File size: 16,757 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 |
// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <algorithm>
#include <cfloat>
#include <cmath>
#include <functional>
#include <limits>
#include <memory>
#include <random>
#include <vector>
#include <xnnpack.h>
#include <benchmark/benchmark.h>
#ifdef BENCHMARK_TENSORFLOW_LITE
#include "flatbuffers/include/flatbuffers/flatbuffers.h"
#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/schema/schema_generated.h"
#include "tensorflow/lite/version.h"
#endif // BENCHMARK_TENSORFLOW_LITE
#include "bench/utils.h"
static void xnnpack_average_pooling_qu8(benchmark::State& state, const char* net) {
const size_t batch_size = state.range(0);
const size_t input_height = state.range(1);
const size_t input_width = state.range(2);
const size_t pooling_size = state.range(3);
const size_t padding_size = state.range(4);
const size_t stride = state.range(5);
const size_t channels = state.range(6);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng));
const size_t output_height = (2 * padding_size + input_height - pooling_size) / stride + 1;
const size_t output_width = (2 * padding_size + input_width - pooling_size) / stride + 1;
std::vector<uint8_t> input(batch_size * input_height * input_width * channels + XNN_EXTRA_BYTES / sizeof(uint8_t));
std::generate(input.begin(), input.end(), std::ref(u8rng));
std::vector<uint8_t> output(batch_size * output_height * output_width * channels);
std::fill(output.begin(), output.end(), 0xA5);
xnn_status status = xnn_initialize(nullptr /* allocator */);
if (status != xnn_status_success) {
state.SkipWithError("failed to initialize XNNPACK");
return;
}
xnn_operator_t pooling_op = nullptr;
status = xnn_create_average_pooling2d_nhwc_qu8(
padding_size, padding_size, padding_size, padding_size,
pooling_size, pooling_size,
stride, stride,
channels, channels /* input pixel stride */, channels /* output pixel stride */,
127 /* input zero point */, 0.75f /* input scale */,
127 /* output zero point */, 1.25f /* output scale */,
0, 255,
0 /* flags */, &pooling_op);
if (status != xnn_status_success) {
state.SkipWithError("failed to create Average Pooling operator");
return;
}
status = xnn_reshape_average_pooling2d_nhwc_qu8(
pooling_op,
batch_size, input_height, input_width,
/*output_height_out=*/nullptr, /*output_width_out=*/nullptr,
nullptr /* thread pool */);
if (status != xnn_status_success) {
state.SkipWithError("failed to reshape Average Pooling operator");
return;
}
status = xnn_setup_average_pooling2d_nhwc_qu8(
pooling_op,
input.data(), output.data());
if (status != xnn_status_success) {
state.SkipWithError("failed to setup Average Pooling operator");
return;
}
for (auto _ : state) {
status = xnn_run_operator(pooling_op, nullptr /* thread pool */);
if (status != xnn_status_success) {
state.SkipWithError("failed to run Average Pooling operator");
return;
}
}
status = xnn_delete_operator(pooling_op);
if (status != xnn_status_success) {
state.SkipWithError("failed to delete Average Pooling operator");
return;
}
pooling_op = nullptr;
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
state.counters["bytes"] = benchmark::Counter(
uint64_t(state.iterations()) *
batch_size * (input_height * input_width + output_height * output_width) * channels * sizeof(uint8_t),
benchmark::Counter::kIsRate);
}
static void xnnpack_average_pooling_f32(benchmark::State& state, const char* net) {
const size_t batch_size = state.range(0);
const size_t input_height = state.range(1);
const size_t input_width = state.range(2);
const size_t pooling_size = state.range(3);
const size_t padding_size = state.range(4);
const size_t stride = state.range(5);
const size_t channels = state.range(6);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng));
const size_t output_height = (2 * padding_size + input_height - pooling_size) / stride + 1;
const size_t output_width = (2 * padding_size + input_width - pooling_size) / stride + 1;
std::vector<float> input(batch_size * input_height * input_width * channels + XNN_EXTRA_BYTES / sizeof(float));
std::generate(input.begin(), input.end(), std::ref(f32rng));
std::vector<float> output(batch_size * output_height * output_width * channels);
std::fill(output.begin(), output.end(), std::nanf(""));
xnn_status status = xnn_initialize(nullptr /* allocator */);
if (status != xnn_status_success) {
state.SkipWithError("failed to initialize XNNPACK");
return;
}
xnn_operator_t pooling_op = nullptr;
status = xnn_create_average_pooling2d_nhwc_f32(
padding_size, padding_size, padding_size, padding_size,
pooling_size, pooling_size,
stride, stride,
channels, channels /* input pixel stride */, channels /* output pixel stride */,
-std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(),
0 /* flags */, &pooling_op);
if (status != xnn_status_success) {
state.SkipWithError("failed to create Average Pooling operator");
return;
}
status = xnn_reshape_average_pooling2d_nhwc_f32(
pooling_op,
batch_size, input_height, input_width,
/*output_height_out=*/nullptr, /*output_width_out=*/nullptr,
nullptr /* thread pool */);
if (status != xnn_status_success) {
state.SkipWithError("failed to reshape Average Pooling operator");
return;
}
status = xnn_setup_average_pooling2d_nhwc_f32(
pooling_op,
input.data(), output.data());
