XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -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 <cfloat> |
| 8 | #include <cmath> |
| 9 | #include <functional> |
| 10 | #include <random> |
| 11 | #include <vector> |
| 12 | |
| 13 | #include <cpuinfo.h> |
| 14 | |
| 15 | #include <benchmark/benchmark.h> |
| 16 | #include "bench/conv.h" |
| 17 | #include "bench/utils.h" |
| 18 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 19 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | #include <xnnpack/igemm.h> |
| 21 | #include <xnnpack/indirection.h> |
| 22 | #include <xnnpack/operator.h> |
| 23 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 24 | #include <xnnpack/params-init.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | |
| 27 | |
| 28 | static void IGEMMBenchmark(benchmark::State& state, |
| 29 | xnn_f32_igemm_ukernel_function f32_igemm, |
| 30 | uint32_t mr, uint32_t nr, uint32_t kr, uint32_t sr) |
| 31 | { |
| 32 | if (!cpuinfo_initialize()) { |
| 33 | state.SkipWithError("cpuinfo initialization failed"); |
| 34 | } |
| 35 | |
| 36 | const size_t input_height = state.range(0); |
| 37 | const size_t input_width = state.range(1); |
| 38 | const size_t kernel_height = state.range(2); |
| 39 | const size_t kernel_width = state.range(3); |
| 40 | const size_t kernel_size = kernel_height * kernel_width; |
| 41 | const size_t padding_height = state.range(4); |
| 42 | const size_t padding_width = state.range(5); |
| 43 | const size_t subsampling = state.range(6); |
| 44 | const size_t dilation = state.range(7); |
| 45 | const size_t group_input_channels = state.range(8); |
| 46 | const size_t group_output_channels = state.range(9); |
| 47 | |
| 48 | std::random_device random_device; |
| 49 | auto rng = std::mt19937(random_device()); |
| 50 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 51 | |
| 52 | const size_t output_pixel_stride = group_output_channels; |
| 53 | const size_t input_pixel_stride = group_input_channels; |
| 54 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 55 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 56 | const size_t padding_left = padding_width / 2; |
| 57 | const size_t padding_top = padding_height / 2; |
| 58 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 59 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 60 | const size_t output_size = output_height * output_width; |
| 61 | |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 62 | const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr); |
| 63 | const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr); |
| 64 | const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 65 | |
| 66 | std::vector<float> a(input_height * input_width * input_pixel_stride); |
| 67 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 68 | std::vector<float> k(group_output_channels * kernel_height * kernel_width * group_input_channels); |
| 69 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 70 | std::vector<float> b(group_output_channels); |
| 71 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 72 | |
| 73 | std::vector<float> z(group_input_channels); |
| 74 | |
| 75 | const size_t w_elements = (kernel_size * kc_stride + 1) * nc_stride; |
| 76 | const size_t i_elements = mc_stride * kernel_size; |
| 77 | const size_t c_elements = output_height * output_width * output_pixel_stride; |
| 78 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 79 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 80 | sizeof(float) * (w_elements + c_elements) + sizeof(void*) * i_elements); |
| 81 | |
| 82 | std::vector<float, AlignedAllocator<float, 32>> w(w_elements * num_buffers); |
| 83 | std::fill(w.begin(), w.end(), 0.0f); |
| 84 | xnn_pack_f32_conv_goki_w( |
| 85 | 1 /* groups */, group_output_channels, kernel_size, group_input_channels, |
| 86 | nr, kr, sr, k.data(), b.data(), w.data()); |
| 87 | for (size_t n = 1; n < num_buffers; n++) { |
| 88 | std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements); |
| 89 | } |
| 90 | |
| 91 | std::vector<const float*> i(i_elements * num_buffers); |
| 92 | xnn_operator convolution_op = { }; |
| 93 | convolution_op.indirection_buffer = reinterpret_cast<const void**>(i.data()); |
| 94 | convolution_op.input = a.data(); |
| 95 | convolution_op.input_pixel_stride = input_pixel_stride; |
| 96 | convolution_op.zero_buffer = z.data(); |
| 97 | convolution_op.groups = 1; |
| 98 | convolution_op.group_input_channels = group_input_channels; |
| 99 | convolution_op.batch_size = 1; |
| 100 | convolution_op.input_height = input_height; |
| 101 | convolution_op.input_width = input_width; |
| 102 | convolution_op.output_height = output_height; |
| 103 | convolution_op.output_width = output_width; |
| 104 | convolution_op.kernel_height = kernel_height; |
| 105 | convolution_op.kernel_width = kernel_width; |
| 106 | convolution_op.stride_height = subsampling; |
| 107 | convolution_op.stride_width = subsampling; |
| 108 | convolution_op.dilation_height = dilation; |
| 109 | convolution_op.dilation_width = dilation; |
| 110 | convolution_op.padding_top = padding_top; |
| 111 | convolution_op.padding_left = padding_left; |
| 112 | xnn_indirection_init_conv2d(&convolution_op, mr, 2 /* log2(sizeof(float)) */); |
| 113 | for (size_t n = 1; n < num_buffers; n++) { |
| 114 | std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements); |
| 115 | } |
| 116 | |
| 117 | std::vector<float> c(c_elements * num_buffers); |
