Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [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" |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 14 | |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 15 | #include <xnnpack/AlignedAllocator.h> |
| 16 | #include <xnnpack/common.h> |
| 17 | #include <xnnpack/params.h> |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 18 | #include <xnnpack/params-init.h> |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 19 | #include <xnnpack/vunary.h> |
| 20 | |
| 21 | |
| 22 | static void f32_vsigmoid( |
| 23 | benchmark::State& state, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 24 | xnn_f32_vsigmoid_ukernel_function sigmoid, |
| 25 | xnn_init_f32_sigmoid_params_fn init_params, |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 26 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 27 | { |
| 28 | if (isa_check && !isa_check(state)) { |
| 29 | return; |
| 30 | } |
| 31 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 32 | const size_t num_elements = state.range(0); |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 33 | |
| 34 | std::random_device random_device; |
| 35 | auto rng = std::mt19937(random_device()); |
| 36 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
| 37 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 38 | std::vector<float, AlignedAllocator<float, 64>> x(num_elements); |
| 39 | std::vector<float, AlignedAllocator<float, 64>> y(num_elements); |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 40 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 41 | std::fill(y.begin(), y.end(), std::nanf("")); |
| 42 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 43 | xnn_f32_sigmoid_params params; |
| 44 | init_params(¶ms); |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 45 | for (auto _ : state) { |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 46 | sigmoid(num_elements * sizeof(float), x.data(), y.data(), ¶ms); |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 47 | } |
| 48 | |
| 49 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 50 | if (cpu_frequency != 0) { |
| 51 | state.counters["cpufreq"] = cpu_frequency; |
| 52 | } |
| 53 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 54 | const size_t elements_per_iteration = num_elements; |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 55 | state.counters["elements"] = |
| 56 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 57 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 58 | const size_t bytes_per_iteration = 2 * num_elements * sizeof(float); |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 59 | state.counters["bytes"] = |
| 60 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 61 | } |
| 62 | |
| 63 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 64 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_div_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 65 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x4, |
| 66 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 67 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 68 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 69 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_div_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 70 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x8, |
| 71 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 72 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 73 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 74 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_div_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 75 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x12, |
| 76 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 77 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 78 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 79 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 80 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16, |
| 81 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 82 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 83 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 84 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_div_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 85 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20, |
| 86 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 87 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 88 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 89 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_div_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 90 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24, |
| 91 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 92 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 93 | ->UseRealTime(); |
| 94 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 95 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_div_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 96 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x4, |
| 97 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 98 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 99 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 100 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_div_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 101 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8, |
| 102 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 103 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 104 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 105 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_div_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 106 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12, |
| 107 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 108 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 109 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 110 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 111 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16, |
| 112 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 113 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 114 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 115 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_div_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 116 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20, |
| 117 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 118 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 119 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 120 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_div_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 121 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24, |
| 122 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 123 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 124 | ->UseRealTime(); |
| 125 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 126 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_div_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 127 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x4, |
| 128 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 129 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 130 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 131 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_div_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 132 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8, |
| 133 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 134 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 135 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 136 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_div_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 137 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12, |
| 138 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 139 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 140 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 141 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 142 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16, |
| 143 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 144 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 145 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 146 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_div_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 147 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20, |
| 148 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 149 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 150 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 151 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_div_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 152 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24, |
| 153 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 154 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 155 | ->UseRealTime(); |
| 156 | #endif // XNN_ARCH_ARM64 |
| 157 | |
| 158 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 159 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2fma_x4, |
| 160 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 161 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 162 | benchmark::utils::CheckNEONFMA) |
| 163 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 164 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 165 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2fma_x8, |
| 166 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 167 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 168 | benchmark::utils::CheckNEONFMA) |
| 169 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 170 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 171 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2fma_x12, |
| 172 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 173 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 174 | benchmark::utils::CheckNEONFMA) |
| 175 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 176 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 177 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2fma_x16, |
| 178 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 179 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 180 | benchmark::utils::CheckNEONFMA) |
| 181 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 182 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 183 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2fma_x20, |
| 184 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 185 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 186 | benchmark::utils::CheckNEONFMA) |
