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 <chrono> |
| 12 | #include <cmath> |
| 13 | #include <functional> |
| 14 | #include <mutex> |
| 15 | #include <random> |
| 16 | #include <vector> |
| 17 | |
| 18 | #include <cpuinfo.h> |
| 19 | |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 20 | #include <benchmark/benchmark.h> |
| 21 | #include "tensorflow/lite/experimental/ruy/ruy.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | #include "bench/gemm.h" |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 23 | #include "bench/utils.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 25 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | #include <xnnpack/gemm.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 27 | #include <xnnpack/pack.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 28 | #include <xnnpack/packx.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | #include <xnnpack/params.h> |
| 30 | #include <xnnpack/ppmm.h> |
| 31 | #include <xnnpack/requantization.h> |
| 32 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 33 | |
| 34 | static void GEMMBenchmark(benchmark::State& state, |
| 35 | xnn_f32_gemm_ukernel_function gemm, |
| 36 | size_t mr, size_t nr, size_t kr, size_t sr) |
| 37 | { |
| 38 | if (!cpuinfo_initialize()) { |
| 39 | state.SkipWithError("cpuinfo initialization failed"); |
| 40 | return; |
| 41 | } |
| 42 | |
| 43 | const size_t mc = state.range(0); |
| 44 | const size_t nc = state.range(1); |
| 45 | const size_t kc = state.range(2); |
| 46 | |
| 47 | const size_t nc_stride = benchmark::utils::roundUp(nc, nr); |
| 48 | const size_t kc_stride = benchmark::utils::roundUp(kc, kr); |
| 49 | |
| 50 | std::random_device random_device; |
| 51 | auto rng = std::mt19937(random_device()); |
| 52 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 53 | |
| 54 | std::vector<float> a(mc * kc); |
| 55 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 56 | std::vector<float> k(nc * kc); |
| 57 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 58 | std::vector<float> b(nc); |
| 59 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 60 | |
| 61 | const size_t w_elements = nc_stride * kc_stride + nc_stride; |
| 62 | const size_t c_elements = mc * nc; |
| 63 | const size_t num_buffers = 1 + |
| 64 | benchmark::utils::divideRoundUp<size_t>(cpuinfo_get_max_cache_size(), |
| 65 | sizeof(float) * (w_elements + c_elements)); |
| 66 | |
| 67 | std::vector<float, AlignedAllocator<float, 32>> w(w_elements * num_buffers); |
| 68 | std::fill(w.begin(), w.end(), 0.0f); |
| 69 | xnn_pack_f32_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data()); |
| 70 | std::vector<float> c(c_elements * num_buffers); |
| 71 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 72 | |
| 73 | xnn_f32_output_params output_params = |
| 74 | xnn_compute_f32_output_params(-std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
| 75 | |
| 76 | size_t buffer_index = 0; |
| 77 | for (auto _ : state) { |
| 78 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 79 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 80 | // - W is not in cache (for any cache level) |
| 81 | // - C is not in cache (for any cache level) |
| 82 | state.PauseTiming(); |
| 83 | benchmark::utils::prefetchToL1(a.data(), a.size() * sizeof(float)); |
| 84 | buffer_index = (buffer_index + 1) % num_buffers; |
| 85 | state.ResumeTiming(); |
| 86 | |
| 87 | for (uint32_t m = 0; m < mc; m += mr) { |
| 88 | const uint32_t mb = min(mc - m, mr); |
| 89 | gemm( |
| 90 | mb, nc, kc * sizeof(float), |
| 91 | a.data() + m * kc, kc * sizeof(float), |
| 92 | w.data() + buffer_index * nc_stride * (kc_stride + 1), |
