XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 4 | // Copyright 2020 Google LLC |
| 5 | // |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 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> |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 10 | #include <array> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 11 | #include <cmath> |
| 12 | #include <functional> |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 13 | #include <limits> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <random> |
| 15 | #include <vector> |
| 16 | |
| 17 | #include <xnnpack.h> |
| 18 | |
| 19 | #include <benchmark/benchmark.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 20 | #include "bench/utils.h" |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 21 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 22 | #include "flatbuffers/include/flatbuffers/flatbuffers.h" |
| 23 | #include "tensorflow/lite/interpreter.h" |
| 24 | #include "tensorflow/lite/kernels/register.h" |
| 25 | #include "tensorflow/lite/model.h" |
| 26 | #include "tensorflow/lite/schema/schema_generated.h" |
| 27 | #include "tensorflow/lite/version.h" |
| 28 | #endif // BENCHMARK_TENSORFLOW_LITE |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 30 | |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 31 | static void xnnpack_sigmoid_f32(benchmark::State& state) { |
| 32 | const size_t batch_size = state.range(0); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 33 | |
| 34 | std::random_device random_device; |
| 35 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 36 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 37 | |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 38 | std::vector<float> input(batch_size + XNN_EXTRA_BYTES / sizeof(float)); |
| 39 | std::vector<float> output(batch_size); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 40 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 41 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 42 | |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 43 | xnn_status status = xnn_initialize(nullptr /* allocator */); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 44 | if (status != xnn_status_success) { |
| 45 | state.SkipWithError("failed to initialize XNNPACK"); |
| 46 | return; |
| 47 | } |
| 48 | |
| 49 | xnn_operator_t sigmoid_op = nullptr; |
| 50 | status = xnn_create_sigmoid_nc_f32( |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 51 | 1 /* channels */, 1 /* input stride */, 1 /* output stride */, |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 52 | 0 /* flags */, &sigmoid_op); |
| 53 | if (status != xnn_status_success || sigmoid_op == nullptr) { |
| 54 | state.SkipWithError("failed to create Sigmoid operator"); |
| 55 | return; |
| 56 | } |
| 57 | |
| 58 | status = xnn_setup_sigmoid_nc_f32( |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 59 | sigmoid_op, batch_size, |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 60 | input.data(), output.data(), |
| 61 | nullptr /* thread pool */); |
| 62 | if (status != xnn_status_success) { |
| 63 | state.SkipWithError("failed to setup Sigmoid operator"); |
| 64 | return; |
| 65 | } |
| 66 | |
| 67 | for (auto _ : state) { |
| 68 | status = xnn_run_operator(sigmoid_op, nullptr /* thread pool */); |
| 69 | if (status != xnn_status_success) { |
| 70 | state.SkipWithError("failed to run Sigmoid operator"); |
| 71 | return; |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | status = xnn_delete_operator(sigmoid_op); |
| 76 | if (status != xnn_status_success) { |
| 77 | state.SkipWithError("failed to delete Sigmoid operator"); |
| 78 | return; |
| 79 | } |
| 80 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 81 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 82 | if (cpu_frequency != 0) { |
| 83 | state.counters["cpufreq"] = cpu_frequency; |
