Gitiles
Code Review Sign In
gerrit-public.fairphone.software / platform / external / XNNPACK / refs/heads/odm/dev/target/13/fp5 / . / bench / convolution.cc
blob: 3d6a587c7572d31ed0c246695051cd4589ddfa22 [file] [log] [blame]
XNNPACK Teamb455b122019-09-27 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>
13#include <ostream>
14#include <random>
15#include <string>
16#include <vector>
17
18#include <cpuinfo.h>
19#include <xnnpack.h>
20
Frank Barchardbb4c18b2019-09-30 11:05:52 -0700[diff] [blame]21#ifdef BENCHMARK_ARM_COMPUTE_LIBRARY
22#include "arm_compute/core/Types.h"
23#include "arm_compute/runtime/Tensor.h"
24#include "arm_compute/runtime/CPP/CPPScheduler.h"
25#include "arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h"
26#include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
27#endif // BENCHMARK_ARM_COMPUTE_LIBRARY
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]28#include <benchmark/benchmark.h>
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]29#ifdef BENCHMARK_TENSORFLOW_LITE
30#include "flatbuffers/include/flatbuffers/flatbuffers.h"
31#include "tensorflow/lite/interpreter.h"
32#include "tensorflow/lite/kernels/register.h"
33#include "tensorflow/lite/model.h"
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]34#include "tensorflow/lite/schema/schema_generated.h"
35#include "tensorflow/lite/version.h"
36#endif // BENCHMARK_TENSORFLOW_LITE
Frank Barchardbb4c18b2019-09-30 11:05:52 -0700[diff] [blame]37#include "bench/utils.h"
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]38
39
40void xnnpack_convolution_q8(benchmark::State& state, const char* net) {
41 const size_t batch_size = state.range(0);
42 const size_t input_height = state.range(1);
43 const size_t input_width = state.range(2);
44 const size_t kernel_height = state.range(3);
45 const size_t kernel_width = state.range(4);
46 const size_t padding_height = state.range(5);
47 const size_t padding_width = state.range(6);
48 const size_t subsampling = state.range(7);
49 const size_t dilation = state.range(8);
50 const size_t groups = state.range(9);
51 const size_t group_input_channels = state.range(10);
52 const size_t group_output_channels = state.range(11);
53
54 std::random_device random_device;
55 auto rng = std::mt19937(random_device());
56 auto s32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng);
57 auto u8rng = std::bind(std::uniform_int_distribution<uint8_t>(), rng);
58
59 const size_t output_pixel_stride = groups * group_output_channels;
60 const size_t input_pixel_stride = groups * group_input_channels;
61 const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
62 const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
63 const size_t padding_left = padding_width / 2;
64 const size_t padding_top = padding_height / 2;
65 const size_t padding_right = padding_width - padding_left;
66 const size_t padding_bottom = padding_height - padding_top;
67 const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
68 const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
69
70 std::vector<uint8_t> input(batch_size * input_height * input_width * input_pixel_stride);
71 std::generate(input.begin(), input.end(), std::ref(u8rng));
72 std::vector<uint8_t> kernel(groups * group_output_channels * kernel_height * kernel_width * group_input_channels);
73 std::generate(kernel.begin(), kernel.end(), std::ref(u8rng));
74 std::vector<int32_t> bias(groups * group_output_channels);
75 std::generate(bias.begin(), bias.end(), std::ref(s32rng));
76 const size_t output_elements = batch_size * output_height * output_width * output_pixel_stride;
77
Marat Dukhan04f03be2019-11-19 12:36:47 -0800[diff] [blame]78 xnn_status status = xnn_initialize(nullptr /* allocator */);
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]79 if (status != xnn_status_success) {
80 state.SkipWithError("failed to initialize XNNPACK");
81 return;
82 }
83
84 if (!cpuinfo_initialize()) {
85 state.SkipWithError("cpuinfo initialization failed");
86 return;
87 }
88 const size_t num_buffers = 1 +
Marat Dukhan42323232019-10-23 02:09:02 -0700[diff] [blame]89 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]90 sizeof(uint8_t) * kernel.size() + sizeof(int32_t) * bias.size() + sizeof(uint8_t) * output_elements);
91 std::vector<uint8_t> output(output_elements * num_buffers);
92
93 std::vector<xnn_operator_t> convolution_operators(num_buffers);
94 for (xnn_operator_t& convolution_op : convolution_operators) {
95 status = xnn_create_convolution2d_nhwc_q8(
96 padding_top, padding_right, padding_bottom, padding_left,
97 kernel_height, kernel_width,
98 subsampling, subsampling,
99 dilation, dilation,
100 groups, group_input_channels, group_output_channels,
101 input_pixel_stride, output_pixel_stride,
102 127, 0.5f,
103 127, 0.5f,
104 kernel.data(), bias.data(),
105 127, 0.5f, 0, 255,
106 0 /* flags */, &convolution_op);
107 if (status != xnn_status_success) {
108 state.SkipWithError("failed to create QINT8 Convolution operator");
109 return;
110 }
111 }
112
113 for (size_t i = 0; i < convolution_operators.size(); i++) {
114 status = xnn_setup_convolution2d_nhwc_q8(
115 convolution_operators[i],
116 batch_size, input_height, input_width,
117 input.data(), output.data() + i * output_elements,
118 nullptr /* thread pool */);
119 if (status != xnn_status_success) {
120 state.SkipWithError("failed to setup QINT8 Convolution operator");
121 return;
122 }
123 }
124
125 size_t buffer_index = 0;
126 for (auto _ : state) {
127 state.PauseTiming();
Marat Dukhan42323232019-10-23 02:09:02 -0700[diff] [blame]128 benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(uint8_t));
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]129 buffer_index = (buffer_index + 1) % num_buffers;
130 state.ResumeTiming();
131
132 status = xnn_run_operator(convolution_operators[buffer_index],
133 nullptr /* thread pool */);
134 if (status != xnn_status_success) {
135 state.SkipWithError("failed to run QINT8 Convolution operator");
136 return;
137 }
138 }
139
140 for (xnn_operator_t& convolution_op : convolution_operators) {
141 status = xnn_delete_operator(convolution_op);
142 if (status != xnn_status_success) {
143 state.SkipWithError("failed to delete QINT8 Convolution operator");
144 return;
145 }
146 convolution_op = nullptr;
147 }
148
Frank Barchardbb4c18b2019-09-30 11:05:52 -0700[diff] [blame]149 state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency();
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]150 state.counters["OPS"] = benchmark::Counter(
151 uint64_t(state.iterations()) * 2 *
152 batch_size * output_height * output_width *
153 groups * group_input_channels * group_output_channels *
154 kernel_height * kernel_width,
155 benchmark::Counter::kIsRate);
156}
157
158void xnnpack_convolution_f32(benchmark::State& state, const char* net) {
159 const size_t batch_size = state.range(0);
160 const size_t input_height = state.range(1);
161 const size_t input_width = state.range(2);
162 const size_t kernel_height = state.range(3);
163 const size_t kernel_width = state.range(4);
164 const size_t padding_height = state.range(5);
165 const size_t padding_width = state.range(6);
166 const size_t subsampling = state.range(7);
167 const size_t dilation = state.range(8);
168 const size_t groups = state.range(9);
169 const size_t group_input_channels = state.range(10);
170 const size_t group_output_channels = state.range(11);
171
172 std::random_device random_device;
173 auto rng = std::mt19937(random_device());
174 auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng);
175
176 const size_t output_pixel_stride = groups * group_output_channels;
177 const size_t input_pixel_stride = groups * group_input_channels;
178 const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
179 const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
180 const size_t padding_left = padding_width / 2;
181 const size_t padding_top = padding_height / 2;
182 const size_t padding_right = padding_width - padding_left;
183 const size_t padding_bottom = padding_height - padding_top;
184 const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
185 const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
186
187 std::vector<float> input(batch_size * input_height * input_width * input_pixel_stride + XNN_EXTRA_BYTES / sizeof(float));
188 std::generate(input.begin(), input.end(), std::ref(f32rng));
189 std::vector<float> kernel(groups * group_output_channels * kernel_height * kernel_width * group_input_channels);
190 std::generate(kernel.begin(), kernel.end(), std::ref(f32rng));
191 std::vector<float> bias(groups * group_output_channels);
192 std::generate(bias.begin(), bias.end(), std::ref(f32rng));
193 const size_t output_elements = batch_size * output_height * output_width * output_pixel_stride;
194
Marat Dukhan04f03be2019-11-19 12:36:47 -0800[diff] [blame]195 xnn_status status = xnn_initialize(nullptr /* allocator */);
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]196 if (status != xnn_status_success) {
197 state.SkipWithError("failed to initialize XNNPACK");
198 return;
199 }
200
201 if (!cpuinfo_initialize()) {
202 state.SkipWithError("cpuinfo initialization failed");
203 return;
204 }
205 const size_t num_buffers = 1 +
Marat Dukhan42323232019-10-23 02:09:02 -0700[diff] [blame]206 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]207 sizeof(float) * (kernel.size() + bias.size() + output_elements));
208 std::vector<float> output(output_elements * num_buffers);
209
210 std::vector<xnn_operator_t> convolution_operators(num_buffers);
211 for (xnn_operator_t& convolution_op : convolution_operators) {
212 status = xnn_create_convolution2d_nhwc_f32(
213 padding_top, padding_right, padding_bottom, padding_left,
214 kernel_height, kernel_width,
215 subsampling, subsampling,
216 dilation, dilation,
217 groups, group_input_channels, group_output_channels,
218 input_pixel_stride, output_pixel_stride,
219 kernel.data(), bias.data(),
220 -std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity(),
221 0 /* flags */, &convolution_op);
222 if (status != xnn_status_success) {
223 state.SkipWithError("failed to create FP32 Convolution operator");
224 return;
225 }
226 }
227
228 for (size_t i = 0; i < convolution_operators.size(); i++) {
229 status = xnn_setup_convolution2d_nhwc_f32(
230 convolution_operators[i],
231 batch_size, input_height, input_width,
232 input.data(), output.data() + i * output_elements,
233 nullptr /* thread pool */);
234 if (status != xnn_status_success) {
235 state.SkipWithError("failed to setup FP32 Convolution operator");
236 return;
237 }
238 }
239
240 size_t buffer_index = 0;
241 for (auto _ : state) {
242 state.PauseTiming();
Marat Dukhan42323232019-10-23 02:09:02 -0700[diff] [blame]243 benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(float));
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]244 buffer_index = (buffer_index + 1) % num_buffers;
245 state.ResumeTiming();
246
247 status = xnn_run_operator(convolution_operators[buffer_index], nullptr /* thread pool */);
