Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <algorithm> |
| 7 | #include <cmath> |
| 8 | #include <functional> |
| 9 | #include <random> |
| 10 | #include <vector> |
| 11 | |
| 12 | #include <benchmark/benchmark.h> |
| 13 | #include "bench/utils.h" |
| 14 | |
| 15 | #include <xnnpack/AlignedAllocator.h> |
| 16 | #include <xnnpack/common.h> |
| 17 | #include <xnnpack/vunary.h> |
| 18 | #include <xnnpack/params.h> |
| 19 | #include <xnnpack/params-init.h> |
| 20 | |
| 21 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 22 | static void f32_velu( |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 23 | benchmark::State& state, |
| 24 | xnn_f32_velu_ukernel_function elu, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 25 | xnn_init_f32_elu_params_fn init_params, |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 26 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 27 | { |
| 28 | if (isa_check && !isa_check(state)) { |
| 29 | return; |
| 30 | } |
| 31 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 32 | const size_t num_elements = state.range(0); |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [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>(-20.0f, 10.0f), std::ref(rng)); |
| 37 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -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 | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 40 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 41 | std::fill(y.begin(), y.end(), std::nanf("")); |
| 42 | |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame] | 43 | union xnn_f32_elu_params params; |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 44 | init_params(¶ms, 1.0f /* prescale */, 1.0f /* alpha */, 1.0f /* beta */); |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 45 | for (auto _ : state) { |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 46 | elu(num_elements * sizeof(float), x.data(), y.data(), ¶ms); |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [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 | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 54 | const size_t elements_per_iteration = num_elements; |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 55 | state.counters["elements"] = |
| 56 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 57 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 58 | const size_t bytes_per_iteration = 2 * num_elements * sizeof(float); |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [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_ARM || XNN_ARCH_ARM64 |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 64 | BENCHMARK_CAPTURE(f32_velu, neonfma_lut16_p3_x4, |
| 65 | xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 66 | xnn_init_f32_elu_neonfma_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 67 | benchmark::utils::CheckNEONFMA) |
| 68 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 69 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 70 | BENCHMARK_CAPTURE(f32_velu, neonfma_lut16_p3_x8, |
| 71 | xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 72 | xnn_init_f32_elu_neonfma_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 73 | benchmark::utils::CheckNEONFMA) |
| 74 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 75 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 76 | BENCHMARK_CAPTURE(f32_velu, neonfma_lut16_p3_x12, |
| 77 | xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 78 | xnn_init_f32_elu_neonfma_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 79 | benchmark::utils::CheckNEONFMA) |
| 80 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 81 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 82 | BENCHMARK_CAPTURE(f32_velu, neonfma_lut16_p3_x16, |
| 83 | xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 84 | xnn_init_f32_elu_neonfma_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 85 | benchmark::utils::CheckNEONFMA) |
| 86 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 87 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 88 | BENCHMARK_CAPTURE(f32_velu, neonfma_lut16_p3_x20, |
| 89 | xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 90 | xnn_init_f32_elu_neonfma_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 91 | benchmark::utils::CheckNEONFMA) |
| 92 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 93 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 94 | BENCHMARK_CAPTURE(f32_velu, neonfma_lut16_p3_x24, |
| 95 | xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 96 | xnn_init_f32_elu_neonfma_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 97 | benchmark::utils::CheckNEONFMA) |
| 98 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 99 | ->UseRealTime(); |
| 100 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 101 | BENCHMARK_CAPTURE(f32_velu, neonfma_p6_x4, |
| 102 | xnn_f32_velu_ukernel__neonfma_rr1_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 103 | xnn_init_f32_elu_neonfma_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 104 | benchmark::utils::CheckNEONFMA) |
| 105 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 106 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 107 | BENCHMARK_CAPTURE(f32_velu, neonfma_p6_x8, |
| 108 | xnn_f32_velu_ukernel__neonfma_rr1_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 109 | xnn_init_f32_elu_neonfma_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 110 | benchmark::utils::CheckNEONFMA) |
| 111 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 112 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 113 | BENCHMARK_CAPTURE(f32_velu, neonfma_p6_x12, |
| 114 | xnn_f32_velu_ukernel__neonfma_rr1_p6_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 115 | xnn_init_f32_elu_neonfma_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 116 | benchmark::utils::CheckNEONFMA) |
| 117 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 118 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 119 | BENCHMARK_CAPTURE(f32_velu, neonfma_p6_x16, |
| 120 | xnn_f32_velu_ukernel__neonfma_rr1_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 121 | xnn_init_f32_elu_neonfma_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 122 | benchmark::utils::CheckNEONFMA) |
| 123 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 124 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 125 | BENCHMARK_CAPTURE(f32_velu, neonfma_p6_x20, |
| 126 | xnn_f32_velu_ukernel__neonfma_rr1_p6_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 127 | xnn_init_f32_elu_neonfma_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 128 | benchmark::utils::CheckNEONFMA) |
| 129 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 130 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 131 | BENCHMARK_CAPTURE(f32_velu, neonfma_p6_x24, |
| 132 | xnn_f32_velu_ukernel__neonfma_rr1_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 133 | xnn_init_f32_elu_neonfma_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 134 | benchmark::utils::CheckNEONFMA) |
| 135 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 136 | ->UseRealTime(); |
| 137 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 138 | BENCHMARK_CAPTURE(f32_velu, neon_lut16_p3_x4, |
| 139 | xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 140 | xnn_init_f32_elu_neon_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 141 | benchmark::utils::CheckNEON) |
| 142 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 143 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 144 | BENCHMARK_CAPTURE(f32_velu, neon_lut16_p3_x8, |
| 145 | xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 146 | xnn_init_f32_elu_neon_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 147 | benchmark::utils::CheckNEON) |
| 148 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 149 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 150 | BENCHMARK_CAPTURE(f32_velu, neon_lut16_p3_x12, |