| 187 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 188 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 189 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2fma_x24, |
| 190 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 191 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 192 | benchmark::utils::CheckNEONFMA) |
| 193 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 194 | ->UseRealTime(); |
| 195 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 196 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr1recps1fma_x4, |
| 197 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 198 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 199 | benchmark::utils::CheckNEONFMA) |
| 200 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 201 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 202 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr1recps1fma_x8, |
| 203 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 204 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 205 | benchmark::utils::CheckNEONFMA) |
| 206 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 207 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 208 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr1recps1fma_x12, |
| 209 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 210 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 211 | benchmark::utils::CheckNEONFMA) |
| 212 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 213 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 214 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr1recps1fma_x16, |
| 215 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 216 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 217 | benchmark::utils::CheckNEONFMA) |
| 218 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 219 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 220 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr1recps1fma_x20, |
| 221 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 222 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 223 | benchmark::utils::CheckNEONFMA) |
| 224 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 225 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 226 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr1recps1fma_x24, |
| 227 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 228 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 229 | benchmark::utils::CheckNEONFMA) |
| 230 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 231 | ->UseRealTime(); |
| 232 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 233 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2recps_x4, |
| 234 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 235 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 236 | benchmark::utils::CheckNEONFMA) |
| 237 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 238 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 239 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2recps_x8, |
| 240 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 241 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 242 | benchmark::utils::CheckNEONFMA) |
| 243 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 244 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 245 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2recps_x12, |
| 246 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 247 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 248 | benchmark::utils::CheckNEONFMA) |
| 249 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 250 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 251 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2recps_x16, |
| 252 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 253 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 254 | benchmark::utils::CheckNEONFMA) |
| 255 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 256 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 257 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2recps_x20, |
| 258 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 259 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 260 | benchmark::utils::CheckNEONFMA) |
| 261 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 262 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 263 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_p5_nr2recps_x24, |
| 264 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 265 | xnn_init_f32_sigmoid_neonfma_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 266 | benchmark::utils::CheckNEONFMA) |
| 267 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 268 | ->UseRealTime(); |
| 269 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 270 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_p5_nr2recps_x4, |
| 271 | xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 272 | xnn_init_f32_sigmoid_scalar_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 273 | benchmark::utils::CheckNEON) |
| 274 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 275 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 276 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_p5_nr2recps_x8, |
| 277 | xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 278 | xnn_init_f32_sigmoid_scalar_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 279 | benchmark::utils::CheckNEON) |
| 280 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 281 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 282 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_p5_nr2recps_x12, |
| 283 | xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 284 | xnn_init_f32_sigmoid_scalar_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 285 | benchmark::utils::CheckNEON) |
| 286 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 287 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 288 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_p5_nr2recps_x16, |
| 289 | xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 290 | xnn_init_f32_sigmoid_scalar_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 291 | benchmark::utils::CheckNEON) |
| 292 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 293 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 294 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_p5_nr2recps_x20, |
| 295 | xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 296 | xnn_init_f32_sigmoid_scalar_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 297 | benchmark::utils::CheckNEON) |
| 298 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 299 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 300 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_p5_nr2recps_x24, |
| 301 | xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 302 | xnn_init_f32_sigmoid_scalar_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 303 | benchmark::utils::CheckNEON) |
| 304 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 305 | ->UseRealTime(); |
| 306 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 307 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2fma_x4, |
| 308 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 309 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 310 | benchmark::utils::CheckNEONFMA) |
| 311 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 312 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 313 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2fma_x8, |
| 314 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 315 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 316 | benchmark::utils::CheckNEONFMA) |
| 317 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 318 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 319 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2fma_x12, |
| 320 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 321 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 322 | benchmark::utils::CheckNEONFMA) |
| 323 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 324 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 325 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2fma_x16, |
| 326 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 327 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 328 | benchmark::utils::CheckNEONFMA) |
| 329 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 330 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 331 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2fma_x20, |
| 332 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 333 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 334 | benchmark::utils::CheckNEONFMA) |
| 335 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 336 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 337 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2fma_x24, |
| 338 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 339 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 340 | benchmark::utils::CheckNEONFMA) |
| 341 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 342 | ->UseRealTime(); |
| 343 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 344 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr1recps1fma_x4, |
| 345 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 346 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 347 | benchmark::utils::CheckNEONFMA) |
| 348 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 349 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 350 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr1recps1fma_x8, |
| 351 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 352 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 353 | benchmark::utils::CheckNEONFMA) |
| 354 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 355 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 356 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr1recps1fma_x12, |
| 357 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 358 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 359 | benchmark::utils::CheckNEONFMA) |
| 360 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 361 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 362 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr1recps1fma_x16, |
| 363 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 364 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 365 | benchmark::utils::CheckNEONFMA) |
| 366 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 367 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 368 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr1recps1fma_x20, |