| 93 | c.data() + (buffer_index * mc + m) * nc, nc * sizeof(float), nr * sizeof(float), |
| 94 | &output_params); |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 99 | state.counters["FLOPS"] = benchmark::Counter( |
| 100 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 101 | } |
| 102 | |
| 103 | static void PPMM1PBenchmark(benchmark::State& state, |
| 104 | xnn_f32_ppmm_ukernel_function ppmm, |
| 105 | xnn_x32_packx_ukernel_function packx, |
| 106 | size_t mr, size_t nr) |
| 107 | { |
| 108 | if (!cpuinfo_initialize()) { |
| 109 | state.SkipWithError("cpuinfo initialization failed"); |
| 110 | return; |
| 111 | } |
| 112 | |
| 113 | const size_t mc = state.range(0); |
| 114 | const size_t nc = state.range(1); |
| 115 | const size_t kc = state.range(2); |
| 116 | |
| 117 | const size_t nc_stride = benchmark::utils::roundUp(nc, nr); |
| 118 | |
| 119 | std::random_device random_device; |
| 120 | auto rng = std::mt19937(random_device()); |
| 121 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 122 | |
| 123 | std::vector<float> a(mc * kc); |
| 124 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 125 | std::vector<float> k(nc * kc); |
| 126 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 127 | std::vector<float> b(nc); |
| 128 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 129 | |
| 130 | std::vector<uint32_t, AlignedAllocator<uint32_t, 32>> t(mr * kc); |
| 131 | |
| 132 | const size_t w_elements = nc_stride * kc + nc_stride; |
| 133 | const size_t c_elements = mc * nc; |
| 134 | const size_t num_buffers = 1 + |
| 135 | benchmark::utils::divideRoundUp<size_t>(cpuinfo_get_max_cache_size(), |
| 136 | sizeof(float) * (w_elements + c_elements)); |
| 137 | |
| 138 | std::vector<float, AlignedAllocator<float, 32>> w(w_elements * num_buffers); |
| 139 | std::fill(w.begin(), w.end(), 0.0f); |
| 140 | xnn_pack_f32_gemm_goi_w(1 /* groups */, nc, kc, nr, 1 /* kr */, 1 /* sr */, k.data(), b.data(), w.data()); |
| 141 | std::vector<float> c(c_elements * num_buffers); |
| 142 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 143 | |
| 144 | xnn_f32_output_params output_params = |
| 145 | xnn_compute_f32_output_params(-std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
| 146 | |
| 147 | size_t buffer_index = 0; |
| 148 | for (auto _ : state) { |
| 149 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 150 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 151 | // - W is not in cache (for any cache level) |
| 152 | // - C is not in cache (for any cache level) |
| 153 | state.PauseTiming(); |
| 154 | benchmark::utils::prefetchToL1(a.data(), a.size() * sizeof(float)); |
| 155 | buffer_index = (buffer_index + 1) % num_buffers; |
| 156 | state.ResumeTiming(); |
| 157 | |
| 158 | for (uint32_t m = 0; m < mc; m += mr) { |
| 159 | const uint32_t mb = min(mc - m, mr); |
| 160 | packx(mb, kc, reinterpret_cast<const uint32_t*>(a.data() + m * kc), kc, t.data()); |
| 161 | ppmm( |
| 162 | mb, nc, kc * sizeof(float), |
| 163 | reinterpret_cast<const float*>(t.data()), |
| 164 | w.data() + nc_stride * buffer_index * (kc + 1), |
| 165 | c.data() + (mc * buffer_index + m) * nc, nc * sizeof(float), nr * sizeof(float), |
| 166 | &output_params); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 171 | state.counters["FLOPS"] = benchmark::Counter( |
| 172 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 173 | } |
| 174 | |
| 175 | static void PPMM2PBenchmark(benchmark::State& state, |
| 176 | xnn_f32_ppmm_ukernel_function ppmm, |
| 177 | xnn_x32_packx_ukernel_function packx, |
| 178 | size_t mr, size_t nr) |
| 179 | { |
| 180 | if (!cpuinfo_initialize()) { |