| 151 | xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 152 | xnn_init_f32_elu_neon_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 153 | benchmark::utils::CheckNEON) |
| 154 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 155 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 156 | BENCHMARK_CAPTURE(f32_velu, neon_lut16_p3_x16, |
| 157 | xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 158 | xnn_init_f32_elu_neon_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 159 | benchmark::utils::CheckNEON) |
| 160 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 161 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 162 | BENCHMARK_CAPTURE(f32_velu, neon_lut16_p3_x20, |
| 163 | xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 164 | xnn_init_f32_elu_neon_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 165 | benchmark::utils::CheckNEON) |
| 166 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 167 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 168 | BENCHMARK_CAPTURE(f32_velu, neon_lut16_p3_x24, |
| 169 | xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 170 | xnn_init_f32_elu_neon_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 171 | benchmark::utils::CheckNEON) |
| 172 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 173 | ->UseRealTime(); |
| 174 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 175 | BENCHMARK_CAPTURE(f32_velu, neon_p6_x4, |
| 176 | xnn_f32_velu_ukernel__neon_rr2_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 177 | xnn_init_f32_elu_neon_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 178 | benchmark::utils::CheckNEON) |
| 179 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 180 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 181 | BENCHMARK_CAPTURE(f32_velu, neon_p6_x8, |
| 182 | xnn_f32_velu_ukernel__neon_rr2_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 183 | xnn_init_f32_elu_neon_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 184 | benchmark::utils::CheckNEON) |
| 185 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 186 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 187 | BENCHMARK_CAPTURE(f32_velu, neon_p6_x12, |
| 188 | xnn_f32_velu_ukernel__neon_rr2_p6_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 189 | xnn_init_f32_elu_neon_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 190 | benchmark::utils::CheckNEON) |
| 191 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 192 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 193 | BENCHMARK_CAPTURE(f32_velu, neon_p6_x16, |
| 194 | xnn_f32_velu_ukernel__neon_rr2_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 195 | xnn_init_f32_elu_neon_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 196 | benchmark::utils::CheckNEON) |
| 197 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 198 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 199 | BENCHMARK_CAPTURE(f32_velu, neon_p6_x20, |
| 200 | xnn_f32_velu_ukernel__neon_rr2_p6_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 201 | xnn_init_f32_elu_neon_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 202 | benchmark::utils::CheckNEON) |
| 203 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 204 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 205 | BENCHMARK_CAPTURE(f32_velu, neon_p6_x24, |
| 206 | xnn_f32_velu_ukernel__neon_rr2_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 207 | xnn_init_f32_elu_neon_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 208 | benchmark::utils::CheckNEON) |
| 209 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 210 | ->UseRealTime(); |
| 211 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 212 | |
| 213 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 214 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x16, |
| 215 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 216 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 217 | benchmark::utils::CheckAVX512F) |
| 218 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 219 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 220 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x32, |
| 221 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 222 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 223 | benchmark::utils::CheckAVX512F) |
| 224 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 225 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 226 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x48, |
| 227 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 228 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 229 | benchmark::utils::CheckAVX512F) |
| 230 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 231 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 232 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x64, |
| 233 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 234 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 235 | benchmark::utils::CheckAVX512F) |
| 236 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 237 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 238 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x80, |
| 239 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 240 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 241 | benchmark::utils::CheckAVX512F) |
| 242 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 243 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 244 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x96, |
| 245 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 246 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 247 | benchmark::utils::CheckAVX512F) |
| 248 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 249 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 250 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x112, |
| 251 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 252 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 253 | benchmark::utils::CheckAVX512F) |
| 254 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 255 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 256 | BENCHMARK_CAPTURE(f32_velu, avx512f_lut16_p3_x128, |
| 257 | xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 258 | xnn_init_f32_elu_avx512_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 259 | benchmark::utils::CheckAVX512F) |
| 260 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 261 | ->UseRealTime(); |
| 262 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 263 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x16, |
| 264 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 265 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 266 | benchmark::utils::CheckAVX512F) |
| 267 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 268 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 269 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x32, |
| 270 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 271 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 272 | benchmark::utils::CheckAVX512F) |
| 273 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 274 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 275 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x48, |
| 276 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 277 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 278 | benchmark::utils::CheckAVX512F) |
| 279 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 280 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 281 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x64, |
| 282 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x64, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 283 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 284 | benchmark::utils::CheckAVX512F) |
| 285 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 286 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 287 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x80, |
| 288 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x80, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 289 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 290 | benchmark::utils::CheckAVX512F) |
