XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #include <algorithm> |
| 10 | #include <cfloat> |
| 11 | #include <cmath> |
| 12 | #include <functional> |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 13 | #include <limits> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <random> |
| 15 | #include <vector> |
| 16 | |
| 17 | #include <xnnpack.h> |
| 18 | |
| 19 | #include <benchmark/benchmark.h> |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 20 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 21 | #include "flatbuffers/include/flatbuffers/flatbuffers.h" |
| 22 | #include "tensorflow/lite/interpreter.h" |
| 23 | #include "tensorflow/lite/kernels/register.h" |
| 24 | #include "tensorflow/lite/model.h" |
| 25 | #include "tensorflow/lite/schema/schema_generated.h" |
| 26 | #include "tensorflow/lite/version.h" |
| 27 | #endif // BENCHMARK_TENSORFLOW_LITE |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 28 | #include "bench/utils.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | |
Chao Mei | c664027 | 2020-07-23 09:35:11 -0700 | [diff] [blame] | 30 | #ifndef XNN_NO_QU8_OPERATORS |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 31 | static void xnnpack_average_pooling_qu8(benchmark::State& state, const char* net) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 32 | const size_t batch_size = state.range(0); |
| 33 | const size_t input_height = state.range(1); |
| 34 | const size_t input_width = state.range(2); |
| 35 | const size_t pooling_size = state.range(3); |
| 36 | const size_t padding_size = state.range(4); |
| 37 | const size_t stride = state.range(5); |
| 38 | const size_t channels = state.range(6); |
| 39 | |
| 40 | std::random_device random_device; |
| 41 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame^] | 42 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 43 | |
| 44 | const size_t output_height = (2 * padding_size + input_height - pooling_size) / stride + 1; |
| 45 | const size_t output_width = (2 * padding_size + input_width - pooling_size) / stride + 1; |
| 46 | |
| 47 | std::vector<uint8_t> input(batch_size * input_height * input_width * channels + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 48 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 49 | std::vector<uint8_t> output(batch_size * output_height * output_width * channels); |
| 50 | std::fill(output.begin(), output.end(), 0xA5); |
| 51 | |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 52 | xnn_status status = xnn_initialize(nullptr /* allocator */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 53 | if (status != xnn_status_success) { |
| 54 | state.SkipWithError("failed to initialize XNNPACK"); |
| 55 | return; |
| 56 | } |
| 57 | |
| 58 | xnn_operator_t pooling_op = nullptr; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 59 | status = xnn_create_average_pooling2d_nhwc_qu8( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 60 | padding_size, padding_size, padding_size, padding_size, |
| 61 | pooling_size, pooling_size, |
| 62 | stride, stride, |
| 63 | channels, channels /* input pixel stride */, channels /* output pixel stride */, |
| 64 | 127 /* input zero point */, 0.75f /* input scale */, |
| 65 | 127 /* output zero point */, 1.25f /* output scale */, |
| 66 | 0, 255, |
| 67 | 0 /* flags */, &pooling_op); |
| 68 | if (status != xnn_status_success) { |
| 69 | state.SkipWithError("failed to create Average Pooling operator"); |
| 70 | return; |
| 71 | } |
| 72 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 73 | status = xnn_setup_average_pooling2d_nhwc_qu8( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 74 | pooling_op, |
| 75 | batch_size, input_height, input_width, |
| 76 | input.data(), output.data(), |
| 77 | nullptr /* thread pool */); |
