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 | |
| 13 | #include <cpuinfo.h> |
| 14 | |
| 15 | #include <benchmark/benchmark.h> |
| 16 | #include "bench/dwconv.h" |
| 17 | #include "bench/utils.h" |
| 18 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 19 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | #include <xnnpack/dwconv.h> |
| 21 | #include <xnnpack/indirection.h> |
| 22 | #include <xnnpack/operator.h> |
| 23 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 24 | #include <xnnpack/params-init.h> |
| 25 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | |
| 27 | |
| 28 | static void DWConvCHWBenchmark(benchmark::State& state, |
| 29 | xnn_f32_dwconv_spchw_ukernel_function dwconv, |
| 30 | uint32_t it, uint32_t ot, uint32_t kh, uint32_t kw, uint32_t pw, uint32_t s) |
| 31 | { |
| 32 | if (!cpuinfo_initialize()) { |
| 33 | state.SkipWithError("cpuinfo initialization failed"); |
| 34 | return; |
| 35 | } |
| 36 | |
| 37 | const size_t input_height = state.range(0); |
| 38 | const size_t input_width = state.range(1); |
| 39 | const size_t kernel_height = state.range(2); |
| 40 | const size_t kernel_width = state.range(3); |
| 41 | const size_t padding_height = state.range(4); |
| 42 | const size_t padding_width = state.range(5); |
| 43 | const size_t subsampling = state.range(6); |
| 44 | const size_t dilation = state.range(7); |
| 45 | const size_t channels = state.range(8); |
| 46 | |
| 47 | if (kernel_height != kh) { |
| 48 | state.SkipWithError("kernel height mismatch"); |
| 49 | return; |
| 50 | } |
| 51 | |
| 52 | if (kernel_width != kw) { |
| 53 | state.SkipWithError("kernel width mismatch"); |
| 54 | return; |
| 55 | } |
| 56 | |
| 57 | if (subsampling != s) { |
| 58 | state.SkipWithError("subsampling mismatch"); |
| 59 | return; |
| 60 | } |
| 61 | |
| 62 | if (padding_width % 2 != 0 || padding_width / 2 != pw) { |
| 63 | state.SkipWithError("padding width mismatch"); |
| 64 | return; |
| 65 | } |
| 66 | |
| 67 | if (dilation != 1) { |
| 68 | state.SkipWithError("unsupported dilation"); |
| 69 | return; |
| 70 | } |
| 71 | |
| 72 | std::random_device random_device; |
| 73 | auto rng = std::mt19937(random_device()); |
| 74 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 75 | |
| 76 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 77 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 78 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 79 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 80 | |
| 81 | const size_t inputSize = (input_height + padding_height) * input_width; |
| 82 | const size_t kernel_size = kernel_height * kernel_width; |
| 83 | const size_t output_size = output_height * output_width; |
| 84 | |
| 85 | std::vector<float> input(inputSize * channels + 2 * it); |
| 86 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 87 | std::vector<float> bias(channels); |
| 88 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 89 | std::vector<float> kernel(channels * kernel_size); |
| 90 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 91 | |
| 92 | const size_t w_elements = (kernel_size + 1) * channels; |
| 93 | const size_t o_elements = output_size * channels; |
| 94 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 95 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 96 | sizeof(float) * (w_elements + o_elements)); |
| 97 | |
| 98 | std::vector<float, AlignedAllocator<float, 32>> packed_weights(w_elements * num_buffers); |
| 99 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 100 | for (size_t c = 0; c < channels; c++) { |
| 101 | packed_weights[c * kernel_size + c] = bias[c]; |
| 102 | for (size_t i = 0; i < kernel_size; i++) { |
| 103 | packed_weights[c * kernel_size + c + 1 + i] = kernel[c * kernel_size + i]; |
| 104 | } |
| 105 | } |
| 106 | for (size_t n = 1; n < num_buffers; n++) { |
| 107 | std::copy(packed_weights.cbegin(), packed_weights.cbegin() + w_elements, packed_weights.begin() + n * w_elements); |
| 108 | } |
| 109 | |
| 110 | std::vector<float> output(o_elements * num_buffers); |
| 111 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 112 | |
| 113 | xnn_f32_spchw_params output_params = |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 114 | xnn_init_f32_spchw_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( |