| 291 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 292 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 293 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x96, |
| 294 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x96, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 295 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 296 | benchmark::utils::CheckAVX512F) |
| 297 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 298 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 299 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x112, |
| 300 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x112, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 301 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 302 | benchmark::utils::CheckAVX512F) |
| 303 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 304 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 305 | BENCHMARK_CAPTURE(f32_velu, avx512f_p6_x128, |
| 306 | xnn_f32_velu_ukernel__avx512f_rr1_p6_x128, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 307 | xnn_init_f32_elu_avx512_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 308 | benchmark::utils::CheckAVX512F) |
| 309 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 310 | ->UseRealTime(); |
| 311 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 312 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x8, |
| 313 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 314 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 315 | benchmark::utils::CheckAVX2) |
| 316 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 317 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 318 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x16, |
| 319 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 320 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 321 | benchmark::utils::CheckAVX2) |
| 322 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 323 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 324 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x24, |
| 325 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 326 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 327 | benchmark::utils::CheckAVX2) |
| 328 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 329 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 330 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x32, |
| 331 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 332 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 333 | benchmark::utils::CheckAVX2) |
| 334 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 335 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 336 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x40, |
| 337 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x40, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 338 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 339 | benchmark::utils::CheckAVX2) |
| 340 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 341 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 342 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x48, |
| 343 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 344 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 345 | benchmark::utils::CheckAVX2) |
| 346 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 347 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 348 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x56, |
| 349 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x56, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 350 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 351 | benchmark::utils::CheckAVX2) |
| 352 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 353 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 354 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x64, |
| 355 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 356 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 357 | benchmark::utils::CheckAVX2) |
| 358 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 359 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 360 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x72, |
| 361 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x72, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 362 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 363 | benchmark::utils::CheckAVX2) |
| 364 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 365 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 366 | BENCHMARK_CAPTURE(f32_velu, avx2_lut4_p4_x80, |
| 367 | xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 368 | xnn_init_f32_elu_avx2_rr1_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 369 | benchmark::utils::CheckAVX2) |
| 370 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 371 | ->UseRealTime(); |
| 372 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 373 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x8, |
| 374 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 375 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 376 | benchmark::utils::CheckAVX2) |
| 377 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 378 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 379 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x16, |
| 380 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 381 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 382 | benchmark::utils::CheckAVX2) |
| 383 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 384 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 385 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x24, |
| 386 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 387 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 388 | benchmark::utils::CheckAVX2) |
| 389 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 390 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 391 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x32, |
| 392 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 393 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 394 | benchmark::utils::CheckAVX2) |
| 395 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 396 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 397 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x40, |
| 398 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x40, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 399 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 400 | benchmark::utils::CheckAVX2) |
| 401 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 402 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 403 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x48, |
| 404 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 405 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 406 | benchmark::utils::CheckAVX2) |
| 407 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 408 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 409 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x56, |
| 410 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x56, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 411 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 412 | benchmark::utils::CheckAVX2) |
| 413 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 414 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 415 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x64, |
| 416 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 417 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 418 | benchmark::utils::CheckAVX2) |
| 419 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 420 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 421 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x72, |
| 422 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x72, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 423 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 424 | benchmark::utils::CheckAVX2) |
