XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <algorithm> |
| 7 | #include <cfloat> |
| 8 | #include <cmath> |
| 9 | #include <functional> |
| 10 | #include <random> |
| 11 | #include <vector> |
| 12 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 13 | #include <benchmark/benchmark.h> |
| 14 | #include "bench/dwconv.h" |
| 15 | #include "bench/utils.h" |
| 16 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 17 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 18 | #include <xnnpack/dwconv.h> |
| 19 | #include <xnnpack/indirection.h> |
| 20 | #include <xnnpack/operator.h> |
| 21 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 22 | #include <xnnpack/params-init.h> |
| 23 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | |
| 25 | |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 26 | static void DWConv2DBenchmark(benchmark::State& state, |
| 27 | xnn_f32_dwconv2d_chw_ukernel_function dwconv, |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 28 | uint32_t kh, uint32_t kw, uint32_t pw, uint32_t s, |
| 29 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 30 | { |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 31 | if (isa_check && !isa_check(state)) { |
| 32 | return; |
| 33 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 34 | |
| 35 | const size_t input_height = state.range(0); |
| 36 | const size_t input_width = state.range(1); |
| 37 | const size_t kernel_height = state.range(2); |
| 38 | const size_t kernel_width = state.range(3); |
| 39 | const size_t padding_height = state.range(4); |
| 40 | const size_t padding_width = state.range(5); |
| 41 | const size_t subsampling = state.range(6); |
| 42 | const size_t dilation = state.range(7); |
| 43 | const size_t channels = state.range(8); |
| 44 | |
| 45 | if (kernel_height != kh) { |
| 46 | state.SkipWithError("kernel height mismatch"); |
| 47 | return; |
| 48 | } |
| 49 | |
| 50 | if (kernel_width != kw) { |
| 51 | state.SkipWithError("kernel width mismatch"); |
| 52 | return; |
| 53 | } |
| 54 | |
| 55 | if (subsampling != s) { |
| 56 | state.SkipWithError("subsampling mismatch"); |
| 57 | return; |
| 58 | } |
| 59 | |
| 60 | if (padding_width % 2 != 0 || padding_width / 2 != pw) { |
| 61 | state.SkipWithError("padding width mismatch"); |
| 62 | return; |
| 63 | } |
| 64 | |
| 65 | if (dilation != 1) { |
| 66 | state.SkipWithError("unsupported dilation"); |
| 67 | return; |
| 68 | } |
| 69 | |
| 70 | std::random_device random_device; |
| 71 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 72 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 73 | |
| 74 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 75 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 76 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 77 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 78 | |
| 79 | const size_t inputSize = (input_height + padding_height) * input_width; |
| 80 | const size_t kernel_size = kernel_height * kernel_width; |
| 81 | const size_t output_size = output_height * output_width; |
| 82 | |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 83 | std::vector<float> input(inputSize * channels + 2 * XNN_EXTRA_BYTES); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 84 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 85 | std::vector<float> bias(channels); |
| 86 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 87 | std::vector<float> kernel(channels * kernel_size); |
| 88 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 89 | std::vector<float> zero(input_width + padding_width); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 90 | |
| 91 | const size_t w_elements = (kernel_size + 1) * channels; |
| 92 | const size_t o_elements = output_size * channels; |
| 93 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 94 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 95 | sizeof(float) * (w_elements + o_elements)); |
| 96 | |
Marat Dukhan | e13e639 | 2021-07-26 22:22:35 -0700 | [diff] [blame] | 97 | std::vector<float, AlignedAllocator<float, 64>> packed_weights(w_elements * num_buffers); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 98 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 99 | for (size_t c = 0; c < channels; c++) { |
| 100 | packed_weights[c * kernel_size + c] = bias[c]; |
| 101 | for (size_t i = 0; i < kernel_size; i++) { |
| 102 | packed_weights[c * kernel_size + c + 1 + i] = kernel[c * kernel_size + i]; |
| 103 | } |
| 104 | } |
| 105 | for (size_t n = 1; n < num_buffers; n++) { |
| 106 | std::copy(packed_weights.cbegin(), packed_weights.cbegin() + w_elements, packed_weights.begin() + n * w_elements); |
| 107 | } |
| 108 | |
| 109 | std::vector<float> output(o_elements * num_buffers); |
| 110 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 111 | |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame] | 112 | xnn_f32_chw_params chw_params; |
| 113 | xnn_init_f32_chw_params( |
| 114 | &chw_params, input_width, -std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 115 | |
| 116 | size_t buffer_index = 0; |
| 117 | for (auto _ : state) { |
| 118 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 119 | benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 120 | buffer_index = (buffer_index + 1) % num_buffers; |
| 121 | state.ResumeTiming(); |
| 122 | |
| 123 | for (uint32_t channel = 0; channel < channels; channel++) { |
| 124 | dwconv( |
Marat Dukhan | 7515777 | 2020-10-21 01:46:28 -0700 | [diff] [blame] | 125 | input_height, input_width * sizeof(float), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 126 | input.data() + channel * inputSize, |
| 127 | packed_weights.data() + channel * (kernel_size + 1) + buffer_index * w_elements, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 128 | zero.data(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 129 | output.data() + channel * output_size + buffer_index * o_elements, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 130 | padding_height / 2, // padding_top |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 131 | &chw_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 132 | } |
| 133 | } |
| 134 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 135 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 136 | if (cpu_frequency != 0) { |
| 137 | state.counters["cpufreq"] = cpu_frequency; |
| 138 | } |
| 139 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 140 | state.counters["FLOPS"] = benchmark::Counter( |
| 141 | uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size, |
| 142 | benchmark::Counter::kIsRate); |
| 143 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 144 | state.counters["bytes"] = benchmark::Counter( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 145 | uint64_t(state.iterations()) * (output_size + inputSize + kernel_size + 1 /* bias */) * channels * sizeof(float), |
| 146 | benchmark::Counter::kIsRate); |
| 147 | } |
| 148 | |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 149 | #if XNN_ARCH_ARM |
| 150 | static void dwconv2d_chw_3x3p1__neon_1x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 151 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 152 | } |
| 153 | static void dwconv2d_chw_3x3p1__neon_2x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 154 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 155 | } |
| 156 | static void dwconv2d_chw_3x3p1__neon_3x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 157 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 158 | } |
| 159 | static void dwconv2d_chw_3x3p1__neon_4x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 160 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 161 | } |
| 162 | static void dwconv2d_chw_3x3p1__neon_5x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 163 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 164 | } |
| 165 | static void dwconv2d_chw_3x3p1__neon_6x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 166 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 167 | } |
| 168 | static void dwconv2d_chw_3x3p1__neon_1x4_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 169 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 170 | } |
| 171 | static void dwconv2d_chw_3x3p1__neon_1x4_acc3(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 172 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 173 | } |
| 174 | static void dwconv2d_chw_3x3p1__neon_1x4_acc4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 175 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 176 | } |
| 177 | static void dwconv2d_chw_3x3p1__neon_2x4_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 178 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
| 179 | } |
| 180 | |
| 181 | static void dwconv2d_chw_3x3s2p1__neon_1x4(benchmark::State& state, const char* net) { |
| 182 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 183 | } |
| 184 | static void dwconv2d_chw_3x3s2p1__neon_2x4(benchmark::State& state, const char* net) { |
| 185 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 186 | } |
| 187 | static void dwconv2d_chw_3x3s2p1__neon_3x4(benchmark::State& state, const char* net) { |
| 188 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 189 | } |
| 190 | static void dwconv2d_chw_3x3s2p1__neon_4x4(benchmark::State& state, const char* net) { |
| 191 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 192 | } |
| 193 | static void dwconv2d_chw_3x3s2p1__neon_1x4_acc2(benchmark::State& state, const char* net) { |
| 194 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 195 | } |
| 196 | static void dwconv2d_chw_3x3s2p1__neon_1x4_acc3(benchmark::State& state, const char* net) { |
| 197 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 198 | } |
| 199 | static void dwconv2d_chw_3x3s2p1__neon_1x4_acc4(benchmark::State& state, const char* net) { |
| 200 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 201 | } |
| 202 | static void dwconv2d_chw_3x3s2p1__neon_2x4_acc2(benchmark::State& state, const char* net) { |
| 203 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 204 | } |
| 205 | |
| 206 | static void dwconv2d_chw_5x5p2__neon_1x4(benchmark::State& state, const char* net) { |
| 207 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 208 | } |
| 209 | static void dwconv2d_chw_5x5p2__neon_2x4(benchmark::State& state, const char* net) { |
| 210 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 211 | } |
| 212 | static void dwconv2d_chw_5x5p2__neon_3x4(benchmark::State& state, const char* net) { |
| 213 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 214 | } |
| 215 | static void dwconv2d_chw_5x5p2__neon_4x4(benchmark::State& state, const char* net) { |
| 216 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 217 | } |
| 218 | static void dwconv2d_chw_5x5p2__neon_5x4(benchmark::State& state, const char* net) { |
| 219 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 220 | } |
| 221 | static void dwconv2d_chw_5x5p2__neon_1x4_acc2(benchmark::State& state, const char* net) { |
| 222 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 223 | } |
| 224 | static void dwconv2d_chw_5x5p2__neon_1x4_acc3(benchmark::State& state, const char* net) { |
| 225 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 226 | } |
| 227 | static void dwconv2d_chw_5x5p2__neon_1x4_acc4(benchmark::State& state, const char* net) { |
| 228 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 229 | } |
| 230 | static void dwconv2d_chw_5x5p2__neon_1x4_acc5(benchmark::State& state, const char* net) { |
| 231 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 232 | } |
| 233 | static void dwconv2d_chw_5x5p2__neon_2x4_acc2(benchmark::State& state, const char* net) { |
| 234 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 235 | } |