| 125 | output_height, input_width, |
| 126 | input.data() + channel * inputSize, |
| 127 | packed_weights.data() + channel * (kernel_size + 1) + buffer_index * w_elements, |
| 128 | output.data() + channel * output_size + buffer_index * o_elements, |
| 129 | it * sizeof(float), ot * sizeof(float), |
| 130 | input_width * sizeof(float), output_width * sizeof(float), |
| 131 | &output_params); |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 136 | state.counters["FLOPS"] = benchmark::Counter( |
| 137 | uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size, |
| 138 | benchmark::Counter::kIsRate); |
| 139 | |
| 140 | state.counters["BYTES"] = benchmark::Counter( |
| 141 | uint64_t(state.iterations()) * (output_size + inputSize + kernel_size + 1 /* bias */) * channels * sizeof(float), |
| 142 | benchmark::Counter::kIsRate); |
| 143 | } |
| 144 | |
| 145 | static void DWConvHWoTCTBenchmark(benchmark::State& state, |
| 146 | xnn_f32_dwconv_spchw_ukernel_function dwconv, |
| 147 | uint32_t it, uint32_t ot, uint32_t kh, uint32_t kw, uint32_t pw, uint32_t s) |
| 148 | { |
| 149 | if (!cpuinfo_initialize()) { |
| 150 | state.SkipWithError("cpuinfo initialization failed"); |
| 151 | return; |
| 152 | } |
| 153 | |
| 154 | const size_t input_height = state.range(0); |
| 155 | const size_t input_width = state.range(1); |
| 156 | const size_t kernel_height = state.range(2); |
| 157 | const size_t kernel_width = state.range(3); |
| 158 | const size_t padding_height = state.range(4); |
| 159 | const size_t padding_width = state.range(5); |
| 160 | const size_t subsampling = state.range(6); |
| 161 | const size_t dilation = state.range(7); |
| 162 | const size_t channels = state.range(8); |
| 163 | |
| 164 | if (kernel_height != kh) { |
| 165 | state.SkipWithError("kernel height mismatch"); |
| 166 | return; |
| 167 | } |
| 168 | |
| 169 | if (kernel_width != kw) { |
| 170 | state.SkipWithError("kernel width mismatch"); |
| 171 | return; |
| 172 | } |
| 173 | |
| 174 | if (subsampling != s) { |
| 175 | state.SkipWithError("subsampling mismatch"); |
| 176 | return; |
| 177 | } |
| 178 | |
| 179 | if (padding_width % 2 != 0 || padding_width / 2 != pw) { |
| 180 | state.SkipWithError("padding width mismatch"); |
| 181 | return; |
| 182 | } |
| 183 | |
| 184 | if (dilation != 1) { |
| 185 | state.SkipWithError("unsupported dilation"); |
| 186 | return; |
| 187 | } |
| 188 | |
| 189 | std::random_device random_device; |
| 190 | auto rng = std::mt19937(random_device()); |
| 191 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 192 | |
| 193 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 194 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 195 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 196 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 197 | |
| 198 | const size_t inputSize = (input_height + padding_height) * input_width; |
| 199 | const size_t kernel_size = kernel_height * kernel_width; |
| 200 | const size_t output_size = output_height * output_width; |
| 201 | |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 202 | std::vector<float> input(input_height * benchmark::utils::RoundUp<size_t>(input_width, it) * channels); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 203 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 204 | std::vector<float> bias(channels); |
| 205 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 206 | std::vector<float> kernel(channels * kernel_size); |
| 207 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 208 | |
| 209 | const size_t w_elements = (kernel_size + 1) * channels; |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 210 | const size_t o_elements = output_height * benchmark::utils::RoundUp<size_t>(output_width, ot) * channels; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 211 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 212 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 213 | sizeof(float) * (w_elements + o_elements)); |
| 214 | |
| 215 | std::vector<float, AlignedAllocator<float, 32>> packed_weights(w_elements * num_buffers); |
| 216 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 217 | for (size_t c = 0; c < channels; c++) { |