| 84 | } |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 85 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 86 | state.counters["elements"] = |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 87 | benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 88 | |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 89 | const size_t bytes_per_iteration = 2 * batch_size * sizeof(float); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 90 | state.counters["bytes"] = |
| 91 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 92 | } |
| 93 | |
Marat Dukhan | 9084fc8 | 2021-12-31 10:16:09 -0800 | [diff] [blame] | 94 | #ifndef XNN_NO_QS8_OPERATORS |
| 95 | static void xnnpack_sigmoid_qs8(benchmark::State& state) { |
| 96 | const size_t batch_size = state.range(0); |
| 97 | |
| 98 | std::random_device random_device; |
| 99 | auto rng = std::mt19937(random_device()); |
| 100 | auto i8rng = std::bind( |
| 101 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), |
| 102 | std::ref(rng)); |
| 103 | |
| 104 | std::vector<int8_t> input(batch_size + XNN_EXTRA_BYTES / sizeof(int8_t)); |
| 105 | std::vector<int8_t> output(batch_size); |
| 106 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 107 | std::fill(output.begin(), output.end(), INT8_C(0xA5)); |
| 108 | |
| 109 | xnn_status status = xnn_initialize(nullptr /* allocator */); |
| 110 | if (status != xnn_status_success) { |
| 111 | state.SkipWithError("failed to initialize XNNPACK"); |
| 112 | return; |
| 113 | } |
| 114 | |
| 115 | xnn_operator_t sigmoid_op = nullptr; |
| 116 | status = xnn_create_sigmoid_nc_qs8( |
| 117 | 1 /* channels */, 1 /* input stride */, 1 /* output stride */, |
| 118 | 1 /* input zero point */, 1.0f /* input scale */, |
| 119 | -128 /* output zero point */, 1.0f / 256.0f /* output scale */, |
| 120 | std::numeric_limits<int8_t>::min() /* output min */, std::numeric_limits<int8_t>::max() /* output max */, |
| 121 | 0 /* flags */, &sigmoid_op); |
| 122 | if (status != xnn_status_success || sigmoid_op == nullptr) { |
| 123 | state.SkipWithError("failed to create Sigmoid operator"); |
| 124 | return; |
| 125 | } |
| 126 | |
| 127 | status = xnn_setup_sigmoid_nc_qs8( |
| 128 | sigmoid_op, batch_size, |
| 129 | input.data(), output.data(), |
| 130 | nullptr /* thread pool */); |
| 131 | if (status != xnn_status_success) { |
| 132 | state.SkipWithError("failed to setup Sigmoid operator"); |
| 133 | return; |
| 134 | } |
| 135 | |
| 136 | for (auto _ : state) { |
| 137 | status = xnn_run_operator(sigmoid_op, nullptr /* thread pool */); |
| 138 | if (status != xnn_status_success) { |
| 139 | state.SkipWithError("failed to run Sigmoid operator"); |
| 140 | return; |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | status = xnn_delete_operator(sigmoid_op); |
| 145 | if (status != xnn_status_success) { |
| 146 | state.SkipWithError("failed to delete Sigmoid operator"); |
| 147 | return; |
| 148 | } |
| 149 | |
| 150 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 151 | if (cpu_frequency != 0) { |
| 152 | state.counters["cpufreq"] = cpu_frequency; |
| 153 | } |
| 154 | |
| 155 | state.counters["elements"] = |
| 156 | benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate); |
| 157 | |
| 158 | const size_t bytes_per_iteration = 2 * batch_size * sizeof(int8_t); |
| 159 | state.counters["bytes"] = |
| 160 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 161 | } |
| 162 | #endif // XNN_NO_QS8_OPERATORS |
| 163 | |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 164 | #ifndef XNN_NO_QU8_OPERATORS |
| 165 | static void xnnpack_sigmoid_qu8(benchmark::State& state) { |
| 166 | const size_t batch_size = state.range(0); |
| 167 | |
| 168 | std::random_device random_device; |
| 169 | auto rng = std::mt19937(random_device()); |
| 170 | auto u8rng = std::bind( |
| 171 | std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
| 172 | |
| 173 | std::vector<uint8_t> input(batch_size + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 174 | std::vector<uint8_t> output(batch_size); |
| 175 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