| 236 | static void dwconv2d_chw_5x5p2__neon_2x4_acc3(benchmark::State& state, const char* net) { |
| 237 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 238 | } |
| 239 | static void dwconv2d_chw_5x5p2__neon_3x4_acc2(benchmark::State& state, const char* net) { |
| 240 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 241 | } |
| 242 | static void dwconv2d_chw_5x5p2__neon_4x4_acc2(benchmark::State& state, const char* net) { |
| 243 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 244 | } |
| 245 | |
| 246 | static void dwconv2d_chw_5x5s2p2__neon_1x4(benchmark::State& state, const char* net) { |
| 247 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 248 | } |
| 249 | static void dwconv2d_chw_5x5s2p2__neon_2x4(benchmark::State& state, const char* net) { |
| 250 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 251 | } |
| 252 | static void dwconv2d_chw_5x5s2p2__neon_3x4(benchmark::State& state, const char* net) { |
| 253 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 254 | } |
| 255 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc2(benchmark::State& state, const char* net) { |
| 256 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 257 | } |
| 258 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc3(benchmark::State& state, const char* net) { |
| 259 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 260 | } |
| 261 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc4(benchmark::State& state, const char* net) { |
| 262 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 263 | } |
| 264 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc5(benchmark::State& state, const char* net) { |
| 265 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 266 | } |
| 267 | static void dwconv2d_chw_5x5s2p2__neon_2x4_acc2(benchmark::State& state, const char* net) { |
| 268 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 269 | } |
| 270 | static void dwconv2d_chw_5x5s2p2__neon_2x4_acc3(benchmark::State& state, const char* net) { |
| 271 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 272 | } |
| 273 | static void dwconv2d_chw_5x5s2p2__neon_3x4_acc2(benchmark::State& state, const char* net) { |
| 274 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 275 | } |
| 276 | |
| 277 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4) |
| 278 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_2x4) |
| 279 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_3x4) |
| 280 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_4x4) |
| 281 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_5x4) |
| 282 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_6x4) |
| 283 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc2) |
| 284 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc3) |
| 285 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc4) |
| 286 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_2x4_acc2) |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 287 | |
| 288 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4) |
| 289 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_2x4) |
| 290 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_3x4) |
| 291 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_4x4) |
| 292 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4_acc2) |
| 293 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4_acc3) |
| 294 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4_acc4) |
| 295 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_2x4_acc2) |
| 296 | |
| 297 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4) |
| 298 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_2x4) |
| 299 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_3x4) |
| 300 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_4x4) |
| 301 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_5x4) |
| 302 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc2) |
| 303 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc3) |
| 304 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc4) |
| 305 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc5) |
| 306 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_2x4_acc2) |
| 307 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_2x4_acc3) |
| 308 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_3x4_acc2) |
| 309 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_4x4_acc2) |
| 310 | |
| 311 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4) |
| 312 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_2x4) |
| 313 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_3x4) |
| 314 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc2) |
| 315 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc3) |
| 316 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc4) |
| 317 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc5) |
| 318 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_2x4_acc2) |
| 319 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_2x4_acc3) |
| 320 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_3x4_acc2) |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 321 | #endif // XNN_ARCH_ARM |
| 322 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 323 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 324 | static void dwconv2d_chw_3x3p1__neonfma_1x4(benchmark::State& state, const char* net) { |
| 325 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4, 3, 3, 1, 1); |
| 326 | } |
| 327 | static void dwconv2d_chw_3x3p1__neonfma_2x4(benchmark::State& state, const char* net) { |
| 328 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4, 3, 3, 1, 1); |
| 329 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 330 | static void dwconv2d_chw_3x3p1__neonfma_3x4(benchmark::State& state, const char* net) { |
| 331 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4, 3, 3, 1, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 332 | } |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 333 | static void dwconv2d_chw_3x3p1__neonfma_4x4(benchmark::State& state, const char* net) { |
| 334 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4, 3, 3, 1, 1); |
| 335 | } |
| 336 | static void dwconv2d_chw_3x3p1__neonfma_5x4(benchmark::State& state, const char* net) { |
| 337 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4, 3, 3, 1, 1); |
| 338 | } |
| 339 | static void dwconv2d_chw_3x3p1__neonfma_6x4(benchmark::State& state, const char* net) { |
| 340 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4, 3, 3, 1, 1); |
| 341 | } |
| 342 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 343 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2, 3, 3, 1, 1); |
| 344 | } |
| 345 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 346 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3, 3, 3, 1, 1); |
| 347 | } |
| 348 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 349 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4, 3, 3, 1, 1); |
| 350 | } |
| 351 | static void dwconv2d_chw_3x3p1__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 352 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2, 3, 3, 1, 1); |
| 353 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 354 | |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 355 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4(benchmark::State& state, const char* net) { |
| 356 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4, 3, 3, 1, 2); |
| 357 | } |
| 358 | static void dwconv2d_chw_3x3s2p1__neonfma_2x4(benchmark::State& state, const char* net) { |
| 359 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4, 3, 3, 1, 2); |
| 360 | } |
| 361 | static void dwconv2d_chw_3x3s2p1__neonfma_3x4(benchmark::State& state, const char* net) { |
| 362 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4, 3, 3, 1, 2); |
| 363 | } |
| 364 | static void dwconv2d_chw_3x3s2p1__neonfma_4x4(benchmark::State& state, const char* net) { |
| 365 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4, 3, 3, 1, 2); |
| 366 | } |
| 367 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 368 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2, 3, 3, 1, 2); |
| 369 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 370 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 371 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3, 3, 3, 1, 2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 372 | } |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 373 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 374 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4, 3, 3, 1, 2); |
| 375 | } |
| 376 | static void dwconv2d_chw_3x3s2p1__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 377 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2, 3, 3, 1, 2); |
| 378 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 379 | |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 380 | static void dwconv2d_chw_5x5p2__neonfma_1x4(benchmark::State& state, const char* net) { |
| 381 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4, 5, 5, 2, 1); |
| 382 | } |
| 383 | static void dwconv2d_chw_5x5p2__neonfma_2x4(benchmark::State& state, const char* net) { |
| 384 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4, 5, 5, 2, 1); |
| 385 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 386 | static void dwconv2d_chw_5x5p2__neonfma_3x4(benchmark::State& state, const char* net) { |
| 387 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4, 5, 5, 2, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 388 | } |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 389 | static void dwconv2d_chw_5x5p2__neonfma_4x4(benchmark::State& state, const char* net) { |
| 390 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4, 5, 5, 2, 1); |
| 391 | } |
| 392 | static void dwconv2d_chw_5x5p2__neonfma_5x4(benchmark::State& state, const char* net) { |
| 393 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4, 5, 5, 2, 1); |
| 394 | } |
| 395 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 396 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2, 5, 5, 2, 1); |
| 397 | } |
| 398 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 399 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3, 5, 5, 2, 1); |
| 400 | } |
| 401 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 402 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4, 5, 5, 2, 1); |
| 403 | } |
| 404 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc5(benchmark::State& state, const char* net) { |
| 405 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5, 5, 5, 2, 1); |
| 406 | } |
| 407 | static void dwconv2d_chw_5x5p2__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 408 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2, 5, 5, 2, 1); |
| 409 | } |
| 410 | static void dwconv2d_chw_5x5p2__neonfma_2x4_acc3(benchmark::State& state, const char* net) { |
| 411 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3, 5, 5, 2, 1); |
| 412 | } |
| 413 | static void dwconv2d_chw_5x5p2__neonfma_3x4_acc2(benchmark::State& state, const char* net) { |
| 414 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2, 5, 5, 2, 1); |
| 415 | } |
| 416 | static void dwconv2d_chw_5x5p2__neonfma_4x4_acc2(benchmark::State& state, const char* net) { |
| 417 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2, 5, 5, 2, 1); |
| 418 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 419 | |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 420 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4(benchmark::State& state, const char* net) { |
| 421 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4, 5, 5, 2, 2); |
| 422 | } |