| 218 | packed_weights[c * kernel_size + c] = bias[c]; |
| 219 | for (size_t i = 0; i < kernel_size; i++) { |
| 220 | packed_weights[c * kernel_size + c + 1 + i] = kernel[c * kernel_size + i]; |
| 221 | } |
| 222 | } |
| 223 | for (size_t n = 1; n < num_buffers; n++) { |
| 224 | std::copy(packed_weights.cbegin(), packed_weights.cbegin() + w_elements, packed_weights.begin() + n * w_elements); |
| 225 | } |
| 226 | |
| 227 | std::vector<float> output(o_elements * num_buffers); |
| 228 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 229 | |
| 230 | xnn_f32_spchw_params output_params = |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 231 | xnn_init_f32_spchw_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] | 232 | |
| 233 | size_t buffer_index = 0; |
| 234 | for (auto _ : state) { |
| 235 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 236 | benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 237 | buffer_index = (buffer_index + 1) % num_buffers; |
| 238 | state.ResumeTiming(); |
| 239 | |
| 240 | for (uint32_t channel = 0; channel < channels; channel++) { |
| 241 | dwconv( |
| 242 | output_height, input_width, |
| 243 | input.data() + channel * it, |
| 244 | packed_weights.data() + channel * (kernel_size + 1) + buffer_index * w_elements, |
| 245 | output.data() + channel * ot + buffer_index * o_elements, |
| 246 | it * channels * sizeof(float), ot * channels * sizeof(float), |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 247 | benchmark::utils::RoundUp<size_t>(input_width, it) * channels * sizeof(float), |
| 248 | benchmark::utils::RoundUp<size_t>(output_width, ot) * channels * sizeof(float), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 249 | &output_params); |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 254 | state.counters["FLOPS"] = benchmark::Counter( |
| 255 | uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size, |
| 256 | benchmark::Counter::kIsRate); |
| 257 | |
| 258 | state.counters["BYTES"] = benchmark::Counter( |
| 259 | uint64_t(state.iterations()) * (output_size + inputSize + kernel_size + 1 /* bias */) * channels * sizeof(float), |
| 260 | benchmark::Counter::kIsRate); |
| 261 | } |
| 262 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 263 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 264 | static void CHW_3x3p1__neonfma(benchmark::State& state, const char* net) { |
| 265 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3p1__neonfma, 4, 4, 3, 3, 1, 1); |
| 266 | } |
| 267 | |
| 268 | static void CHW_5x5p2__neonfma(benchmark::State& state, const char* net) { |
| 269 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma, 4, 4, 5, 5, 2, 1); |
| 270 | } |
| 271 | |
| 272 | static void CHW_3x3s2p1__neonfma(benchmark::State& state, const char* net) { |
| 273 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma, 4, 4, 3, 3, 1, 2); |
| 274 | } |
| 275 | |
| 276 | static void CHW_5x5s2p2__neonfma(benchmark::State& state, const char* net) { |
| 277 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma, 4, 4, 5, 5, 2, 2); |
| 278 | } |
| 279 | |
| 280 | static void HWo4C4_3x3p1__neonfma(benchmark::State& state, const char* net) { |
| 281 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3p1__neonfma, 4, 4, 3, 3, 1, 1); |
| 282 | } |
| 283 | |
| 284 | static void HWo4C4_5x5p2__neonfma(benchmark::State& state, const char* net) { |
| 285 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma, 4, 4, 5, 5, 2, 1); |
| 286 | } |
| 287 | |
| 288 | static void HWo4C4_3x3s2p1__neonfma(benchmark::State& state, const char* net) { |
| 289 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma, 4, 4, 3, 3, 1, 2); |
| 290 | } |
| 291 | |
| 292 | static void HWo4C4_5x5s2p2__neonfma(benchmark::State& state, const char* net) { |
| 293 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma, 4, 4, 5, 5, 2, 2); |
| 294 | } |
| 295 | |
| 296 | BENCHMARK_DWCONV(CHW_3x3p1__neonfma) |
| 297 | BENCHMARK_DWCONV(CHW_5x5p2__neonfma) |
| 298 | BENCHMARK_DWCONV(CHW_3x3s2p1__neonfma) |
| 299 | BENCHMARK_DWCONV(CHW_5x5s2p2__neonfma) |
| 300 | BENCHMARK_DWCONV(HWo4C4_3x3p1__neonfma) |
| 301 | BENCHMARK_DWCONV(HWo4C4_5x5p2__neonfma) |
| 302 | BENCHMARK_DWCONV(HWo4C4_3x3s2p1__neonfma) |