248 if (status != xnn_status_success) {
249 state.SkipWithError("failed to run FP32 Convolution operator");
250 return;
251 }
252 }
253
254 for (xnn_operator_t& convolution_op : convolution_operators) {
255 status = xnn_delete_operator(convolution_op);
256 if (status != xnn_status_success) {
257 state.SkipWithError("failed to delete FP32 Convolution operator");
258 return;
259 }
260 convolution_op = nullptr;
261 }
262
263 state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency();
264 state.counters["FLOPS"] = benchmark::Counter(
265 uint64_t(state.iterations()) * 2 *
266 batch_size * output_height * output_width *
267 groups * group_input_channels * group_output_channels *
268 kernel_height * kernel_width,
269 benchmark::Counter::kIsRate);
270}
271
272#ifdef BENCHMARK_TENSORFLOW_LITE
273void tflite_convolution_f32(benchmark::State& state, const char* net) {
274 const size_t batch_size = state.range(0);
275 const size_t input_height = state.range(1);
276 const size_t input_width = state.range(2);
277 const size_t kernel_height = state.range(3);
278 const size_t kernel_width = state.range(4);
279 const size_t padding_height = state.range(5);
280 const size_t padding_width = state.range(6);
281 const size_t subsampling = state.range(7);
282 const size_t dilation = state.range(8);
283 const size_t groups = state.range(9);
284 const size_t group_input_channels = state.range(10);
285 const size_t group_output_channels = state.range(11);
286
287 bool is_depthwise = false;
288 if (groups != 1) {
289 if (group_input_channels == 1) {
290 is_depthwise = true;
291 } else {
292 state.SkipWithError("grouped convolution is not supported");
293 return;
294 }
295 }
296
297 std::random_device random_device;
298 auto rng = std::mt19937(random_device());
299 auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng);
300
301 const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
302 const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
303
304 tflite::Padding padding = tflite::Padding_VALID;
305 if (padding_width == (effective_kernel_width - 1) && padding_height == (effective_kernel_height - 1)) {
306 padding = tflite::Padding_SAME;
307 } else if (padding_width == 0 && padding_height == 0) {
308 padding = tflite::Padding_VALID;
309 } else {
310 state.SkipWithError("unsupported padding");
311 return;
312 }
313
314 const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
315 const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
316
317 std::vector<float> kernel(groups * group_output_channels * kernel_height * kernel_width * group_input_channels);
318 std::generate(kernel.begin(), kernel.end(), std::ref(f32rng));
319 std::vector<float> bias(groups * group_output_channels);
320 std::generate(bias.begin(), bias.end(), std::ref(f32rng));
321
322 flatbuffers::FlatBufferBuilder builder;
323 flatbuffers::Offset<tflite::OperatorCode> operator_code =
324 CreateOperatorCode(
325 builder,
326 is_depthwise ? tflite::BuiltinOperator_DEPTHWISE_CONV_2D : tflite::BuiltinOperator_CONV_2D,
327 0);
328
329 flatbuffers::Offset<tflite::Conv2DOptions> conv2d_options = CreateConv2DOptions(
330 builder,
331 padding,
332 static_cast<int32_t>(subsampling), static_cast<int32_t>(subsampling),
333 tflite::ActivationFunctionType_NONE,
334 static_cast<int32_t>(dilation), static_cast<int32_t>(dilation));
335
336 flatbuffers::Offset<tflite::DepthwiseConv2DOptions> dwconv2d_options = CreateDepthwiseConv2DOptions(
337 builder,
338 padding,
339 static_cast<int32_t>(subsampling), static_cast<int32_t>(subsampling),
340 static_cast<int32_t>(group_output_channels),
341 tflite::ActivationFunctionType_NONE,
342 static_cast<int32_t>(dilation), static_cast<int32_t>(dilation));
343
344 flatbuffers::Offset<tflite::Buffer> buffers[3] = {
345 tflite::CreateBuffer(builder, builder.CreateVector({})),
346 tflite::CreateBuffer(builder, builder.CreateVector(
347 reinterpret_cast<const uint8_t*>(kernel.data()),
348 sizeof(float) * kernel.size())),
349 tflite::CreateBuffer(builder, builder.CreateVector(
350 reinterpret_cast<const uint8_t*>(bias.data()),
351 sizeof(float) * bias.size())),
352 };
353
354 const int32_t input_shape[4] = {
355 static_cast<int32_t>(batch_size),
356 static_cast<int32_t>(input_height),
357 static_cast<int32_t>(input_width),
358 static_cast<int32_t>(groups * group_input_channels)
359 };
360 const int32_t output_shape[4] = {
361 static_cast<int32_t>(batch_size),
362 static_cast<int32_t>(output_height),
363 static_cast<int32_t>(output_width),
364 static_cast<int32_t>(groups * group_output_channels)
365 };
366 const int32_t filter_shape[4] = {
367 static_cast<int32_t>(group_output_channels),
368 static_cast<int32_t>(kernel_height),
369 static_cast<int32_t>(kernel_width),
370 static_cast<int32_t>(groups * group_input_channels)
371 };
372 const int32_t bias_shape[1] = {
373 static_cast<int32_t>(groups * group_output_channels)
374 };
375
376 flatbuffers::Offset<tflite::Tensor> tensors[4] = {
377 tflite::CreateTensor(builder,
378 builder.CreateVector<int32_t>(input_shape, 4),
379 tflite::TensorType_FLOAT32,
380 0 /* buffer id */,
381 builder.CreateString("input")),
382 tflite::CreateTensor(builder,
383 builder.CreateVector<int32_t>(filter_shape, 4),
384 tflite::TensorType_FLOAT32,
385 1 /* buffer id */,
386 builder.CreateString("filter")),
387 tflite::CreateTensor(builder,
388 builder.CreateVector<int32_t>(bias_shape, 1),
389 tflite::TensorType_FLOAT32,
390 2 /* buffer id */,
391 builder.CreateString("bias")),
392 tflite::CreateTensor(builder,
393 builder.CreateVector<int32_t>(output_shape, 4),
394 tflite::TensorType_FLOAT32,
395 0 /* buffer id */,
396 builder.CreateString("output")),
397 };
398
399 const int32_t op_inputs[3] = { 0, 1, 2 };
400 const int32_t op_outputs[1] = { 3 };
401 flatbuffers::Offset<tflite::Operator> op = CreateOperator(
402 builder,
403 0 /* opcode_index */,
404 builder.CreateVector<int32_t>(op_inputs, 3),
405 builder.CreateVector<int32_t>(op_outputs, 1),
406 is_depthwise ? tflite::BuiltinOptions_DepthwiseConv2DOptions : tflite::BuiltinOptions_Conv2DOptions,
407 is_depthwise ? dwconv2d_options.Union() : conv2d_options.Union(),
408 /*custom_options */ 0,
409 tflite::CustomOptionsFormat_FLEXBUFFERS);
410
411 const int32_t graph_inputs[1] = { 0 };
412 const int32_t graph_outputs[1] = { 3 };
413 flatbuffers::Offset<tflite::SubGraph> subgraph = CreateSubGraph(
414 builder,
415 builder.CreateVector(tensors, 4),
416 builder.CreateVector<int32_t>(graph_inputs, 1),
417 builder.CreateVector<int32_t>(graph_outputs, 1),
418 builder.CreateVector(&op, 1),
419 builder.CreateString("Conv2D subgraph"));
420
421 flatbuffers::Offset<flatbuffers::String> description = builder.CreateString("Conv2D model");
422
423 flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder,
424 TFLITE_SCHEMA_VERSION,
425 builder.CreateVector(&operator_code, 1),
426 builder.CreateVector(&subgraph, 1),
427 description,
428 builder.CreateVector(buffers, 3));
429
430 builder.Finish(model_buffer);
431
432 const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer());
433 tflite::ops::builtin::BuiltinOpResolver resolver;
434 tflite::InterpreterBuilder interpreterBuilder(model, resolver);
435 std::unique_ptr<tflite::Interpreter> interpreter;
436 if (interpreterBuilder(&interpreter) != kTfLiteOk) {
437 state.SkipWithError("failed to create TFLite interpreter");
438 return;
439 }
440 if (interpreter == nullptr) {
441 state.SkipWithError("TFLite interpreter is null");
442 return;
443 }
444 interpreter->SetNumThreads(1);
445
446 if (interpreter->AllocateTensors() != kTfLiteOk) {
447 state.SkipWithError("failed to allocate tensors");
448 return;
449 }
450
451 std::generate(
452 interpreter->typed_tensor<float>(0),
453 interpreter->typed_tensor<float>(0) + batch_size * groups * group_input_channels * input_height * input_width,
454 std::ref(f32rng));
455
456 for (auto _ : state) {
457 state.PauseTiming();
Marat Dukhan42323232019-10-23 02:09:02 -0700[diff] [blame]458 benchmark::utils::WipeCache();
459 benchmark::utils::PrefetchToL1(
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]460 interpreter->typed_tensor<float>(0),
461 batch_size * groups * group_input_channels * input_height * input_width * sizeof(float));
462 state.ResumeTiming();
463
464 if (interpreter->Invoke() != kTfLiteOk) {
465 state.SkipWithError("failed to invoke TFLite interpreter");
466 return;
467 }
468 }
469
470 state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency();
471 state.counters["FLOPS"] = benchmark::Counter(
472 uint64_t(state.iterations()) * 2 *
473 batch_size * output_height * output_width *
474 groups * group_input_channels * group_output_channels *
475 kernel_height * kernel_width,
476 benchmark::Counter::kIsRate);
477
478 interpreter.reset();
479}
480#endif // BENCHMARK_TENSORFLOW_LITE
481
482#ifdef BENCHMARK_ARM_COMPUTE_LIBRARY
483static std::string compare_with_convolution_f32_reference_output(
484 const benchmark::State& state, const float* input, size_t input_size,
485 const float* kernel, size_t kernel_size, const float* bias, size_t bias_size,
486 const float* output, size_t output_size)
487{
488 const size_t batch_size = state.range(0);
489 const size_t input_height = state.range(1);
490 const size_t input_width = state.range(2);
491 const size_t kernel_height = state.range(3);
492 const size_t kernel_width = state.range(4);
493 const size_t padding_height = state.range(5);
494 const size_t padding_width = state.range(6);
495 const size_t subsampling = state.range(7);
496 const size_t dilation = state.range(8);
497 const size_t groups = state.range(9);
498 const size_t group_input_channels = state.range(10);
499 const size_t group_output_channels = state.range(11);
500
501 const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
502 const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
503 const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
504 const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
505 const size_t input_pixel_stride = groups * group_input_channels;
506 const size_t padding_left = padding_width / 2;
507 const size_t padding_top = padding_height / 2;
508
509 assert(input_size == batch_size * input_height * input_width * groups * group_input_channels);
510
511 assert(kernel_size == group_output_channels * kernel_height * kernel_width * groups * group_input_channels);
512
513 assert(bias_size == groups * group_output_channels);
514
515 assert(output_size == batch_size * output_height * output_width * groups * group_output_channels);
516
517 std::vector<float> output_ref(output_size);
518 for (size_t i = 0; i < batch_size; i++) {
519 for (size_t oy = 0; oy < output_height; oy++) {
520 for (size_t ox = 0; ox < output_width; ox++) {
521 for (size_t g = 0; g < groups; g++) {
522 for (size_t oc = 0; oc < group_output_channels; oc++) {
523 output_ref[(((i * output_height + oy) * output_width + ox) * groups + g) * group_output_channels + oc] =
524 bias[g * group_output_channels + oc];
525 }
526 }
527 }
528 }
529 }
530 for (size_t i = 0; i < batch_size; i++) {