| 423 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4(benchmark::State& state, const char* net) { |
| 424 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4, 5, 5, 2, 2); |
| 425 | } |
| 426 | static void dwconv2d_chw_5x5s2p2__neonfma_3x4(benchmark::State& state, const char* net) { |
| 427 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4, 5, 5, 2, 2); |
| 428 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 429 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 430 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2, 5, 5, 2, 2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 431 | } |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 432 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 433 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3, 5, 5, 2, 2); |
| 434 | } |
| 435 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 436 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4, 5, 5, 2, 2); |
| 437 | } |
| 438 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc5(benchmark::State& state, const char* net) { |
| 439 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5, 5, 5, 2, 2); |
| 440 | } |
| 441 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 442 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2, 5, 5, 2, 2); |
| 443 | } |
| 444 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4_acc3(benchmark::State& state, const char* net) { |
| 445 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3, 5, 5, 2, 2); |
| 446 | } |
| 447 | static void dwconv2d_chw_5x5s2p2__neonfma_3x4_acc2(benchmark::State& state, const char* net) { |
| 448 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2, 5, 5, 2, 2); |
| 449 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 450 | |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 451 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4) |
| 452 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_2x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 453 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_3x4) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 454 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_4x4) |
| 455 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_5x4) |
| 456 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_6x4) |
| 457 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc2) |
| 458 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc3) |
| 459 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc4) |
| 460 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_2x4_acc2) |
| 461 | |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 462 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4) |
| 463 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_2x4) |
| 464 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_3x4) |
| 465 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_4x4) |
| 466 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 467 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc3) |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 468 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc4) |
| 469 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_2x4_acc2) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 470 | |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 471 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4) |
| 472 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 473 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_3x4) |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 474 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_4x4) |
| 475 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_5x4) |
| 476 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc2) |
| 477 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc3) |
| 478 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc4) |
| 479 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc5) |
| 480 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4_acc2) |
| 481 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4_acc3) |
| 482 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_3x4_acc2) |
| 483 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_4x4_acc2) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 484 | |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 485 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4) |
| 486 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4) |
| 487 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_3x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 488 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc2) |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 489 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc3) |
| 490 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc4) |
| 491 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc5) |
| 492 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4_acc2) |
| 493 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4_acc3) |
| 494 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_3x4_acc2) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 495 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 496 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 497 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 498 | static void dwconv2d_chw_3x3p1__sse_1x4(benchmark::State& state, const char* net) { |
| 499 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4, 3, 3, 1, 1); |
| 500 | } |
| 501 | static void dwconv2d_chw_3x3p1__sse_2x4(benchmark::State& state, const char* net) { |
| 502 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4, 3, 3, 1, 1); |
| 503 | } |
| 504 | static void dwconv2d_chw_3x3p1__sse_3x4(benchmark::State& state, const char* net) { |
| 505 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4, 3, 3, 1, 1); |
| 506 | } |
| 507 | static void dwconv2d_chw_3x3p1__sse_4x4(benchmark::State& state, const char* net) { |
| 508 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4, 3, 3, 1, 1); |
| 509 | } |
| 510 | static void dwconv2d_chw_3x3p1__sse_5x4(benchmark::State& state, const char* net) { |
| 511 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4, 3, 3, 1, 1); |
| 512 | } |
| 513 | static void dwconv2d_chw_3x3p1__sse_6x4(benchmark::State& state, const char* net) { |
| 514 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4, 3, 3, 1, 1); |
| 515 | } |
| 516 | static void dwconv2d_chw_3x3p1__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 517 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2, 3, 3, 1, 1); |
| 518 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 519 | static void dwconv2d_chw_3x3p1__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 520 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3, 3, 3, 1, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 521 | } |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 522 | static void dwconv2d_chw_3x3p1__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 523 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4, 3, 3, 1, 1); |
| 524 | } |
| 525 | static void dwconv2d_chw_3x3p1__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 526 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2, 3, 3, 1, 1); |
| 527 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 528 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 529 | static void dwconv2d_chw_3x3p1__ssse3_1x4(benchmark::State& state, const char* net) { |
| 530 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 531 | } |
| 532 | static void dwconv2d_chw_3x3p1__ssse3_2x4(benchmark::State& state, const char* net) { |
| 533 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 534 | } |
| 535 | static void dwconv2d_chw_3x3p1__ssse3_3x4(benchmark::State& state, const char* net) { |
| 536 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 537 | } |
| 538 | static void dwconv2d_chw_3x3p1__ssse3_4x4(benchmark::State& state, const char* net) { |
| 539 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 540 | } |
| 541 | static void dwconv2d_chw_3x3p1__ssse3_5x4(benchmark::State& state, const char* net) { |
| 542 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 543 | } |
| 544 | static void dwconv2d_chw_3x3p1__ssse3_6x4(benchmark::State& state, const char* net) { |
| 545 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 546 | } |
| 547 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc2(benchmark::State& state, const char* net) { |
| 548 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 549 | } |
| 550 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc3(benchmark::State& state, const char* net) { |
| 551 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 552 | } |
| 553 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc4(benchmark::State& state, const char* net) { |
| 554 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 555 | } |
| 556 | static void dwconv2d_chw_3x3p1__ssse3_2x4_acc2(benchmark::State& state, const char* net) { |
| 557 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 558 | } |
| 559 | |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 560 | static void dwconv2d_chw_3x3s2p1__sse_1x4(benchmark::State& state, const char* net) { |
| 561 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4, 3, 3, 1, 2); |
| 562 | } |
| 563 | static void dwconv2d_chw_3x3s2p1__sse_2x4(benchmark::State& state, const char* net) { |
| 564 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4, 3, 3, 1, 2); |
| 565 | } |
| 566 | static void dwconv2d_chw_3x3s2p1__sse_3x4(benchmark::State& state, const char* net) { |
| 567 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4, 3, 3, 1, 2); |
| 568 | } |
| 569 | static void dwconv2d_chw_3x3s2p1__sse_4x4(benchmark::State& state, const char* net) { |
| 570 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4, 3, 3, 1, 2); |
| 571 | } |
| 572 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 573 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2, 3, 3, 1, 2); |
| 574 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 575 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 576 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3, 3, 3, 1, 2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 577 | } |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 578 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 579 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4, 3, 3, 1, 2); |
| 580 | } |
| 581 | static void dwconv2d_chw_3x3s2p1__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 582 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2, 3, 3, 1, 2); |
| 583 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 584 | |
Marat Dukhan | d050389 | 2020-10-30 08:22:04 -0700 | [diff] [blame] | 585 | static void dwconv2d_chw_5x5p2__sse_1x4(benchmark::State& state, const char* net) { |
| 586 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4, 5, 5, 2, 1); |
| 587 | } |
| 588 | static void dwconv2d_chw_5x5p2__sse_2x4(benchmark::State& state, const char* net) { |
| 589 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4, 5, 5, 2, 1); |
| 590 | } |
| 591 | static void dwconv2d_chw_5x5p2__sse_3x4(benchmark::State& state, const char* net) { |
| 592 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4, 5, 5, 2, 1); |
| 593 | } |
| 594 | static void dwconv2d_chw_5x5p2__sse_4x4(benchmark::State& state, const char* net) { |
| 595 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4, 5, 5, 2, 1); |
| 596 | } |
| 597 | static void dwconv2d_chw_5x5p2__sse_5x4(benchmark::State& state, const char* net) { |
| 598 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4, 5, 5, 2, 1); |
| 599 | } |
| 600 | static void dwconv2d_chw_5x5p2__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 601 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2, 5, 5, 2, 1); |