| 303 | BENCHMARK_DWCONV(HWo4C4_5x5s2p2__neonfma) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 304 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 305 | |
| 306 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 307 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 308 | static void CHW_3x3p1__sse(benchmark::State& state, const char* net) { |
| 309 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3p1__sse, 4, 4, 3, 3, 1, 1); |
| 310 | } |
| 311 | |
| 312 | static void CHW_3x3s2p1__sse(benchmark::State& state, const char* net) { |
| 313 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3s2p1__sse, 4, 4, 3, 3, 1, 2); |
| 314 | } |
| 315 | |
| 316 | static void HWo4C4_3x3p1__sse(benchmark::State& state, const char* net) { |
| 317 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3p1__sse, 4, 4, 3, 3, 1, 1); |
| 318 | } |
| 319 | |
| 320 | static void HWo4C4_3x3s2p1__sse(benchmark::State& state, const char* net) { |
| 321 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3s2p1__sse, 4, 4, 3, 3, 1, 2); |
| 322 | } |
| 323 | |
| 324 | BENCHMARK_DWCONV(CHW_3x3p1__sse) |
| 325 | BENCHMARK_DWCONV(CHW_3x3s2p1__sse) |
| 326 | BENCHMARK_DWCONV(HWo4C4_3x3p1__sse) |
| 327 | BENCHMARK_DWCONV(HWo4C4_3x3s2p1__sse) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 328 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 329 | |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 330 | static void CHW_3x3p1__scalar(benchmark::State& state, const char* net) { |
| 331 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar, 1, 1, 3, 3, 1, 1); |
| 332 | } |
| 333 | |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 334 | static void CHW_5x5p2__scalar(benchmark::State& state, const char* net) { |
| 335 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar, 1, 1, 5, 5, 2, 1); |
| 336 | } |
| 337 | |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 338 | static void CHW_3x3s2p1__scalar(benchmark::State& state, const char* net) { |
| 339 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar, 1, 1, 3, 3, 1, 2); |
| 340 | } |
| 341 | |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 342 | static void CHW_5x5s2p2__scalar(benchmark::State& state, const char* net) { |
| 343 | DWConvCHWBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar, 1, 1, 5, 5, 2, 2); |
| 344 | } |
| 345 | |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 346 | static void HWC_3x3p1__scalar(benchmark::State& state, const char* net) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 347 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar, 1, 1, 3, 3, 1, 1); |
| 348 | } |
| 349 | |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 350 | static void HWC_5x5p2__scalar(benchmark::State& state, const char* net) { |
| 351 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar, 1, 1, 5, 5, 2, 1); |
| 352 | } |
| 353 | |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 354 | static void HWC_3x3s2p1__scalar(benchmark::State& state, const char* net) { |
| 355 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar, 1, 1, 3, 3, 1, 2); |
| 356 | } |
| 357 | |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 358 | static void HWC_5x5s2p2__scalar(benchmark::State& state, const char* net) { |
| 359 | DWConvHWoTCTBenchmark(state, xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar, 1, 1, 5, 5, 2, 2); |
| 360 | } |
| 361 | |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 362 | |
| 363 | BENCHMARK_DWCONV(CHW_3x3p1__scalar) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 364 | BENCHMARK_DWCONV(CHW_5x5p2__scalar) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 365 | BENCHMARK_DWCONV(CHW_3x3s2p1__scalar) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 366 | BENCHMARK_DWCONV(CHW_5x5s2p2__scalar) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 367 | BENCHMARK_DWCONV(HWC_3x3p1__scalar) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 368 | BENCHMARK_DWCONV(HWC_5x5p2__scalar) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 369 | BENCHMARK_DWCONV(HWC_3x3s2p1__scalar) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 370 | BENCHMARK_DWCONV(HWC_5x5s2p2__scalar) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 371 | |
| 372 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 373 | BENCHMARK_MAIN(); |
| 374 | #endif |