| 181 | state.SkipWithError("cpuinfo initialization failed"); |
| 182 | return; |
| 183 | } |
| 184 | |
| 185 | const size_t mc = state.range(0); |
| 186 | const size_t nc = state.range(1); |
| 187 | const size_t kc = state.range(2); |
| 188 | |
| 189 | const size_t mc_stride = benchmark::utils::roundUp(mc, mr); |
| 190 | const size_t nc_stride = benchmark::utils::roundUp(nc, nr); |
| 191 | |
| 192 | std::random_device random_device; |
| 193 | auto rng = std::mt19937(random_device()); |
| 194 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 195 | |
| 196 | std::vector<float> a(mc * kc); |
| 197 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 198 | std::vector<float> k(nc * kc); |
| 199 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 200 | std::vector<float> b(nc); |
| 201 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 202 | |
| 203 | std::vector<uint32_t, AlignedAllocator<uint32_t, 32>> t(mc_stride * kc); |
| 204 | |
| 205 | const size_t w_elements = nc_stride * kc + nc_stride; |
| 206 | const size_t c_elements = mc * nc; |
| 207 | const size_t num_buffers = 1 + |
| 208 | benchmark::utils::divideRoundUp<size_t>(cpuinfo_get_max_cache_size(), |
| 209 | sizeof(float) * (w_elements + c_elements)); |
| 210 | |
| 211 | std::vector<float, AlignedAllocator<float, 32>> w(w_elements * num_buffers); |
| 212 | std::fill(w.begin(), w.end(), 0.0f); |
| 213 | xnn_pack_f32_gemm_goi_w(1 /* groups */, nc, kc, nr, 1 /* kr */, 1 /* sr */, k.data(), b.data(), w.data()); |
| 214 | std::vector<float> c(c_elements * num_buffers); |
| 215 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 216 | |
| 217 | xnn_f32_output_params output_params = |
| 218 | xnn_compute_f32_output_params(-std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
| 219 | |
| 220 | size_t buffer_index = 0; |
| 221 | for (auto _ : state) { |
| 222 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 223 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 224 | // - W is not in cache (for any cache level) |
| 225 | // - C is not in cache (for any cache level) |
| 226 | state.PauseTiming(); |
| 227 | benchmark::utils::prefetchToL1(a.data(), a.size() * sizeof(float)); |
| 228 | buffer_index = (buffer_index + 1) % num_buffers; |
| 229 | state.ResumeTiming(); |
| 230 | |
| 231 | for (uint32_t m = 0; m < mc; m += mr) { |
| 232 | const uint32_t mb = min(mc - m, mr); |
| 233 | packx(mb, kc, reinterpret_cast<const uint32_t*>(a.data() + m * kc), kc, t.data() + m * kc); |
| 234 | } |
| 235 | for (uint32_t m = 0; m < mc; m += mr) { |
| 236 | const uint32_t mb = min(mc - m, mr); |
| 237 | ppmm( |
| 238 | mb, nc, kc * sizeof(float), |
| 239 | reinterpret_cast<const float*>(t.data() + m * kc), |
| 240 | w.data() + nc_stride * buffer_index * (kc + 1), |
| 241 | c.data() + (mc * buffer_index + m) * nc, nc * sizeof(float), nr * sizeof(float), |
| 242 | &output_params); |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 247 | state.counters["FLOPS"] = benchmark::Counter( |
| 248 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 249 | } |
| 250 | |
| 251 | static void RuyBenchmark(benchmark::State& state, uint32_t threads) |
| 252 | { |
| 253 | std::random_device random_device; |
| 254 | auto rng = std::mt19937(random_device()); |
| 255 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 256 | |
| 257 | const size_t mc = state.range(0); |
| 258 | const size_t nc = state.range(1); |
| 259 | const size_t kc = state.range(2); |
| 260 | |
| 261 | const size_t num_buffers = 1 + |
| 262 | benchmark::utils::divideRoundUp<size_t>(cpuinfo_get_max_cache_size(), |
| 263 | sizeof(float) * (nc * (mc + kc + 1))); |
| 264 | |