| 602 | } |
| 603 | static void dwconv2d_chw_5x5p2__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 604 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3, 5, 5, 2, 1); |
| 605 | } |
| 606 | static void dwconv2d_chw_5x5p2__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 607 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4, 5, 5, 2, 1); |
| 608 | } |
| 609 | static void dwconv2d_chw_5x5p2__sse_1x4_acc5(benchmark::State& state, const char* net) { |
| 610 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5, 5, 5, 2, 1); |
| 611 | } |
| 612 | static void dwconv2d_chw_5x5p2__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 613 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2, 5, 5, 2, 1); |
| 614 | } |
| 615 | static void dwconv2d_chw_5x5p2__sse_2x4_acc3(benchmark::State& state, const char* net) { |
| 616 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3, 5, 5, 2, 1); |
| 617 | } |
| 618 | static void dwconv2d_chw_5x5p2__sse_3x4_acc2(benchmark::State& state, const char* net) { |
| 619 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2, 5, 5, 2, 1); |
| 620 | } |
| 621 | static void dwconv2d_chw_5x5p2__sse_4x4_acc2(benchmark::State& state, const char* net) { |
| 622 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2, 5, 5, 2, 1); |
| 623 | } |
| 624 | |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 625 | static void dwconv2d_chw_5x5s2p2__sse_1x4(benchmark::State& state, const char* net) { |
| 626 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4, 5, 5, 2, 2); |
| 627 | } |
| 628 | static void dwconv2d_chw_5x5s2p2__sse_2x4(benchmark::State& state, const char* net) { |
| 629 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4, 5, 5, 2, 2); |
| 630 | } |
| 631 | static void dwconv2d_chw_5x5s2p2__sse_3x4(benchmark::State& state, const char* net) { |
| 632 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4, 5, 5, 2, 2); |
| 633 | } |
| 634 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 635 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2, 5, 5, 2, 2); |
| 636 | } |
| 637 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 638 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3, 5, 5, 2, 2); |
| 639 | } |
| 640 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 641 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4, 5, 5, 2, 2); |
| 642 | } |
| 643 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc5(benchmark::State& state, const char* net) { |
| 644 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5, 5, 5, 2, 2); |
| 645 | } |
| 646 | static void dwconv2d_chw_5x5s2p2__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 647 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc2, 5, 5, 2, 2); |
| 648 | } |
| 649 | static void dwconv2d_chw_5x5s2p2__sse_2x4_acc3(benchmark::State& state, const char* net) { |
| 650 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc3, 5, 5, 2, 2); |
| 651 | } |
| 652 | static void dwconv2d_chw_5x5s2p2__sse_3x4_acc2(benchmark::State& state, const char* net) { |
| 653 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2, 5, 5, 2, 2); |
| 654 | } |
| 655 | |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 656 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4) |
| 657 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_2x4) |
| 658 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_3x4) |
| 659 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_4x4) |
| 660 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_5x4) |
| 661 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_6x4) |
| 662 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 663 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc3) |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 664 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc4) |
| 665 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_2x4_acc2) |
| 666 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 667 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4) |
| 668 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_2x4) |
| 669 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_3x4) |
| 670 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_4x4) |
| 671 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_5x4) |
| 672 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_6x4) |
| 673 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc2) |
| 674 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc3) |
| 675 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc4) |
| 676 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_2x4_acc2) |
| 677 | |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 678 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4) |
| 679 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_2x4) |
| 680 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_3x4) |
| 681 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_4x4) |
| 682 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 683 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc3) |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 684 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc4) |
| 685 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_2x4_acc2) |
Marat Dukhan | d050389 | 2020-10-30 08:22:04 -0700 | [diff] [blame] | 686 | |
| 687 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4) |
| 688 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4) |
| 689 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_3x4) |
| 690 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_4x4) |
| 691 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_5x4) |
| 692 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc2) |
| 693 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc3) |
| 694 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc4) |
| 695 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc5) |
| 696 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4_acc2) |
| 697 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4_acc3) |
| 698 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_3x4_acc2) |
| 699 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_4x4_acc2) |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 700 | |
| 701 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4) |
| 702 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4) |
| 703 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_3x4) |
| 704 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc2) |
| 705 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc3) |
| 706 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc4) |
| 707 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc5) |
| 708 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4_acc2) |
| 709 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4_acc3) |
| 710 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_3x4_acc2) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 711 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 712 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 713 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 714 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 715 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 716 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 717 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 718 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 719 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 720 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 721 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 722 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 723 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 724 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_4x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 725 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 726 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 727 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 728 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 729 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_6x4(benchmark::State& state, const char* net) { |
| 730 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 731 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 732 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 733 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 734 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 735 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 736 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3, 3, 3, 1, 1); |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 737 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 738 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 739 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4, 3, 3, 1, 1); |
Erich Elsen | fd7a6e3 | 2020-06-11 12:04:44 -0700 | [diff] [blame] | 740 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 741 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 742 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2, 3, 3, 1, 1); |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 743 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 744 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 745 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4, 3, 3, 1, 1); |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 746 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 747 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 748 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 749 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 750 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 751 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 752 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 753 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 754 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_4x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 755 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 756 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 757 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 758 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 759 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_6x4(benchmark::State& state, const char* net) { |
| 760 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 761 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 762 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 763 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc2, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 764 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 765 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 766 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc3, 3, 3, 1, 1); |
Frank Barchard | db5c32d | 2020-11-16 23:58:42 -0800 | [diff] [blame] | 767 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 768 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 769 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 770 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 771 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 772 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 773 | } |
| 774 | |
Frank Barchard | 02bb429 | 2020-12-15 18:25:32 -0800 | [diff] [blame] | 775 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 776 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4, 3, 3, 1, 1); |
| 777 | } |
| 778 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 779 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4, 3, 3, 1, 1); |
| 780 | } |
| 781 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 782 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4, 3, 3, 1, 1); |