Marat Dukhan | 9084fc8 | 2021-12-31 10:16:09 -0800 | [diff] [blame] | 176 | std::fill(output.begin(), output.end(), UINT8_C(0xA5)); |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 177 | |
| 178 | xnn_status status = xnn_initialize(nullptr /* allocator */); |
| 179 | if (status != xnn_status_success) { |
| 180 | state.SkipWithError("failed to initialize XNNPACK"); |
| 181 | return; |
| 182 | } |
| 183 | |
| 184 | xnn_operator_t sigmoid_op = nullptr; |
| 185 | status = xnn_create_sigmoid_nc_qu8( |
| 186 | 1 /* channels */, 1 /* input stride */, 1 /* output stride */, |
Marat Dukhan | 9084fc8 | 2021-12-31 10:16:09 -0800 | [diff] [blame] | 187 | 128 /* input zero point */, 1.0f /* input scale */, |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 188 | 0 /* output zero point */, 1.0f / 256.0f /* output scale */, |
Marat Dukhan | 9084fc8 | 2021-12-31 10:16:09 -0800 | [diff] [blame] | 189 | std::numeric_limits<uint8_t>::min() /* output min */, std::numeric_limits<uint8_t>::max() /* output max */, |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 190 | 0 /* flags */, &sigmoid_op); |
| 191 | if (status != xnn_status_success || sigmoid_op == nullptr) { |
| 192 | state.SkipWithError("failed to create Sigmoid operator"); |
| 193 | return; |
| 194 | } |
| 195 | |
| 196 | status = xnn_setup_sigmoid_nc_qu8( |
| 197 | sigmoid_op, batch_size, |
| 198 | input.data(), output.data(), |
| 199 | nullptr /* thread pool */); |
| 200 | if (status != xnn_status_success) { |
| 201 | state.SkipWithError("failed to setup Sigmoid operator"); |
| 202 | return; |
| 203 | } |
| 204 | |
| 205 | for (auto _ : state) { |
| 206 | status = xnn_run_operator(sigmoid_op, nullptr /* thread pool */); |
| 207 | if (status != xnn_status_success) { |
| 208 | state.SkipWithError("failed to run Sigmoid operator"); |
| 209 | return; |
| 210 | } |
| 211 | } |
| 212 | |
| 213 | status = xnn_delete_operator(sigmoid_op); |
| 214 | if (status != xnn_status_success) { |
| 215 | state.SkipWithError("failed to delete Sigmoid operator"); |
| 216 | return; |
| 217 | } |
| 218 | |
| 219 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 220 | if (cpu_frequency != 0) { |
| 221 | state.counters["cpufreq"] = cpu_frequency; |
| 222 | } |
| 223 | |
| 224 | state.counters["elements"] = |
| 225 | benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate); |
| 226 | |
| 227 | const size_t bytes_per_iteration = 2 * batch_size * sizeof(uint8_t); |
| 228 | state.counters["bytes"] = |
| 229 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 230 | } |
| 231 | #endif // XNN_NO_QU8_OPERATORS |
| 232 | |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 233 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 234 | static void tflite_sigmoid_f32(benchmark::State& state) { |
| 235 | const size_t batch_size = state.range(0); |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 236 | |
| 237 | std::random_device random_device; |
| 238 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 239 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 240 | |
| 241 | flatbuffers::FlatBufferBuilder builder; |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 242 | const flatbuffers::Offset<tflite::OperatorCode> operator_code = |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 243 | CreateOperatorCode(builder, tflite::BuiltinOperator_LOGISTIC); |
| 244 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 245 | const std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers{{ |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 246 | tflite::CreateBuffer(builder, builder.CreateVector({})), |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 247 | }}; |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 248 | |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 249 | const std::array<int32_t, 1> shape{{ |
