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 | // |
| 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 <random> |
| 14 | #include <vector> |
| 15 | |
| 16 | #include <xnnpack.h> |
| 17 | |
| 18 | #include <benchmark/benchmark.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame^] | 19 | #include "bench/utils.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | |
| 21 | |
| 22 | static void average_pooling_q8(benchmark::State& state, const char* net) { |
| 23 | const size_t batch_size = state.range(0); |
| 24 | const size_t input_height = state.range(1); |
| 25 | const size_t input_width = state.range(2); |
| 26 | const size_t pooling_size = state.range(3); |
| 27 | const size_t padding_size = state.range(4); |
| 28 | const size_t stride = state.range(5); |
| 29 | const size_t channels = state.range(6); |
| 30 | |
| 31 | std::random_device random_device; |
| 32 | auto rng = std::mt19937(random_device()); |
| 33 | auto u8rng = std::bind(std::uniform_int_distribution<uint8_t>(), rng); |
| 34 | |
| 35 | const size_t output_height = (2 * padding_size + input_height - pooling_size) / stride + 1; |
| 36 | const size_t output_width = (2 * padding_size + input_width - pooling_size) / stride + 1; |
| 37 | |
| 38 | std::vector<uint8_t> input(batch_size * input_height * input_width * channels + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 39 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 40 | std::vector<uint8_t> output(batch_size * output_height * output_width * channels); |
| 41 | std::fill(output.begin(), output.end(), 0xA5); |
| 42 | |
| 43 | xnn_status status = xnn_initialize(); |
| 44 | if (status != xnn_status_success) { |
| 45 | state.SkipWithError("failed to initialize XNNPACK"); |
| 46 | return; |
| 47 | } |
| 48 | |
| 49 | xnn_operator_t pooling_op = nullptr; |
| 50 | status = xnn_create_average_pooling2d_nhwc_q8( |
| 51 | padding_size, padding_size, padding_size, padding_size, |
| 52 | pooling_size, pooling_size, |
| 53 | stride, stride, |
| 54 | channels, channels /* input pixel stride */, channels /* output pixel stride */, |
| 55 | 127 /* input zero point */, 0.75f /* input scale */, |
| 56 | 127 /* output zero point */, 1.25f /* output scale */, |
| 57 | 0, 255, |
| 58 | 0 /* flags */, &pooling_op); |
| 59 | if (status != xnn_status_success) { |
| 60 | state.SkipWithError("failed to create Average Pooling operator"); |
| 61 | return; |
| 62 | } |
| 63 | |
| 64 | status = xnn_setup_average_pooling2d_nhwc_q8( |
| 65 | pooling_op, |
| 66 | batch_size, input_height, input_width, |
| 67 | input.data(), output.data(), |
| 68 | nullptr /* thread pool */); |
| 69 | if (status != xnn_status_success) { |
| 70 | state.SkipWithError("failed to setup Average Pooling operator"); |
| 71 | return; |
| 72 | } |
| 73 | |
| 74 | for (auto _ : state) { |
| 75 | status = xnn_run_operator(pooling_op, nullptr /* thread pool */); |
| 76 | if (status != xnn_status_success) { |
| 77 | state.SkipWithError("failed to run Average Pooling operator"); |
| 78 | return; |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | status = xnn_delete_operator(pooling_op); |
| 83 | if (status != xnn_status_success) { |
| 84 | state.SkipWithError("failed to delete Average Pooling operator"); |
| 85 | return; |
| 86 | } |
| 87 | pooling_op = nullptr; |