| 783 | } |
| 784 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_4x4(benchmark::State& state, const char* net) { |
| 785 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4, 3, 3, 1, 1); |
| 786 | } |
| 787 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_5x4(benchmark::State& state, const char* net) { |
| 788 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4, 3, 3, 1, 1); |
| 789 | } |
| 790 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_6x4(benchmark::State& state, const char* net) { |
| 791 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4, 3, 3, 1, 1); |
| 792 | } |
| 793 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 794 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2, 3, 3, 1, 1); |
| 795 | } |
| 796 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 797 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3, 3, 3, 1, 1); |
| 798 | } |
| 799 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 800 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4, 3, 3, 1, 1); |
| 801 | } |
| 802 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 803 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2, 3, 3, 1, 1); |
| 804 | } |
| 805 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 806 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4, 3, 3, 1, 1); |
| 807 | } |
| 808 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 809 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4, 3, 3, 1, 1); |
| 810 | } |
| 811 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 812 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4, 3, 3, 1, 1); |
| 813 | } |
| 814 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_4x4(benchmark::State& state, const char* net) { |
| 815 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4, 3, 3, 1, 1); |
| 816 | } |
| 817 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_5x4(benchmark::State& state, const char* net) { |
| 818 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4, 3, 3, 1, 1); |
| 819 | } |
| 820 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_6x4(benchmark::State& state, const char* net) { |
| 821 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4, 3, 3, 1, 1); |
| 822 | } |
| 823 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 824 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2, 3, 3, 1, 1); |
| 825 | } |
| 826 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 827 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3, 3, 3, 1, 1); |
| 828 | } |
| 829 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 830 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4, 3, 3, 1, 1); |
| 831 | } |
| 832 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 833 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2, 3, 3, 1, 1); |
| 834 | } |
| 835 | |
Frank Barchard | c5704bf | 2020-12-21 23:09:00 -0800 | [diff] [blame] | 836 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 837 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4, 3, 3, 1, 2); |
| 838 | } |
| 839 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 840 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4, 3, 3, 1, 2); |
| 841 | } |
| 842 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 843 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4, 3, 3, 1, 2); |
| 844 | } |
| 845 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 846 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4, 3, 3, 1, 2); |
| 847 | } |
| 848 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 849 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2, 3, 3, 1, 2); |
| 850 | } |
| 851 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 852 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3, 3, 3, 1, 2); |
| 853 | } |
| 854 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 855 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4, 3, 3, 1, 2); |
| 856 | } |
| 857 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 858 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2, 3, 3, 1, 2); |
| 859 | } |
| 860 | |
| 861 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 862 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4, 3, 3, 1, 2); |
| 863 | } |
| 864 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 865 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4, 3, 3, 1, 2); |
| 866 | } |
| 867 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 868 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4, 3, 3, 1, 2); |
| 869 | } |
| 870 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 871 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4, 3, 3, 1, 2); |
| 872 | } |
| 873 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 874 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2, 3, 3, 1, 2); |
| 875 | } |
| 876 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 877 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3, 3, 3, 1, 2); |
| 878 | } |
| 879 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 880 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4, 3, 3, 1, 2); |
| 881 | } |
| 882 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 883 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2, 3, 3, 1, 2); |
| 884 | } |
| 885 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 886 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 887 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 888 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 889 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 890 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 891 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 892 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 893 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 894 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 895 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_4x4(benchmark::State& state, const char* net) { |
| 896 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 897 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 898 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 899 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 900 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 901 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 902 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 903 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 904 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 905 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 906 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 907 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 908 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 909 | } |
| 910 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 911 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 912 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 913 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 914 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 915 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 916 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 917 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 918 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 919 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 920 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_4x4(benchmark::State& state, const char* net) { |
| 921 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 922 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 923 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 924 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 925 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 926 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 927 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 928 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 929 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 930 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 931 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 932 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 933 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 934 | } |
| 935 | |
Frank Barchard | b20dcd6 | 2020-12-15 16:46:14 -0800 | [diff] [blame] | 936 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 937 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4, 5, 5, 2, 1); |
| 938 | } |
| 939 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 940 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4, 5, 5, 2, 1); |
| 941 | } |
| 942 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 943 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4, 5, 5, 2, 1); |
| 944 | } |
| 945 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 946 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4, 5, 5, 2, 1); |
| 947 | } |
| 948 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 949 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4, 5, 5, 2, 1); |
| 950 | } |
| 951 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 952 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2, 5, 5, 2, 1); |
| 953 | } |
| 954 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 955 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3, 5, 5, 2, 1); |
| 956 | } |
| 957 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 958 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4, 5, 5, 2, 1); |
| 959 | } |
| 960 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 961 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5, 5, 5, 2, 1); |
| 962 | } |
| 963 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 964 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2, 5, 5, 2, 1); |
| 965 | } |
| 966 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 967 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3, 5, 5, 2, 1); |
| 968 | } |
| 969 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 970 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2, 5, 5, 2, 1); |
| 971 | } |
| 972 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2(benchmark::State& state, const char* net) { |
| 973 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2, 5, 5, 2, 1); |
| 974 | } |
| 975 | |
| 976 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 977 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4, 5, 5, 2, 1); |
| 978 | } |
| 979 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 980 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4, 5, 5, 2, 1); |
| 981 | } |
| 982 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 983 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4, 5, 5, 2, 1); |
| 984 | } |
| 985 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 986 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4, 5, 5, 2, 1); |
| 987 | } |
| 988 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 989 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4, 5, 5, 2, 1); |
| 990 | } |
| 991 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 992 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2, 5, 5, 2, 1); |
| 993 | } |
| 994 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 995 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3, 5, 5, 2, 1); |
| 996 | } |
| 997 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 998 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4, 5, 5, 2, 1); |
| 999 | } |
| 1000 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1001 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5, 5, 5, 2, 1); |
| 1002 | } |
| 1003 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1004 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2, 5, 5, 2, 1); |
| 1005 | } |