531 for (size_t oy = 0; oy < output_height; oy++) {
532 for (size_t ox = 0; ox < output_width; ox++) {
533 for (size_t ky = 0; ky < kernel_height; ky++) {
534 const size_t iy = oy * subsampling + ky * dilation - padding_top;
535 if (iy < input_height) {
536 for (size_t kx = 0; kx < kernel_width; kx++) {
537 const size_t ix = ox * subsampling + kx * dilation - padding_left;
538 if (ix < input_width) {
539 for (size_t g = 0; g < groups; g++) {
540 for (size_t oc = 0; oc < group_output_channels; oc++) {
541 for (size_t ic = 0; ic < group_input_channels; ic++) {
542 output_ref[(((i * output_height + oy) * output_width + ox) * groups + g) * group_output_channels + oc] +=
543 input[((i * input_height + iy) * input_width + ix) * input_pixel_stride + g * group_input_channels + ic] *
544 kernel[(((oc * kernel_height + ky) * kernel_width + kx) * groups + g) * group_input_channels + ic];
545 } // group_input_channels loop
546 } // group_output_channels loop
547 } // groups loop
548 }
549 } // kernel_width loop
550 }
551 } // kernel_height loop
552 } // output_width loop
553 } // output_height loop
554 } // batch_size loop
555
556 const float relative_error_tolerance = 1e-4;
557 for (size_t i = 0; i < batch_size; i++) {
558 for (size_t y = 0; y < output_height; y++) {
559 for (size_t x = 0; x < output_width; x++) {
560 for (size_t g = 0; g < groups; g++) {
561 for (size_t c = 0; c < group_output_channels; c++) {
562 const size_t idx = (((i * output_height + y) * output_width + x) * groups + g) * group_output_channels + c;
563 const float value_ref = output_ref[idx];
564 const float value = output[idx];
565 if (std::abs(value - value_ref) > std::max(std::abs(value_ref) * relative_error_tolerance, std::numeric_limits<float>::epsilon())) {
566 std::ostringstream error_stream;
567 error_stream << "(x, y) = (" << x << ", " << y << "), group = " << g
568 << ", channel = " << c << ", refValue = " << value_ref
569 << ", actualValue = " << value
570 << ", absDiff=" << std::abs(value - value_ref);
571 return error_stream.str();
572 }
573 }
574 }
575 }
576 }
577 }
578 return "";
579}
580
581void armcl_convolution_f32(benchmark::State& state, const char* net) {
582 const size_t batch_size = state.range(0);
583 const size_t input_height = state.range(1);
584 const size_t input_width = state.range(2);
585 const size_t kernel_height = state.range(3);
586 const size_t kernel_width = state.range(4);
587 const size_t padding_height = state.range(5);
588 const size_t padding_width = state.range(6);
589 const size_t subsampling = state.range(7);
590 const size_t dilation = state.range(8);
591 const size_t groups = state.range(9);
592 const size_t group_input_channels = state.range(10);
593 const size_t group_output_channels = state.range(11);
594
595 const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
596 const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
597 const size_t padding_left = padding_width / 2;
598 const size_t padding_top = padding_height / 2;
599 const size_t padding_right = padding_width - padding_left;
600 const size_t padding_bottom = padding_height - padding_top;
601 const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
602 const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
603
604 arm_compute::PadStrideInfo pad_stride_info(
605 subsampling /* stride height */,
606 subsampling /* stride width */,
607 padding_left, padding_right, padding_top, padding_bottom,
608 arm_compute::DimensionRoundingType::FLOOR);
609 arm_compute::Size2D dilation_info(dilation, dilation);
610 // Note: activation is disabled by default.
611 arm_compute::ActivationLayerInfo activation_info;
612
613 // Note: no batch size and reverse order of dimensions, i.e. CWHN for NHWC.
614 arm_compute::TensorShape input_shape(
615 /* C */ groups * group_input_channels,
616 /* W */ input_width,
617 /* H */ input_height,
618 /* N */ batch_size);
619 arm_compute::TensorInfo input_info(
620 input_shape,
621 1 /* number of channels per element (!) */,
622 arm_compute::DataType::F32);
623 input_info.set_data_layout(arm_compute::DataLayout::NHWC);
624 arm_compute::Tensor input_tensor;
625 input_tensor.allocator()->init(input_info);
626 input_tensor.allocator()->allocate();
627
628 // Note: reverse order of dimensions, i.e. for IWHO for OHWI.
629 arm_compute::TensorShape kernel_shape(
630 /* I */ groups * group_input_channels,
631 /* W */ kernel_width,
632 /* H */ kernel_height,
633 /* O */ group_output_channels);
634 arm_compute::TensorInfo kernel_info(
635 kernel_shape,
636 1 /* number of channels per element (!) */,
637 arm_compute::DataType::F32);
638 kernel_info.set_data_layout(arm_compute::DataLayout::NHWC);
639 arm_compute::Tensor kernelTensor;
640 kernelTensor.allocator()->init(kernel_info);
641 kernelTensor.allocator()->allocate();
642
643 arm_compute::TensorShape bias_shape(groups * group_output_channels);
644 arm_compute::TensorInfo bias_info(
645 bias_shape,
646 1 /* number of channels per element (!) */,
647 arm_compute::DataType::F32);
648 bias_info.set_data_layout(arm_compute::DataLayout::NHWC);
649 arm_compute::Tensor bias_tensor;
650 bias_tensor.allocator()->init(bias_info);
651 bias_tensor.allocator()->allocate();
652
653 // Note: no batch size and reverse order of dimensions, i.e. CWHN for NHWC.
654 arm_compute::TensorShape output_shape(
655 /* C */ groups * group_output_channels,
656 /* W */ output_width,
657 /* H */ output_height,
658 /* N */ batch_size);
659 arm_compute::TensorInfo output_info(
660 output_shape,
661 1 /* number of channels per element (!) */,
662 arm_compute::DataType::F32);
663 output_info.set_data_layout(arm_compute::DataLayout::NHWC);
664 arm_compute::Tensor output_tensor;
665 output_tensor.allocator()->init(output_info);
666 output_tensor.allocator()->allocate();
667
668 std::random_device random_device;
669 auto rng = std::mt19937(random_device());
670 auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng);
671
672 std::generate(
673 reinterpret_cast<float*>(input_tensor.buffer()),
674 reinterpret_cast<float*>(input_tensor.buffer()) + input_shape.total_size(),
675 std::ref(f32rng));
676 std::generate(
677 reinterpret_cast<float*>(kernelTensor.buffer()),
678 reinterpret_cast<float*>(kernelTensor.buffer()) + kernel_shape.total_size(),
679 std::ref(f32rng));
680 std::generate(
681 reinterpret_cast<float*>(bias_tensor.buffer()),
682 reinterpret_cast<float*>(bias_tensor.buffer()) + bias_shape.total_size(),
683 std::ref(f32rng));
684 std::generate(
685 reinterpret_cast<float*>(output_tensor.buffer()),
686 reinterpret_cast<float*>(output_tensor.buffer()) + output_shape.total_size(),
687 std::ref(f32rng));
688
689 bool is_depthwise = false;
690 if (groups != 1) {
691 // NEConvolutionLayer uses NEGEMMConvolutionLayer by default, which doesn't support grouped convolution.
692 // However, depthwise convolution is supported via NEDepthwiseConvolutionLayer.
693 if (group_input_channels == 1) {
694 is_depthwise = true;
695 } else {
696 state.SkipWithError("grouped convolution is not supported");
697 return;
698 }
699 }
700
701 std::shared_ptr<arm_compute::IFunction> layer;
702 if (is_depthwise) {
703 if (dilation != 1) {
704 state.SkipWithError("dilated depthwise convolution is not supported");
705 return;
706 }
707
708 // Avoid NEDepthwiseConvolutionLayer3x3 when stride isn't 2 in order to pass the output verification.
709 // TODO(b/130206370) This looks like a bug and needs further investigation.
710 if (kernel_height == 3 && kernel_width == 3 && subsampling == 2) {
711 auto* depthwise_3x3_convolution_layer = new arm_compute::NEDepthwiseConvolutionLayer3x3();
712 layer.reset(depthwise_3x3_convolution_layer);
713 depthwise_3x3_convolution_layer->configure(
714 &input_tensor, &kernelTensor, &bias_tensor, &output_tensor,
715 pad_stride_info, group_output_channels, activation_info);
716
717 if (!depthwise_3x3_convolution_layer->validate(
718 &input_info, &kernel_info, &bias_info, &output_info,
719 pad_stride_info, group_output_channels, activation_info))
720 {
721 state.SkipWithError("validation failed");
722 return;
723 }
724 } else {
725 auto* depthwise_convolution_layer = new arm_compute::NEDepthwiseConvolutionLayer();
726 layer.reset(depthwise_convolution_layer);
727 depthwise_convolution_layer->configure(
728 &input_tensor, &kernelTensor, &bias_tensor, &output_tensor,
729 pad_stride_info, group_output_channels, activation_info);
730
731 if (!depthwise_convolution_layer->validate(
732 &input_info, &kernel_info, &bias_info, &output_info,
733 pad_stride_info, group_output_channels, activation_info))
734 {
735 state.SkipWithError("validation failed");
736 return;
737 }
738 }
739 } else {
740 auto* convolution_layer = new arm_compute::NEConvolutionLayer();
741 layer.reset(convolution_layer);
742 convolution_layer->configure(
743 &input_tensor, &kernelTensor, &bias_tensor, &output_tensor,
744 pad_stride_info, arm_compute::WeightsInfo(), dilation_info, activation_info,
745 true /* enable fast math */, groups);
746
747 if (!convolution_layer->validate(
748 &input_info, &kernel_info, &bias_info, &output_info,
749 pad_stride_info, arm_compute::WeightsInfo(), dilation_info, activation_info,
750 true /* enable fast math */, groups))
751 {
752 state.SkipWithError("validation failed");
753 return;
754 }
755 }
756
757 // Dry run to let ACL do one-time initializations.
758 arm_compute::CPPScheduler::get().set_num_threads(1);
759 layer->run();
760
761 for (auto _ : state) {
762 state.PauseTiming();
Marat Dukhan42323232019-10-23 02:09:02 -0700[diff] [blame]763 benchmark::utils::WipeCache();
764 benchmark::utils::PrefetchToL1(
XNNPACK Teamb455b122019-09-27 18:10:33 -0700[diff] [blame]765 input_tensor.buffer(),
766 batch_size * groups * group_input_channels * input_height * input_width * sizeof(float));
767 state.ResumeTiming();
768
769 layer->run();
770 }
771
772 // Validate outputs.
773 const std::string error_string = compare_with_convolution_f32_reference_output(
774 state, reinterpret_cast<const float*>(input_tensor.buffer()),
775 input_shape.total_size(),
776 reinterpret_cast<const float*>(kernelTensor.buffer()),
777 kernel_shape.total_size(),
778 reinterpret_cast<const float*>(bias_tensor.buffer()),
779 bias_shape.total_size(),
780 reinterpret_cast<const float*>(output_tensor.buffer()),
781 output_shape.total_size());
782
783 if (!error_string.empty()) {
784 state.SkipWithError(("validation failed: " + error_string).c_str());
785 return;
786 }
787
788 input_tensor.allocator()->free();
789 kernelTensor.allocator()->free();
790 bias_tensor.allocator()->free();
791 output_tensor.allocator()->free();
792
793 state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency();
794 state.counters["FLOPS"] = benchmark::Counter(
795 uint64_t(state.iterations()) * 2 *
796 batch_size * output_height * output_width *
797 groups * group_input_channels * group_output_channels *
798 kernel_height * kernel_width,
799 benchmark::Counter::kIsRate);
800}
801#endif // BENCHMARK_ARM_COMPUTE_LIBRARY
802
803// ShuffleNet v1 with 1 group.