| 369 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 370 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 371 | benchmark::utils::CheckNEONFMA) |
| 372 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 373 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 374 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr1recps1fma_x24, |
| 375 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 376 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 377 | benchmark::utils::CheckNEONFMA) |
| 378 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 379 | ->UseRealTime(); |
| 380 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 381 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2recps_x4, |
| 382 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 383 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 384 | benchmark::utils::CheckNEONFMA) |
| 385 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 386 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 387 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2recps_x8, |
| 388 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 389 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 390 | benchmark::utils::CheckNEONFMA) |
| 391 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 392 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 393 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2recps_x12, |
| 394 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 395 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 396 | benchmark::utils::CheckNEONFMA) |
| 397 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 398 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 399 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2recps_x16, |
| 400 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 401 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 402 | benchmark::utils::CheckNEONFMA) |
| 403 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 404 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 405 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2recps_x20, |
| 406 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 407 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 408 | benchmark::utils::CheckNEONFMA) |
| 409 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 410 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 411 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut64_p2_nr2recps_x24, |
| 412 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 413 | xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 414 | benchmark::utils::CheckNEONFMA) |
| 415 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 416 | ->UseRealTime(); |
| 417 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 418 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut64_p2_nr2recps_x4, |
| 419 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 420 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 421 | benchmark::utils::CheckNEON) |
| 422 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 423 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 424 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut64_p2_nr2recps_x8, |
| 425 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 426 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 427 | benchmark::utils::CheckNEON) |
| 428 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 429 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 430 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut64_p2_nr2recps_x12, |
| 431 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 432 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 433 | benchmark::utils::CheckNEON) |
| 434 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 435 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 436 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut64_p2_nr2recps_x16, |
| 437 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 438 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 439 | benchmark::utils::CheckNEON) |
| 440 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 441 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 442 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut64_p2_nr2recps_x20, |
| 443 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 444 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 445 | benchmark::utils::CheckNEON) |
| 446 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 447 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 448 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut64_p2_nr2recps_x24, |
| 449 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 450 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 451 | benchmark::utils::CheckNEON) |
| 452 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 453 | ->UseRealTime(); |
| 454 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 455 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2fma_x4, |
| 456 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 457 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 458 | benchmark::utils::CheckNEONFMA) |
| 459 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 460 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 461 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2fma_x8, |
| 462 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 463 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 464 | benchmark::utils::CheckNEONFMA) |
| 465 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 466 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 467 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2fma_x12, |
| 468 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 469 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 470 | benchmark::utils::CheckNEONFMA) |
| 471 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 472 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 473 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2fma_x16, |
| 474 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 475 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 476 | benchmark::utils::CheckNEONFMA) |
| 477 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 478 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 479 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2fma_x20, |
| 480 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 481 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 482 | benchmark::utils::CheckNEONFMA) |
| 483 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 484 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 485 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2fma_x24, |
| 486 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 487 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 488 | benchmark::utils::CheckNEONFMA) |
| 489 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 490 | ->UseRealTime(); |
| 491 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 492 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr1recps1fma_x4, |
| 493 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 494 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 495 | benchmark::utils::CheckNEONFMA) |
| 496 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 497 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 498 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr1recps1fma_x8, |
| 499 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 500 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 501 | benchmark::utils::CheckNEONFMA) |
| 502 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 503 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 504 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr1recps1fma_x12, |
| 505 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 506 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 507 | benchmark::utils::CheckNEONFMA) |
| 508 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 509 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 510 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr1recps1fma_x16, |
| 511 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 512 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 513 | benchmark::utils::CheckNEONFMA) |
| 514 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 515 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 516 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr1recps1fma_x20, |
| 517 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 518 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 519 | benchmark::utils::CheckNEONFMA) |
| 520 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 521 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 522 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr1recps1fma_x24, |
| 523 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 524 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 525 | benchmark::utils::CheckNEONFMA) |
| 526 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 527 | ->UseRealTime(); |
| 528 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 529 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2recps_x4, |
| 530 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 531 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 532 | benchmark::utils::CheckNEONFMA) |
| 533 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 534 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 535 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2recps_x8, |
| 536 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 537 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 538 | benchmark::utils::CheckNEONFMA) |
| 539 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 540 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 541 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2recps_x12, |
| 542 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 543 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 544 | benchmark::utils::CheckNEONFMA) |
| 545 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 546 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 547 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2recps_x16, |
| 548 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 549 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 550 | benchmark::utils::CheckNEONFMA) |
| 551 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 552 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 553 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2recps_x20, |