| 250 | static_cast<int32_t>(batch_size) |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 251 | }}; |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 252 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 253 | const std::array<flatbuffers::Offset<tflite::Tensor>, 2> tensors{{ |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 254 | tflite::CreateTensor(builder, |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 255 | builder.CreateVector<int32_t>(shape.data(), shape.size()), |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 256 | tflite::TensorType_FLOAT32), |
| 257 | tflite::CreateTensor(builder, |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 258 | builder.CreateVector<int32_t>(shape.data(), shape.size()), |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 259 | tflite::TensorType_FLOAT32), |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 260 | }}; |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 261 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 262 | const std::array<int32_t, 1> op_inputs{{ 0 }}; |
| 263 | const std::array<int32_t, 1> op_outputs{{ 1 }}; |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 264 | flatbuffers::Offset<tflite::Operator> op = tflite::CreateOperator( |
| 265 | builder, |
| 266 | 0 /* opcode_index */, |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 267 | builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()), |
| 268 | builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size())); |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 269 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 270 | const std::array<int32_t, 1> graph_inputs{{ 0 }}; |
| 271 | const std::array<int32_t, 1> graph_outputs{{ 1 }}; |
| 272 | const flatbuffers::Offset<tflite::SubGraph> subgraph = tflite::CreateSubGraph( |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 273 | builder, |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 274 | builder.CreateVector(tensors.data(), tensors.size()), |
| 275 | builder.CreateVector<int32_t>(graph_inputs.data(), graph_inputs.size()), |
| 276 | builder.CreateVector<int32_t>(graph_outputs.data(), graph_outputs.size()), |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 277 | builder.CreateVector(&op, 1)); |
| 278 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 279 | const flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder, |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 280 | TFLITE_SCHEMA_VERSION, |
| 281 | builder.CreateVector(&operator_code, 1), |
| 282 | builder.CreateVector(&subgraph, 1), |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 283 | builder.CreateString("Sigmoid model"), |
| 284 | builder.CreateVector(buffers.data(), buffers.size())); |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 285 | |
| 286 | builder.Finish(model_buffer); |
| 287 | |
| 288 | const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer()); |
Chao Mei | f9fdaa7 | 2021-05-18 23:04:34 -0700 | [diff] [blame] | 289 | tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates resolver; |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 290 | tflite::InterpreterBuilder interpreterBuilder(model, resolver); |
| 291 | std::unique_ptr<tflite::Interpreter> interpreter; |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 292 | if (interpreterBuilder(&interpreter) != kTfLiteOk || interpreter == nullptr) { |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 293 | state.SkipWithError("failed to create TFLite interpreter"); |
| 294 | return; |
| 295 | } |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 296 | interpreter->SetNumThreads(1); |
| 297 | |
| 298 | if (interpreter->AllocateTensors() != kTfLiteOk) { |
| 299 | state.SkipWithError("failed to allocate tensors"); |
| 300 | return; |
| 301 | } |
| 302 | |
| 303 | std::generate( |
| 304 | interpreter->typed_tensor<float>(0), |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 305 | interpreter->typed_tensor<float>(0) + batch_size, |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 306 | std::ref(f32rng)); |
| 307 | |
| 308 | for (auto _ : state) { |
| 309 | if (interpreter->Invoke() != kTfLiteOk) { |
| 310 | state.SkipWithError("failed to invoke TFLite interpreter"); |
| 311 | return; |
| 312 | } |
