Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1 | // Copyright 2019 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 <cmath> |
| 8 | #include <functional> |
| 9 | #include <random> |
| 10 | #include <vector> |
| 11 | |
| 12 | #include <xnnpack.h> |
| 13 | |
| 14 | #include <benchmark/benchmark.h> |
| 15 | |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 16 | #include "bench/utils.h" |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 17 | #include "models/models.h" |
| 18 | |
| 19 | |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 20 | static void End2EndBenchmark( |
| 21 | benchmark::State& state, |
| 22 | models::ExecutionPlanFactory model_factory) |
| 23 | { |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 24 | if (xnn_initialize(nullptr /* allocator */) != xnn_status_success) { |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 25 | state.SkipWithError("failed to initialize XNNPACK"); |
| 26 | return; |
| 27 | } |
| 28 | |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 29 | const size_t num_threads = state.range(0); |
| 30 | std::unique_ptr<pthreadpool, decltype(&pthreadpool_destroy)> threadpool( |
| 31 | pthreadpool_create(num_threads), pthreadpool_destroy); |
| 32 | |
| 33 | auto execution_plan = model_factory(threadpool.get()); |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 34 | if (execution_plan.empty()) { |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 35 | state.SkipWithError("failed to create a model"); |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 36 | return; |
| 37 | } |
| 38 | |
| 39 | for (auto _ : state) { |
| 40 | for (const std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>& op : execution_plan) { |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 41 | xnn_status status = xnn_run_operator(op.get(), threadpool.get()); |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 42 | if (status != xnn_status_success) { |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 43 | state.SkipWithError("failed to run a model"); |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 44 | return; |
| 45 | } |
| 46 | } |
| 47 | } |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 48 | |
| 49 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 50 | if (cpu_frequency != 0) { |
| 51 | state.counters["cpufreq"] = cpu_frequency; |
| 52 | } |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 53 | } |
| 54 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 55 | static void FP32MobileNetV1(benchmark::State& state) { |
| 56 | End2EndBenchmark(state, models::FP32MobileNetV1); |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 57 | } |
| 58 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 59 | static void FP32MobileNetV2(benchmark::State& state) { |
| 60 | End2EndBenchmark(state, models::FP32MobileNetV2); |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 61 | } |
| 62 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 63 | static void FP32MobileNetV3Large(benchmark::State& state) { |
| 64 | End2EndBenchmark(state, models::FP32MobileNetV3Large); |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 65 | } |
| 66 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 67 | static void FP32MobileNetV3Small(benchmark::State& state) { |
| 68 | End2EndBenchmark(state, models::FP32MobileNetV3Small); |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 69 | } |
| 70 | |
Marat Dukhan | 4cea232 | 2021-03-09 09:35:36 -0800 | [diff] [blame] | 71 | static void FP32Sparse80MobileNetV1(benchmark::State& state) { |
| 72 | End2EndBenchmark(state, [](pthreadpool_t threadpool) { |
| 73 | return models::FP32SparseMobileNetV1(0.8f, threadpool); |
| 74 | }); |
| 75 | } |
| 76 | |
| 77 | static void FP32Sparse80MobileNetV2(benchmark::State& state) { |
| 78 | End2EndBenchmark(state, [](pthreadpool_t threadpool) { |
| 79 | return models::FP32SparseMobileNetV2(0.8f, threadpool); |
| 80 | }); |
| 81 | } |
| 82 | |
| 83 | static void FP32Sparse80MobileNetV3Large(benchmark::State& state) { |
| 84 | End2EndBenchmark(state, [](pthreadpool_t threadpool) { |
| 85 | return models::FP32SparseMobileNetV3Large(0.8f, threadpool); |
| 86 | }); |
| 87 | } |
| 88 | |
| 89 | static void FP32Sparse80MobileNetV3Small(benchmark::State& state) { |
| 90 | End2EndBenchmark(state, [](pthreadpool_t threadpool) { |
| 91 | return models::FP32SparseMobileNetV3Small(0.8f, threadpool); |
| 92 | }); |
| 93 | } |
| 94 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 95 | static void FP16MobileNetV1(benchmark::State& state) { |
| 96 | End2EndBenchmark(state, models::FP16MobileNetV1); |
| 97 | } |
| 98 | |
| 99 | static void FP16MobileNetV2(benchmark::State& state) { |
| 100 | End2EndBenchmark(state, models::FP16MobileNetV2); |
| 101 | } |
| 102 | |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 103 | static void FP16MobileNetV3Large(benchmark::State& state) { |
| 104 | End2EndBenchmark(state, models::FP16MobileNetV3Large); |
| 105 | } |
| 106 | |
| 107 | static void FP16MobileNetV3Small(benchmark::State& state) { |
| 108 | End2EndBenchmark(state, models::FP16MobileNetV3Small); |
| 109 | } |
| 110 | |
Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 111 | static void QS8MobileNetV1(benchmark::State& state) { |
| 112 | End2EndBenchmark(state, models::QS8MobileNetV1); |
| 113 | } |
| 114 | |
Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 115 | static void QS8MobileNetV2(benchmark::State& state) { |
| 116 | End2EndBenchmark(state, models::QS8MobileNetV2); |
| 117 | } |
| 118 | |
Marat Dukhan | 12a23bb | 2021-03-08 08:13:21 -0800 | [diff] [blame] | 119 | static void QU8MobileNetV1(benchmark::State& state) { |
| 120 | End2EndBenchmark(state, models::QU8MobileNetV1); |
| 121 | } |
| 122 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 123 | BENCHMARK(FP32MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 124 | BENCHMARK(FP32MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 125 | BENCHMARK(FP32MobileNetV3Large)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 126 | BENCHMARK(FP32MobileNetV3Small)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 127 | |
Marat Dukhan | 4cea232 | 2021-03-09 09:35:36 -0800 | [diff] [blame] | 128 | BENCHMARK(FP32Sparse80MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 129 | BENCHMARK(FP32Sparse80MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 130 | BENCHMARK(FP32Sparse80MobileNetV3Large)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 131 | BENCHMARK(FP32Sparse80MobileNetV3Small)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 132 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 133 | BENCHMARK(FP16MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 134 | BENCHMARK(FP16MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 135 | BENCHMARK(FP16MobileNetV3Large)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 136 | BENCHMARK(FP16MobileNetV3Small)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 137 | |
Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 138 | BENCHMARK(QS8MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 139 | BENCHMARK(QS8MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 140 | |
Marat Dukhan | 12a23bb | 2021-03-08 08:13:21 -0800 | [diff] [blame] | 141 | BENCHMARK(QU8MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime(); |
| 142 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 143 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 144 | BENCHMARK_MAIN(); |
| 145 | #endif |