| 425 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 426 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 427 | BENCHMARK_CAPTURE(f32_velu, avx2_lut8_p4_x80, |
| 428 | xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 429 | xnn_init_f32_elu_avx2_rr1_lut8_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 430 | benchmark::utils::CheckAVX2) |
| 431 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 432 | ->UseRealTime(); |
| 433 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 434 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x8, |
| 435 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 436 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 437 | benchmark::utils::CheckAVX2) |
| 438 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 439 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 440 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x16, |
| 441 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 442 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 443 | benchmark::utils::CheckAVX2) |
| 444 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 445 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 446 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x24, |
| 447 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 448 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 449 | benchmark::utils::CheckAVX2) |
| 450 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 451 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 452 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x32, |
| 453 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 454 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 455 | benchmark::utils::CheckAVX2) |
| 456 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 457 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 458 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x40, |
| 459 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x40, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 460 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 461 | benchmark::utils::CheckAVX2) |
| 462 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 463 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 464 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x48, |
| 465 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 466 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 467 | benchmark::utils::CheckAVX2) |
| 468 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 469 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 470 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x56, |
| 471 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x56, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 472 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 473 | benchmark::utils::CheckAVX2) |
| 474 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 475 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 476 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x64, |
| 477 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 478 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 479 | benchmark::utils::CheckAVX2) |
| 480 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 481 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 482 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x72, |
| 483 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x72, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 484 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 485 | benchmark::utils::CheckAVX2) |
| 486 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 487 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 488 | BENCHMARK_CAPTURE(f32_velu, avx2_lut16_p3_x80, |
| 489 | xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 490 | xnn_init_f32_elu_avx2_rr1_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 491 | benchmark::utils::CheckAVX2) |
| 492 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 493 | ->UseRealTime(); |
| 494 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 495 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x8, |
| 496 | xnn_f32_velu_ukernel__avx2_rr1_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 497 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 498 | benchmark::utils::CheckAVX2) |
| 499 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 500 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 501 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x16, |
| 502 | xnn_f32_velu_ukernel__avx2_rr1_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 503 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 504 | benchmark::utils::CheckAVX2) |
| 505 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 506 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 507 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x24, |
| 508 | xnn_f32_velu_ukernel__avx2_rr1_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 509 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 510 | benchmark::utils::CheckAVX2) |
| 511 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 512 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 513 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x32, |
| 514 | xnn_f32_velu_ukernel__avx2_rr1_p6_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 515 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 516 | benchmark::utils::CheckAVX2) |
| 517 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 518 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 519 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x40, |
| 520 | xnn_f32_velu_ukernel__avx2_rr1_p6_x40, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 521 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 522 | benchmark::utils::CheckAVX2) |
| 523 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 524 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 525 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x48, |
| 526 | xnn_f32_velu_ukernel__avx2_rr1_p6_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 527 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 528 | benchmark::utils::CheckAVX2) |
| 529 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 530 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 531 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x56, |
| 532 | xnn_f32_velu_ukernel__avx2_rr1_p6_x56, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 533 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 534 | benchmark::utils::CheckAVX2) |
| 535 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 536 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 537 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x64, |
| 538 | xnn_f32_velu_ukernel__avx2_rr1_p6_x64, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 539 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 540 | benchmark::utils::CheckAVX2) |
| 541 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 542 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 543 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x72, |
| 544 | xnn_f32_velu_ukernel__avx2_rr1_p6_x72, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 545 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 546 | benchmark::utils::CheckAVX2) |
| 547 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 548 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 549 | BENCHMARK_CAPTURE(f32_velu, avx2_p6_x80, |
| 550 | xnn_f32_velu_ukernel__avx2_rr1_p6_x80, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 551 | xnn_init_f32_elu_avx2_rr1_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 552 | benchmark::utils::CheckAVX2) |
| 553 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 554 | ->UseRealTime(); |
| 555 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 556 | BENCHMARK_CAPTURE(f32_velu, avx_lut4_p4_x8, |