| 88 | |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame^] | 89 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 90 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 91 | state.counters["bytes"] = benchmark::Counter( |
| 92 | uint64_t(state.iterations()) * |
| 93 | batch_size * (input_height * input_width + output_height * output_width) * channels * sizeof(uint8_t), |
| 94 | benchmark::Counter::kIsRate); |
| 95 | } |
| 96 | |
| 97 | // ShuffleNet v1 with 1 group. |
| 98 | static void ShuffleNetV1G1(benchmark::internal::Benchmark* b) { |
| 99 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 100 | |
| 101 | /* N H W K P S C */ |
| 102 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 103 | b->Args({1, 28, 28, 3, 1, 2, 144}); |
| 104 | b->Args({1, 14, 14, 3, 1, 2, 288}); |
| 105 | b->Args({1, 7, 7, 3, 1, 2, 576}); |
| 106 | } |
| 107 | |
| 108 | // ShuffleNet v1 with 2 groups. |
| 109 | static void ShuffleNetV1G2(benchmark::internal::Benchmark* b) { |
| 110 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 111 | |
| 112 | /* N H W K P S C */ |
| 113 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 114 | b->Args({1, 28, 28, 3, 1, 2, 200}); |
| 115 | b->Args({1, 14, 14, 3, 1, 2, 400}); |
| 116 | b->Args({1, 7, 7, 3, 1, 2, 800}); |
| 117 | } |
| 118 | |
| 119 | // ShuffleNet v1 with 3 groups. |
| 120 | static void ShuffleNetV1G3(benchmark::internal::Benchmark* b) { |
| 121 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 122 | |
| 123 | /* N H W K P S C */ |
| 124 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 125 | b->Args({1, 28, 28, 3, 1, 2, 240}); |
| 126 | b->Args({1, 14, 14, 3, 1, 2, 480}); |
| 127 | b->Args({1, 7, 7, 3, 1, 2, 960}); |
| 128 | } |
| 129 | |
| 130 | // ShuffleNet v1 with 4 groups. |
| 131 | static void ShuffleNetV1G4(benchmark::internal::Benchmark* b) { |
| 132 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 133 | |
| 134 | /* N H W K P S C */ |
| 135 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 136 | b->Args({1, 28, 28, 3, 1, 2, 272}); |
| 137 | b->Args({1, 14, 14, 3, 1, 2, 576}); |
| 138 | b->Args({1, 7, 7, 3, 1, 2, 1088}); |
| 139 | } |
| 140 | |
| 141 | // ShuffleNet v1 with 8 groups. |
| 142 | static void ShuffleNetV1G8(benchmark::internal::Benchmark* b) { |
| 143 | b->ArgNames({"N", "H", "W", "K", "P", "S", "C"}); |
| 144 | |
| 145 | /* N H W K P S C */ |
| 146 | b->Args({1, 56, 56, 3, 1, 2, 24}); |
| 147 | b->Args({1, 28, 28, 3, 1, 2, 384}); |
| 148 | b->Args({1, 14, 14, 3, 1, 2, 768}); |
| 149 | b->Args({1, 7, 7, 3, 1, 2, 1536}); |
| 150 | } |
| 151 | |
| 152 | BENCHMARK_CAPTURE(average_pooling_q8, shufflenet_v1_g1, "ShuffleNet v1 (1 group)")->Apply(ShuffleNetV1G1)->UseRealTime(); |
| 153 | BENCHMARK_CAPTURE(average_pooling_q8, shufflenet_v1_g2, "ShuffleNet v1 (2 groups)")->Apply(ShuffleNetV1G2)->UseRealTime(); |
| 154 | BENCHMARK_CAPTURE(average_pooling_q8, shufflenet_v1_g3, "ShuffleNet v1 (3 groups)")->Apply(ShuffleNetV1G3)->UseRealTime(); |
| 155 | BENCHMARK_CAPTURE(average_pooling_q8, shufflenet_v1_g4, "ShuffleNet v1 (4 groups)")->Apply(ShuffleNetV1G4)->UseRealTime(); |
| 156 | BENCHMARK_CAPTURE(average_pooling_q8, shufflenet_v1_g8, "ShuffleNet v1 (8 groups)")->Apply(ShuffleNetV1G8)->UseRealTime(); |
| 157 | |
| 158 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 159 | BENCHMARK_MAIN(); |
| 160 | #endif |