804static void ShuffleNetV1G1(benchmark::internal::Benchmark* b) {
805 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
806
807 /*************************** Conv 1 **************************/
808 /* N H W KH KW PH PW S D G GCin GCout */
809 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
810 /******************* Stage 2: stride-2 unit ******************/
811 /* N H W KH KW PH PW S D G GCin GCout */
812 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 36});
813 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 36, 1, 1});
814 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 36, 120});
815 /******************* Stage 2: stride-1 units *****************/
816 /* N H W KH KW PH PW S D G GCin GCout */
817 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 144, 36});
818 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 36, 1, 1});
819 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 36, 144});
820 /******************* Stage 3: stride-2 unit ******************/
821 /* N H W KH KW PH PW S D G GCin GCout */
822 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 144, 72});
823 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 72, 1, 1});
824 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 72, 144});
825 /******************* Stage 3: stride-1 units *****************/
826 /* N H W KH KW PH PW S D G GCin GCout */
827 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 288, 72});
828 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 72, 1, 1});
829 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 72, 288});
830 /******************* Stage 4: stride-2 unit ******************/
831 /* N H W KH KW PH PW S D G GCin GCout */
832 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 288, 144});
833 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 144, 1, 1});
834 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 144, 288});
835 /******************* Stage 4: stride-1 units *****************/
836 /* N H W KH KW PH PW S D G GCin GCout */
837 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 576, 144});
838 b->Args({1, 7, 7, 3, 3, 2, 2, 2, 1, 144, 1, 1});
839 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 144, 576});
840}
841
842// ShuffleNet v1 with 2 groups.
843static void ShuffleNetV1G2(benchmark::internal::Benchmark* b) {
844 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
845
846 /*************************** Conv 1 **************************/
847 /* N H W KH KW PH PW S D G GCin GCout */
848 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
849 /******************* Stage 2: stride-2 unit ******************/
850 /* N H W KH KW PH PW S D G GCin GCout */
851 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 50});
852 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 50, 1, 1});
853 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 2, 25, 88});
854 /******************* Stage 2: stride-1 units *****************/
855 /* N H W KH KW PH PW S D G GCin GCout */
856 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 2, 100, 25});
857 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 50, 1, 1});
858 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 2, 25, 100});
859 /******************* Stage 3: stride-2 unit ******************/
860 /* N H W KH KW PH PW S D G GCin GCout */
861 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 2, 100, 50});
862 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 100, 1, 1});
863 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 2, 50, 100});
864 /******************* Stage 3: stride-1 units *****************/
865 /* N H W KH KW PH PW S D G GCin GCout */
866 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 2, 200, 50});
867 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 100, 1, 1});
868 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 2, 50, 200});
869 /******************* Stage 4: stride-2 unit ******************/
870 /* N H W KH KW PH PW S D G GCin GCout */
871 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 2, 200, 100});
872 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 200, 1, 1});
873 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 2, 100, 200});
874 /******************* Stage 4: stride-1 units *****************/
875 /* N H W KH KW PH PW S D G GCin GCout */
876 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 2, 400, 100});
877 b->Args({1, 7, 7, 3, 3, 2, 2, 2, 1, 200, 1, 1});
878 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 2, 100, 400});
879}
880
881// ShuffleNet v1 with 3 groups.
882static void ShuffleNetV1G3(benchmark::internal::Benchmark* b) {
883 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
884
885 /*************************** Conv 1 **************************/
886 /* N H W KH KW PH PW S D G GCin GCout */
887 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
888 /******************* Stage 2: stride-2 unit ******************/
889 /* N H W KH KW PH PW S D G GCin GCout */
890 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 60});
891 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 60, 1, 1});
892 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 3, 20, 72});
893 /******************* Stage 2: stride-1 units *****************/
894 /* N H W KH KW PH PW S D G GCin GCout */
895 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 3, 80, 20});
896 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 60, 1, 1});
897 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 3, 20, 80});
898 /******************* Stage 3: stride-2 unit ******************/
899 /* N H W KH KW PH PW S D G GCin GCout */
900 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 3, 80, 40});
901 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 120, 1, 1});
902 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 3, 40, 80});
903 /******************* Stage 3: stride-1 units *****************/
904 /* N H W KH KW PH PW S D G GCin GCout */
905 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 3, 160, 40});
906 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 120, 1, 1});
907 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 3, 40, 160});
908 /******************* Stage 4: stride-2 unit ******************/
909 /* N H W KH KW PH PW S D G GCin GCout */
910 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 3, 160, 80});
911 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 240, 1, 1});
912 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 3, 80, 160});
913 /******************* Stage 4: stride-1 units *****************/
914 /* N H W KH KW PH PW S D G GCin GCout */
915 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 3, 320, 80});
916 b->Args({1, 7, 7, 3, 3, 2, 2, 2, 1, 240, 1, 1});
917 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 3, 80, 320});
918}
919
920// ShuffleNet v1 with 4 groups.
921static void ShuffleNetV1G4(benchmark::internal::Benchmark* b) {
922 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
923
924 /*************************** Conv 1 **************************/
925 /* N H W KH KW PH PW S D G GCin GCout */
926 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
927 /******************* Stage 2: stride-2 unit ******************/
928 /* N H W KH KW PH PW S D G GCin GCout */
929 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 68});
930 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 68, 1, 1});
931 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 4, 17, 62});
932 /******************* Stage 2: stride-1 units *****************/
933 /* N H W KH KW PH PW S D G GCin GCout */
934 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 4, 68, 17});
935 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 68, 1, 1});
936 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 4, 17, 68});
937 /******************* Stage 3: stride-2 unit ******************/
938 /* N H W KH KW PH PW S D G GCin GCout */
939 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 4, 68, 34});
940 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 136, 1, 1});
941 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 4, 34, 68});
942 /******************* Stage 3: stride-1 units *****************/
943 /* N H W KH KW PH PW S D G GCin GCout */
944 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 4, 136, 34});
945 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 136, 1, 1});
946 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 4, 34, 136});
947 /******************* Stage 4: stride-2 unit ******************/
948 /* N H W KH KW PH PW S D G GCin GCout */
949 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 4, 136, 68});
950 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 272, 1, 1});
951 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 4, 68, 136});
952 /******************* Stage 4: stride-1 units *****************/
953 /* N H W KH KW PH PW S D G GCin GCout */
954 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 4, 272, 68});
955 b->Args({1, 7, 7, 3, 3, 2, 2, 2, 1, 272, 1, 1});
956 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 4, 68, 272});
957}
958
959// ShuffleNet v1 with 8 groups.
960static void ShuffleNetV1G8(benchmark::internal::Benchmark* b) {
961 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
962
963 /*************************** Conv 1 **************************/
964 /* N H W KH KW PH PW S D G GCin GCout */
965 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
966 /******************* Stage 2: stride-2 unit ******************/
967 /* N H W KH KW PH PW S D G GCin GCout */
968 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 96});
969 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 96, 1, 1});
970 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 8, 12, 45});
971 /******************* Stage 2: stride-1 units *****************/
972 /* N H W KH KW PH PW S D G GCin GCout */
973 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 8, 48, 12});
974 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 96, 1, 1});
975 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 8, 12, 48});
976 /******************* Stage 3: stride-2 unit ******************/
977 /* N H W KH KW PH PW S D G GCin GCout */
978 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 8, 48, 24});
979 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 192, 1, 1});
980 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 8, 24, 48});
981 /******************* Stage 3: stride-1 units *****************/
982 /* N H W KH KW PH PW S D G GCin GCout */
983 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 8, 96, 24});
984 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 192, 1, 1});
985 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 8, 24, 96});
986 /******************* Stage 4: stride-2 unit ******************/
987 /* N H W KH KW PH PW S D G GCin GCout */
988 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 8, 96, 48});
989 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 384, 1, 1});
990 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 8, 48, 96});
991 /******************* Stage 4: stride-1 units *****************/
992 /* N H W KH KW PH PW S D G GCin GCout */
993 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 8, 192, 48});
994 b->Args({1, 7, 7, 3, 3, 2, 2, 2, 1, 384, 1, 1});
995 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 8, 48, 192});
996}
997
998// ShuffleNet v2 (0.5X scale)
999static void ShuffleNetV2X05(benchmark::internal::Benchmark* b) {
1000 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1001
1002 /*************************** Conv 1 **************************/
1003 /* N H W KH KW PH PW S D G GCin GCout */
1004 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
1005 /************************** Stage 2 **************************/
1006 /* N H W KH KW PH PW S D G GCin GCout */
1007 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 24, 1, 1});
1008 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 24, 24});
1009 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 24});
1010 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 24, 1, 1});
1011 /************************** Stage 3 **************************/
1012 /* N H W KH KW PH PW S D G GCin GCout */
1013 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 48, 1, 1});
1014 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 48, 48});
1015 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 48, 48});
1016 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 48, 1, 1});
1017 /************************** Stage 4 **************************/