| 557 | xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 558 | xnn_init_f32_elu_avx_rr2_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 559 | benchmark::utils::CheckAVX) |
| 560 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 561 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 562 | BENCHMARK_CAPTURE(f32_velu, avx_lut4_p4_x16, |
| 563 | xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 564 | xnn_init_f32_elu_avx_rr2_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 565 | benchmark::utils::CheckAVX) |
| 566 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 567 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 568 | BENCHMARK_CAPTURE(f32_velu, avx_lut4_p4_x24, |
| 569 | xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 570 | xnn_init_f32_elu_avx_rr2_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 571 | benchmark::utils::CheckAVX) |
| 572 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 573 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 574 | BENCHMARK_CAPTURE(f32_velu, avx_lut4_p4_x32, |
| 575 | xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 576 | xnn_init_f32_elu_avx_rr2_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 577 | benchmark::utils::CheckAVX) |
| 578 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 579 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 580 | BENCHMARK_CAPTURE(f32_velu, avx_lut4_p4_x40, |
| 581 | xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x40, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 582 | xnn_init_f32_elu_avx_rr2_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 583 | benchmark::utils::CheckAVX) |
| 584 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 585 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 586 | BENCHMARK_CAPTURE(f32_velu, avx_lut4_p4_x48, |
| 587 | xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 588 | xnn_init_f32_elu_avx_rr2_lut4_p4_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 589 | benchmark::utils::CheckAVX) |
| 590 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 591 | ->UseRealTime(); |
| 592 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 593 | BENCHMARK_CAPTURE(f32_velu, avx_lut16_p3_x8, |
| 594 | xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 595 | xnn_init_f32_elu_avx_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 596 | benchmark::utils::CheckAVX) |
| 597 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 598 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 599 | BENCHMARK_CAPTURE(f32_velu, avx_lut16_p3_x16, |
| 600 | xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 601 | xnn_init_f32_elu_avx_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 602 | benchmark::utils::CheckAVX) |
| 603 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 604 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 605 | BENCHMARK_CAPTURE(f32_velu, avx_lut16_p3_x24, |
| 606 | xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 607 | xnn_init_f32_elu_avx_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 608 | benchmark::utils::CheckAVX) |
| 609 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 610 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 611 | BENCHMARK_CAPTURE(f32_velu, avx_lut16_p3_x32, |
| 612 | xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 613 | xnn_init_f32_elu_avx_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 614 | benchmark::utils::CheckAVX) |
| 615 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 616 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 617 | BENCHMARK_CAPTURE(f32_velu, avx_lut16_p3_x40, |
| 618 | xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 619 | xnn_init_f32_elu_avx_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 620 | benchmark::utils::CheckAVX) |
| 621 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 622 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 623 | BENCHMARK_CAPTURE(f32_velu, avx_lut16_p3_x48, |
| 624 | xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 625 | xnn_init_f32_elu_avx_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 626 | benchmark::utils::CheckAVX) |
| 627 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 628 | ->UseRealTime(); |
| 629 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 630 | BENCHMARK_CAPTURE(f32_velu, avx_p6_x8, |
| 631 | xnn_f32_velu_ukernel__avx_rr2_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 632 | xnn_init_f32_elu_avx_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 633 | benchmark::utils::CheckAVX) |
| 634 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 635 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 636 | BENCHMARK_CAPTURE(f32_velu, avx_p6_x16, |
| 637 | xnn_f32_velu_ukernel__avx_rr2_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 638 | xnn_init_f32_elu_avx_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 639 | benchmark::utils::CheckAVX) |
| 640 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 641 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 642 | BENCHMARK_CAPTURE(f32_velu, avx_p6_x24, |
| 643 | xnn_f32_velu_ukernel__avx_rr2_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 644 | xnn_init_f32_elu_avx_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 645 | benchmark::utils::CheckAVX) |
| 646 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 647 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 648 | BENCHMARK_CAPTURE(f32_velu, avx_p6_x32, |
| 649 | xnn_f32_velu_ukernel__avx_rr2_p6_x32, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 650 | xnn_init_f32_elu_avx_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 651 | benchmark::utils::CheckAVX) |
| 652 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 653 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 654 | BENCHMARK_CAPTURE(f32_velu, avx_p6_x40, |
| 655 | xnn_f32_velu_ukernel__avx_rr2_p6_x40, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 656 | xnn_init_f32_elu_avx_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 657 | benchmark::utils::CheckAVX) |
| 658 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 659 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 660 | BENCHMARK_CAPTURE(f32_velu, avx_p6_x48, |
| 661 | xnn_f32_velu_ukernel__avx_rr2_p6_x48, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 662 | xnn_init_f32_elu_avx_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 663 | benchmark::utils::CheckAVX) |
| 664 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 665 | ->UseRealTime(); |
| 666 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 667 | BENCHMARK_CAPTURE(f32_velu, sse41_lut16_p3_x4, |
| 668 | xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 669 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 670 | benchmark::utils::CheckSSE41) |
| 671 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 672 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 673 | BENCHMARK_CAPTURE(f32_velu, sse41_lut16_p3_x8, |
| 674 | xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 675 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 676 | benchmark::utils::CheckSSE41) |
| 677 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 678 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 679 | BENCHMARK_CAPTURE(f32_velu, sse41_lut16_p3_x12, |
| 680 | xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 681 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 682 | benchmark::utils::CheckSSE41) |
| 683 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 684 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 685 | BENCHMARK_CAPTURE(f32_velu, sse41_lut16_p3_x16, |
| 686 | xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 687 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 688 | benchmark::utils::CheckSSE41) |
| 689 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 690 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 691 | BENCHMARK_CAPTURE(f32_velu, sse41_lut16_p3_x20, |
| 692 | xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 693 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 694 | benchmark::utils::CheckSSE41) |