if (status != xnn_status_success) {
state.SkipWithError("failed to setup Average Pooling operator");
return;
}
for (auto _ : state) {
status = xnn_run_operator(pooling_op, nullptr /* thread pool */);
if (status != xnn_status_success) {
state.SkipWithError("failed to run Average Pooling operator");
return;
}
}
status = xnn_delete_operator(pooling_op);
if (status != xnn_status_success) {
state.SkipWithError("failed to delete Average Pooling operator");
return;
}
pooling_op = nullptr;
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
state.counters["bytes"] = benchmark::Counter(
uint64_t(state.iterations()) *
batch_size * (input_height * input_width + output_height * output_width) * channels * sizeof(float),
benchmark::Counter::kIsRate);
}
#ifdef BENCHMARK_TENSORFLOW_LITE
void tflite_average_pooling_f32(benchmark::State& state, const char* net) {
const size_t batch_size = state.range(0);
const size_t input_height = state.range(1);
const size_t input_width = state.range(2);
const size_t pooling_size = state.range(3);
const size_t padding_size = state.range(4);
const size_t stride = state.range(5);
const size_t channels = state.range(6);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng));
tflite::Padding padding = tflite::Padding_VALID;
if (2 * padding_size == (pooling_size - 1)) {
padding = tflite::Padding_SAME;
} else if (padding_size == 0) {
padding = tflite::Padding_VALID;
} else {
state.SkipWithError("unsupported padding");
return;
}
const size_t output_height = (2 * padding_size + input_height - pooling_size) / stride + 1;
const size_t output_width = (2 * padding_size + input_width - pooling_size) / stride + 1;
std::vector<float> input(batch_size * input_height * input_width * channels + XNN_EXTRA_BYTES / sizeof(float));
std::generate(input.begin(), input.end(), std::ref(f32rng));
std::vector<float> output(batch_size * output_height * output_width * channels);
std::fill(output.begin(), output.end(), std::nanf(""));
flatbuffers::FlatBufferBuilder builder;
flatbuffers::Offset<tflite::OperatorCode> operator_code =
CreateOperatorCode(builder, tflite::BuiltinOperator_AVERAGE_POOL_2D);
flatbuffers::Offset<tflite::Pool2DOptions> pool2d_options = CreatePool2DOptions(
builder, padding,
stride /* stride_w */, stride /* stride_h */,
pooling_size /* filter_width */, pooling_size /* filter_height */,
tflite::ActivationFunctionType_NONE);
flatbuffers::Offset<tflite::Buffer> buffers[1] = {
tflite::CreateBuffer(builder, builder.CreateVector({})),
};
const int32_t input_shape[4] = {
static_cast<int32_t>(batch_size),
static_cast<int32_t>(input_height),
static_cast<int32_t>(input_width),
static_cast<int32_t>(channels)
};
const int32_t output_shape[4] = {
static_cast<int32_t>(batch_size),
static_cast<int32_t>(output_height),
static_cast<int32_t>(output_width),
static_cast<int32_t>(channels)
};
flatbuffers::Offset<tflite::Tensor> tensors[2] = {
tflite::CreateTensor(builder,
builder.CreateVector<int32_t>(input_shape, 4),
tflite::TensorType_FLOAT32),
tflite::CreateTensor(builder,
builder.CreateVector<int32_t>(output_shape, 4),
tflite::TensorType_FLOAT32),
};
const int32_t op_inputs[1] = { 0 };
const int32_t op_outputs[1] = { 1 };
flatbuffers::Offset<tflite::Operator> op = CreateOperator(
builder,
0 /* opcode_index */,
builder.CreateVector<int32_t>(op_inputs, 1),
builder.CreateVector<int32_t>(op_outputs, 1),
tflite::BuiltinOptions_Pool2DOptions,
pool2d_options.Union());
const int32_t graph_inputs[1] = { 0 };
const int32_t graph_outputs[1] = { 1 };
flatbuffers::Offset<tflite::SubGraph> subgraph = CreateSubGraph(
builder,
builder.CreateVector(tensors, 2),
builder.CreateVector<int32_t>(graph_inputs, 1),
builder.CreateVector<int32_t>(graph_outputs, 1),
builder.CreateVector(&op, 1));
flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder,
TFLITE_SCHEMA_VERSION,
builder.CreateVector(&operator_code, 1),
builder.CreateVector(&subgraph, 1),
builder.CreateString("AVERAGE_POOL_2D model"),
builder.CreateVector(buffers, 1));
builder.Finish(model_buffer);
const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer());
tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates resolver;
tflite::InterpreterBuilder interpreterBuilder(model, resolver);
std::unique_ptr<tflite::Interpreter> interpreter;
if (interpreterBuilder(&interpreter) != kTfLiteOk) {
state.SkipWithError("failed to create TFLite interpreter");
return;
}
if (interpreter == nullptr) {
state.SkipWithError("TFLite interpreter is null");
return;
}
interpreter->SetNumThreads(1);
if (interpreter->AllocateTensors() != kTfLiteOk) {
state.SkipWithError("failed to allocate tensors");
return;
}
std::generate(
interpreter->typed_tensor<float>(0),
interpreter->typed_tensor<float>(0) + batch_size * input_height * input_width * channels,
std::ref(f32rng));
for (auto _ : state) {
if (interpreter->Invoke() != kTfLiteOk) {
state.SkipWithError("failed to invoke TFLite interpreter");
return;
}
}
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
state.counters["bytes"] = benchmark::Counter(
uint64_t(state.iterations()) *
batch_size * (input_height * input_width + output_height * output_width) * channels * sizeof(float),
benchmark::Counter::kIsRate);
}
#endif // BENCHMARK_TENSORFLOW_LITE
// Final global average pooling in ImageNet classification models.