1018 /* N H W KH KW PH PW S D G GCin GCout */
1019 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 96, 1, 1});
1020 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 96, 96});
1021 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 96, 96});
1022 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 96, 1, 1});
1023 /*************************** Conv 5 **************************/
1024 /* N H W KH KW PH PW S D G GCin GCout */
1025 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 192, 1024});
1026}
1027
1028// ShuffleNet v2 (1.0X scale)
1029static void ShuffleNetV2X10(benchmark::internal::Benchmark* b) {
1030 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1031
1032 /*************************** Conv 1 **************************/
1033 /* N H W KH KW PH PW S D G GCin GCout */
1034 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
1035 /************************** Stage 2 **************************/
1036 /* N H W KH KW PH PW S D G GCin GCout */
1037 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 24, 1, 1});
1038 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 24, 58});
1039 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 58});
1040 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 58, 1, 1});
1041 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 58, 58});
1042 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 58, 1, 1});
1043 /************************** Stage 3 **************************/
1044 /* N H W KH KW PH PW S D G GCin GCout */
1045 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 116, 1, 1});
1046 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 116, 116});
1047 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 116, 116});
1048 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 116, 1, 1});
1049 /************************** Stage 4 **************************/
1050 /* N H W KH KW PH PW S D G GCin GCout */
1051 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 232, 1, 1});
1052 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 232, 232});
1053 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 232, 232});
1054 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 232, 1, 1});
1055 /*************************** Conv 5 **************************/
1056 /* N H W KH KW PH PW S D G GCin GCout */
1057 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 464, 1024});
1058}
1059
1060// ShuffleNet v2 (1.5X scale)
1061static void ShuffleNetV2X15(benchmark::internal::Benchmark* b) {
1062 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1063
1064 /*************************** Conv 1 **************************/
1065 /* N H W KH KW PH PW S D G GCin GCout */
1066 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
1067 /************************** Stage 2 **************************/
1068 /* N H W KH KW PH PW S D G GCin GCout */
1069 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 24, 1, 1});
1070 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 24, 88});
1071 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 88});
1072 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 88, 1, 1});
1073 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 88, 88});
1074 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 88, 1, 1});
1075 /************************** Stage 3 **************************/
1076 /* N H W KH KW PH PW S D G GCin GCout */
1077 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 176, 1, 1});
1078 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 176, 176});
1079 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 176, 176});
1080 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 176, 1, 1});
1081 /************************** Stage 4 **************************/
1082 /* N H W KH KW PH PW S D G GCin GCout */
1083 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 352, 1, 1});
1084 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 352, 352});
1085 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 352, 352});
1086 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 352, 1, 1});
1087 /*************************** Conv 5 **************************/
1088 /* N H W KH KW PH PW S D G GCin GCout */
1089 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 704, 1024});
1090}
1091
1092// ShuffleNet v2 (2.0X scale)
1093static void ShuffleNetV2X20(benchmark::internal::Benchmark* b) {
1094 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1095
1096 /*************************** Conv 1 **************************/
1097 /* N H W KH KW PH PW S D G GCin GCout */
1098 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 24});
1099 /************************** Stage 2 **************************/
1100 /* N H W KH KW PH PW S D G GCin GCout */
1101 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 24, 1, 1});
1102 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 24, 122});
1103 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 122});
1104 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 122, 1, 1});
1105 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 122, 122});
1106 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 122, 1, 1});
1107 /************************** Stage 3 **************************/
1108 /* N H W KH KW PH PW S D G GCin GCout */
1109 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 244, 1, 1});
1110 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 244, 244});
1111 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 244, 244});
1112 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 244, 1, 1});
1113 /************************** Stage 4 **************************/
1114 /* N H W KH KW PH PW S D G GCin GCout */
1115 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 488, 1, 1});
1116 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 488, 488});
1117 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 488, 488});
1118 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 488, 1, 1});
1119 /*************************** Conv 5 **************************/
1120 /* N H W KH KW PH PW S D G GCin GCout */
1121 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 976, 2048});
1122}
1123
1124static void MobileNetV1(benchmark::internal::Benchmark* b) {
1125 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1126
1127 /* N H W KH KW PH PW S D G GCin GCout */
1128 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 32});
1129 b->Args({1, 112, 112, 3, 3, 2, 2, 1, 1, 32, 1, 1});
1130 b->Args({1, 112, 112, 1, 1, 0, 0, 1, 1, 1, 32, 64});
1131 b->Args({1, 112, 112, 3, 3, 2, 2, 2, 1, 64, 1, 1});
1132 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 64, 128});
1133 b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 128, 1, 1});
1134 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 128, 128});
1135 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 128, 1, 1});
1136 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 128, 256});
1137 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 256, 1, 1});
1138 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 256, 256});
1139 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 256, 1, 1});
1140 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 256, 512});
1141 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 512, 1, 1});
1142 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 512, 512});
1143 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 512, 1, 1});
1144 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 512, 1024});
1145 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 1024, 1, 1});
1146 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 1024, 1024});
1147}
1148
1149static void MobileNetV2(benchmark::internal::Benchmark* b) {
1150 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1151
1152 /* N H W KH KW PH PW S D G GCin GCout */
1153 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 32});
1154
1155 /************************ Bottleneck 1 ***********************/
1156 /* N H W KH KW PH PW S D G GCin GCout */
1157 b->Args({1, 112, 112, 3, 3, 2, 2, 1, 1, 32, 1, 1});
1158 b->Args({1, 112, 112, 1, 1, 0, 0, 1, 1, 1, 32, 16});
1159
1160 /************************ Bottleneck 2 ***********************/
1161 /* N H W KH KW PH PW S D G GCin GCout */
1162 b->Args({1, 112, 112, 1, 1, 0, 0, 1, 1, 1, 16, 96});
1163 b->Args({1, 112, 112, 3, 3, 2, 2, 2, 1, 96, 1, 1});
1164 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 96, 24});
1165 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 144});
1166 b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 144, 1, 1});
1167 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 144, 24});
1168
1169 /************************ Bottleneck 3 ***********************/
1170 /* N H W KH KW PH PW S D G GCin GCout */
1171//b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 144});
1172 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 144, 1, 1});
1173 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 144, 32});
1174 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 32, 192});
1175 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 192, 1, 1});
1176 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 192, 32});
1177//b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 32, 192});
1178//b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 192, 1, 1});
1179//b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 192, 32});
1180
1181 /************************ Bottleneck 4 ***********************/
1182 /* N H W KH KW PH PW S D G GCin GCout */
1183//b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 32, 192});
1184 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 192, 1, 1});
1185 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 192, 64});
1186 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 64, 384});
1187 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 384, 1, 1});
1188 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 384, 64});
1189//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 64, 384});
1190//b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 384, 1, 1});
1191//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 384, 64});
1192//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 64, 384});
1193//b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 384, 1, 1});
1194//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 384, 64});
1195
1196 /************************ Bottleneck 5 ***********************/
1197 /* N H W KH KW PH PW S D G GCin GCout */
1198//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 64, 384});
1199//b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 384, 1, 1});
1200 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 384, 96});
1201 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 96, 576});
1202 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 576, 1, 1});
1203 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 576, 96});
1204//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 96, 576});
1205//b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 576, 1, 1});
1206//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 576, 96});
1207
1208 /************************ Bottleneck 6 ***********************/
1209 /* N H W KH KW PH PW S D G GCin GCout */
1210//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 96, 576});
1211 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 576, 1, 1});
1212 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 576, 160});
1213 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 160, 960});
1214 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 960, 1, 1});
1215 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 960, 160});
1216//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 160, 960});
1217//b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 960, 1, 1});
1218//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 960, 160});
1219
1220 /************************ Bottleneck 7 ***********************/
1221 /* N H W KH KW PH PW S D G GCin GCout */
1222//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 160, 960});
1223//b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 960, 1, 1});
1224 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 960, 320});
1225
1226 /******************** Pre-pooling Conv2D *********************/
1227 /* N H W KH KW PH PW S D G GCin GCout */
1228 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 320, 1280});
1229 /******************** Post-pooling Conv2D ********************/
1230 /* N H W KH KW PH PW S D G GCin GCout */