| 695 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 696 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 697 | BENCHMARK_CAPTURE(f32_velu, sse41_lut16_p3_x24, |
| 698 | xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 699 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 700 | benchmark::utils::CheckSSE41) |
| 701 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 702 | ->UseRealTime(); |
| 703 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 704 | BENCHMARK_CAPTURE(f32_velu, sse41_p6_x4, |
| 705 | xnn_f32_velu_ukernel__sse41_rr2_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 706 | xnn_init_f32_elu_sse2_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 707 | benchmark::utils::CheckSSE41) |
| 708 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 709 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 710 | BENCHMARK_CAPTURE(f32_velu, sse41_p6_x8, |
| 711 | xnn_f32_velu_ukernel__sse41_rr2_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 712 | xnn_init_f32_elu_sse2_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 713 | benchmark::utils::CheckSSE41) |
| 714 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 715 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 716 | BENCHMARK_CAPTURE(f32_velu, sse41_p6_x12, |
| 717 | xnn_f32_velu_ukernel__sse41_rr2_p6_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 718 | xnn_init_f32_elu_sse2_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 719 | benchmark::utils::CheckSSE41) |
| 720 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 721 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 722 | BENCHMARK_CAPTURE(f32_velu, sse41_p6_x16, |
| 723 | xnn_f32_velu_ukernel__sse41_rr2_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 724 | xnn_init_f32_elu_sse2_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 725 | benchmark::utils::CheckSSE41) |
| 726 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 727 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 728 | BENCHMARK_CAPTURE(f32_velu, sse41_p6_x20, |
| 729 | xnn_f32_velu_ukernel__sse41_rr2_p6_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 730 | xnn_init_f32_elu_sse2_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 731 | benchmark::utils::CheckSSE41) |
| 732 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 733 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 734 | BENCHMARK_CAPTURE(f32_velu, sse41_p6_x24, |
| 735 | xnn_f32_velu_ukernel__sse41_rr2_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 736 | xnn_init_f32_elu_sse2_rr2_p6_params, |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 737 | benchmark::utils::CheckSSE41) |
| 738 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 739 | ->UseRealTime(); |
| 740 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 741 | BENCHMARK_CAPTURE(f32_velu, sse2_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 742 | xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4, |
| 743 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 744 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 745 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 746 | BENCHMARK_CAPTURE(f32_velu, sse2_lut16_p3_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 747 | xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8, |
| 748 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 749 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 750 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 751 | BENCHMARK_CAPTURE(f32_velu, sse2_lut16_p3_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 752 | xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12, |
| 753 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 754 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 755 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 756 | BENCHMARK_CAPTURE(f32_velu, sse2_lut16_p3_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 757 | xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16, |
| 758 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 759 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 760 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 761 | BENCHMARK_CAPTURE(f32_velu, sse2_lut16_p3_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 762 | xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20, |
| 763 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 764 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 765 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 766 | BENCHMARK_CAPTURE(f32_velu, sse2_lut16_p3_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 767 | xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24, |
| 768 | xnn_init_f32_elu_sse2_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 769 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 770 | ->UseRealTime(); |
| 771 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 772 | BENCHMARK_CAPTURE(f32_velu, sse2_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 773 | xnn_f32_velu_ukernel__sse2_rr2_p6_x4, |
| 774 | xnn_init_f32_elu_sse2_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 775 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 776 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 777 | BENCHMARK_CAPTURE(f32_velu, sse2_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 778 | xnn_f32_velu_ukernel__sse2_rr2_p6_x8, |
| 779 | xnn_init_f32_elu_sse2_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 780 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 781 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 782 | BENCHMARK_CAPTURE(f32_velu, sse2_p6_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 783 | xnn_f32_velu_ukernel__sse2_rr2_p6_x12, |
| 784 | xnn_init_f32_elu_sse2_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 785 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 786 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 787 | BENCHMARK_CAPTURE(f32_velu, sse2_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 788 | xnn_f32_velu_ukernel__sse2_rr2_p6_x16, |
| 789 | xnn_init_f32_elu_sse2_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 790 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 791 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 792 | BENCHMARK_CAPTURE(f32_velu, sse2_p6_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 793 | xnn_f32_velu_ukernel__sse2_rr2_p6_x20, |
| 794 | xnn_init_f32_elu_sse2_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 795 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 796 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 797 | BENCHMARK_CAPTURE(f32_velu, sse2_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 798 | xnn_f32_velu_ukernel__sse2_rr2_p6_x24, |
| 799 | xnn_init_f32_elu_sse2_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 800 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 801 | ->UseRealTime(); |
| 802 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 803 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 804 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 805 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 806 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x4, |
| 807 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 808 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 809 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 810 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_lut16_p3_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 811 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x8, |
| 812 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 813 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 814 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 815 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_lut16_p3_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 816 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x12, |
