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 <cpuinfo.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 17 | |
| 18 | #include <benchmark/benchmark.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 19 | #include "bench/utils.h" |
| 20 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 21 | #include <xnnpack/common.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 22 | #include <xnnpack/requantization-stubs.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 23 | |
| 24 | inline uint32_t divideRoundUp(uint32_t x, uint32_t q) |
| 25 | { |
| 26 | return x / q + uint32_t(x % q != 0); |
| 27 | } |
| 28 | |
| 29 | inline uint32_t roundUp(uint32_t x, uint32_t q) |
| 30 | { |
| 31 | return q * divideRoundUp(x, q); |
| 32 | } |
| 33 | |
| 34 | inline uint32_t min(uint32_t a, uint32_t b) |
| 35 | { |
| 36 | return a < b ? a : b; |
| 37 | } |
| 38 | |
| 39 | class Requantization : public benchmark::Fixture { |
| 40 | public: |
| 41 | inline Requantization() |
| 42 | { |
| 43 | cpuinfo_initialize(); |
| 44 | const size_t l1d_size = cpuinfo_get_l1d_cache(0)->size; |
| 45 | const size_t l1d_reserve = 1024; |
| 46 | n_ = (l1d_size - l1d_reserve) / (sizeof(int32_t) + sizeof(uint8_t)); |
| 47 | n_ = n_ / 16 * 16; |
| 48 | } |
| 49 | |
| 50 | virtual void SetUp(const benchmark::State&) override |
| 51 | { |
| 52 | std::random_device random_device; |
| 53 | auto rng = std::mt19937(random_device()); |
| 54 | auto s32rng = std::bind(std::uniform_int_distribution<int32_t>(), rng); |
| 55 | |
| 56 | input_.resize(n()); |
| 57 | std::generate(input_.begin(), input_.end(), std::ref(s32rng)); |
| 58 | output_.resize(n()); |
| 59 | std::fill(output_.begin(), output_.end(), 0xA5); |
| 60 | } |
| 61 | |
| 62 | virtual void TearDown(benchmark::State& state) override |
| 63 | { |
| 64 | state.SetItemsProcessed(uint64_t(state.iterations()) * n()); |
| 65 | state.SetBytesProcessed(uint64_t(state.iterations()) * n() * (sizeof(int32_t) + sizeof(uint8_t))); |
| 66 | input_.clear(); |
| 67 | output_.clear(); |
| 68 | } |
| 69 | |
| 70 | inline const int32_t* input() const |
| 71 | { |
| 72 | return input_.data(); |
| 73 | } |
| 74 | |
| 75 | inline uint8_t* output() |
| 76 | { |
| 77 | return output_.data(); |
| 78 | } |
| 79 | |
| 80 | inline size_t n() const |
| 81 | { |
| 82 | return n_; |
| 83 | } |
| 84 | |
| 85 | protected: |
| 86 | std::vector<int32_t, AlignedAllocator<int32_t, 32>> input_; |
| 87 | std::vector<uint8_t> output_; |
| 88 | size_t n_; |
| 89 | }; |
| 90 | |
| 91 | BENCHMARK_F(Requantization, precise__scalar_unsigned32)(benchmark::State& state) |
| 92 | { |
| 93 | for (auto _ : state) { |
| 94 | xnn_requantize_precise__scalar_unsigned32( |
| 95 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 96 | } |
| 97 | } |
| 98 | |
| 99 | BENCHMARK_F(Requantization, precise__scalar_unsigned64)(benchmark::State& state) |
| 100 | { |
| 101 | for (auto _ : state) { |
| 102 | xnn_requantize_precise__scalar_unsigned64( |
| 103 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | BENCHMARK_F(Requantization, precise__scalar_signed64)(benchmark::State& state) |
| 108 | { |
| 109 | for (auto _ : state) { |
| 110 | xnn_requantize_precise__scalar_signed64( |
| 111 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | BENCHMARK_F(Requantization, fp32__scalar_lrintf)(benchmark::State& state) |
| 116 | { |
| 117 | for (auto _ : state) { |
| 118 | xnn_requantize_fp32__scalar_lrintf( |
| 119 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | BENCHMARK_F(Requantization, fp32__scalar_magic)(benchmark::State& state) |
| 124 | { |
| 125 | for (auto _ : state) { |
| 126 | xnn_requantize_fp32__scalar_magic( |
| 127 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | BENCHMARK_F(Requantization, gemmlowp__scalar)(benchmark::State& state) |
| 132 | { |
| 133 | for (auto _ : state) { |
| 134 | xnn_requantize_gemmlowp__scalar( |
| 135 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 136 | } |
| 137 | } |
| 138 | |
| 139 | BENCHMARK_F(Requantization, precise__psimd)(benchmark::State& state) |