| 554 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 555 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 556 | benchmark::utils::CheckNEONFMA) |
| 557 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 558 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 559 | BENCHMARK_CAPTURE(f32_vsigmoid, neonfma_rr1_lut2048_p1_nr2recps_x24, |
| 560 | xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 561 | xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 562 | benchmark::utils::CheckNEONFMA) |
| 563 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 564 | ->UseRealTime(); |
| 565 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 566 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut2048_p1_nr2recps_x4, |
| 567 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x4, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 568 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 569 | benchmark::utils::CheckNEON) |
| 570 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 571 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 572 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut2048_p1_nr2recps_x8, |
| 573 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 574 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 575 | benchmark::utils::CheckNEON) |
| 576 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 577 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 578 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut2048_p1_nr2recps_x12, |
| 579 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 580 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 581 | benchmark::utils::CheckNEON) |
| 582 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 583 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 584 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut2048_p1_nr2recps_x16, |
| 585 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 586 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 587 | benchmark::utils::CheckNEON) |
| 588 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 589 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 590 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut2048_p1_nr2recps_x20, |
| 591 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 592 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 593 | benchmark::utils::CheckNEON) |
| 594 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 595 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 596 | BENCHMARK_CAPTURE(f32_vsigmoid, neon_rr2_lut2048_p1_nr2recps_x24, |
| 597 | xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 598 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 599 | benchmark::utils::CheckNEON) |
| 600 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 601 | ->UseRealTime(); |
| 602 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 603 | |
| 604 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 605 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x16, |
| 606 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 607 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 608 | benchmark::utils::CheckAVX512F) |
| 609 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 610 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 611 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x32, |
| 612 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 613 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 614 | benchmark::utils::CheckAVX512F) |
| 615 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 616 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 617 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x48, |
| 618 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 619 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 620 | benchmark::utils::CheckAVX512F) |
| 621 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 622 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 623 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x64, |
| 624 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 625 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 626 | benchmark::utils::CheckAVX512F) |
| 627 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 628 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 629 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x80, |
| 630 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 631 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 632 | benchmark::utils::CheckAVX512F) |
| 633 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 634 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 635 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x96, |
| 636 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x96, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 637 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 638 | benchmark::utils::CheckAVX512F) |
| 639 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 640 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 641 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x112, |
| 642 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 643 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 644 | benchmark::utils::CheckAVX512F) |
| 645 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 646 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 647 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_div_x128, |
| 648 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 649 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 650 | benchmark::utils::CheckAVX512F) |
| 651 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 652 | ->UseRealTime(); |
| 653 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 654 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x16, |
| 655 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 656 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 657 | benchmark::utils::CheckAVX512F) |
| 658 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 659 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 660 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x32, |
| 661 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 662 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 663 | benchmark::utils::CheckAVX512F) |
| 664 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 665 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 666 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x48, |
| 667 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 668 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 669 | benchmark::utils::CheckAVX512F) |
| 670 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 671 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 672 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x64, |
| 673 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 674 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 675 | benchmark::utils::CheckAVX512F) |
| 676 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 677 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 678 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x80, |
| 679 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 680 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 681 | benchmark::utils::CheckAVX512F) |
| 682 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 683 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 684 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x96, |
| 685 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 686 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 687 | benchmark::utils::CheckAVX512F) |
| 688 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 689 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 690 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x112, |
| 691 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 692 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 693 | benchmark::utils::CheckAVX512F) |
| 694 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 695 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 696 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_p5_scalef_nr1fma_x128, |
| 697 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 698 | xnn_init_f32_sigmoid_avx512_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 699 | benchmark::utils::CheckAVX512F) |
| 700 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 701 | ->UseRealTime(); |
| 702 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 703 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x16, |
| 704 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 705 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 706 | benchmark::utils::CheckAVX512F) |
| 707 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 708 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 709 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x32, |
| 710 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 711 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 712 | benchmark::utils::CheckAVX512F) |
| 713 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 714 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 715 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x48, |
| 716 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 717 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 718 | benchmark::utils::CheckAVX512F) |
| 719 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 720 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 721 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x64, |
| 722 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 723 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 724 | benchmark::utils::CheckAVX512F) |
| 725 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 726 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 727 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x80, |
| 728 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 729 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 730 | benchmark::utils::CheckAVX512F) |
| 731 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 732 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 733 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x96, |
| 734 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 735 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 736 | benchmark::utils::CheckAVX512F) |
| 737 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 738 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 739 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x112, |
| 740 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 741 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 742 | benchmark::utils::CheckAVX512F) |