| 817 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 818 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 819 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 820 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_lut16_p3_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 821 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x16, |
| 822 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 823 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 824 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 825 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_lut16_p3_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 826 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20, |
| 827 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 828 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 829 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 830 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_lut16_p3_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 831 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24, |
| 832 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 833 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 834 | ->UseRealTime(); |
| 835 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 836 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 837 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x4, |
| 838 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 839 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 840 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 841 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_lut16_p3_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 842 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x8, |
| 843 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 844 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 845 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 846 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_lut16_p3_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 847 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12, |
| 848 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 849 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 850 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 851 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_lut16_p3_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 852 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16, |
| 853 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 854 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 855 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 856 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_lut16_p3_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 857 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20, |
| 858 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 859 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 860 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 861 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_lut16_p3_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 862 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24, |
| 863 | xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 864 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 865 | ->UseRealTime(); |
| 866 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 867 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 868 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x4, |
| 869 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 870 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 871 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 872 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 873 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8, |
| 874 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 875 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 876 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 877 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_p6_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 878 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12, |
| 879 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 880 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 881 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 882 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 883 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16, |
| 884 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 885 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 886 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 887 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_p6_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 888 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20, |
| 889 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 890 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 891 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 892 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_arm_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 893 | xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24, |
| 894 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 895 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 896 | ->UseRealTime(); |
| 897 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 898 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 899 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x4, |
| 900 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 901 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 902 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 903 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_p6_x8, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 904 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8, |
| 905 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 906 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 907 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 908 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_p6_x12, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 909 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12, |
| 910 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 911 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 912 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 913 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_p6_x16, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 914 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16, |
| 915 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 916 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 917 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 918 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_p6_x20, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 919 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20, |
| 920 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 921 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 922 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 923 | BENCHMARK_CAPTURE(f32_velu, wasmsimd_x86_p6_x24, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 924 | xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24, |
| 925 | xnn_init_f32_elu_wasmsimd_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 926 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 927 | ->UseRealTime(); |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 928 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 929 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 930 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 931 | BENCHMARK_CAPTURE(f32_velu, wasm_lut16_p3_x1, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 932 | xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x1, |
| 933 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 934 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 935 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 936 | BENCHMARK_CAPTURE(f32_velu, wasm_lut16_p3_x2, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 937 | xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x2, |