| 78 | if (status != xnn_status_success) { |
| 79 | state.SkipWithError("failed to setup Average Pooling operator"); |
| 80 | return; |
| 81 | } |
| 82 | |
| 83 | for (auto _ : state) { |
| 84 | status = xnn_run_operator(pooling_op, nullptr /* thread pool */); |
| 85 | if (status != xnn_status_success) { |
| 86 | state.SkipWithError("failed to run Average Pooling operator"); |
| 87 | return; |
| 88 | } |
| 89 | } |
| 90 | |
| 91 | status = xnn_delete_operator(pooling_op); |
| 92 | if (status != xnn_status_success) { |
| 93 | state.SkipWithError("failed to delete Average Pooling operator"); |
| 94 | return; |
| 95 | } |
| 96 | pooling_op = nullptr; |
| 97 | |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 98 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 99 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 100 | state.counters["bytes"] = benchmark::Counter( |
| 101 | uint64_t(state.iterations()) * |
| 102 | batch_size * (input_height * input_width + output_height * output_width) * channels * sizeof(uint8_t), |
| 103 | benchmark::Counter::kIsRate); |
| 104 | } |
Chao Mei | c664027 | 2020-07-23 09:35:11 -0700 | [diff] [blame] | 105 | #endif // XNN_NO_QU8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 106 | |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 107 | static void xnnpack_average_pooling_f32(benchmark::State& state, const char* net) { |
| 108 | const size_t batch_size = state.range(0); |
| 109 | const size_t input_height = state.range(1); |
| 110 | const size_t input_width = state.range(2); |
| 111 | const size_t pooling_size = state.range(3); |
| 112 | const size_t padding_size = state.range(4); |
| 113 | const size_t stride = state.range(5); |
| 114 | const size_t channels = state.range(6); |
| 115 | |
| 116 | std::random_device random_device; |
| 117 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame^] | 118 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng)); |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 119 | |
| 120 | const size_t output_height = (2 * padding_size + input_height - pooling_size) / stride + 1; |
| 121 | const size_t output_width = (2 * padding_size + input_width - pooling_size) / stride + 1; |
| 122 | |
| 123 | std::vector<float> input(batch_size * input_height * input_width * channels + XNN_EXTRA_BYTES / sizeof(float)); |
| 124 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 125 | std::vector<float> output(batch_size * output_height * output_width * channels); |
| 126 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 127 | |
| 128 | xnn_status status = xnn_initialize(nullptr /* allocator */); |
| 129 | if (status != xnn_status_success) { |
| 130 | state.SkipWithError("failed to initialize XNNPACK"); |
| 131 | return; |
| 132 | } |
| 133 | |
| 134 | xnn_operator_t pooling_op = nullptr; |
| 135 | status = xnn_create_average_pooling2d_nhwc_f32( |
| 136 | padding_size, padding_size, padding_size, padding_size, |
| 137 | pooling_size, pooling_size, |
| 138 | stride, stride, |
| 139 | channels, channels /* input pixel stride */, channels /* output pixel stride */, |
| 140 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 141 | 0 /* flags */, &pooling_op); |
| 142 | if (status != xnn_status_success) { |
| 143 | state.SkipWithError("failed to create Average Pooling operator"); |
| 144 | return; |
| 145 | } |
| 146 | |
| 147 | status = xnn_setup_average_pooling2d_nhwc_f32( |
| 148 | pooling_op, |
| 149 | batch_size, input_height, input_width, |
| 150 | input.data(), output.data(), |
| 151 | nullptr /* thread pool */); |
| 152 | if (status != xnn_status_success) { |
| 153 | state.SkipWithError("failed to setup Average Pooling operator"); |
| 154 | return; |
| 155 | } |
| 156 | |
| 157 | for (auto _ : state) { |
