x24, |
| 107 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 108 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 109 | benchmark::utils::CheckNEONFMA) |
| 110 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 111 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 112 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x28, |
| 113 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 114 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 115 | benchmark::utils::CheckNEONFMA) |
| 116 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 117 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 118 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x32, |
| 119 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 120 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 121 | benchmark::utils::CheckNEONFMA) |
| 122 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 123 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 124 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x36, |
| 125 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 126 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 127 | benchmark::utils::CheckNEONFMA) |
| 128 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 129 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 130 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x40, |
| 131 | xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 132 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 133 | benchmark::utils::CheckNEONFMA) |
| 134 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 135 | ->UseRealTime(); |
| 136 | |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 137 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x4, |
| 138 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x4, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 139 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 140 | benchmark::utils::CheckNEONFMA) |
| 141 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 142 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 143 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x8, |
| 144 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 145 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 146 | benchmark::utils::CheckNEONFMA) |
| 147 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 148 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 149 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x12, |
| 150 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 151 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 152 | benchmark::utils::CheckNEONFMA) |
| 153 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 154 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 155 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x16, |
| 156 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 157 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 158 | benchmark::utils::CheckNEONFMA) |
| 159 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 160 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 161 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x20, |
| 162 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 163 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 164 | benchmark::utils::CheckNEONFMA) |
| 165 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 166 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 167 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x24, |
| 168 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 169 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 170 | benchmark::utils::CheckNEONFMA) |
| 171 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 172 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 173 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x28, |
| 174 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 175 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 176 | benchmark::utils::CheckNEONFMA) |
| 177 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 178 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 179 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x32, |
| 180 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 181 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 182 | benchmark::utils::CheckNEONFMA) |
| 183 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 184 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 185 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x36, |
| 186 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 187 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 188 | benchmark::utils::CheckNEONFMA) |
| 189 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 190 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 191 | BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x40, |
| 192 | xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40, |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 193 | nullptr /* init params */, |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 194 | benchmark::utils::CheckNEONFMA) |
| 195 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 196 | ->UseRealTime(); |
| 197 | #endif // XNN_ARCH_ARM64 || XNN_ARCH_ARM64 |
| 198 | |
| 199 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 200 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x16, |
| 201 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16, |
| 202 | xnn_init_f32_sqrt_avx512_params, |
| 203 | benchmark::utils::CheckAVX512F) |
| 204 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 205 | ->UseRealTime(); |
| 206 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x32, |
| 207 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32, |
| 208 | xnn_init_f32_sqrt_avx512_params, |
| 209 | benchmark::utils::CheckAVX512F) |
| 210 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 211 | ->UseRealTime(); |
| 212 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x48, |
| 213 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48, |
| 214 | xnn_init_f32_sqrt_avx512_params, |
| 215 | benchmark::utils::CheckAVX512F) |
| 216 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 217 | ->UseRealTime(); |
| 218 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x64, |
| 219 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64, |