| 265 | std::vector<float> a(mc * kc); |
| 266 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 267 | std::vector<float> k(num_buffers * nc * kc); |
| 268 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 269 | std::vector<float> b(num_buffers * nc); |
| 270 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 271 | std::vector<float> c(num_buffers * nc * mc); |
| 272 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 273 | |
| 274 | // Note: context must be static to avoid the cost of re-creating it for each benchmark. |
| 275 | static ruy::Context context; |
| 276 | context.max_num_threads = threads; |
| 277 | |
| 278 | ruy::Matrix<float> ruy_a; |
| 279 | ruy::MakeSimpleLayout(nc, kc, ruy::Order::kRowMajor, &ruy_a.layout); |
| 280 | ruy::Matrix<float> ruy_b; |
| 281 | ruy::MakeSimpleLayout(kc, mc, ruy::Order::kColMajor, &ruy_b.layout); |
| 282 | ruy_b.data = a.data(); |
| 283 | ruy::Matrix<float> ruy_c; |
| 284 | ruy::MakeSimpleLayout(nc, mc, ruy::Order::kColMajor, &ruy_c.layout); |
| 285 | |
| 286 | ruy::BasicSpec<float, float> spec; |
| 287 | |
| 288 | // ruy::Context uses deferred initialization, which affects percieved GEMM performance. Initialization happens during |
| 289 | // the first GEMM calls, and per Benoit Jacob it takes up to ~250 milliseconds for performance to stabilize. |
| 290 | // Thus, on the first benchmark, we compute GEMM for 500 milliseconds (to be safe) without recording performance, and |
| 291 | // keep the ruy::Context object initialized (by being static) between subsequent benchmarks. |
| 292 | static std::once_flag warmup; |
| 293 | std::call_once(warmup, [&](){ |
| 294 | auto start = std::chrono::steady_clock::now(); |
| 295 | do { |
| 296 | ruy_a.data = k.data(); |
| 297 | ruy_c.data = c.data(); |
| 298 | spec.bias = b.data(); |
| 299 | |
| 300 | ruy::Mul<ruy::kAllPaths>(ruy_a, ruy_b, spec, &context, &ruy_c); |
| 301 | } while (std::chrono::duration<double>(std::chrono::steady_clock::now() - start).count() < 0.5); |
| 302 | }); |
| 303 | |
| 304 | size_t buffer_index = 0; |
| 305 | for (auto _ : state) { |
| 306 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 307 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 308 | // - K is not in cache (for any cache level) |
| 309 | // - B is not in cache (for any cache level) |
| 310 | // - C is not in cache (for any cache level) |
| 311 | state.PauseTiming(); |
| 312 | benchmark::utils::prefetchToL1(a.data(), a.size() * sizeof(float)); |
| 313 | buffer_index = (buffer_index + 1) % num_buffers; |
| 314 | state.ResumeTiming(); |
| 315 | |
| 316 | ruy_a.data = k.data() + buffer_index * nc * kc; |
| 317 | ruy_c.data = c.data() + buffer_index * mc * nc; |
| 318 | spec.bias = b.data() + buffer_index * nc; |
| 319 | |
| 320 | ruy::Mul<ruy::kAllPaths>(ruy_a, ruy_b, spec, &context, &ruy_c); |
| 321 | } |
| 322 | |
| 323 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 324 | state.counters["FLOPS"] = benchmark::Counter( |
| 325 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 326 | } |
| 327 | |
| 328 | static void ruy_st(benchmark::State& state, const char* net) |
| 329 | { |
| 330 | RuyBenchmark(state, 1); |
| 331 | } |
| 332 | |
| 333 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 334 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 335 | static void sgemm_1x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
| 336 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x12__aarch64_neonfma_cortex_a53, 1, 12, 1, 1); |
| 337 | } |
| 338 | static void sgemm_1x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { |
| 339 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a57, 1, 8, 1, 1); |
