Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <algorithm> |
| 7 | #include <array> |
| 8 | #include <cmath> |
| 9 | #include <functional> |
| 10 | #include <limits> |
| 11 | #include <random> |
| 12 | #include <vector> |
| 13 | |
| 14 | #include <xnnpack.h> |
| 15 | |
| 16 | #include <benchmark/benchmark.h> |
| 17 | #include "bench/utils.h" |
| 18 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 19 | #include "flatbuffers/include/flatbuffers/flatbuffers.h" |
| 20 | #include "tensorflow/lite/interpreter.h" |
| 21 | #include "tensorflow/lite/kernels/register.h" |
| 22 | #include "tensorflow/lite/model.h" |
| 23 | #include "tensorflow/lite/schema/schema_generated.h" |
| 24 | #include "tensorflow/lite/version.h" |
| 25 | #endif // BENCHMARK_TENSORFLOW_LITE |
| 26 | |
| 27 | |
| 28 | static void xnnpack_hardswish_f32(benchmark::State& state) { |
| 29 | const size_t batch_size = state.range(0); |
| 30 | const size_t channels = state.range(1); |
| 31 | |
| 32 | std::random_device random_device; |
| 33 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | a11ca34 | 2020-06-25 23:45:07 -0700 | [diff] [blame] | 34 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 35 | |
| 36 | std::vector<float> input(batch_size * channels); |
| 37 | std::vector<float> output(batch_size * channels); |
| 38 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 39 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 40 | |
| 41 | xnn_status status = xnn_initialize(nullptr /* allocator */); |
| 42 | if (status != xnn_status_success) { |
| 43 | state.SkipWithError("failed to initialize XNNPACK"); |
| 44 | return; |
| 45 | } |
| 46 | |
| 47 | xnn_operator_t hardswish_op = nullptr; |
| 48 | status = xnn_create_hardswish_nc_f32( |
| 49 | channels, channels /* input stride */, channels /* output stride */, |
| 50 | 0 /* flags */, &hardswish_op); |
| 51 | if (status != xnn_status_success || hardswish_op == nullptr) { |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 52 | state.SkipWithError("failed to create HardSwish operator"); |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 53 | return; |
| 54 | } |
| 55 | |
| 56 | status = xnn_setup_hardswish_nc_f32( |
| 57 | hardswish_op, |
| 58 | batch_size, |
| 59 | input.data(), output.data(), |
| 60 | nullptr /* thread pool */); |
| 61 | if (status != xnn_status_success) { |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 62 | state.SkipWithError("failed to setup HardSwish operator"); |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 63 | return; |
| 64 | } |
| 65 | |
| 66 | for (auto _ : state) { |
| 67 | status = xnn_run_operator(hardswish_op, nullptr /* thread pool */); |
| 68 | if (status != xnn_status_success) { |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 69 | state.SkipWithError("failed to run HardSwish operator"); |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 70 | return; |
| 71 | } |
| 72 | } |
| 73 | |
| 74 | status = xnn_delete_operator(hardswish_op); |
| 75 | if (status != xnn_status_success) { |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 76 | state.SkipWithError("failed to delete HardSwish operator"); |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 77 | return; |
| 78 | } |
| 79 | |
| 80 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 81 | |
| 82 | const size_t elements_per_iteration = batch_size * channels; |
| 83 | state.counters["elements"] = |
| 84 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 85 | |
| 86 | const size_t bytes_per_iteration = 2 * elements_per_iteration * sizeof(float); |
| 87 | state.counters["bytes"] = |
| 88 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 89 | } |
| 90 | |
| 91 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 92 | static void tflite_hardswish_f32(benchmark::State& state) { |
| 93 | const size_t batch_size = state.range(0); |
| 94 | const size_t channels = state.range(1); |
| 95 | |
| 96 | std::random_device random_device; |
| 97 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | a11ca34 | 2020-06-25 23:45:07 -0700 | [diff] [blame] | 98 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 99 | |
| 100 | flatbuffers::FlatBufferBuilder builder; |
| 101 | const flatbuffers::Offset<tflite::OperatorCode> operator_code = |
| 102 | CreateOperatorCode(builder, tflite::BuiltinOperator_HARD_SWISH); |
| 103 | |
| 104 | const std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers{{ |
| 105 | tflite::CreateBuffer(builder, builder.CreateVector({})), |
| 106 | }}; |
| 107 | |
| 108 | const std::array<int32_t, 4> input_shape{{ |