1231 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1280, 1000});
1232}
1233
1234static void MobileNetV3Small(benchmark::internal::Benchmark* b) {
1235 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1236
1237 /*********************** Initial Stage ***********************/
1238 /* N H W KH KW PH PW S D G GCin GCout */
1239 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 16});
1240 /*********************** Bottleneck 1 ************************/
1241 /* N H W KH KW PH PW S D G GCin GCout */
1242 b->Args({1, 112, 112, 3, 3, 2, 2, 2, 1, 16, 1, 1});
1243 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 16, 8});
1244 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 8, 16});
1245 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 16, 16});
1246 /*********************** Bottleneck 2 ************************/
1247 /* N H W KH KW PH PW S D G GCin GCout */
1248 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 16, 72});
1249 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 72, 1, 1});
1250 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 72, 24});
1251 /*********************** Bottleneck 3 ************************/
1252 /* N H W KH KW PH PW S D G GCin GCout */
1253 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 24, 88});
1254 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 88, 1, 1});
1255 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 88, 24});
1256 /*********************** Bottleneck 4 ************************/
1257 /* N H W KH KW PH PW S D G GCin GCout */
1258 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 24, 96});
1259 b->Args({1, 28, 28, 5, 5, 4, 4, 2, 1, 96, 1, 1});
1260 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 96, 24});
1261 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 24, 96});
1262 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 96, 40});
1263 /*********************** Bottleneck 5 ************************/
1264 /* N H W KH KW PH PW S D G GCin GCout */
1265 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 40, 240});
1266 b->Args({1, 14, 14, 5, 5, 4, 4, 1, 1, 240, 1, 1});
1267 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 240, 64});
1268 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 64, 240});
1269 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 240, 40});
1270 /*********************** Bottleneck 6 ************************/
1271 /* N H W KH KW PH PW S D G GCin GCout */
1272//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 40, 240});
1273//b->Args({1, 14, 14, 5, 5, 4, 4, 1, 1, 240, 1, 1});
1274//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 240, 64});
1275//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 64, 240});
1276//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 240, 40});
1277 /*********************** Bottleneck 7 ************************/
1278 /* N H W KH KW PH PW S D G GCin GCout */
1279 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 40, 120});
1280 b->Args({1, 14, 14, 5, 5, 4, 4, 1, 1, 120, 1, 1});
1281 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 120, 32});
1282 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 32, 120});
1283 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 120, 48});
1284 /*********************** Bottleneck 8 ************************/
1285 /* N H W KH KW PH PW S D G GCin GCout */
1286 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 48, 144});
1287 b->Args({1, 14, 14, 5, 5, 4, 4, 1, 1, 144, 1, 1});
1288 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 144, 40});
1289 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 40, 144});
1290 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 144, 48});
1291 /*********************** Bottleneck 9 ************************/
1292 /* N H W KH KW PH PW S D G GCin GCout */
1293 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 48, 288});
1294 b->Args({1, 14, 14, 5, 5, 4, 4, 2, 1, 288, 1, 1});
1295 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 288, 72});
1296 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 72, 288});
1297 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 288, 96});
1298 /*********************** Bottleneck 10 ***********************/
1299 /* N H W KH KW PH PW S D G GCin GCout */
1300 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 96, 576});
1301 b->Args({1, 7, 7, 5, 5, 4, 4, 1, 1, 576, 1, 1});
1302 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 576, 144});
1303 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 144, 576});
1304 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 576, 96});
1305 /*********************** Bottleneck 11 ***********************/
1306 /* N H W KH KW PH PW S D G GCin GCout */
1307//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 96, 576});
1308//b->Args({1, 7, 7, 5, 5, 4, 4, 1, 1, 576, 1, 1});
1309//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 576, 144});
1310//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 144, 576});
1311//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 576, 96});
1312 /************************ Last Stage ************************/
1313 /* N H W KH KW PH PW S D G GCin GCout */
1314//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 96, 576});
1315 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 576, 1024});
1316 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1024, 1001});
1317}
1318
1319static void MobileNetV3Large(benchmark::internal::Benchmark* b) {
1320 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1321
1322 /*********************** Initial Stage ***********************/
1323 /* N H W KH KW PH PW S D G GCin GCout */
1324 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 16});
1325 /*********************** Bottleneck 1 ************************/
1326 /* N H W KH KW PH PW S D G GCin GCout */
1327 b->Args({1, 112, 112, 3, 3, 2, 2, 1, 1, 16, 1, 1});
1328 b->Args({1, 112, 112, 1, 1, 0, 0, 1, 1, 1, 16, 16});
1329 /*********************** Bottleneck 2 ************************/
1330 /* N H W KH KW PH PW S D G GCin GCout */
1331 b->Args({1, 112, 112, 1, 1, 0, 0, 1, 1, 1, 16, 64});
1332 b->Args({1, 112, 112, 3, 3, 2, 2, 2, 1, 64, 1, 1});
1333 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 64, 24});
1334 /*********************** Bottleneck 3 ************************/
1335 /* N H W KH KW PH PW S D G GCin GCout */
1336 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 72});
1337 b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 72, 1, 1});
1338 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 72, 24});
1339 /*********************** Bottleneck 4 ************************/
1340 /* N H W KH KW PH PW S D G GCin GCout */
1341//b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 24, 72});
1342 b->Args({1, 56, 56, 5, 5, 4, 4, 2, 1, 72, 1, 1});
1343 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 72, 24});
1344 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 24, 72});
1345 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 72, 40});
1346 /*********************** Bottleneck 5 ************************/
1347 /* N H W KH KW PH PW S D G GCin GCout */
1348 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 40, 120});
1349 b->Args({1, 28, 28, 5, 5, 4, 4, 1, 1, 120, 1, 1});
1350 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 120, 32});
1351 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 32, 120});
1352 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 120, 40});
1353 /*********************** Bottleneck 6 ************************/
1354 /* N H W KH KW PH PW S D G GCin GCout */
1355//b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 40, 120});
1356//b->Args({1, 28, 28, 5, 5, 4, 4, 1, 1, 120, 1, 1});
1357//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 120, 32});
1358//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 32, 120});
1359//b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 120, 40});
1360 /*********************** Bottleneck 7 ************************/
1361 /* N H W KH KW PH PW S D G GCin GCout */
1362 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 40, 240});
1363 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 240, 1, 1});
1364 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 240, 80});
1365 /*********************** Bottleneck 8 ************************/
1366 /* N H W KH KW PH PW S D G GCin GCout */
1367 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 80, 200});
1368 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 200, 1, 1});
1369 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 200, 80});
1370 /*********************** Bottleneck 9 ************************/
1371 /* N H W KH KW PH PW S D G GCin GCout */
1372 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 80, 184});
1373 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 184, 1, 1});
1374 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 184, 80});
1375 /********************** Bottleneck 10 ***********************/
1376 /* N H W KH KW PH PW S D G GCin GCout */
1377//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 80, 184});
1378//b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 184, 1, 1});
1379//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 184, 80});
1380 /********************** Bottleneck 11 ***********************/
1381 /* N H W KH KW PH PW S D G GCin GCout */
1382 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 80, 480});
1383 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 480, 1, 1});
1384 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 480, 120});
1385 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 120, 480});
1386 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 480, 112});
1387 /********************** Bottleneck 12 ***********************/
1388 /* N H W KH KW PH PW S D G GCin GCout */
1389 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 112, 672});
1390 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 672, 1, 1});
1391 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 672, 168});
1392 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 168, 672});
1393 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 672, 112});
1394 /********************** Bottleneck 13 ***********************/
1395 /* N H W KH KW PH PW S D G GCin GCout */
1396//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 112, 672});
1397 b->Args({1, 14, 14, 5, 5, 4, 4, 2, 1, 672, 1, 1});
1398 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 672, 160});
1399 /********************** Bottleneck 14 ***********************/
1400 /* N H W KH KW PH PW S D G GCin GCout */
1401 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 160, 960});
1402 b->Args({1, 7, 7, 5, 5, 4, 4, 1, 1, 960, 1, 1});
1403 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 960, 240});
1404 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 240, 960});
1405 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 960, 160});
1406 /********************** Bottleneck 15 ***********************/
1407 /* N H W KH KW PH PW S D G GCin GCout */
1408//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 160, 960});
1409//b->Args({1, 7, 7, 5, 5, 4, 4, 1, 1, 960, 1, 1});
1410//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 960, 240});
1411//b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 240, 960});
1412//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 960, 160});
1413 /************************ Last Stage ***********************/
1414 /* N H W KH KW PH PW S D G GCin GCout */
1415//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 160, 960});
1416 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 960, 1280});
1417 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1280, 1001});
1418}
1419
1420// SqueezeNet 1.0
1421static void SqueezeNetV10(benchmark::internal::Benchmark* b) {
1422 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1423
1424 /************************** Conv 1 *************************/
1425 /* N H W KH KW PH PW S D G GCin GCout */
1426 b->Args({1, 224, 224, 7, 7, 6, 6, 2, 1, 1, 3, 96});
1427 /************************** Fire 2 *************************/
1428 /* N H W KH KW PH PW S D G GCin GCout */
1429 b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 96, 16});
1430 b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 16, 64});
1431 b->Args({1, 55, 55, 3, 3, 2, 2, 1, 1, 1, 16, 64});
1432 /************************** Fire 3 *************************/
1433 /* N H W KH KW PH PW S D G GCin GCout */
1434 b->Args({1, 56, 55, 1, 1, 0, 0, 1, 1, 1, 128, 16});