| 340 | } |
| 341 | static void sgemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 342 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75, 1, 8, 1, 1); |
| 343 | } |
| 344 | static void sgemm_4x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
| 345 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x12__aarch64_neonfma_cortex_a53, 4, 12, 1, 1); |
| 346 | } |
| 347 | static void sgemm_4x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { |
| 348 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a57, 4, 8, 1, 1); |
| 349 | } |
| 350 | static void sgemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 351 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a75, 4, 8, 1, 1); |
| 352 | } |
| 353 | static void sgemm_4x8__aarch64_neonfma_ld64(benchmark::State& state, const char* net) { |
| 354 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_ld64, 4, 8, 1, 1); |
| 355 | } |
| 356 | static void sgemm_4x8__aarch64_neonfma_ld128(benchmark::State& state, const char* net) { |
| 357 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_ld128, 4, 8, 1, 1); |
| 358 | } |
| 359 | static void sgemm_5x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 360 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_5x8__aarch64_neonfma_cortex_a75, 5, 8, 1, 1); |
| 361 | } |
| 362 | static void sgemm_6x8__aarch64_neonfma_ld64(benchmark::State& state, const char* net) { |
| 363 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_ld64, 6, 8, 1, 1); |
| 364 | } |
| 365 | static void sgemm_6x8__aarch64_neonfma_ld128(benchmark::State& state, const char* net) { |
| 366 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_ld128, 6, 8, 1, 1); |
| 367 | } |
| 368 | static void sgemm_6x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { |
| 369 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a57, 6, 8, 1, 1); |
| 370 | } |
| 371 | static void sgemm_6x8__aarch64_neonfma_cortex_a73(benchmark::State& state, const char* net) { |
| 372 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a73, 6, 8, 1, 1); |
| 373 | } |
| 374 | static void sgemm_6x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 375 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a75, 6, 8, 1, 1); |
| 376 | } |
| 377 | |
| 378 | BENCHMARK_GEMM(sgemm_1x12__aarch64_neonfma_cortex_a53) |
| 379 | BENCHMARK_GEMM(sgemm_1x8__aarch64_neonfma_cortex_a57) |
| 380 | BENCHMARK_GEMM(sgemm_1x8__aarch64_neonfma_cortex_a75) |
| 381 | BENCHMARK_GEMM(sgemm_4x12__aarch64_neonfma_cortex_a53) |
| 382 | BENCHMARK_GEMM(sgemm_4x8__aarch64_neonfma_cortex_a57) |
| 383 | BENCHMARK_GEMM(sgemm_4x8__aarch64_neonfma_cortex_a75) |
| 384 | BENCHMARK_GEMM(sgemm_4x8__aarch64_neonfma_ld128) |
| 385 | BENCHMARK_GEMM(sgemm_4x8__aarch64_neonfma_ld64) |
| 386 | BENCHMARK_GEMM(sgemm_5x8__aarch64_neonfma_cortex_a75) |
| 387 | BENCHMARK_GEMM(sgemm_6x8__aarch64_neonfma_cortex_a57) |
| 388 | BENCHMARK_GEMM(sgemm_6x8__aarch64_neonfma_cortex_a73) |
| 389 | BENCHMARK_GEMM(sgemm_6x8__aarch64_neonfma_cortex_a75) |
| 390 | BENCHMARK_GEMM(sgemm_6x8__aarch64_neonfma_ld64) |
| 391 | BENCHMARK_GEMM(sgemm_6x8__aarch64_neonfma_ld128) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 392 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 393 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 394 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 395 | static void sgemm_4x12__neon_ld64(benchmark::State& state, const char* net) { |
| 396 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x12__neon_ld64, 4, 12, 1, 1); |
| 397 | } |
| 398 | |
| 399 | static void sgemm_1x8__neon_ld64(benchmark::State& state, const char* net) { |