| 158 | status = xnn_run_operator(pooling_op, nullptr /* thread pool */); |
| 159 | if (status != xnn_status_success) { |
| 160 | state.SkipWithError("failed to run Average Pooling operator"); |
| 161 | return; |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | status = xnn_delete_operator(pooling_op); |
| 166 | if (status != xnn_status_success) { |
| 167 | state.SkipWithError("failed to delete Average Pooling operator"); |
| 168 | return; |
| 169 | } |
| 170 | pooling_op = nullptr; |
| 171 | |
| 172 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 173 | |
| 174 | state.counters["bytes"] = benchmark::Counter( |
| 175 | uint64_t(state.iterations()) * |
| 176 | batch_size * (input_height * input_width + output_height * output_width) * channels * sizeof(float), |
| 177 | benchmark::Counter::kIsRate); |
| 178 | } |
| 179 | |
| 180 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 181 | void tflite_average_pooling_f32(benchmark::State& state, const char* net) { |
| 182 | const size_t batch_size = state.range(0); |
| 183 | const size_t input_height = state.range(1); |
| 184 | const size_t input_width = state.range(2); |
| 185 | const size_t pooling_size = state.range(3); |
| 186 | const size_t padding_size = state.range(4); |
| 187 | const size_t stride = state.range(5); |
| 188 | const size_t channels = state.range(6); |
| 189 | |
| 190 | std::random_device random_device; |
| 191 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame^] | 192 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng)); |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 193 | |
| 194 | tflite::Padding padding = tflite::Padding_VALID; |
| 195 | if (2 * padding_size == (pooling_size - 1)) { |
| 196 | padding = tflite::Padding_SAME; |
| 197 | } else if (padding_size == 0) { |
| 198 | padding = tflite::Padding_VALID; |
| 199 | } else { |
| 200 | state.SkipWithError("unsupported padding"); |
| 201 | return; |
| 202 | } |
| 203 | |
| 204 | const size_t output_height = (2 * padding_size + input_height - pooling_size) / stride + 1; |
| 205 | const size_t output_width = (2 * padding_size + input_width - pooling_size) / stride + 1; |
| 206 | |
| 207 | std::vector<float> input(batch_size * input_height * input_width * channels + XNN_EXTRA_BYTES / sizeof(float)); |
| 208 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 209 | std::vector<float> output(batch_size * output_height * output_width * channels); |
| 210 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 211 | |
| 212 | flatbuffers::FlatBufferBuilder builder; |
| 213 | flatbuffers::Offset<tflite::OperatorCode> operator_code = |
| 214 | CreateOperatorCode(builder, tflite::BuiltinOperator_AVERAGE_POOL_2D); |
| 215 | |
| 216 | flatbuffers::Offset<tflite::Pool2DOptions> pool2d_options = CreatePool2DOptions( |
| 217 | builder, padding, |
| 218 | stride /* stride_w */, stride /* stride_h */, |
| 219 | pooling_size /* filter_width */, pooling_size /* filter_height */, |
| 220 | tflite::ActivationFunctionType_NONE); |
| 221 | |
| 222 | flatbuffers::Offset<tflite::Buffer> buffers[1] = { |
| 223 | tflite::CreateBuffer(builder, builder.CreateVector({})), |
| 224 | }; |
| 225 | |
| 226 | const int32_t input_shape[4] = { |
| 227 | static_cast<int32_t>(batch_size), |
| 228 | static_cast<int32_t>(input_height), |
| 229 | static_cast<int32_t>(input_width), |
| 230 | static_cast<int32_t>(channels) |
| 231 | }; |
| 232 | const int32_t output_shape[4] = { |
| 233 | static_cast<int32_t>(batch_size), |
| 234 | static_cast<int32_t>(output_height), |
| 235 | static_cast<int32_t>(output_width), |
| 236 | static_cast<int32_t>(channels) |
| 237 | }; |
| 238 | |
| 239 | flatbuffers::Offset<tflite::Tensor> tensors[2] = { |
| 240 | tflite::CreateTensor(builder, |