| 1006 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1007 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc3, 5, 5, 2, 1); |
| 1008 | } |
| 1009 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1010 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2, 5, 5, 2, 1); |
| 1011 | } |
| 1012 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2(benchmark::State& state, const char* net) { |
| 1013 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2, 5, 5, 2, 1); |
| 1014 | } |
| 1015 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1016 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 1017 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1018 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1019 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 1020 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4, 5, 5, 2, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1021 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1022 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 1023 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4, 5, 5, 2, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1024 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1025 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4(benchmark::State& state, const char* net) { |
| 1026 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1027 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1028 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_5x4(benchmark::State& state, const char* net) { |
| 1029 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1030 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1031 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1032 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1033 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1034 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1035 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1036 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1037 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1038 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1039 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1040 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1041 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1042 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1043 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1044 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1045 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1046 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1047 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1048 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1049 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1050 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1051 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1052 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4_acc2(benchmark::State& state, const char* net) { |
| 1053 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1054 | } |
| 1055 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1056 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 1057 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1058 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1059 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 1060 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1061 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1062 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 1063 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1064 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1065 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4(benchmark::State& state, const char* net) { |
| 1066 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1067 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1068 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_5x4(benchmark::State& state, const char* net) { |
| 1069 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1070 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1071 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1072 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1073 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1074 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1075 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1076 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1077 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1078 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1079 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1080 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1081 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1082 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1083 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1084 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1085 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1086 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1087 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1088 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1089 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1090 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1091 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1092 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4_acc2(benchmark::State& state, const char* net) { |
| 1093 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1094 | } |
| 1095 | |
Frank Barchard | c6889b3 | 2020-12-21 11:27:22 -0800 | [diff] [blame] | 1096 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 1097 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4, 5, 5, 2, 2); |
| 1098 | } |
| 1099 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 1100 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4, 5, 5, 2, 2); |
| 1101 | } |
| 1102 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 1103 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4, 5, 5, 2, 2); |
| 1104 | } |
| 1105 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1106 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2, 5, 5, 2, 2); |
| 1107 | } |
| 1108 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1109 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3, 5, 5, 2, 2); |
| 1110 | } |
| 1111 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1112 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4, 5, 5, 2, 2); |
| 1113 | } |
| 1114 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1115 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5, 5, 5, 2, 2); |
| 1116 | } |
| 1117 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1118 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2, 5, 5, 2, 2); |
| 1119 | } |
| 1120 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1121 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3, 5, 5, 2, 2); |
| 1122 | } |
| 1123 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1124 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2, 5, 5, 2, 2); |
| 1125 | } |
| 1126 | |
| 1127 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 1128 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4, 5, 5, 2, 2); |
| 1129 | } |
| 1130 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 1131 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4, 5, 5, 2, 2); |
| 1132 | } |
| 1133 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 1134 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4, 5, 5, 2, 2); |
| 1135 | } |
| 1136 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1137 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2, 5, 5, 2, 2); |
| 1138 | } |
| 1139 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1140 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3, 5, 5, 2, 2); |
| 1141 | } |
| 1142 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1143 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4, 5, 5, 2, 2); |
| 1144 | } |
| 1145 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1146 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5, 5, 5, 2, 2); |
| 1147 | } |
| 1148 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1149 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2, 5, 5, 2, 2); |
| 1150 | } |
| 1151 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1152 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3, 5, 5, 2, 2); |
| 1153 | } |
| 1154 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1155 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2, 5, 5, 2, 2); |
| 1156 | } |
| 1157 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1158 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 1159 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1160 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1161 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 1162 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_2x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1163 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1164 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 1165 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1166 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1167 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1168 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2, 5, 5, 2, 2); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1169 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1170 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1171 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1172 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1173 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1174 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1175 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1176 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1177 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1178 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1179 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1180 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_2x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1181 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1182 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1183 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_2x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1184 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1185 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1186 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1187 | } |
| 1188 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1189 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 1190 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1191 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1192 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 1193 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1194 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1195 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 1196 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1197 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1198 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1199 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2, 5, 5, 2, 2); |
