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
16 #include "bench/utils.h"
17 #include "models/models.h"
18
19
End2EndBenchmark(benchmark::State & state,models::ExecutionPlanFactory model_factory)20 static void End2EndBenchmark(
21 benchmark::State& state,
22 models::ExecutionPlanFactory model_factory)
23 {
24 if (xnn_initialize(nullptr /* allocator */) != xnn_status_success) {
25 state.SkipWithError("failed to initialize XNNPACK");
26 return;
27 }
28
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());
34 if (execution_plan.empty()) {
35 state.SkipWithError("failed to create a model");
36 return;
37 }
38
39 for (auto _ : state) {
40 for (const std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>& op : execution_plan) {
41 xnn_status status = xnn_run_operator(op.get(), threadpool.get());
42 if (status != xnn_status_success) {
43 state.SkipWithError("failed to run a model");
44 return;
45 }
46 }
47 }
48
49 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
50 if (cpu_frequency != 0) {
51 state.counters["cpufreq"] = cpu_frequency;
52 }
53 }
54
FP32MobileNetV1(benchmark::State & state)55 static void FP32MobileNetV1(benchmark::State& state) {
56 End2EndBenchmark(state, models::FP32MobileNetV1);
57 }
58
FP32MobileNetV2(benchmark::State & state)59 static void FP32MobileNetV2(benchmark::State& state) {
60 End2EndBenchmark(state, models::FP32MobileNetV2);
61 }
62
FP32MobileNetV3Large(benchmark::State & state)63 static void FP32MobileNetV3Large(benchmark::State& state) {
64 End2EndBenchmark(state, models::FP32MobileNetV3Large);
65 }
66
FP32MobileNetV3Small(benchmark::State & state)67 static void FP32MobileNetV3Small(benchmark::State& state) {
68 End2EndBenchmark(state, models::FP32MobileNetV3Small);
69 }
70
FP32Sparse80MobileNetV1(benchmark::State & state)71 static void FP32Sparse80MobileNetV1(benchmark::State& state) {
72 End2EndBenchmark(state, [](pthreadpool_t threadpool) {
73 return models::FP32SparseMobileNetV1(0.8f, threadpool);
74 });
75 }
76
FP32Sparse80MobileNetV2(benchmark::State & state)77 static void FP32Sparse80MobileNetV2(benchmark::State& state) {
78 End2EndBenchmark(state, [](pthreadpool_t threadpool) {
79 return models::FP32SparseMobileNetV2(0.8f, threadpool);
80 });
81 }
82
FP32Sparse80MobileNetV3Large(benchmark::State & state)83 static void FP32Sparse80MobileNetV3Large(benchmark::State& state) {
84 End2EndBenchmark(state, [](pthreadpool_t threadpool) {
85 return models::FP32SparseMobileNetV3Large(0.8f, threadpool);
86 });
87 }
88
FP32Sparse80MobileNetV3Small(benchmark::State & state)89 static void FP32Sparse80MobileNetV3Small(benchmark::State& state) {
90 End2EndBenchmark(state, [](pthreadpool_t threadpool) {
91 return models::FP32SparseMobileNetV3Small(0.8f, threadpool);
92 });
93 }
94
FP16MobileNetV1(benchmark::State & state)95 static void FP16MobileNetV1(benchmark::State& state) {
96 End2EndBenchmark(state, models::FP16MobileNetV1);
97 }
98
FP16MobileNetV2(benchmark::State & state)99 static void FP16MobileNetV2(benchmark::State& state) {
100 End2EndBenchmark(state, models::FP16MobileNetV2);
101 }
102
FP16MobileNetV3Large(benchmark::State & state)103 static void FP16MobileNetV3Large(benchmark::State& state) {
104 End2EndBenchmark(state, models::FP16MobileNetV3Large);
105 }
106
FP16MobileNetV3Small(benchmark::State & state)107 static void FP16MobileNetV3Small(benchmark::State& state) {
108 End2EndBenchmark(state, models::FP16MobileNetV3Small);
109 }
110
QC8MobileNetV1(benchmark::State & state)111 static void QC8MobileNetV1(benchmark::State& state) {
112 End2EndBenchmark(state, models::QC8MobileNetV1);
113 }
114
QC8MobileNetV2(benchmark::State & state)115 static void QC8MobileNetV2(benchmark::State& state) {
116 End2EndBenchmark(state, models::QC8MobileNetV2);
117 }
118
QS8MobileNetV1(benchmark::State & state)119 static void QS8MobileNetV1(benchmark::State& state) {
120 End2EndBenchmark(state, models::QS8MobileNetV1);
121 }
122
QS8MobileNetV2(benchmark::State & state)123 static void QS8MobileNetV2(benchmark::State& state) {
124 End2EndBenchmark(state, models::QS8MobileNetV2);
125 }
126
QU8MobileNetV1(benchmark::State & state)127 static void QU8MobileNetV1(benchmark::State& state) {
128 End2EndBenchmark(state, models::QU8MobileNetV1);
129 }
130
QU8MobileNetV2(benchmark::State & state)131 static void QU8MobileNetV2(benchmark::State& state) {
132 End2EndBenchmark(state, models::QU8MobileNetV2);
133 }
134
135 BENCHMARK(FP32MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
136 BENCHMARK(FP32MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
137 BENCHMARK(FP32MobileNetV3Large)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
138 BENCHMARK(FP32MobileNetV3Small)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
139
140 BENCHMARK(FP32Sparse80MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
141 BENCHMARK(FP32Sparse80MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
142 BENCHMARK(FP32Sparse80MobileNetV3Large)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
143 BENCHMARK(FP32Sparse80MobileNetV3Small)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
144
145 BENCHMARK(FP16MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
146 BENCHMARK(FP16MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
147 BENCHMARK(FP16MobileNetV3Large)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
148 BENCHMARK(FP16MobileNetV3Small)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
149
150 BENCHMARK(QC8MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
151 BENCHMARK(QC8MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
152
153 BENCHMARK(QS8MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
154 BENCHMARK(QS8MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
155
156 BENCHMARK(QU8MobileNetV1)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
157 BENCHMARK(QU8MobileNetV2)->Apply(benchmark::utils::MultiThreadingParameters)->Unit(benchmark::kMicrosecond)->UseRealTime();
158
159 #ifndef XNNPACK_BENCHMARK_NO_MAIN
160 BENCHMARK_MAIN();
161 #endif
162