| 313 | } |
| 314 | |
Marat Dukhan | 401d97b | 2020-12-02 12:32:09 -0800 | [diff] [blame] | 315 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 316 | if (cpu_frequency != 0) { |
| 317 | state.counters["cpufreq"] = cpu_frequency; |
| 318 | } |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 319 | |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 320 | state.counters["elements"] = |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 321 | benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate); |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 322 | |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 323 | const size_t bytes_per_iteration = 2 * batch_size * sizeof(float); |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 324 | state.counters["bytes"] = |
| 325 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 326 | |
| 327 | interpreter.reset(); |
| 328 | } |
Marat Dukhan | 9084fc8 | 2021-12-31 10:16:09 -0800 | [diff] [blame] | 329 | |
| 330 | static void tflite_sigmoid_qs8(benchmark::State& state) { |
| 331 | const size_t batch_size = state.range(0); |
| 332 | |
| 333 | std::random_device random_device; |
| 334 | auto rng = std::mt19937(random_device()); |
| 335 | auto i8rng = std::bind( |
| 336 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), |
| 337 | std::ref(rng)); |
| 338 | |
| 339 | flatbuffers::FlatBufferBuilder builder; |
| 340 | const flatbuffers::Offset<tflite::OperatorCode> operator_code = |
| 341 | CreateOperatorCode(builder, tflite::BuiltinOperator_LOGISTIC); |
| 342 | |
| 343 | const std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers{{ |
| 344 | tflite::CreateBuffer(builder, builder.CreateVector({})), |
| 345 | }}; |
| 346 | |
| 347 | const std::array<int32_t, 1> shape{{ |
| 348 | static_cast<int32_t>(batch_size) |
| 349 | }}; |
| 350 | |
| 351 | const std::array<flatbuffers::Offset<tflite::Tensor>, 2> tensors{{ |
| 352 | tflite::CreateTensor(builder, |
| 353 | builder.CreateVector<int32_t>(shape.data(), shape.size()), |
| 354 | tflite::TensorType_INT8, 0 /* buffer */, 0 /* name */, |
| 355 | tflite::CreateQuantizationParameters(builder, |
| 356 | 0 /*min*/, 0 /*max*/, |
| 357 | builder.CreateVector<float>({1.0f /* scale */}), |
| 358 | builder.CreateVector<int64_t>({1 /* zero point */}))), |
| 359 | tflite::CreateTensor(builder, |
| 360 | builder.CreateVector<int32_t>(shape.data(), shape.size()), |
| 361 | tflite::TensorType_INT8, 0 /* buffer */, 0 /* name */, |
| 362 | tflite::CreateQuantizationParameters(builder, |
| 363 | 0 /*min*/, 0 /*max*/, |
| 364 | builder.CreateVector<float>({1.0f / 256.0f /* scale */}), |
| 365 | builder.CreateVector<int64_t>({-128 /* zero point */}))), |
| 366 | }}; |
| 367 | |
| 368 | const std::array<int32_t, 1> op_inputs{{ 0 }}; |
| 369 | const std::array<int32_t, 1> op_outputs{{ 1 }}; |
| 370 | flatbuffers::Offset<tflite::Operator> op = tflite::CreateOperator( |
| 371 | builder, |
| 372 | 0 /* opcode_index */, |
| 373 | builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()), |
| 374 | builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size())); |
| 375 | |
| 376 | const std::array<int32_t, 1> graph_inputs{{ 0 }}; |
| 377 | const std::array<int32_t, 1> graph_outputs{{ 1 }}; |
| 378 | const flatbuffers::Offset<tflite::SubGraph> subgraph = tflite::CreateSubGraph( |
| 379 | builder, |
| 380 | builder.CreateVector(tensors.data(), tensors.size()), |
| 381 | builder.CreateVector<int32_t>(graph_inputs.data(), graph_inputs.size()), |
| 382 | builder.CreateVector<int32_t>(graph_outputs.data(), graph_outputs.size()), |
| 383 | builder.CreateVector(&op, 1)); |
| 384 | |
| 385 | const flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder, |
| 386 | TFLITE_SCHEMA_VERSION, |
| 387 | builder.CreateVector(&operator_code, 1), |
| 388 | builder.CreateVector(&subgraph, 1), |
| 389 | builder.CreateString("Sigmoid model"), |
| 390 | builder.CreateVector(buffers.data(), buffers.size())); |
| 391 | |
| 392 | builder.Finish(model_buffer); |
| 393 | |