Frank Barchard | db5c32d | 2020-11-16 23:58:42 -0800 | [diff] [blame] | 1200 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1201 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1202 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1203 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1204 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1205 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1206 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1207 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1208 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1209 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1210 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1211 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1212 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1213 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1214 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1215 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1216 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1217 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1218 | } |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 1219 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1220 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4) |
| 1221 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4) |
| 1222 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_3x4) |
| 1223 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_4x4) |
| 1224 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_5x4) |
| 1225 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_6x4) |
| 1226 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1227 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1228 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1229 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2) |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1230 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1231 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4) |
| 1232 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4) |
| 1233 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_3x4) |
| 1234 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_4x4) |
| 1235 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_5x4) |
| 1236 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_6x4) |
| 1237 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1238 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1239 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1240 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2) |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1241 | |
Frank Barchard | 02bb429 | 2020-12-15 18:25:32 -0800 | [diff] [blame] | 1242 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4) |
| 1243 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4) |
| 1244 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_3x4) |
| 1245 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_4x4) |
| 1246 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_5x4) |
| 1247 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_6x4) |
| 1248 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc2) |
| 1249 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc3) |
| 1250 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc4) |
| 1251 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4_acc2) |
| 1252 | |
| 1253 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4) |
| 1254 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4) |
| 1255 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_3x4) |
| 1256 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_4x4) |
| 1257 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_5x4) |
| 1258 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_6x4) |
| 1259 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc2) |
| 1260 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc3) |
| 1261 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc4) |
| 1262 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4_acc2) |
| 1263 | |
Frank Barchard | c5704bf | 2020-12-21 23:09:00 -0800 | [diff] [blame] | 1264 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4) |
| 1265 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4) |
| 1266 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_3x4) |
| 1267 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_4x4) |
| 1268 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1269 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1270 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1271 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2) |
| 1272 | |
| 1273 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4) |
| 1274 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4) |
| 1275 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_3x4) |
| 1276 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_4x4) |
| 1277 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1278 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1279 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1280 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2) |
| 1281 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1282 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4) |
| 1283 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4) |
| 1284 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_3x4) |
| 1285 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_4x4) |
| 1286 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc2) |
| 1287 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc3) |
| 1288 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc4) |
| 1289 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4_acc2) |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 1290 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1291 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4) |
| 1292 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4) |
| 1293 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_3x4) |
| 1294 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_4x4) |
| 1295 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc2) |
| 1296 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc3) |
| 1297 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc4) |
| 1298 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4_acc2) |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 1299 | |
Frank Barchard | b20dcd6 | 2020-12-15 16:46:14 -0800 | [diff] [blame] | 1300 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4) |
| 1301 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4) |
| 1302 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4) |
| 1303 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4) |
| 1304 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_5x4) |
| 1305 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1306 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1307 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1308 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5) |
| 1309 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2) |
| 1310 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3) |
| 1311 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2) |
| 1312 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2) |
| 1313 | |
| 1314 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4) |
| 1315 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4) |
| 1316 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4) |
| 1317 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4) |
| 1318 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_5x4) |
| 1319 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1320 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1321 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1322 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5) |
| 1323 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2) |
| 1324 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc3) |
| 1325 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2) |
| 1326 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2) |
| 1327 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1328 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4) |
| 1329 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4) |
| 1330 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4) |
| 1331 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4) |
| 1332 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_5x4) |
| 1333 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc2) |
| 1334 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc3) |
| 1335 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc4) |
| 1336 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc5) |
| 1337 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc2) |
| 1338 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc3) |
| 1339 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4_acc2) |
| 1340 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4_acc2) |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1341 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1342 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4) |
| 1343 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4) |
| 1344 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4) |
| 1345 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4) |
| 1346 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_5x4) |
| 1347 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc2) |
| 1348 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc3) |
| 1349 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc4) |
| 1350 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc5) |
| 1351 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc2) |
| 1352 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc3) |
| 1353 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4_acc2) |
| 1354 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4_acc2) |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1355 | |
Frank Barchard | c6889b3 | 2020-12-21 11:27:22 -0800 | [diff] [blame] | 1356 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4) |
| 1357 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4) |
| 1358 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4) |
| 1359 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1360 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1361 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1362 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5) |
| 1363 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2) |
| 1364 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3) |
| 1365 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2) |
| 1366 | |
| 1367 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4) |
| 1368 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4) |
| 1369 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4) |
| 1370 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1371 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1372 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1373 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5) |
| 1374 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2) |
| 1375 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3) |
| 1376 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2) |
| 1377 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1378 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4) |
| 1379 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4) |
| 1380 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4) |
| 1381 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc2) |
| 1382 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc3) |
| 1383 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc4) |
| 1384 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc5) |
| 1385 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc2) |
| 1386 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc3) |