| 400 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x8__neon_ld64, 1, 8, 1, 1); |
| 401 | } |
| 402 | |
| 403 | static void sgemm_1x8__neonfma_ld64(benchmark::State& state, const char* net) { |
| 404 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x8__neonfma_ld64, 1, 8, 1, 1); |
| 405 | } |
| 406 | |
| 407 | static void sgemm_4x8__neon_ld64(benchmark::State& state, const char* net) { |
| 408 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__neon_ld64, 4, 8, 1, 1); |
| 409 | } |
| 410 | |
| 411 | static void sgemm_4x8__neon_ld128(benchmark::State& state, const char* net) { |
| 412 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__neon_ld128, 4, 8, 1, 1); |
| 413 | } |
| 414 | |
| 415 | static void sgemm_5x8__neon_ld64(benchmark::State& state, const char* net) { |
| 416 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_5x8__neon_ld64, 5, 8, 1, 1); |
| 417 | } |
| 418 | |
| 419 | static void sgemm_6x8__neon_ld64(benchmark::State& state, const char* net) { |
| 420 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__neon_ld64, 6, 8, 1, 1); |
| 421 | } |
| 422 | |
| 423 | static void sgemm_4x12__neonfma_ld64(benchmark::State& state, const char* net) { |
| 424 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x12__neonfma_ld64, 4, 12, 1, 1); |
| 425 | } |
| 426 | |
| 427 | static void sgemm_4x8__neonfma_ld64(benchmark::State& state, const char* net) { |
| 428 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__neonfma_ld64, 4, 8, 1, 1); |
| 429 | } |
| 430 | |
| 431 | static void sgemm_4x8__neonfma_ld128(benchmark::State& state, const char* net) { |
| 432 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__neonfma_ld128, 4, 8, 1, 1); |
| 433 | } |
| 434 | |
| 435 | static void sgemm_5x8__neonfma_ld64(benchmark::State& state, const char* net) { |
| 436 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_5x8__neonfma_ld64, 5, 8, 1, 1); |
| 437 | } |
| 438 | |
| 439 | static void sgemm_6x8__neonfma_ld64(benchmark::State& state, const char* net) { |
| 440 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__neonfma_ld64, 6, 8, 1, 1); |
| 441 | } |
| 442 | |
| 443 | static void sppmm_4x8_unipass__neonfma(benchmark::State& state, const char* net) { |
| 444 | PPMM1PBenchmark(state, xnn_f32_ppmm_ukernel_4x8__neonfma, xnn_x32_packx_ukernel_4x__neon_st4, 4, 8); |
| 445 | } |
| 446 | |
| 447 | static void sppmm_4x8_twopass__neonfma(benchmark::State& state, const char* net) { |
| 448 | PPMM2PBenchmark(state, xnn_f32_ppmm_ukernel_4x8__neonfma, xnn_x32_packx_ukernel_4x__neon_st4, 4, 8); |
| 449 | } |
| 450 | |
| 451 | BENCHMARK_GEMM(sgemm_4x12__neon_ld64) |
| 452 | BENCHMARK_GEMM(sgemm_4x12__neonfma_ld64) |
| 453 | BENCHMARK_GEMM(sgemm_1x8__neon_ld64) |
| 454 | BENCHMARK_GEMM(sgemm_1x8__neonfma_ld64) |
| 455 | BENCHMARK_GEMM(sgemm_4x8__neon_ld128) |
| 456 | BENCHMARK_GEMM(sgemm_4x8__neon_ld64) |
| 457 | BENCHMARK_GEMM(sgemm_4x8__neonfma_ld128) |
| 458 | BENCHMARK_GEMM(sgemm_4x8__neonfma_ld64) |
| 459 | BENCHMARK_GEMM(sgemm_5x8__neon_ld64) |
| 460 | BENCHMARK_GEMM(sgemm_5x8__neonfma_ld64) |
| 461 | BENCHMARK_GEMM(sgemm_6x8__neon_ld64) |
| 462 | BENCHMARK_GEMM(sgemm_6x8__neonfma_ld64) |
| 463 | |
| 464 | BENCHMARK_GEMM(sppmm_4x8_unipass__neonfma) |
| 465 | BENCHMARK_GEMM(sppmm_4x8_twopass__neonfma) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 466 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 467 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 468 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 469 | static void sgemm_1x8__sse_load1(benchmark::State& state, const char* net) { |
| 470 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x8__sse_load1, 1, 8, 1, 1); |
| 471 | } |