| 394 | const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer()); |
| 395 | tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates resolver; |
| 396 | tflite::InterpreterBuilder interpreterBuilder(model, resolver); |
| 397 | std::unique_ptr<tflite::Interpreter> interpreter; |
| 398 | if (interpreterBuilder(&interpreter) != kTfLiteOk || interpreter == nullptr) { |
| 399 | state.SkipWithError("failed to create TFLite interpreter"); |
| 400 | return; |
| 401 | } |
| 402 | interpreter->SetNumThreads(1); |
| 403 | |
| 404 | if (interpreter->AllocateTensors() != kTfLiteOk) { |
| 405 | state.SkipWithError("failed to allocate tensors"); |
| 406 | return; |
| 407 | } |
| 408 | |
| 409 | std::generate( |
| 410 | interpreter->typed_tensor<int8_t>(0), |
| 411 | interpreter->typed_tensor<int8_t>(0) + batch_size, |
| 412 | std::ref(i8rng)); |
| 413 | |
| 414 | for (auto _ : state) { |
| 415 | if (interpreter->Invoke() != kTfLiteOk) { |
| 416 | state.SkipWithError("failed to invoke TFLite interpreter"); |
| 417 | return; |
| 418 | } |
| 419 | } |
| 420 | |
| 421 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 422 | if (cpu_frequency != 0) { |
| 423 | state.counters["cpufreq"] = cpu_frequency; |
| 424 | } |
| 425 | |
| 426 | state.counters["elements"] = |
| 427 | benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate); |
| 428 | |
| 429 | const size_t bytes_per_iteration = 2 * batch_size * sizeof(int8_t); |
| 430 | state.counters["bytes"] = |
| 431 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 432 | |
| 433 | interpreter.reset(); |
| 434 | } |
| 435 | |
| 436 | static void tflite_sigmoid_qu8(benchmark::State& state) { |
| 437 | const size_t batch_size = state.range(0); |
| 438 | |
| 439 | std::random_device random_device; |
| 440 | auto rng = std::mt19937(random_device()); |
| 441 | auto u8rng = std::bind( |
| 442 | std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), |
| 443 | std::ref(rng)); |
| 444 | |
| 445 | flatbuffers::FlatBufferBuilder builder; |
| 446 | const flatbuffers::Offset<tflite::OperatorCode> operator_code = |
| 447 | CreateOperatorCode(builder, tflite::BuiltinOperator_LOGISTIC); |
| 448 | |
| 449 | const std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers{{ |
| 450 | tflite::CreateBuffer(builder, builder.CreateVector({})), |
| 451 | }}; |
| 452 | |
| 453 | const std::array<int32_t, 1> shape{{ |
| 454 | static_cast<int32_t>(batch_size) |
| 455 | }}; |
| 456 | |
| 457 | const std::array<flatbuffers::Offset<tflite::Tensor>, 2> tensors{{ |
| 458 | tflite::CreateTensor(builder, |
| 459 | builder.CreateVector<int32_t>(shape.data(), shape.size()), |
| 460 | tflite::TensorType_UINT8, 0 /* buffer */, 0 /* name */, |
| 461 | tflite::CreateQuantizationParameters(builder, |
| 462 | 0 /*min*/, 0 /*max*/, |
| 463 | builder.CreateVector<float>({1.0f /* scale */}), |
| 464 | builder.CreateVector<int64_t>({128 /* zero point */}))), |
| 465 | tflite::CreateTensor(builder, |
| 466 | builder.CreateVector<int32_t>(shape.data(), shape.size()), |
| 467 | tflite::TensorType_UINT8, 0 /* buffer */, 0 /* name */, |
| 468 | tflite::CreateQuantizationParameters(builder, |
| 469 | 0 /*min*/, 0 /*max*/, |
| 470 | builder.CreateVector<float>({1.0f / 256.0f /* scale */}), |
| 471 | builder.CreateVector<int64_t>({0 /* zero point */}))), |
| 472 | }}; |
| 473 | |
| 474 | const std::array<int32_t, 1> op_inputs{{ 0 }}; |
| 475 | const std::array<int32_t, 1> op_outputs{{ 1 }}; |
| 476 | flatbuffers::Offset<tflite::Operator> op = tflite::CreateOperator( |
| 477 | builder, |
| 478 | 0 /* opcode_index */, |
| 479 | builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()), |
| 480 | builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size())); |
| 481 | |
| 482 | const std::array<int32_t, 1> graph_inputs{{ 0 }}; |
| 483 | const std::array<int32_t, 1> graph_outputs{{ 1 }}; |
| 484 | const flatbuffers::Offset<tflite::SubGraph> subgraph = tflite::CreateSubGraph( |
| 485 | builder, |
| 486 | builder.CreateVector(tensors.data(), tensors.size()), |