| 938 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 939 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 940 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 941 | BENCHMARK_CAPTURE(f32_velu, wasm_lut16_p3_x3, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 942 | xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x3, |
| 943 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 944 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 945 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 946 | BENCHMARK_CAPTURE(f32_velu, wasm_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 947 | xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x4, |
| 948 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 949 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 950 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 951 | BENCHMARK_CAPTURE(f32_velu, wasm_lut16_p3_x5, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 952 | xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x5, |
| 953 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 954 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 955 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 956 | BENCHMARK_CAPTURE(f32_velu, wasm_lut16_p3_x6, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 957 | xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6, |
| 958 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 959 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 960 | ->UseRealTime(); |
| 961 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 962 | BENCHMARK_CAPTURE(f32_velu, wasm_p6_x1, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 963 | xnn_f32_velu_ukernel__wasm_rr2_p6_x1, |
| 964 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 965 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 966 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 967 | BENCHMARK_CAPTURE(f32_velu, wasm_p6_x2, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 968 | xnn_f32_velu_ukernel__wasm_rr2_p6_x2, |
| 969 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 970 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 971 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 972 | BENCHMARK_CAPTURE(f32_velu, wasm_p6_x3, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 973 | xnn_f32_velu_ukernel__wasm_rr2_p6_x3, |
| 974 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 975 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 976 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 977 | BENCHMARK_CAPTURE(f32_velu, wasm_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 978 | xnn_f32_velu_ukernel__wasm_rr2_p6_x4, |
| 979 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 980 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 981 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 982 | BENCHMARK_CAPTURE(f32_velu, wasm_p6_x5, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 983 | xnn_f32_velu_ukernel__wasm_rr2_p6_x5, |
| 984 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 985 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 986 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 987 | BENCHMARK_CAPTURE(f32_velu, wasm_p6_x6, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 988 | xnn_f32_velu_ukernel__wasm_rr2_p6_x6, |
| 989 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 990 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 991 | ->UseRealTime(); |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 992 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 993 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 994 | BENCHMARK_CAPTURE(f32_velu, scalar_lut16_p3_x1, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 995 | xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x1, |
| 996 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 997 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 998 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 999 | BENCHMARK_CAPTURE(f32_velu, scalar_lut16_p3_x2, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1000 | xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x2, |
| 1001 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1002 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1003 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1004 | BENCHMARK_CAPTURE(f32_velu, scalar_lut16_p3_x3, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1005 | xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x3, |
| 1006 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1007 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1008 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1009 | BENCHMARK_CAPTURE(f32_velu, scalar_lut16_p3_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1010 | xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x4, |
| 1011 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1012 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1013 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1014 | BENCHMARK_CAPTURE(f32_velu, scalar_lut16_p3_x5, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1015 | xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x5, |
| 1016 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1017 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1018 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1019 | BENCHMARK_CAPTURE(f32_velu, scalar_lut16_p3_x6, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1020 | xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6, |
| 1021 | xnn_init_f32_elu_scalar_rr2_lut16_p3_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1022 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1023 | ->UseRealTime(); |
| 1024 | |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1025 | BENCHMARK_CAPTURE(f32_velu, scalar_p6_x1, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1026 | xnn_f32_velu_ukernel__scalar_rr2_p6_x1, |
| 1027 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1028 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1029 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1030 | BENCHMARK_CAPTURE(f32_velu, scalar_p6_x2, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1031 | xnn_f32_velu_ukernel__scalar_rr2_p6_x2, |
| 1032 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1033 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1034 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1035 | BENCHMARK_CAPTURE(f32_velu, scalar_p6_x3, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1036 | xnn_f32_velu_ukernel__scalar_rr2_p6_x3, |
| 1037 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1038 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1039 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1040 | BENCHMARK_CAPTURE(f32_velu, scalar_p6_x4, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1041 | xnn_f32_velu_ukernel__scalar_rr2_p6_x4, |
| 1042 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1043 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1044 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1045 | BENCHMARK_CAPTURE(f32_velu, scalar_p6_x5, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1046 | xnn_f32_velu_ukernel__scalar_rr2_p6_x5, |
| 1047 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1048 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1049 | ->UseRealTime(); |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1050 | BENCHMARK_CAPTURE(f32_velu, scalar_p6_x6, |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 1051 | xnn_f32_velu_ukernel__scalar_rr2_p6_x6, |
| 1052 | xnn_init_f32_elu_scalar_rr2_p6_params) |
Marat Dukhan | 12e426c | 2021-08-06 22:09:44 -0700 | [diff] [blame] | 1053 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 1054 | ->UseRealTime(); |
| 1055 | |
| 1056 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 1057 | BENCHMARK_MAIN(); |
| 1058 | #endif |