| 472 | |
| 473 | static void sgemm_4x8__sse_load1(benchmark::State& state, const char* net) { |
| 474 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__sse_load1, 4, 8, 1, 1); |
| 475 | } |
| 476 | |
| 477 | static void sgemm_1x8__sse_dup(benchmark::State& state, const char* net) { |
| 478 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x8__sse_dup, 1, 8, 1, 1); |
| 479 | } |
| 480 | |
| 481 | static void sgemm_4x8__sse_dup(benchmark::State& state, const char* net) { |
| 482 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__sse_dup, 4, 8, 1, 1); |
| 483 | } |
| 484 | |
| 485 | static void sgemm_1x8s4__sse(benchmark::State& state, const char* net) { |
| 486 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x8s4__sse, 1, 8, 1, 4); |
| 487 | } |
| 488 | |
| 489 | static void sgemm_4x8s4__sse(benchmark::State& state, const char* net) { |
| 490 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8s4__sse, 4, 8, 1, 4); |
| 491 | } |
| 492 | |
| 493 | static void sppmm_4x8_unipass__sse(benchmark::State& state, const char* net) { |
| 494 | PPMM1PBenchmark(state, xnn_f32_ppmm_ukernel_4x8__sse, xnn_x32_packx_ukernel_4x__sse, 4, 8); |
| 495 | } |
| 496 | |
| 497 | static void sppmm_4x8_twopass__sse(benchmark::State& state, const char* net) { |
| 498 | PPMM2PBenchmark(state, xnn_f32_ppmm_ukernel_4x8__sse, xnn_x32_packx_ukernel_4x__sse, 4, 8); |
| 499 | } |
| 500 | |
| 501 | BENCHMARK_GEMM(sgemm_1x8__sse_load1) |
| 502 | BENCHMARK_GEMM(sgemm_4x8__sse_load1) |
| 503 | BENCHMARK_GEMM(sgemm_1x8__sse_dup) |
| 504 | BENCHMARK_GEMM(sgemm_4x8__sse_dup) |
| 505 | BENCHMARK_GEMM(sgemm_1x8s4__sse) |
| 506 | BENCHMARK_GEMM(sgemm_4x8s4__sse) |
| 507 | BENCHMARK_GEMM(sppmm_4x8_unipass__sse) |
| 508 | BENCHMARK_GEMM(sppmm_4x8_twopass__sse) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 509 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 510 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 511 | #if !XNN_ARCH_WASM && !XNN_ARCH_ASMJS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 512 | static void sgemm_4x8__psimd_loadsplat(benchmark::State& state, const char* net) { |
| 513 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__psimd_loadsplat, 4, 8, 1, 1); |
| 514 | } |
| 515 | |
| 516 | static void sgemm_6x8__psimd_loadsplat(benchmark::State& state, const char* net) { |
| 517 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__psimd_loadsplat, 6, 8, 1, 1); |
| 518 | } |
| 519 | |
| 520 | static void sgemm_4x8__psimd_splat(benchmark::State& state, const char* net) { |
| 521 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8__psimd_splat, 4, 8, 1, 1); |
| 522 | } |
| 523 | |
| 524 | static void sgemm_6x8__psimd_splat(benchmark::State& state, const char* net) { |
| 525 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8__psimd_splat, 6, 8, 1, 1); |
| 526 | } |
| 527 | |
| 528 | static void sgemm_4x8s4__psimd(benchmark::State& state, const char* net) { |
| 529 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x8s4__psimd, 4, 8, 1, 4); |
| 530 | } |
| 531 | |
| 532 | static void sgemm_6x8s4__psimd(benchmark::State& state, const char* net) { |
| 533 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_6x8s4__psimd, 6, 8, 1, 4); |
| 534 | } |
| 535 | |
| 536 | static void sppmm_4x8_unipass__psimd(benchmark::State& state, const char* net) { |
| 537 | PPMM1PBenchmark(state, xnn_f32_ppmm_ukernel_4x8__psimd, xnn_x32_packx_ukernel_4x__psimd, 4, 8); |
| 538 | } |
| 539 | |
| 540 | static void sppmm_4x8_twopass__psimd(benchmark::State& state, const char* net) { |
| 541 | PPMM2PBenchmark(state, xnn_f32_ppmm_ukernel_4x8__psimd, xnn_x32_packx_ukernel_4x__psimd, 4, 8); |
| 542 | } |