| 743 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 744 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 745 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_div_x128, |
| 746 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 747 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 748 | benchmark::utils::CheckAVX512F) |
| 749 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 750 | ->UseRealTime(); |
| 751 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 752 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x16, |
| 753 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 754 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 755 | benchmark::utils::CheckAVX512F) |
| 756 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 757 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 758 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x32, |
| 759 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 760 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 761 | benchmark::utils::CheckAVX512F) |
| 762 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 763 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 764 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x48, |
| 765 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 766 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 767 | benchmark::utils::CheckAVX512F) |
| 768 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 769 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 770 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x64, |
| 771 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 772 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 773 | benchmark::utils::CheckAVX512F) |
| 774 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 775 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 776 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x80, |
| 777 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 778 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 779 | benchmark::utils::CheckAVX512F) |
| 780 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 781 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 782 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x96, |
| 783 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 784 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 785 | benchmark::utils::CheckAVX512F) |
| 786 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 787 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 788 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x112, |
| 789 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 790 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 791 | benchmark::utils::CheckAVX512F) |
| 792 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 793 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 794 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut16_p3_perm_scalef_nr1fma_x128, |
| 795 | xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 796 | xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 797 | benchmark::utils::CheckAVX512F) |
| 798 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 799 | ->UseRealTime(); |
| 800 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 801 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x16, |
| 802 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 803 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 804 | benchmark::utils::CheckAVX512F) |
| 805 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 806 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 807 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x32, |
| 808 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 809 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 810 | benchmark::utils::CheckAVX512F) |
| 811 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 812 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 813 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x48, |
| 814 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 815 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 816 | benchmark::utils::CheckAVX512F) |
| 817 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 818 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 819 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x64, |
| 820 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 821 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 822 | benchmark::utils::CheckAVX512F) |
| 823 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 824 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 825 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x80, |
| 826 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 827 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 828 | benchmark::utils::CheckAVX512F) |
| 829 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 830 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 831 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x96, |
| 832 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 833 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 834 | benchmark::utils::CheckAVX512F) |
| 835 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 836 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 837 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x112, |
| 838 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 839 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 840 | benchmark::utils::CheckAVX512F) |
| 841 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 842 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 843 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_div_x128, |
| 844 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 845 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 846 | benchmark::utils::CheckAVX512F) |
| 847 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 848 | ->UseRealTime(); |
| 849 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 850 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x16, |
| 851 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 852 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 853 | benchmark::utils::CheckAVX512F) |
| 854 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 855 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 856 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x32, |
| 857 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 858 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 859 | benchmark::utils::CheckAVX512F) |
| 860 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 861 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 862 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x48, |
| 863 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 864 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 865 | benchmark::utils::CheckAVX512F) |
| 866 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 867 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 868 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x64, |
| 869 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 870 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 871 | benchmark::utils::CheckAVX512F) |
| 872 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 873 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 874 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x80, |
| 875 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 876 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 877 | benchmark::utils::CheckAVX512F) |
| 878 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 879 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 880 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x96, |
| 881 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 882 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 883 | benchmark::utils::CheckAVX512F) |
| 884 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 885 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 886 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x112, |
| 887 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 888 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 889 | benchmark::utils::CheckAVX512F) |
| 890 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 891 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 892 | BENCHMARK_CAPTURE(f32_vsigmoid, avx512f_lut32_p2_perm2_scalef_nr1fma_x128, |
| 893 | xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 894 | xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 895 | benchmark::utils::CheckAVX512F) |
| 896 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 897 | ->UseRealTime(); |
| 898 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 899 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x8, |
| 900 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 901 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 902 | benchmark::utils::CheckAVX2) |
| 903 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 904 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 905 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x16, |
| 906 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 907 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 908 | benchmark::utils::CheckAVX2) |
| 909 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 910 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 911 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x24, |
| 912 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 913 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 914 | benchmark::utils::CheckAVX2) |
| 915 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 916 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 917 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x32, |
| 918 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 919 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 920 | benchmark::utils::CheckAVX2) |
| 921 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 922 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 923 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x40, |
| 924 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x40, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 925 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 926 | benchmark::utils::CheckAVX2) |
| 927 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 928 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 929 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x48, |