| 109 | static_cast<int32_t>(batch_size), |
| 110 | static_cast<int32_t>(1 /* height */), |
| 111 | static_cast<int32_t>(1 /* width */), |
| 112 | static_cast<int32_t>(channels) |
| 113 | }}; |
| 114 | const std::array<int32_t, 4> output_shape{{ |
| 115 | static_cast<int32_t>(batch_size), |
| 116 | static_cast<int32_t>(1 /* height */), |
| 117 | static_cast<int32_t>(1 /* width */), |
| 118 | static_cast<int32_t>(channels) |
| 119 | }}; |
| 120 | |
| 121 | const std::array<flatbuffers::Offset<tflite::Tensor>, 2> tensors{{ |
| 122 | tflite::CreateTensor(builder, |
| 123 | builder.CreateVector<int32_t>(input_shape.data(), input_shape.size()), |
| 124 | tflite::TensorType_FLOAT32), |
| 125 | tflite::CreateTensor(builder, |
| 126 | builder.CreateVector<int32_t>(output_shape.data(), output_shape.size()), |
| 127 | tflite::TensorType_FLOAT32), |
| 128 | }}; |
| 129 | |
| 130 | const std::array<int32_t, 1> op_inputs{{ 0 }}; |
| 131 | const std::array<int32_t, 1> op_outputs{{ 1 }}; |
| 132 | flatbuffers::Offset<tflite::Operator> op = tflite::CreateOperator( |
| 133 | builder, |
| 134 | 0 /* opcode_index */, |
| 135 | builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()), |
| 136 | builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size())); |
| 137 | |
| 138 | const std::array<int32_t, 1> graph_inputs{{ 0 }}; |
| 139 | const std::array<int32_t, 1> graph_outputs{{ 1 }}; |
| 140 | const flatbuffers::Offset<tflite::SubGraph> subgraph = tflite::CreateSubGraph( |
| 141 | builder, |
| 142 | builder.CreateVector(tensors.data(), tensors.size()), |
| 143 | builder.CreateVector<int32_t>(graph_inputs.data(), graph_inputs.size()), |
| 144 | builder.CreateVector<int32_t>(graph_outputs.data(), graph_outputs.size()), |
| 145 | builder.CreateVector(&op, 1)); |
| 146 | |
| 147 | const flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder, |
| 148 | TFLITE_SCHEMA_VERSION, |
| 149 | builder.CreateVector(&operator_code, 1), |
| 150 | builder.CreateVector(&subgraph, 1), |
| 151 | builder.CreateString("HardSwish model"), |
| 152 | builder.CreateVector(buffers.data(), buffers.size())); |
| 153 | |
| 154 | builder.Finish(model_buffer); |
| 155 | |
| 156 | const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer()); |
| 157 | tflite::ops::builtin::BuiltinOpResolver resolver; |
| 158 | tflite::InterpreterBuilder interpreterBuilder(model, resolver); |
| 159 | std::unique_ptr<tflite::Interpreter> interpreter; |
| 160 | if (interpreterBuilder(&interpreter) != kTfLiteOk) { |
| 161 | state.SkipWithError("failed to create TFLite interpreter"); |
| 162 | return; |
| 163 | } |
| 164 | if (interpreter == nullptr) { |
| 165 | state.SkipWithError("TFLite interpreter is null"); |
| 166 | return; |
| 167 | } |
| 168 | interpreter->SetNumThreads(1); |
| 169 | |
| 170 | if (interpreter->AllocateTensors() != kTfLiteOk) { |
| 171 | state.SkipWithError("failed to allocate tensors"); |
| 172 | return; |
| 173 | } |
| 174 | |
| 175 | std::generate( |
| 176 | interpreter->typed_tensor<float>(0), |
| 177 | interpreter->typed_tensor<float>(0) + batch_size * channels, |
| 178 | std::ref(f32rng)); |
| 179 | |
| 180 | for (auto _ : state) { |
| 181 | if (interpreter->Invoke() != kTfLiteOk) { |
| 182 | state.SkipWithError("failed to invoke TFLite interpreter"); |
| 183 | return; |
| 184 | } |
| 185 | } |
| 186 | |
| 187 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 188 | |
| 189 | const size_t elements_per_iteration = batch_size * channels; |
| 190 | state.counters["elements"] = |
| 191 | benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| 192 | |
| 193 | const size_t bytes_per_iteration = 2 * elements_per_iteration * sizeof(float); |
| 194 | state.counters["bytes"] = |
| 195 | benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| 196 | |
| 197 | interpreter.reset(); |
| 198 | } |
| 199 | #endif // BENCHMARK_TENSORFLOW_LITE |
| 200 | |
| 201 | static void CharacteristicArguments(benchmark::internal::Benchmark* b) |
| 202 | { |
| 203 | b->ArgNames({"N", "C"}); |
| 204 | |
| 205 | int32_t c = 16; |
| 206 | for (int32_t n = 224; n >= 7; n /= 2) { |
| 207 | b->Args({n * n, c}); |
| 208 | c *= 2; |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | BENCHMARK(xnnpack_hardswish_f32)->Apply(CharacteristicArguments)->UseRealTime(); |
| 213 | |
| 214 | #ifdef BENCHMARK_TENSORFLOW_LITE |
| 215 | BENCHMARK(tflite_hardswish_f32)->Apply(CharacteristicArguments)->UseRealTime(); |
| 216 | #endif // BENCHMARK_TENSORFLOW_LITE |
| 217 | |
| 218 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 219 | BENCHMARK_MAIN(); |
| 220 | #endif |