static void ImageNet(benchmark::internal::Benchmark* b) {
b->ArgNames({"N", "H", "W", "K", "P", "S", "C"});
/* N H W K P S C */
b->Args({1, 13, 13, 13, 0, 1, 1000});
b->Args({1, 7, 7, 7, 0, 1, 1000});
}
// ShuffleNet v1 with 1 group.
static void ShuffleNetV1G1(benchmark::internal::Benchmark* b) {
b->ArgNames({"N", "H", "W", "K", "P", "S", "C"});
/* N H W K P S C */
b->Args({1, 56, 56, 3, 1, 2, 24});
b->Args({1, 28, 28, 3, 1, 2, 144});
b->Args({1, 14, 14, 3, 1, 2, 288});
b->Args({1, 7, 7, 3, 1, 2, 576});
}
// ShuffleNet v1 with 2 groups.
static void ShuffleNetV1G2(benchmark::internal::Benchmark* b) {
b->ArgNames({"N", "H", "W", "K", "P", "S", "C"});
/* N H W K P S C */
b->Args({1, 56, 56, 3, 1, 2, 24});
b->Args({1, 28, 28, 3, 1, 2, 200});
b->Args({1, 14, 14, 3, 1, 2, 400});
b->Args({1, 7, 7, 3, 1, 2, 800});
}
// ShuffleNet v1 with 3 groups.
static void ShuffleNetV1G3(benchmark::internal::Benchmark* b) {
b->ArgNames({"N", "H", "W", "K", "P", "S", "C"});
/* N H W K P S C */
b->Args({1, 56, 56, 3, 1, 2, 24});
b->Args({1, 28, 28, 3, 1, 2, 240});
b->Args({1, 14, 14, 3, 1, 2, 480});
b->Args({1, 7, 7, 3, 1, 2, 960});
}
// ShuffleNet v1 with 4 groups.
static void ShuffleNetV1G4(benchmark::internal::Benchmark* b) {
b->ArgNames({"N", "H", "W", "K", "P", "S", "C"});
/* N H W K P S C */
b->Args({1, 56, 56, 3, 1, 2, 24});
b->Args({1, 28, 28, 3, 1, 2, 272});
b->Args({1, 14, 14, 3, 1, 2, 576});
b->Args({1, 7, 7, 3, 1, 2, 1088});
}
// ShuffleNet v1 with 8 groups.
static void ShuffleNetV1G8(benchmark::internal::Benchmark* b) {
b->ArgNames({"N", "H", "W", "K", "P", "S", "C"});
/* N H W K P S C */
b->Args({1, 56, 56, 3, 1, 2, 24});
b->Args({1, 28, 28, 3, 1, 2, 384});
b->Args({1, 14, 14, 3, 1, 2, 768});
b->Args({1, 7, 7, 3, 1, 2, 1536});
}
BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, imagenet, "ImageNet")->Apply(ImageNet)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_f32, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime();
#ifdef BENCHMARK_TENSORFLOW_LITE
BENCHMARK_CAPTURE(tflite_average_pooling_f32, imagenet, "ImageNet")->Apply(ImageNet)->UseRealTime();
BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime();
BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime();
BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime();
BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime();
BENCHMARK_CAPTURE(tflite_average_pooling_f32, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime();
#endif // BENCHMARK_TENSORFLOW_LITE
BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, imagenet, "ImageNet")->Apply(ImageNet)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime();
BENCHMARK_CAPTURE(xnnpack_average_pooling_qu8, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime();
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif
|