| 241 | builder.CreateVector<int32_t>(input_shape, 4), |
| 242 | tflite::TensorType_FLOAT32), |
| 243 | tflite::CreateTensor(builder, |
| 244 | builder.CreateVector<int32_t>(output_shape, 4), |
| 245 | tflite::TensorType_FLOAT32), |
| 246 | }; |
| 247 | |
| 248 | const int32_t op_inputs[1] = { 0 }; |
| 249 | const int32_t op_outputs[1] = { 1 }; |
| 250 | flatbuffers::Offset<tflite::Operator> op = CreateOperator( |
| 251 | builder, |
| 252 | 0 /* opcode_index */, |
| 253 | builder.CreateVector<int32_t>(op_inputs, 1), |
| 254 | builder.CreateVector<int32_t>(op_outputs, 1), |
| 255 | tflite::BuiltinOptions_Pool2DOptions, |
| 256 | pool2d_options.Union()); |
| 257 | |
| 258 | const int32_t graph_inputs[1] = { 0 }; |
| 259 | const int32_t graph_outputs[1] = { 1 }; |
| 260 | flatbuffers::Offset<tflite::SubGraph> subgraph = CreateSubGraph( |
| 261 | builder, |
| 262 | builder.CreateVector(tensors, 2), |
| 263 | builder.CreateVector<int32_t>(graph_inputs, 1), |
| 264 | builder.CreateVector<int32_t>(graph_outputs, 1), |
| 265 | builder.CreateVector(&op, 1)); |
| 266 | |
| 267 | flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder, |
| 268 | TFLITE_SCHEMA_VERSION, |
| 269 | builder.CreateVector(&operator_code, 1), |
| 270 | builder.CreateVector(&subgraph, 1), |
| 271 | builder.CreateString("AVERAGE_POOL_2D model"), |
| 272 | builder.CreateVector(buffers, 1)); |
| 273 | |
| 274 | builder.Finish(model_buffer); |
| 275 | |
| 276 | const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer()); |
| 277 | tflite::ops::builtin::BuiltinOpResolver resolver; |
| 278 | tflite::InterpreterBuilder interpreterBuilder(model, resolver); |
| 279 | std::unique_ptr<tflite::Interpreter> interpreter; |
| 280 | if (interpreterBuilder(&interpreter) != kTfLiteOk) { |
| 281 | state.SkipWithError("failed to create TFLite interpreter"); |
| 282 | return; |
| 283 | } |
| 284 | if (interpreter == nullptr) { |
| 285 | state.SkipWithError("TFLite interpreter is null"); |
| 286 | return; |
| 287 | } |
| 288 | interpreter->SetNumThreads(1); |
| 289 | |
| 290 | if (interpreter->AllocateTensors() != kTfLiteOk) { |
| 291 | state.SkipWithError("failed to allocate tensors"); |
| 292 | return; |
| 293 | } |
| 294 | |
| 295 | std::generate( |
| 296 | interpreter->typed_tensor<float>(0), |
| 297 | interpreter->typed_tensor<float>(0) + batch_size * input_height * input_width * channels, |
| 298 | std::ref(f32rng)); |
| 299 | |
| 300 | for (auto _ : state) { |
| 301 | if (interpreter->Invoke() != kTfLiteOk) { |
| 302 | state.SkipWithError("failed to invoke TFLite interpreter"); |
| 303 | return; |
| 304 | } |
| 305 | } |
| 306 | |
| 307 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 308 | |
| 309 | state.counters["bytes"] = benchmark::Counter( |
| 310 | uint64_t(state.iterations()) * |
| 311 | batch_size * (input_height * input_width + output_height * output_width) * channels * sizeof(float), |
| 312 | benchmark::Counter::kIsRate); |
| 313 | } |
| 314 | #endif // BENCHMARK_TENSORFLOW_LITE |
| 315 | |
| 316 | // Final global average pooling in ImageNet classification models. |
| 317 | static void ImageNet(benchmark::internal::Benchmark* b) { |
| 318 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 319 | |
| 320 | /* N H W K P S C */ |
| 321 | b->Args({1, 13, 13, 13, 0, 1, 1000}); |
| 322 | b->Args({1, 7, 7, 7, 0, 1, 1000}); |
| 323 | } |
| 324 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 325 | // ShuffleNet v1 with 1 group. |
| 326 | static void ShuffleNetV1G1(benchmark::internal::Benchmark* b) { |
| 327 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 328 | |
| 329 | /* N H W K P S C */ |
| 330 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 331 | b->Args({1, 28, 28, 3, 1, 2, 144}); |
| 332 | b->Args({1, 14, 14, 3, 1, 2, 288}); |