| 1387 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4_acc2) |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1388 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1389 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4) |
| 1390 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4) |
| 1391 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4) |
| 1392 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc2) |
| 1393 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc3) |
| 1394 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc4) |
| 1395 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc5) |
| 1396 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc2) |
| 1397 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc3) |
| 1398 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4_acc2) |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1399 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 1400 | |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1401 | static void dwconv2d_chw_3x3p1__scalar_1x1(benchmark::State& state, const char* net) { |
| 1402 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1, 3, 3, 1, 1); |
| 1403 | } |
| 1404 | static void dwconv2d_chw_3x3p1__scalar_2x1(benchmark::State& state, const char* net) { |
| 1405 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1, 3, 3, 1, 1); |
| 1406 | } |
| 1407 | static void dwconv2d_chw_3x3p1__scalar_3x1(benchmark::State& state, const char* net) { |
| 1408 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1, 3, 3, 1, 1); |
| 1409 | } |
| 1410 | static void dwconv2d_chw_3x3p1__scalar_4x1(benchmark::State& state, const char* net) { |
| 1411 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1, 3, 3, 1, 1); |
| 1412 | } |
| 1413 | static void dwconv2d_chw_3x3p1__scalar_5x1(benchmark::State& state, const char* net) { |
| 1414 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1, 3, 3, 1, 1); |
| 1415 | } |
| 1416 | static void dwconv2d_chw_3x3p1__scalar_6x1(benchmark::State& state, const char* net) { |
| 1417 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1, 3, 3, 1, 1); |
| 1418 | } |
| 1419 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1420 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2, 3, 3, 1, 1); |
| 1421 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1422 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1423 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3, 3, 3, 1, 1); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 1424 | } |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1425 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1426 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4, 3, 3, 1, 1); |
| 1427 | } |
| 1428 | static void dwconv2d_chw_3x3p1__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1429 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2, 3, 3, 1, 1); |
| 1430 | } |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1431 | |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1432 | static void dwconv2d_chw_3x3s2p1__scalar_1x1(benchmark::State& state, const char* net) { |
| 1433 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1, 3, 3, 1, 2); |
| 1434 | } |
| 1435 | static void dwconv2d_chw_3x3s2p1__scalar_2x1(benchmark::State& state, const char* net) { |
| 1436 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1, 3, 3, 1, 2); |
| 1437 | } |
| 1438 | static void dwconv2d_chw_3x3s2p1__scalar_3x1(benchmark::State& state, const char* net) { |
| 1439 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1, 3, 3, 1, 2); |
| 1440 | } |
| 1441 | static void dwconv2d_chw_3x3s2p1__scalar_4x1(benchmark::State& state, const char* net) { |
| 1442 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1, 3, 3, 1, 2); |
| 1443 | } |
| 1444 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1445 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2, 3, 3, 1, 2); |
| 1446 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1447 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1448 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3, 3, 3, 1, 2); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 1449 | } |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1450 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1451 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4, 3, 3, 1, 2); |
| 1452 | } |
| 1453 | static void dwconv2d_chw_3x3s2p1__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1454 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2, 3, 3, 1, 2); |
| 1455 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 1456 | |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1457 | static void dwconv2d_chw_5x5p2__scalar_1x1(benchmark::State& state, const char* net) { |
| 1458 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1, 5, 5, 2, 1); |
| 1459 | } |
| 1460 | static void dwconv2d_chw_5x5p2__scalar_2x1(benchmark::State& state, const char* net) { |
| 1461 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1, 5, 5, 2, 1); |
| 1462 | } |
| 1463 | static void dwconv2d_chw_5x5p2__scalar_3x1(benchmark::State& state, const char* net) { |
| 1464 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1, 5, 5, 2, 1); |
| 1465 | } |
| 1466 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1467 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2, 5, 5, 2, 1); |
| 1468 | } |
| 1469 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1470 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3, 5, 5, 2, 1); |
| 1471 | } |
| 1472 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1473 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4, 5, 5, 2, 1); |
| 1474 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1475 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc5(benchmark::State& state, const char* net) { |
| 1476 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5, 5, 5, 2, 1); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 1477 | } |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1478 | static void dwconv2d_chw_5x5p2__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1479 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2, 5, 5, 2, 1); |
| 1480 | } |
| 1481 | static void dwconv2d_chw_5x5p2__scalar_2x1_acc3(benchmark::State& state, const char* net) { |
| 1482 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3, 5, 5, 2, 1); |
| 1483 | } |
| 1484 | static void dwconv2d_chw_5x5p2__scalar_3x1_acc2(benchmark::State& state, const char* net) { |
| 1485 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2, 5, 5, 2, 1); |
| 1486 | } |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1487 | |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1488 | static void dwconv2d_chw_5x5s2p2__scalar_1x1(benchmark::State& state, const char* net) { |
| 1489 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1, 5, 5, 2, 2); |
| 1490 | } |
| 1491 | static void dwconv2d_chw_5x5s2p2__scalar_2x1(benchmark::State& state, const char* net) { |
| 1492 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1, 5, 5, 2, 2); |
| 1493 | } |
| 1494 | static void dwconv2d_chw_5x5s2p2__scalar_3x1(benchmark::State& state, const char* net) { |
| 1495 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1, 5, 5, 2, 2); |
| 1496 | } |
| 1497 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1498 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2, 5, 5, 2, 2); |
| 1499 | } |
| 1500 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1501 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3, 5, 5, 2, 2); |
| 1502 | } |
| 1503 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1504 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4, 5, 5, 2, 2); |
| 1505 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1506 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc5(benchmark::State& state, const char* net) { |
| 1507 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5, 5, 5, 2, 2); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 1508 | } |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1509 | static void dwconv2d_chw_5x5s2p2__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1510 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2, 5, 5, 2, 2); |
| 1511 | } |
| 1512 | static void dwconv2d_chw_5x5s2p2__scalar_2x1_acc3(benchmark::State& state, const char* net) { |
| 1513 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3, 5, 5, 2, 2); |
| 1514 | } |
| 1515 | static void dwconv2d_chw_5x5s2p2__scalar_3x1_acc2(benchmark::State& state, const char* net) { |
| 1516 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2, 5, 5, 2, 2); |
| 1517 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 1518 | |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1519 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1) |
| 1520 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_2x1) |
| 1521 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_3x1) |
| 1522 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_4x1) |
| 1523 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_5x1) |
| 1524 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_6x1) |
| 1525 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1526 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc3) |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1527 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc4) |
| 1528 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_2x1_acc2) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1529 | |
| 1530 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1) |
| 1531 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_2x1) |
| 1532 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_3x1) |
| 1533 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_4x1) |
| 1534 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1535 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc3) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1536 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc4) |
| 1537 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_2x1_acc2) |
| 1538 | |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1539 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1) |
| 1540 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1) |
| 1541 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_3x1) |
| 1542 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc2) |
| 1543 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc3) |
| 1544 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1545 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc5) |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1546 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1_acc2) |
| 1547 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1_acc3) |
| 1548 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_3x1_acc2) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1549 | |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1550 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1) |
| 1551 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1) |
| 1552 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_3x1) |
| 1553 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc2) |
| 1554 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc3) |
| 1555 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1556 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc5) |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1557 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1_acc2) |
| 1558 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1_acc3) |
| 1559 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_3x1_acc2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1560 | |
| 1561 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 1562 | BENCHMARK_MAIN(); |
| 1563 | #endif |