| 930 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 931 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 932 | benchmark::utils::CheckAVX2) |
| 933 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 934 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 935 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x56, |
| 936 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x56, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 937 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 938 | benchmark::utils::CheckAVX2) |
| 939 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 940 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 941 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x64, |
| 942 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 943 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 944 | benchmark::utils::CheckAVX2) |
| 945 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 946 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 947 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x72, |
| 948 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x72, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 949 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 950 | benchmark::utils::CheckAVX2) |
| 951 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 952 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 953 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_div_x80, |
| 954 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 955 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 956 | benchmark::utils::CheckAVX2) |
| 957 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 958 | ->UseRealTime(); |
| 959 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 960 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x8, |
| 961 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 962 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 963 | benchmark::utils::CheckAVX2) |
| 964 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 965 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 966 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x16, |
| 967 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 968 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 969 | benchmark::utils::CheckAVX2) |
| 970 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 971 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 972 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x24, |
| 973 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 974 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 975 | benchmark::utils::CheckAVX2) |
| 976 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 977 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 978 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x32, |
| 979 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 980 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 981 | benchmark::utils::CheckAVX2) |
| 982 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 983 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 984 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x40, |
| 985 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x40, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 986 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 987 | benchmark::utils::CheckAVX2) |
| 988 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 989 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 990 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x48, |
| 991 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 992 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 993 | benchmark::utils::CheckAVX2) |
| 994 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 995 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 996 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x56, |
| 997 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x56, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 998 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 999 | benchmark::utils::CheckAVX2) |
| 1000 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1001 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1002 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x64, |
| 1003 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1004 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1005 | benchmark::utils::CheckAVX2) |
| 1006 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1007 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1008 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x72, |
| 1009 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x72, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1010 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1011 | benchmark::utils::CheckAVX2) |
| 1012 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1013 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1014 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr1fma_x80, |
| 1015 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1016 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1017 | benchmark::utils::CheckAVX2) |
| 1018 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1019 | ->UseRealTime(); |
| 1020 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1021 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x8, |
| 1022 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1023 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1024 | benchmark::utils::CheckAVX2) |
| 1025 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1026 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1027 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x16, |
| 1028 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1029 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1030 | benchmark::utils::CheckAVX2) |
| 1031 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1032 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1033 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x24, |
| 1034 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1035 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1036 | benchmark::utils::CheckAVX2) |
| 1037 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1038 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1039 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x32, |
| 1040 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1041 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1042 | benchmark::utils::CheckAVX2) |
| 1043 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1044 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1045 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x40, |
| 1046 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x40, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1047 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1048 | benchmark::utils::CheckAVX2) |
| 1049 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1050 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1051 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x48, |
| 1052 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1053 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1054 | benchmark::utils::CheckAVX2) |
| 1055 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1056 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1057 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x56, |
| 1058 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x56, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1059 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1060 | benchmark::utils::CheckAVX2) |
| 1061 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1062 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1063 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x64, |
| 1064 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1065 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1066 | benchmark::utils::CheckAVX2) |
| 1067 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1068 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1069 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x72, |
| 1070 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x72, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1071 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1072 | benchmark::utils::CheckAVX2) |
| 1073 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1074 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1075 | BENCHMARK_CAPTURE(f32_vsigmoid, avx2_p5_nr2fma_x80, |
| 1076 | xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1077 | xnn_init_f32_sigmoid_avx2_rr1_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1078 | benchmark::utils::CheckAVX2) |
| 1079 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1080 | ->UseRealTime(); |
| 1081 | |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1082 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x8, |
| 1083 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1084 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1085 | benchmark::utils::CheckAVX) |
| 1086 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1087 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1088 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x16, |
| 1089 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1090 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1091 | benchmark::utils::CheckAVX) |
| 1092 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1093 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1094 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x24, |
| 1095 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1096 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1097 | benchmark::utils::CheckAVX) |
| 1098 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1099 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1100 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x32, |
| 1101 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1102 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1103 | benchmark::utils::CheckAVX) |
| 1104 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1105 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1106 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x40, |
| 1107 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x40, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1108 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1109 | benchmark::utils::CheckAVX) |
| 1110 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1111 | ->UseRealTime(); |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1112 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x48, |