| 487 | builder.CreateVector<int32_t>(graph_inputs.data(), graph_inputs.size()), |
| 488 | builder.CreateVector<int32_t>(graph_outputs.data(), graph_outputs.size()), |
| 489 | builder.CreateVector(&op, 1)); |
| 490 | |
| 491 | const flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder, |
| 492 | TFLITE_SCHEMA_VERSION, |
| 493 | builder.CreateVector(&operator_code, 1), |
| 494 | builder.CreateVector(&subgraph, 1), |
| 495 | builder.CreateString("Sigmoid model"), |
| 496 | builder.CreateVector(buffers.data(), buffers.size())); |
| 497 | |
| 498 | builder.Finish(model_buffer); |
| 499 | |
| 500 | const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer()); |
| 501 | tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates resolver; |
| 502 | tflite::InterpreterBuilder interpreterBuilder(model, resolver); |
| 503 | std::unique_ptr<tflite::Interpreter> interpreter; |
| 504 | if (interpreterBuilder(&interpreter) != kTfLiteOk || interpreter == nullptr) { |
| 505 | state.SkipWithError("failed to create TFLite interpreter"); |
| 506 | return; |
| 507 | } |
| 508 | interpreter->SetNumThreads(1); |
| 509 | |
| 510 | if (interpreter->AllocateTensors() != kTfLiteOk) { |
| 511 | state.SkipWithError("failed to allocate tensors"); |
| 512 | return; |
| 513 | } |
| 514 | |
| 515 | std::generate( |
| 516 | interpreter->typed_tensor<uint8_t>(0), |
| 517 | interpreter->typed_tensor<uint8_t>(0) + batch_size, |
| 518 | std::ref(u8rng)); |
| 519 | |
| 520 | for (auto _ : state) { |
| 521 | if (interpreter->Invoke() != kTfLiteOk) { |
| 522 | state.SkipWithError("failed to invoke TFLite interpreter"); |
| 523 | return; |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 528 | if (cpu_frequency != 0) { |
| 529 | state.counters["cpufreq"] = cpu_frequency; |
| 530 | } |
| 531 | |
| 532 | state.counters["elements"] = |
| 533 | benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate); |
| 534 | |
| 535 | const size_t bytes_per_iteration = 2 * batch_size * sizeof(uint8_t); |
| 536 | state.counters["bytes"] = |
| 537 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 538 | |
| 539 | interpreter.reset(); |
| 540 | } |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 541 | #endif // BENCHMARK_TENSORFLOW_LITE |
| 542 | |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 543 | BENCHMARK(xnnpack_sigmoid_f32) |
| 544 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 545 | ->UseRealTime(); |
Marat Dukhan | 9084fc8 | 2021-12-31 10:16:09 -0800 | [diff] [blame] | 546 | #ifndef XNN_NO_QS8_OPERATORS |
| 547 | BENCHMARK(xnnpack_sigmoid_qs8) |
| 548 | ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>) |
| 549 | ->UseRealTime(); |
| 550 | #endif // XNN_NO_QS8_OPERATORS |
Chao Mei | c664027 | 2020-07-23 09:35:11 -0700 | [diff] [blame] | 551 | #ifndef XNN_NO_QU8_OPERATORS |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 552 | BENCHMARK(xnnpack_sigmoid_qu8) |
| 553 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>) |
| 554 | ->UseRealTime(); |
Chao Mei | c664027 | 2020-07-23 09:35:11 -0700 | [diff] [blame] | 555 | #endif // XNN_NO_QU8_OPERATORS |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 556 | |
| 557 | #ifdef BENCHMARK_TENSORFLOW_LITE |
Marat Dukhan | a0129e9 | 2021-12-30 15:59:28 -0800 | [diff] [blame] | 558 | BENCHMARK(tflite_sigmoid_f32) |
| 559 | ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| 560 | ->UseRealTime(); |
Marat Dukhan | 9084fc8 | 2021-12-31 10:16:09 -0800 | [diff] [blame] | 561 | BENCHMARK(tflite_sigmoid_qs8) |
| 562 | ->Apply(benchmark::utils::UnaryElementwiseParameters<int8_t, int8_t>) |
| 563 | ->UseRealTime(); |
| 564 | BENCHMARK(tflite_sigmoid_qu8) |
| 565 | ->Apply(benchmark::utils::UnaryElementwiseParameters<uint8_t, uint8_t>) |
| 566 | ->UseRealTime(); |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 567 | #endif // BENCHMARK_TENSORFLOW_LITE |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 568 | |
| 569 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 570 | BENCHMARK_MAIN(); |
| 571 | #endif |