1435//b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 16, 64});
1436//b->Args({1, 55, 55, 3, 3, 2, 2, 1, 1, 1, 16, 64});
1437 /************************** Fire 4 *************************/
1438 /* N H W KH KW PH PW S D G GCin GCout */
1439 b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 128, 32});
1440 b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 32, 128});
1441 b->Args({1, 55, 55, 3, 3, 2, 2, 1, 1, 1, 32, 128});
1442 /************************** Fire 5 *************************/
1443 /* N H W KH KW PH PW S D G GCin GCout */
1444 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 256, 32});
1445 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 32, 128});
1446 b->Args({1, 27, 27, 3, 3, 2, 2, 1, 1, 1, 32, 128});
1447 /************************** Fire 6 *************************/
1448 /* N H W KH KW PH PW S D G GCin GCout */
1449 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 256, 48});
1450 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 48, 192});
1451 b->Args({1, 27, 27, 3, 3, 2, 2, 1, 1, 1, 48, 192});
1452 /************************** Fire 7 *************************/
1453 /* N H W KH KW PH PW S D G GCin GCout */
1454 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 384, 48});
1455//b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 48, 192});
1456//b->Args({1, 27, 27, 3, 3, 2, 2, 1, 1, 1, 48, 192});
1457 /************************** Fire 8 *************************/
1458 /* N H W KH KW PH PW S D G GCin GCout */
1459 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 384, 64});
1460 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 64, 256});
1461 b->Args({1, 27, 27, 3, 3, 2, 2, 1, 1, 1, 64, 256});
1462 /************************** Fire 9 *************************/
1463 /* N H W KH KW PH PW S D G GCin GCout */
1464 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 512, 64});
1465 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 64, 256});
1466 b->Args({1, 13, 13, 3, 3, 2, 2, 1, 1, 1, 64, 256});
1467 /************************* Conv 10 *************************/
1468 /* N H W KH KW PH PW S D G GCin GCout */
1469 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 512, 1000});
1470}
1471
1472// SqueezeNet 1.1
1473static void SqueezeNetV11(benchmark::internal::Benchmark* b) {
1474 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1475
1476 /************************** Conv 1 *************************/
1477 /* N H W KH KW PH PW S D G GCin GCout */
1478 b->Args({1, 224, 224, 3, 3, 2, 2, 2, 1, 1, 3, 64});
1479 /************************** Fire 2 *************************/
1480 /* N H W KH KW PH PW S D G GCin GCout */
1481 b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 64, 16});
1482 b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 16, 64});
1483 b->Args({1, 55, 55, 3, 3, 2, 2, 1, 1, 1, 16, 64});
1484 /************************** Fire 3 *************************/
1485 /* N H W KH KW PH PW S D G GCin GCout */
1486 b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 128, 16});
1487//b->Args({1, 55, 55, 1, 1, 0, 0, 1, 1, 1, 16, 64});
1488//b->Args({1, 55, 55, 3, 3, 2, 2, 1, 1, 1, 16, 64});
1489 /************************** Fire 4 *************************/
1490 /* N H W KH KW PH PW S D G GCin GCout */
1491 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 128, 32});
1492 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 32, 128});
1493 b->Args({1, 27, 27, 3, 3, 2, 2, 1, 1, 1, 32, 128});
1494 /************************** Fire 5 *************************/
1495 /* N H W KH KW PH PW S D G GCin GCout */
1496 b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 256, 32});
1497//b->Args({1, 27, 27, 1, 1, 0, 0, 1, 1, 1, 32, 128});
1498//b->Args({1, 27, 27, 3, 3, 2, 2, 1, 1, 1, 32, 128});
1499 /************************** Fire 6 *************************/
1500 /* N H W KH KW PH PW S D G GCin GCout */
1501 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 256, 48});
1502 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 48, 192});
1503 b->Args({1, 13, 13, 3, 3, 2, 2, 1, 1, 1, 48, 192});
1504 /************************** Fire 7 *************************/
1505 /* N H W KH KW PH PW S D G GCin GCout */
1506 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 384, 48});
1507//b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 48, 192});
1508//b->Args({1, 13, 13, 3, 3, 2, 2, 1, 1, 1, 48, 192});
1509 /************************** Fire 8 *************************/
1510 /* N H W KH KW PH PW S D G GCin GCout */
1511 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 384, 64});
1512 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 64, 256});
1513 b->Args({1, 13, 13, 3, 3, 2, 2, 1, 1, 1, 64, 256});
1514 /************************** Fire 9 *************************/
1515 /* N H W KH KW PH PW S D G GCin GCout */
1516 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 512, 64});
1517//b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 64, 256});
1518//b->Args({1, 13, 13, 3, 3, 2, 2, 1, 1, 1, 64, 256});
1519 /************************* Conv 10 *************************/
1520 /* N H W KH KW PH PW S D G GCin GCout */
1521 b->Args({1, 13, 13, 1, 1, 0, 0, 1, 1, 1, 512, 1000});
1522}
1523
1524static void InceptionV3(benchmark::internal::Benchmark* b) {
1525 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1526
1527 /* N H W KH KW PH PW S D G GCin GCout */
1528 b->Args({1, 299, 299, 3, 3, 0, 0, 2, 1, 1, 3, 32});
1529 b->Args({1, 149, 149, 3, 3, 0, 0, 1, 1, 1, 32, 32});
1530 b->Args({1, 147, 147, 3, 3, 2, 2, 1, 1, 1, 32, 64});
1531 b->Args({1, 73, 73, 1, 1, 0, 0, 1, 1, 1, 64, 80});
1532 b->Args({1, 73, 73, 3, 3, 0, 0, 1, 1, 1, 80, 192});
1533 b->Args({1, 35, 35, 1, 1, 0, 0, 1, 1, 1, 192, 64});
1534 b->Args({1, 35, 35, 1, 1, 0, 0, 1, 1, 1, 192, 48});
1535 b->Args({1, 35, 35, 5, 5, 4, 4, 1, 1, 1, 48, 64});
1536 b->Args({1, 35, 35, 3, 3, 2, 2, 1, 1, 1, 64, 96});
1537 b->Args({1, 35, 35, 3, 3, 2, 2, 1, 1, 1, 96, 96});
1538 b->Args({1, 35, 35, 1, 1, 0, 0, 1, 1, 1, 192, 32});
1539 b->Args({1, 35, 35, 1, 1, 0, 0, 1, 1, 1, 256, 64});
1540 b->Args({1, 35, 35, 1, 1, 0, 0, 1, 1, 1, 256, 48});
1541 b->Args({1, 35, 35, 1, 1, 0, 0, 1, 1, 1, 288, 64});
1542 b->Args({1, 35, 35, 1, 1, 0, 0, 1, 1, 1, 288, 48});
1543 b->Args({1, 35, 35, 3, 3, 0, 0, 2, 1, 1, 288, 384});
1544 b->Args({1, 35, 35, 3, 3, 0, 0, 2, 1, 1, 96, 96});
1545 b->Args({1, 17, 17, 1, 1, 0, 0, 1, 1, 1, 768, 192});
1546 b->Args({1, 17, 17, 1, 1, 0, 0, 1, 1, 1, 768, 128});
1547 b->Args({1, 17, 17, 1, 7, 0, 6, 1, 1, 1, 128, 128});
1548 b->Args({1, 17, 17, 7, 1, 6, 0, 1, 1, 1, 128, 192});
1549 b->Args({1, 17, 17, 7, 1, 6, 0, 1, 1, 1, 128, 128});
1550 b->Args({1, 17, 17, 1, 7, 0, 6, 1, 1, 1, 128, 192});
1551 b->Args({1, 17, 17, 1, 1, 0, 0, 1, 1, 1, 768, 160});
1552 b->Args({1, 17, 17, 1, 7, 0, 6, 1, 1, 1, 160, 160});
1553 b->Args({1, 17, 17, 7, 1, 6, 0, 1, 1, 1, 160, 192});
1554 b->Args({1, 17, 17, 7, 1, 6, 0, 1, 1, 1, 160, 160});
1555 b->Args({1, 17, 17, 1, 7, 0, 6, 1, 1, 1, 160, 192});
1556 b->Args({1, 17, 17, 1, 7, 0, 6, 1, 1, 1, 192, 192});
1557 b->Args({1, 17, 17, 7, 1, 6, 0, 1, 1, 1, 192, 192});
1558 b->Args({1, 17, 17, 3, 3, 0, 0, 2, 1, 1, 192, 320});
1559 b->Args({1, 17, 17, 3, 3, 0, 0, 2, 1, 1, 192, 192});
1560 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 1280, 320});
1561 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 1280, 384});
1562 b->Args({1, 8, 8, 1, 3, 0, 2, 1, 1, 1, 384, 384});
1563 b->Args({1, 8, 8, 3, 1, 2, 0, 1, 1, 1, 384, 384});
1564 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 1280, 448});
1565 b->Args({1, 8, 8, 3, 3, 2, 2, 1, 1, 1, 448, 384});
1566 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 1280, 192});
1567 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 2048, 320});
1568 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 2048, 384});
1569 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 2048, 448});
1570 b->Args({1, 8, 8, 1, 1, 0, 0, 1, 1, 1, 2048, 192});
1571 b->Args({1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 2048, 1001});
1572}
1573
1574static void ResNet18(benchmark::internal::Benchmark* b) {
1575 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1576
1577 /************************* Conv 1 *************************/
1578 /* N H W KH KW PH PW S D G GCin GCout */
1579 b->Args({1, 224, 224, 7, 7, 6, 6, 2, 1, 1, 3, 64});
1580 /************************ Conv 2.X ************************/
1581 /* N H W KH KW PH PW S D G GCin GCout */
1582 b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 1, 64, 64});
1583 /************************ Conv 3.X ************************/
1584 /* N H W KH KW PH PW S D G GCin GCout */
1585 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 1, 64, 128});
1586 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 1, 128, 128});
1587 b->Args({1, 56, 56, 1, 1, 0, 0, 2, 1, 1, 64, 128});
1588 /************************ Conv 4.X ************************/
1589 /* N H W KH KW PH PW S D G GCin GCout */
1590 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 1, 128, 256});
1591 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 1, 256, 256});
1592 b->Args({1, 28, 28, 1, 1, 0, 0, 2, 1, 1, 128, 256});
1593 /************************ Conv 5.X ************************/
1594 /* N H W KH KW PH PW S D G GCin GCout */
1595 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 1, 256, 512});
1596 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 1, 512, 512});
1597 b->Args({1, 14, 14, 1, 1, 0, 0, 2, 1, 1, 256, 512});
1598}
1599
1600static void ResNet50(benchmark::internal::Benchmark* b) {
1601 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1602
1603 /************************* Conv 1 *************************/
1604 /* N H W KH KW PH PW S D G GCin GCout */
1605 b->Args({1, 224, 224, 7, 7, 6, 6, 2, 1, 1, 3, 64});
1606 /************************ Conv 2.1 ************************/
1607 /* N H W KH KW PH PW S D G GCin GCout */
1608 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 64, 64});
1609 b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 1, 64, 64});
1610 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 64, 256});
1611//b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 64, 256});
1612 /************************ Conv 2.X ************************/
1613 /* N H W KH KW PH PW S D G GCin GCout */
1614 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 256, 64});
1615//b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 1, 64, 64});
1616//b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 64, 256});
1617 /************************ Conv 3.1 ************************/
1618 /* N H W KH KW PH PW S D G GCin GCout */
1619 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 256, 128});
1620 b->Args({1, 56, 56, 3, 3, 2, 2, 2, 1, 1, 128, 128});
1621 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 128, 512});
1622 b->Args({1, 56, 56, 1, 1, 0, 0, 2, 1, 1, 256, 512});
1623 /************************ Conv 3.X ************************/
1624 /* N H W KH KW PH PW S D G GCin GCout */
1625 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 512, 128});
1626 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 1, 128, 128});
1627//b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 128, 512});
1628 /************************ Conv 4.1 ************************/
1629 /* N H W KH KW PH PW S D G GCin GCout */
1630 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 512, 256});
1631 b->Args({1, 28, 28, 3, 3, 2, 2, 2, 1, 1, 256, 256});
1632 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 256, 1024});
1633 b->Args({1, 28, 28, 1, 1, 0, 0, 2, 1, 1, 512, 1024});
1634 /************************ Conv 4.X ************************/
1635 /* N H W KH KW PH PW S D G GCin GCout */
1636 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 1024, 256});
1637 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 1, 256, 256});
1638//b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 256, 1024});
1639 /************************ Conv 5.1 ************************/
1640 /* N H W KH KW PH PW S D G GCin GCout */
1641 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 1024, 512});
1642 b->Args({1, 14, 14, 3, 3, 2, 2, 2, 1, 1, 512, 512});
1643 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 512, 2048});
1644 b->Args({1, 14, 14, 1, 1, 0, 0, 2, 1, 1, 1024, 2048});
1645 /************************ Conv 5.X ************************/
1646 /* N H W KH KW PH PW S D G GCin GCout */
1647 b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 2048, 512});
1648 b->Args({1, 7, 7, 3, 3, 2, 2, 1, 1, 1, 512, 512});
1649//b->Args({1, 7, 7, 1, 1, 0, 0, 1, 1, 1, 512, 2048});