| 140 | { |
| 141 | for (auto _ : state) { |
| 142 | xnn_requantize_precise__psimd( |
| 143 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | BENCHMARK_F(Requantization, fp32__psimd)(benchmark::State& state) |
| 148 | { |
| 149 | for (auto _ : state) { |
| 150 | xnn_requantize_fp32__psimd( |
| 151 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 152 | } |
| 153 | } |
| 154 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 155 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 156 | BENCHMARK_F(Requantization, precise__neon)(benchmark::State& state) |
| 157 | { |
| 158 | for (auto _ : state) { |
| 159 | xnn_requantize_precise__neon( |
| 160 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 161 | } |
| 162 | } |
| 163 | |
| 164 | BENCHMARK_F(Requantization, fp32__neon)(benchmark::State& state) |
| 165 | { |
| 166 | for (auto _ : state) { |
| 167 | xnn_requantize_fp32__neon( |
| 168 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | BENCHMARK_F(Requantization, q31__neon)(benchmark::State& state) |
| 173 | { |
| 174 | for (auto _ : state) { |
| 175 | xnn_requantize_q31__neon( |
| 176 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | BENCHMARK_F(Requantization, gemmlowp__neon)(benchmark::State& state) |
| 181 | { |
| 182 | for (auto _ : state) { |
| 183 | xnn_requantize_gemmlowp__neon( |
| 184 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 185 | } |
| 186 | } |
| 187 | #endif |
| 188 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 189 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 190 | BENCHMARK_F(Requantization, precise__sse2)(benchmark::State& state) |
| 191 | { |
| 192 | for (auto _ : state) { |
| 193 | xnn_requantize_precise__sse2( |
| 194 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | BENCHMARK_F(Requantization, precise__ssse3)(benchmark::State& state) |
| 199 | { |
| 200 | for (auto _ : state) { |
| 201 | xnn_requantize_precise__ssse3( |
| 202 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | BENCHMARK_F(Requantization, precise__sse4)(benchmark::State& state) |
| 207 | { |
| 208 | for (auto _ : state) { |
| 209 | xnn_requantize_precise__sse4( |
| 210 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | BENCHMARK_F(Requantization, fp32__sse2)(benchmark::State& state) |
| 215 | { |
| 216 | for (auto _ : state) { |
| 217 | xnn_requantize_fp32__sse2( |
| 218 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 219 | } |
| 220 | } |
| 221 | |
| 222 | BENCHMARK_F(Requantization, q31__sse2)(benchmark::State& state) |
| 223 | { |
| 224 | for (auto _ : state) { |
| 225 | xnn_requantize_q31__sse2( |
| 226 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | BENCHMARK_F(Requantization, q31__ssse3)(benchmark::State& state) |
| 231 | { |
| 232 | for (auto _ : state) { |
| 233 | xnn_requantize_q31__ssse3( |
| 234 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 235 | } |
| 236 | } |
| 237 | |
| 238 | BENCHMARK_F(Requantization, q31__sse4)(benchmark::State& state) |
| 239 | { |
| 240 | for (auto _ : state) { |
| 241 | xnn_requantize_q31__sse4( |
| 242 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | BENCHMARK_F(Requantization, gemmlowp__sse2)(benchmark::State& state) |
| 247 | { |
| 248 | for (auto _ : state) { |
| 249 | xnn_requantize_gemmlowp__sse2( |
| 250 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | BENCHMARK_F(Requantization, gemmlowp__ssse3)(benchmark::State& state) |
| 255 | { |
| 256 | for (auto _ : state) { |
| 257 | xnn_requantize_gemmlowp__ssse3( |
| 258 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | BENCHMARK_F(Requantization, gemmlowp__sse4)(benchmark::State& state) |
| 263 | { |
| 264 | for (auto _ : state) { |
| 265 | xnn_requantize_gemmlowp__sse4( |
| 266 | n(), input(), 0x1.0p-12f /* scale */, 128 /* zero point */, 1 /* qmin */, 254 /* qmax */, output()); |
| 267 | } |
| 268 | } |
| 269 | #endif |
| 270 | |
| 271 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 272 | BENCHMARK_MAIN(); |
| 273 | #endif |