| 220 | xnn_init_f32_sqrt_avx512_params, |
| 221 | benchmark::utils::CheckAVX512F) |
| 222 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 223 | ->UseRealTime(); |
| 224 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x80, |
| 225 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80, |
| 226 | xnn_init_f32_sqrt_avx512_params, |
| 227 | benchmark::utils::CheckAVX512F) |
| 228 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 229 | ->UseRealTime(); |
| 230 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x96, |
| 231 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96, |
| 232 | xnn_init_f32_sqrt_avx512_params, |
| 233 | benchmark::utils::CheckAVX512F) |
| 234 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 235 | ->UseRealTime(); |
| 236 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x112, |
| 237 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112, |
| 238 | xnn_init_f32_sqrt_avx512_params, |
| 239 | benchmark::utils::CheckAVX512F) |
| 240 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 241 | ->UseRealTime(); |
| 242 | BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x128, |
| 243 | xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128, |
| 244 | xnn_init_f32_sqrt_avx512_params, |
| 245 | benchmark::utils::CheckAVX512F) |
| 246 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 247 | ->UseRealTime(); |
| 248 | |
| 249 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x8, |
| 250 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8, |
| 251 | xnn_init_f32_sqrt_fma_params, |
| 252 | benchmark::utils::CheckFMA3) |
| 253 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 254 | ->UseRealTime(); |
| 255 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x16, |
| 256 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16, |
| 257 | xnn_init_f32_sqrt_fma_params, |
| 258 | benchmark::utils::CheckFMA3) |
| 259 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 260 | ->UseRealTime(); |
| 261 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x24, |
| 262 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24, |
| 263 | xnn_init_f32_sqrt_fma_params, |
| 264 | benchmark::utils::CheckFMA3) |
| 265 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 266 | ->UseRealTime(); |
| 267 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x32, |
| 268 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32, |
| 269 | xnn_init_f32_sqrt_fma_params, |
| 270 | benchmark::utils::CheckFMA3) |
| 271 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 272 | ->UseRealTime(); |
| 273 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x40, |
| 274 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40, |
| 275 | xnn_init_f32_sqrt_fma_params, |
| 276 | benchmark::utils::CheckFMA3) |
| 277 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 278 | ->UseRealTime(); |
| 279 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x48, |
| 280 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48, |
| 281 | xnn_init_f32_sqrt_fma_params, |
| 282 | benchmark::utils::CheckFMA3) |
| 283 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 284 | ->UseRealTime(); |
| 285 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x56, |
| 286 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56, |
| 287 | xnn_init_f32_sqrt_fma_params, |
| 288 | benchmark::utils::CheckFMA3) |
| 289 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 290 | ->UseRealTime(); |
| 291 | BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x64, |
| 292 | xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64, |
| 293 | xnn_init_f32_sqrt_fma_params, |
| 294 | benchmark::utils::CheckFMA3) |
| 295 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 296 | ->UseRealTime(); |
| 297 | |
| 298 | BENCHMARK_CAPTURE(f32_vsqrt, avx_sqrt_x8, |
| 299 | xnn_f32_vsqrt_ukernel__avx_sqrt_x8, |
| 300 | xnn_init_f32_sqrt_avx_params, |
| 301 | benchmark::utils::CheckAVX) |
| 302 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 303 | ->UseRealTime(); |
| 304 | BENCHMARK_CAPTURE(f32_vsqrt, avx_sqrt_x16, |
| 305 | xnn_f32_vsqrt_ukernel__avx_sqrt_x16, |
| 306 | xnn_init_f32_sqrt_avx_params, |
| 307 | benchmark::utils::CheckAVX) |
| 308 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 309 | ->UseRealTime(); |
| 310 | |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 311 | BENCHMARK_CAPTURE(f32_vsqrt, sse_sqrt_x4, |
| 312 | xnn_f32_vsqrt_ukernel__sse_sqrt_x4) |
| 313 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 314 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 315 | BENCHMARK_CAPTURE(f32_vsqrt, sse_sqrt_x8, |
| 316 | xnn_f32_vsqrt_ukernel__sse_sqrt_x8) |
| 317 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 318 | ->UseRealTime(); |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 319 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 320 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 321 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 322 | BENCHMARK_CAPTURE(f32_vsqrt, wasmsimd_sqrt_x4, |
| 323 | xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4) |
| 324 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 325 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 326 | BENCHMARK_CAPTURE(f32_vsqrt, wasmsimd_sqrt_x8, |
| 327 | xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8) |
| 328 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 329 | ->UseRealTime(); |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 330 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 331 | |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 332 | BENCHMARK_CAPTURE(f32_vsqrt, scalar_sqrt_x1, |
| 333 | xnn_f32_vsqrt_ukernel__scalar_sqrt_x1) |
| 334 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 335 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 336 | BENCHMARK_CAPTURE(f32_vsqrt, scalar_sqrt_x2, |
| 337 | xnn_f32_vsqrt_ukernel__scalar_sqrt_x2) |
| 338 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 339 | ->UseRealTime(); |
Marat Dukhan | 5aeb32b | 2021-08-06 22:08:39 -0700 | [diff] [blame] | 340 | BENCHMARK_CAPTURE(f32_vsqrt, scalar_sqrt_x4, |
| 341 | xnn_f32_vsqrt_ukernel__scalar_sqrt_x4) |
| 342 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 343 | ->UseRealTime(); |
| 344 | |
| 345 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 346 | BENCHMARK_MAIN(); |
| 347 | #endif |