| 1113 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1114 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1115 | benchmark::utils::CheckAVX) |
| 1116 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1117 | ->UseRealTime(); |
| 1118 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x56, |
| 1119 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x56, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1120 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1121 | benchmark::utils::CheckAVX) |
| 1122 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1123 | ->UseRealTime(); |
| 1124 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x64, |
| 1125 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1126 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1127 | benchmark::utils::CheckAVX) |
| 1128 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1129 | ->UseRealTime(); |
| 1130 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x72, |
| 1131 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x72, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1132 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1133 | benchmark::utils::CheckAVX) |
| 1134 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1135 | ->UseRealTime(); |
| 1136 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_div_x80, |
| 1137 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1138 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1139 | benchmark::utils::CheckAVX) |
| 1140 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1141 | ->UseRealTime(); |
| 1142 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x8, |
| 1143 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x8, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1144 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1145 | benchmark::utils::CheckAVX) |
| 1146 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1147 | ->UseRealTime(); |
| 1148 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x16, |
| 1149 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x16, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1150 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1151 | benchmark::utils::CheckAVX) |
| 1152 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1153 | ->UseRealTime(); |
| 1154 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x24, |
| 1155 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x24, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1156 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1157 | benchmark::utils::CheckAVX) |
| 1158 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1159 | ->UseRealTime(); |
| 1160 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x32, |
| 1161 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x32, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1162 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1163 | benchmark::utils::CheckAVX) |
| 1164 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1165 | ->UseRealTime(); |
| 1166 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x40, |
| 1167 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x40, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1168 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1169 | benchmark::utils::CheckAVX) |
| 1170 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1171 | ->UseRealTime(); |
| 1172 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x48, |
| 1173 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x48, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1174 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1175 | benchmark::utils::CheckAVX) |
| 1176 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1177 | ->UseRealTime(); |
| 1178 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x56, |
| 1179 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x56, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1180 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1181 | benchmark::utils::CheckAVX) |
| 1182 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1183 | ->UseRealTime(); |
| 1184 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x64, |
| 1185 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x64, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1186 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1187 | benchmark::utils::CheckAVX) |
| 1188 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1189 | ->UseRealTime(); |
| 1190 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x72, |
| 1191 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x72, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1192 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1193 | benchmark::utils::CheckAVX) |
| 1194 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1195 | ->UseRealTime(); |
| 1196 | BENCHMARK_CAPTURE(f32_vsigmoid, avx_p5_nr2_x80, |
| 1197 | xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x80, |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1198 | xnn_init_f32_sigmoid_avx_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1199 | benchmark::utils::CheckAVX) |
| 1200 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1201 | ->UseRealTime(); |
| 1202 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1203 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_p5_div_x4, |
| 1204 | xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x4, |
| 1205 | xnn_init_f32_sigmoid_sse2_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1206 | benchmark::utils::CheckSSE41) |
| 1207 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1208 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1209 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_p5_div_x8, |
| 1210 | xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x8, |
| 1211 | xnn_init_f32_sigmoid_sse2_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1212 | benchmark::utils::CheckSSE41) |
| 1213 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1214 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1215 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_p5_div_x12, |
| 1216 | xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x12, |
| 1217 | xnn_init_f32_sigmoid_sse2_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1218 | benchmark::utils::CheckSSE41) |
| 1219 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1220 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1221 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_p5_div_x16, |
| 1222 | xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x16, |
| 1223 | xnn_init_f32_sigmoid_sse2_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1224 | benchmark::utils::CheckSSE41) |
| 1225 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1226 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1227 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_p5_div_x20, |
| 1228 | xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x20, |
| 1229 | xnn_init_f32_sigmoid_sse2_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1230 | benchmark::utils::CheckSSE41) |
| 1231 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1232 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1233 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_p5_div_x24, |
| 1234 | xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x24, |
| 1235 | xnn_init_f32_sigmoid_sse2_rr2_p5_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1236 | benchmark::utils::CheckSSE41) |
| 1237 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1238 | ->UseRealTime(); |
| 1239 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1240 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_lut64_p2_div_x4, |
| 1241 | xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x4, |
| 1242 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1243 | benchmark::utils::CheckSSE41) |
| 1244 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1245 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1246 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_lut64_p2_div_x8, |
| 1247 | xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x8, |
| 1248 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1249 | benchmark::utils::CheckSSE41) |
| 1250 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1251 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1252 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_lut64_p2_div_x12, |
| 1253 | xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x12, |
| 1254 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1255 | benchmark::utils::CheckSSE41) |
| 1256 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1257 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1258 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_lut64_p2_div_x16, |
| 1259 | xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x16, |
| 1260 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1261 | benchmark::utils::CheckSSE41) |
| 1262 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1263 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1264 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_lut64_p2_div_x20, |
| 1265 | xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x20, |
| 1266 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1267 | benchmark::utils::CheckSSE41) |
| 1268 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1269 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1270 | BENCHMARK_CAPTURE(f32_vsigmoid, sse41_rr2_lut64_p2_div_x24, |
| 1271 | xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x24, |
| 1272 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1273 | benchmark::utils::CheckSSE41) |
| 1274 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1275 | ->UseRealTime(); |
| 1276 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1277 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_p5_div_x4, |
| 1278 | xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x4, |
| 1279 | xnn_init_f32_sigmoid_sse2_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1280 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1281 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1282 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_p5_div_x8, |
| 1283 | xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x8, |
| 1284 | xnn_init_f32_sigmoid_sse2_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1285 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1286 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1287 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_p5_div_x12, |