| 333 | b->Args({1, 7, 7, 3, 1, 2, 576}); |
| 334 | } |
| 335 | |
| 336 | // ShuffleNet v1 with 2 groups. |
| 337 | static void ShuffleNetV1G2(benchmark::internal::Benchmark* b) { |
| 338 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 339 | |
| 340 | /* N H W K P S C */ |
| 341 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 342 | b->Args({1, 28, 28, 3, 1, 2, 200}); |
| 343 | b->Args({1, 14, 14, 3, 1, 2, 400}); |
| 344 | b->Args({1, 7, 7, 3, 1, 2, 800}); |
| 345 | } |
| 346 | |
| 347 | // ShuffleNet v1 with 3 groups. |
| 348 | static void ShuffleNetV1G3(benchmark::internal::Benchmark* b) { |
| 349 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 350 | |
| 351 | /* N H W K P S C */ |
| 352 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 353 | b->Args({1, 28, 28, 3, 1, 2, 240}); |
| 354 | b->Args({1, 14, 14, 3, 1, 2, 480}); |
| 355 | b->Args({1, 7, 7, 3, 1, 2, 960}); |
| 356 | } |
| 357 | |
| 358 | // ShuffleNet v1 with 4 groups. |
| 359 | static void ShuffleNetV1G4(benchmark::internal::Benchmark* b) { |
| 360 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 361 | |
| 362 | /* N H W K P S C */ |
| 363 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 364 | b->Args({1, 28, 28, 3, 1, 2, 272}); |
| 365 | b->Args({1, 14, 14, 3, 1, 2, 576}); |
| 366 | b->Args({1, 7, 7, 3, 1, 2, 1088}); |
| 367 | } |
| 368 | |
| 369 | // ShuffleNet v1 with 8 groups. |
| 370 | static void ShuffleNetV1G8(benchmark::internal::Benchmark* b) { |
| 371 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 372 | |
| 373 | /* N H W K P S C */ |
| 374 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 375 | b->Args({1, 28, 28, 3, 1, 2, 384}); |
| 376 | b->Args({1, 14, 14, 3, 1, 2, 768}); |
| 377 | b->Args({1, 7, 7, 3, 1, 2, 1536}); |
| 378 | } |
| 379 | |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 380 | BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, imagenet, "ImageNet")->Apply(ImageNet)->UseRealTime(); |
| 381 | BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime(); |
| 382 | BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime(); |
| 383 | BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime(); |
| 384 | BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime(); |
| 385 | BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime(); |
| 386 | |
| 387 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 388 | BENCHMARK_CAPTURE(tflite_average_pooling_f32, imagenet, "ImageNet")->Apply(ImageNet)->UseRealTime(); |
| 389 | BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime(); |
| 390 | BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime(); |
| 391 | BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime(); |
| 392 | BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime(); |
| 393 | BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime(); |
| 394 | #endif // BENCHMARK_TENSORFLOW_LITE |
| 395 | |
Chao Mei | c664027 | 2020-07-23 09:35:11 -0700 | [diff] [blame] | 396 | #ifndef XNN_NO_QU8_OPERATORS |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 397 | BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, imagenet, "ImageNet")->Apply(ImageNet)->UseRealTime(); |
| 398 | BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime(); |
| 399 | BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime(); |
| 400 | BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime(); |
| 401 | BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime(); |
| 402 | BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime(); |
Chao Mei | c664027 | 2020-07-23 09:35:11 -0700 | [diff] [blame] | 403 | #endif // XNN_NO_QU8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 404 | |
| 405 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 406 | BENCHMARK_MAIN(); |
| 407 | #endif |