1650}
1651
1652static void VGG(benchmark::internal::Benchmark* b) {
1653 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1654
1655 /************************* Conv 1.1 ************************/
1656 /* N H W KH KW PH PW S D G GCin GCout */
1657 b->Args({1, 224, 224, 3, 3, 2, 2, 1, 1, 1, 3, 64});
1658 /************************* Conv 1.2 ************************/
1659 /* N H W KH KW PH PW S D G GCin GCout */
1660 b->Args({1, 224, 224, 3, 3, 2, 2, 1, 1, 1, 64, 64});
1661
1662 /************************* Conv 2.1 ************************/
1663 /* N H W KH KW PH PW S D G GCin GCout */
1664 b->Args({1, 112, 112, 3, 3, 2, 2, 1, 1, 1, 64, 128});
1665 /************************* Conv 2.2 ************************/
1666 /* N H W KH KW PH PW S D G GCin GCout */
1667 b->Args({1, 112, 112, 3, 3, 2, 2, 1, 1, 1, 128, 128});
1668
1669 /************************* Conv 3.1 ************************/
1670 /* N H W KH KW PH PW S D G GCin GCout */
1671 b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 1, 128, 256});
1672 /************************* Conv 3.2 ************************/
1673 /* N H W KH KW PH PW S D G GCin GCout */
1674 b->Args({1, 56, 56, 3, 3, 2, 2, 1, 1, 1, 256, 256});
1675 /************************* Conv 3.3 ************************/
1676 /* N H W KH KW PH PW S D G GCin GCout */
1677 b->Args({1, 56, 56, 1, 1, 0, 0, 1, 1, 1, 256, 256});
1678
1679 /************************* Conv 4.1 ************************/
1680 /* N H W KH KW PH PW S D G GCin GCout */
1681 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 1, 256, 512});
1682 /************************* Conv 4.2 ************************/
1683 /* N H W KH KW PH PW S D G GCin GCout */
1684 b->Args({1, 28, 28, 3, 3, 2, 2, 1, 1, 1, 512, 512});
1685 /************************* Conv 4.3 ************************/
1686 /* N H W KH KW PH PW S D G GCin GCout */
1687 b->Args({1, 28, 28, 1, 1, 0, 0, 1, 1, 1, 512, 512});
1688
1689 /************************* Conv 5.X ************************/
1690 /* N H W KH KW PH PW S D G GCin GCout */
1691 b->Args({1, 14, 14, 3, 3, 2, 2, 1, 1, 1, 512, 512});
1692 /************************* Conv 5.3 ************************/
1693 /* N H W KH KW PH PW S D G GCin GCout */
1694 b->Args({1, 14, 14, 1, 1, 0, 0, 1, 1, 1, 512, 512});
1695}
1696
1697// SRCNN (9-1-5)
1698static void SRCNN915(benchmark::internal::Benchmark* b) {
1699 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1700
1701 /* N H W KH KW PH PW S D G GCin GCout */
1702 b->Args({1, 384, 384, 9, 9, 0, 0, 1, 1, 1, 1, 64});
1703 b->Args({1, 376, 376, 1, 1, 0, 0, 1, 1, 1, 64, 32});
1704 b->Args({1, 376, 376, 5, 5, 0, 0, 1, 1, 1, 32, 1});
1705}
1706
1707// SRCNN (9-3-5)
1708static void SRCNN935(benchmark::internal::Benchmark* b) {
1709 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1710
1711 /* N H W KH KW PH PW S D G GCin GCout */
1712 b->Args({1, 384, 384, 9, 9, 0, 0, 1, 1, 1, 1, 64});
1713 b->Args({1, 376, 376, 3, 3, 0, 0, 1, 1, 1, 64, 32});
1714 b->Args({1, 374, 374, 5, 5, 0, 0, 1, 1, 1, 32, 1});
1715}
1716
1717// SRCNN (9-5-5)
1718static void SRCNN955(benchmark::internal::Benchmark* b) {
1719 b->ArgNames({"N", "H", "W", "KH", "KW", "PH", "PW", "S", "D", "G", "GCin", "GCout"});
1720
1721 /* N H W KH KW PH PW S D G GCin GCout */
1722 b->Args({1, 384, 384, 9, 9, 0, 0, 1, 1, 1, 1, 64});
1723 b->Args({1, 376, 376, 5, 5, 0, 0, 1, 1, 1, 64, 32});
1724 b->Args({1, 372, 372, 5, 5, 0, 0, 1, 1, 1, 32, 1});
1725}
1726
1727BENCHMARK_CAPTURE(xnnpack_convolution_f32, mobilenet_v1, "MobileNet v1")->Apply(MobileNetV1)->UseRealTime();
1728BENCHMARK_CAPTURE(xnnpack_convolution_f32, mobilenet_v2, "MobileNet v2")->Apply(MobileNetV2)->UseRealTime();
1729BENCHMARK_CAPTURE(xnnpack_convolution_f32, mobilenet_v3_small, "MobileNet v3 Small")->Apply(MobileNetV3Small)->UseRealTime();
1730BENCHMARK_CAPTURE(xnnpack_convolution_f32, mobilenet_v3_large, "MobileNet v3 Large")->Apply(MobileNetV3Large)->UseRealTime();
1731BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime();
1732BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime();
1733BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime();
1734BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime();
1735BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime();
1736BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v2_x05, "ShuffleNet v2 0.5X")->Apply(ShuffleNetV2X05)->UseRealTime();
1737BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v2_x10, "ShuffleNet v2 1.0X")->Apply(ShuffleNetV2X10)->UseRealTime();
1738BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v2_x15, "ShuffleNet v2 1.5X")->Apply(ShuffleNetV2X15)->UseRealTime();
1739BENCHMARK_CAPTURE(xnnpack_convolution_f32, shufflenet_v2_x20, "ShuffleNet v2 2.0X")->Apply(ShuffleNetV2X20)->UseRealTime();
1740BENCHMARK_CAPTURE(xnnpack_convolution_f32, squeezenet_v10, "SqueezeNet 1.0")->Apply(SqueezeNetV10)->UseRealTime();
1741BENCHMARK_CAPTURE(xnnpack_convolution_f32, squeezenet_v11, "SqueezeNet 1.1")->Apply(SqueezeNetV11)->UseRealTime();
1742BENCHMARK_CAPTURE(xnnpack_convolution_f32, inception_v3, "Inception v3")->Apply(InceptionV3)->UseRealTime();
1743BENCHMARK_CAPTURE(xnnpack_convolution_f32, resnet18, "ResNet-18")->Apply(ResNet18)->UseRealTime();
1744BENCHMARK_CAPTURE(xnnpack_convolution_f32, resnet50, "ResNet-50")->Apply(ResNet50)->UseRealTime();
1745BENCHMARK_CAPTURE(xnnpack_convolution_f32, vgg, "VGG")->Apply(VGG)->UseRealTime();
1746BENCHMARK_CAPTURE(xnnpack_convolution_f32, srcnn915, "SRCNN (9-1-5)")->Apply(SRCNN915)->UseRealTime();
1747BENCHMARK_CAPTURE(xnnpack_convolution_f32, srcnn935, "SRCNN (9-3-5)")->Apply(SRCNN935)->UseRealTime();
1748BENCHMARK_CAPTURE(xnnpack_convolution_f32, srcnn955, "SRCNN (9-5-5)")->Apply(SRCNN955)->UseRealTime();
1749
1750BENCHMARK_CAPTURE(xnnpack_convolution_q8, mobilenet_v1, "MobileNet v1")->Apply(MobileNetV1)->UseRealTime();
1751BENCHMARK_CAPTURE(xnnpack_convolution_q8, mobilenet_v2, "MobileNet v2")->Apply(MobileNetV2)->UseRealTime();
1752BENCHMARK_CAPTURE(xnnpack_convolution_q8, mobilenet_v3_small, "MobileNet v3 Small")->Apply(MobileNetV3Small)->UseRealTime();
1753BENCHMARK_CAPTURE(xnnpack_convolution_q8, mobilenet_v3_large, "MobileNet v3 Large")->Apply(MobileNetV3Large)->UseRealTime();
1754BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime();
1755BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime();
1756BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime();
1757BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime();
1758BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime();
1759BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v2_x05, "ShuffleNet v2 0.5X")->Apply(ShuffleNetV2X05)->UseRealTime();
1760BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v2_x10, "ShuffleNet v2 1.0X")->Apply(ShuffleNetV2X10)->UseRealTime();
1761BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v2_x15, "ShuffleNet v2 1.5X")->Apply(ShuffleNetV2X15)->UseRealTime();
1762BENCHMARK_CAPTURE(xnnpack_convolution_q8, shufflenet_v2_x20, "ShuffleNet v2 2.0X")->Apply(ShuffleNetV2X20)->UseRealTime();
1763BENCHMARK_CAPTURE(xnnpack_convolution_q8, squeezenet_v10, "SqueezeNet 1.0")->Apply(SqueezeNetV10)->UseRealTime();
1764BENCHMARK_CAPTURE(xnnpack_convolution_q8, squeezenet_v11, "SqueezeNet 1.1")->Apply(SqueezeNetV11)->UseRealTime();
1765BENCHMARK_CAPTURE(xnnpack_convolution_q8, inception_v3, "Inception v3")->Apply(InceptionV3)->UseRealTime();
1766BENCHMARK_CAPTURE(xnnpack_convolution_q8, resnet18, "ResNet-18")->Apply(ResNet18)->UseRealTime();
1767BENCHMARK_CAPTURE(xnnpack_convolution_q8, resnet50, "ResNet-50")->Apply(ResNet50)->UseRealTime();
1768BENCHMARK_CAPTURE(xnnpack_convolution_q8, vgg, "VGG")->Apply(VGG)->UseRealTime();
1769BENCHMARK_CAPTURE(xnnpack_convolution_q8, srcnn915, "SRCNN (9-1-5)")->Apply(SRCNN915)->UseRealTime();
1770BENCHMARK_CAPTURE(xnnpack_convolution_q8, srcnn935, "SRCNN (9-3-5)")->Apply(SRCNN935)->UseRealTime();
1771BENCHMARK_CAPTURE(xnnpack_convolution_q8, srcnn955, "SRCNN (9-5-5)")->Apply(SRCNN955)->UseRealTime();
1772
1773#ifdef BENCHMARK_TENSORFLOW_LITE
1774 BENCHMARK_CAPTURE(tflite_convolution_f32, mobilenet_v1, "MobileNet v1")->Apply(MobileNetV1)->UseRealTime();
1775 BENCHMARK_CAPTURE(tflite_convolution_f32, mobilenet_v2, "MobileNet v2")->Apply(MobileNetV2)->UseRealTime();
1776 BENCHMARK_CAPTURE(tflite_convolution_f32, mobilenet_v3_small, "MobileNet v3 Small")->Apply(MobileNetV3Small)->UseRealTime();
1777 BENCHMARK_CAPTURE(tflite_convolution_f32, mobilenet_v3_large, "MobileNet v3 Large")->Apply(MobileNetV3Large)->UseRealTime();
1778 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime();
1779 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime();
1780 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime();
1781 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime();
1782 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime();
1783 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v2_x05, "ShuffleNet v2 0.5X")->Apply(ShuffleNetV2X05)->UseRealTime();
1784 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v2_x10, "ShuffleNet v2 1.0X")->Apply(ShuffleNetV2X10)->UseRealTime();
1785 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v2_x15, "ShuffleNet v2 1.5X")->Apply(ShuffleNetV2X15)->UseRealTime();
1786 BENCHMARK_CAPTURE(tflite_convolution_f32, shufflenet_v2_x20, "ShuffleNet v2 2.0X")->Apply(ShuffleNetV2X20)->UseRealTime();
1787 BENCHMARK_CAPTURE(tflite_convolution_f32, squeezenet_v10, "SqueezeNet 1.0")->Apply(SqueezeNetV10)->UseRealTime();
1788 BENCHMARK_CAPTURE(tflite_convolution_f32, squeezenet_v11, "SqueezeNet 1.1")->Apply(SqueezeNetV11)->UseRealTime();
1789 BENCHMARK_CAPTURE(tflite_convolution_f32, inception_v3, "Inception v3")->Apply(InceptionV3)->UseRealTime();
1790 BENCHMARK_CAPTURE(tflite_convolution_f32, resnet18, "ResNet-18")->Apply(ResNet18)->UseRealTime();
1791 BENCHMARK_CAPTURE(tflite_convolution_f32, resnet50, "ResNet-50")->Apply(ResNet50)->UseRealTime();
1792 BENCHMARK_CAPTURE(tflite_convolution_f32, vgg, "VGG")->Apply(VGG)->UseRealTime();
1793 BENCHMARK_CAPTURE(tflite_convolution_f32, srcnn915, "SRCNN (9-1-5)")->Apply(SRCNN915)->UseRealTime();
1794 BENCHMARK_CAPTURE(tflite_convolution_f32, srcnn935, "SRCNN (9-3-5)")->Apply(SRCNN935)->UseRealTime();
1795 BENCHMARK_CAPTURE(tflite_convolution_f32, srcnn955, "SRCNN (9-5-5)")->Apply(SRCNN955)->UseRealTime();
1796#endif // BENCHMARK_TENSORFLOW_LITE
1797
1798#ifdef BENCHMARK_ARM_COMPUTE_LIBRARY
1799 BENCHMARK_CAPTURE(armcl_convolution_f32, mobilenet_v1, "MobileNet v1")->Apply(MobileNetV1)->UseRealTime();
1800 BENCHMARK_CAPTURE(armcl_convolution_f32, mobilenet_v2, "MobileNet v2")->Apply(MobileNetV2)->UseRealTime();
1801 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime();
1802 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime();
1803 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime();
1804 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime();
1805 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime();
1806 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v2_x05, "ShuffleNet v2 0.5X")->Apply(ShuffleNetV2X05)->UseRealTime();
1807 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v2_x10, "ShuffleNet v2 1.0X")->Apply(ShuffleNetV2X10)->UseRealTime();
1808 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v2_x15, "ShuffleNet v2 1.5X")->Apply(ShuffleNetV2X15)->UseRealTime();
1809 BENCHMARK_CAPTURE(armcl_convolution_f32, shufflenet_v2_x20, "ShuffleNet v2 2.0X")->Apply(ShuffleNetV2X20)->UseRealTime();
1810 BENCHMARK_CAPTURE(armcl_convolution_f32, squeezenet_v10, "SqueezeNet 1.0")->Apply(SqueezeNetV10)->UseRealTime();
1811 BENCHMARK_CAPTURE(armcl_convolution_f32, squeezenet_v11, "SqueezeNet 1.1")->Apply(SqueezeNetV11)->UseRealTime();
1812 BENCHMARK_CAPTURE(armcl_convolution_f32, inception_v3, "Inception v3")->Apply(InceptionV3)->UseRealTime();
1813 BENCHMARK_CAPTURE(armcl_convolution_f32, resnet18, "ResNet-18")->Apply(ResNet18)->UseRealTime();
1814 BENCHMARK_CAPTURE(armcl_convolution_f32, resnet50, "ResNet-50")->Apply(ResNet50)->UseRealTime();
1815 BENCHMARK_CAPTURE(armcl_convolution_f32, vgg, "VGG")->Apply(VGG)->UseRealTime();
1816 BENCHMARK_CAPTURE(armcl_convolution_f32, srcnn915, "SRCNN (9-1-5)")->Apply(SRCNN915)->UseRealTime();
1817 BENCHMARK_CAPTURE(armcl_convolution_f32, srcnn935, "SRCNN (9-3-5)")->Apply(SRCNN935)->UseRealTime();
1818 BENCHMARK_CAPTURE(armcl_convolution_f32, srcnn955, "SRCNN (9-5-5)")->Apply(SRCNN955)->UseRealTime();
1819#endif // BENCHMARK_ARM_COMPUTE_LIBRARY
1820
1821#ifndef XNNPACK_BENCHMARK_NO_MAIN
1822BENCHMARK_MAIN();
1823#endif