| 543 | |
| 544 | BENCHMARK_GEMM(sgemm_4x8__psimd_loadsplat) |
| 545 | BENCHMARK_GEMM(sgemm_6x8__psimd_loadsplat) |
| 546 | BENCHMARK_GEMM(sgemm_4x8__psimd_splat) |
| 547 | BENCHMARK_GEMM(sgemm_6x8__psimd_splat) |
| 548 | BENCHMARK_GEMM(sgemm_4x8s4__psimd) |
| 549 | BENCHMARK_GEMM(sgemm_6x8s4__psimd) |
| 550 | BENCHMARK_GEMM(sppmm_4x8_unipass__psimd) |
| 551 | BENCHMARK_GEMM(sppmm_4x8_twopass__psimd) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame^] | 552 | #endif // !XNN_ARCH_WASM && !XNN_ARCH_ASMJS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 553 | |
| 554 | static void sgemm_1x4__scalar(benchmark::State& state, const char* net) { |
| 555 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_1x4__scalar, 1, 4, 1, 1); |
| 556 | } |
| 557 | |
| 558 | static void sgemm_2x4__scalar(benchmark::State& state, const char* net) { |
| 559 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_2x4__scalar, 2, 4, 1, 1); |
| 560 | } |
| 561 | |
| 562 | static void sgemm_4x4__scalar(benchmark::State& state, const char* net) { |
| 563 | GEMMBenchmark(state, xnn_f32_gemm_ukernel_4x4__scalar, 4, 4, 1, 1); |
| 564 | } |
| 565 | |
| 566 | static void sppmm_2x4_unipass__scalar(benchmark::State& state, const char* net) { |
| 567 | PPMM1PBenchmark(state, xnn_f32_ppmm_ukernel_2x4__scalar, xnn_x32_packx_ukernel_2x__scalar, 2, 4); |
| 568 | } |
| 569 | |
| 570 | static void sppmm_4x2_unipass__scalar(benchmark::State& state, const char* net) { |
| 571 | PPMM1PBenchmark(state, xnn_f32_ppmm_ukernel_4x2__scalar, xnn_x32_packx_ukernel_4x__scalar, 4, 2); |
| 572 | } |
| 573 | |
| 574 | static void sppmm_4x4_unipass__scalar(benchmark::State& state, const char* net) { |
| 575 | PPMM1PBenchmark(state, xnn_f32_ppmm_ukernel_4x4__scalar, xnn_x32_packx_ukernel_4x__scalar, 4, 4); |
| 576 | } |
| 577 | |
| 578 | static void sppmm_3x3_unipass__scalar(benchmark::State& state, const char* net) { |
| 579 | PPMM1PBenchmark(state, xnn_f32_ppmm_ukernel_3x3__scalar, xnn_x32_packx_ukernel_3x__scalar, 3, 3); |
| 580 | } |
| 581 | |
| 582 | static void sppmm_2x4_twopass__scalar(benchmark::State& state, const char* net) { |
| 583 | PPMM2PBenchmark(state, xnn_f32_ppmm_ukernel_2x4__scalar, xnn_x32_packx_ukernel_2x__scalar, 2, 4); |
| 584 | } |
| 585 | |
| 586 | static void sppmm_4x2_twopass__scalar(benchmark::State& state, const char* net) { |
| 587 | PPMM2PBenchmark(state, xnn_f32_ppmm_ukernel_4x2__scalar, xnn_x32_packx_ukernel_4x__scalar, 4, 2); |
| 588 | } |
| 589 | |
| 590 | static void sppmm_4x4_twopass__scalar(benchmark::State& state, const char* net) { |
| 591 | PPMM2PBenchmark(state, xnn_f32_ppmm_ukernel_4x4__scalar, xnn_x32_packx_ukernel_4x__scalar, 4, 4); |
| 592 | } |
| 593 | |
| 594 | static void sppmm_3x3_twopass__scalar(benchmark::State& state, const char* net) { |
| 595 | PPMM2PBenchmark(state, xnn_f32_ppmm_ukernel_3x3__scalar, xnn_x32_packx_ukernel_3x__scalar, 3, 3); |
| 596 | } |
| 597 | |
| 598 | BENCHMARK_GEMM(sgemm_1x4__scalar) |
| 599 | BENCHMARK_GEMM(sgemm_2x4__scalar) |
| 600 | BENCHMARK_GEMM(sgemm_4x4__scalar) |
| 601 | |
| 602 | BENCHMARK_GEMM(sppmm_2x4_unipass__scalar) |
| 603 | BENCHMARK_GEMM(sppmm_4x2_unipass__scalar) |
| 604 | BENCHMARK_GEMM(sppmm_4x4_unipass__scalar) |
| 605 | BENCHMARK_GEMM(sppmm_3x3_unipass__scalar) |
| 606 | |
| 607 | BENCHMARK_GEMM(sppmm_2x4_twopass__scalar) |
| 608 | BENCHMARK_GEMM(sppmm_4x2_twopass__scalar) |
| 609 | BENCHMARK_GEMM(sppmm_4x4_twopass__scalar) |
| 610 | BENCHMARK_GEMM(sppmm_3x3_twopass__scalar) |
| 611 | |
| 612 | |
| 613 | BENCHMARK_GEMM(ruy_st) |
| 614 | |
| 615 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 616 | BENCHMARK_MAIN(); |
| 617 | #endif |