| 1288 | xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x12, |
| 1289 | xnn_init_f32_sigmoid_sse2_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1290 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1291 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1292 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_p5_div_x16, |
| 1293 | xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x16, |
| 1294 | xnn_init_f32_sigmoid_sse2_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1295 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1296 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1297 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_p5_div_x20, |
| 1298 | xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x20, |
| 1299 | xnn_init_f32_sigmoid_sse2_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1300 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1301 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1302 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_p5_div_x24, |
| 1303 | xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x24, |
| 1304 | xnn_init_f32_sigmoid_sse2_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1305 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1306 | ->UseRealTime(); |
| 1307 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1308 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_lut64_p2_div_x4, |
| 1309 | xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x4, |
| 1310 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1311 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1312 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1313 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_lut64_p2_div_x8, |
| 1314 | xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x8, |
| 1315 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1316 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1317 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1318 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_lut64_p2_div_x12, |
| 1319 | xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x12, |
| 1320 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1321 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1322 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1323 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_lut64_p2_div_x16, |
| 1324 | xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x16, |
| 1325 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1326 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1327 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1328 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_lut64_p2_div_x20, |
| 1329 | xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x20, |
| 1330 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1331 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 1332 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1333 | BENCHMARK_CAPTURE(f32_vsigmoid, sse2_rr2_lut64_p2_div_x24, |
| 1334 | xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x24, |
| 1335 | xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1336 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1337 | ->UseRealTime(); |
| 1338 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1339 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1340 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1341 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_lut64_p2_div_x4, |
| 1342 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x4, |
| 1343 | xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1344 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1345 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1346 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_lut64_p2_div_x8, |
| 1347 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x8, |
| 1348 | xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1349 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1350 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1351 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_lut64_p2_div_x12, |
| 1352 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x12, |
| 1353 | xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1354 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1355 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1356 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_lut64_p2_div_x16, |
| 1357 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x16, |
| 1358 | xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1359 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1360 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1361 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_lut64_p2_div_x20, |
| 1362 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x20, |
| 1363 | xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1364 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1365 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1366 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_lut64_p2_div_x24, |
| 1367 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x24, |
| 1368 | xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1369 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1370 | ->UseRealTime(); |
| 1371 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1372 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_p5_div_x4, |
| 1373 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x4, |
| 1374 | xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1375 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1376 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1377 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_p5_div_x8, |
| 1378 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x8, |
| 1379 | xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1380 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1381 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1382 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_p5_div_x12, |
| 1383 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x12, |
| 1384 | xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1385 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1386 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1387 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_p5_div_x16, |
| 1388 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x16, |
| 1389 | xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1390 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1391 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1392 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_p5_div_x20, |
| 1393 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20, |
| 1394 | xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1395 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1396 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1397 | BENCHMARK_CAPTURE(f32_vsigmoid, wasmsimd_rr2_p5_div_x24, |
| 1398 | xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x24, |
| 1399 | xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1400 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1401 | ->UseRealTime(); |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1402 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1403 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1404 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_lut2048_p1_div_x1, |
| 1405 | xnn_f32_vsigmoid_ukernel__scalar_rr2_lut2048_p1_div_x1, |
| 1406 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1407 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1408 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1409 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_lut2048_p1_div_x2, |
| 1410 | xnn_f32_vsigmoid_ukernel__scalar_rr2_lut2048_p1_div_x2, |
| 1411 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1412 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1413 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1414 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_lut2048_p1_div_x4, |
| 1415 | xnn_f32_vsigmoid_ukernel__scalar_rr2_lut2048_p1_div_x4, |
| 1416 | xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1417 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1418 | ->UseRealTime(); |
| 1419 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1420 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_lut64_p2_div_x1, |
| 1421 | xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x1, |
| 1422 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1423 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1424 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1425 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_lut64_p2_div_x2, |
| 1426 | xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x2, |
| 1427 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1428 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1429 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1430 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_lut64_p2_div_x4, |
| 1431 | xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x4, |
| 1432 | xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1433 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1434 | ->UseRealTime(); |
| 1435 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1436 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_p5_div_x1, |
| 1437 | xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x1, |
| 1438 | xnn_init_f32_sigmoid_scalar_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1439 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1440 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1441 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_p5_div_x2, |
| 1442 | xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x2, |
| 1443 | xnn_init_f32_sigmoid_scalar_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1444 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1445 | ->UseRealTime(); |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 1446 | BENCHMARK_CAPTURE(f32_vsigmoid, scalar_rr2_p5_div_x4, |
| 1447 | xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x4, |
| 1448 | xnn_init_f32_sigmoid_scalar_rr2_p5_params) |
Marat Dukhan | 9f8ea9b | 2021-08-06 22:09:37 -0700 | [diff] [blame] | 1449 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | 6674d69 | 2021-05-05 22:27:00 -0700 | [diff] [blame] | 1450 | ->UseRealTime(); |
| 1451 | |
| 1452 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 1453 | BENCHMARK_MAIN(); |
| 1454 | #endif |