| 118 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 119 | |
| 120 | xnn_f32_output_params output_params = |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 121 | xnn_init_f32_output_params(-std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 122 | |
| 123 | size_t buffer_index = 0; |
| 124 | for (auto _ : state) { |
| 125 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 126 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 127 | buffer_index = (buffer_index + 1) % num_buffers; |
| 128 | state.ResumeTiming(); |
| 129 | |
| 130 | for (uint32_t m = 0; m < output_size; m += mr) { |
| 131 | const uint32_t mb = min(output_size - m, mr); |
| 132 | for (uint32_t n = 0; n < group_output_channels; n += nr) { |
| 133 | const uint32_t nb = min(group_output_channels - n, nr); |
| 134 | f32_igemm( |
| 135 | mb, nb, group_input_channels * sizeof(float), kernel_size * mr * sizeof(void*), |
| 136 | i.data() + buffer_index * i_elements + m, |
| 137 | w.data() + buffer_index * w_elements + n * (kc_stride * kernel_size + 1), |
| 138 | c.data() + buffer_index * c_elements + m * group_output_channels + n, group_output_channels * sizeof(float), nr * sizeof(float), |
| 139 | 0, z.data(), &output_params); |
| 140 | } |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 145 | state.counters["FLOPS"] = benchmark::Counter( |
| 146 | uint64_t(state.iterations()) * 2 * |
| 147 | output_height * output_width * |
| 148 | group_input_channels * group_output_channels * |
| 149 | kernel_height * kernel_width, |
| 150 | benchmark::Counter::kIsRate); |
| 151 | } |
| 152 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 153 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Frank Barchard | db45b6a | 2019-10-09 16:42:45 -0700 | [diff] [blame] | 154 | static void f32_igemm_1x8__neon_ld64(benchmark::State& state, const char* net) { |
| 155 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__neon_ld64, 1, 8, 1, 1); |
| 156 | } |
| 157 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 158 | static void f32_igemm_4x2__neon_ld64(benchmark::State& state, const char* net) { |
| 159 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x2__neon_ld64, 4, 2, 1, 1); |
| 160 | } |
| 161 | |
| 162 | static void f32_igemm_4x4__neon_ld64(benchmark::State& state, const char* net) { |
| 163 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x4__neon_ld64, 4, 4, 1, 1); |
| 164 | } |
| 165 | |
| 166 | static void f32_igemm_4x8__neon_ld128(benchmark::State& state, const char* net) { |
| 167 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neon_ld128, 4, 8, 1, 1); |
| 168 | } |
| 169 | |
| 170 | static void f32_igemm_4x8__neon_ld64(benchmark::State& state, const char* net) { |
| 171 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neon_ld64, 4, 8, 1, 1); |
| 172 | } |
| 173 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 174 | static void f32_igemm_6x8__neon_ld64(benchmark::State& state, const char* net) { |
| 175 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__neon_ld64, 6, 8, 1, 1); |
| 176 | } |
| 177 | |
Frank Barchard | db45b6a | 2019-10-09 16:42:45 -0700 | [diff] [blame] | 178 | static void f32_igemm_1x8__neonfma_ld64(benchmark::State& state, const char* net) { |
| 179 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__neonfma_ld64, 1, 8, 1, 1); |
| 180 | } |
| 181 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 182 | static void f32_igemm_4x2__neonfma_ld64(benchmark::State& state, const char* net) { |
| 183 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x2__neonfma_ld64, 4, 2, 1, 1); |
| 184 | } |
| 185 | |
| 186 | static void f32_igemm_4x4__neonfma_ld64(benchmark::State& state, const char* net) { |
| 187 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x4__neonfma_ld64, 4, 4, 1, 1); |
| 188 | } |
| 189 | |
| 190 | static void f32_igemm_4x8__neonfma_ld128(benchmark::State& state, const char* net) { |
| 191 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neonfma_ld128, 4, 8, 1, 1); |
| 192 | } |
| 193 | |
| 194 | static void f32_igemm_4x8__neonfma_ld64(benchmark::State& state, const char* net) { |
| 195 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neonfma_ld64, 4, 8, 1, 1); |
| 196 | } |
| 197 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 198 | static void f32_igemm_6x8__neonfma_ld64(benchmark::State& state, const char* net) { |
| 199 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__neonfma_ld64, 6, 8, 1, 1); |
| 200 | } |
| 201 | |
Frank Barchard | df06d80 | 2019-11-20 15:53:46 -0800 | [diff] [blame] | 202 | static void f32_igemm_1x8s4__neon(benchmark::State& state, const char* net) { |
| 203 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8s4__neon, 1, 8, 1, 4); |
| 204 | } |
| 205 | |
| 206 | static void f32_igemm_1x8s4__neonfma(benchmark::State& state, const char* net) { |
| 207 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8s4__neonfma, 1, 8, 1, 4); |
| 208 | } |
| 209 | |
| 210 | static void f32_igemm_4x8s4__neon(benchmark::State& state, const char* net) { |
| 211 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8s4__neon, 4, 8, 1, 4); |
| 212 | } |
| 213 | |
| 214 | static void f32_igemm_4x8s4__neonfma(benchmark::State& state, const char* net) { |
| 215 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8s4__neonfma, 4, 8, 1, 4); |
| 216 | } |
| 217 | |
| 218 | static void f32_igemm_6x8s4__neon(benchmark::State& state, const char* net) { |
| 219 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8s4__neon, 6, 8, 1, 4); |
| 220 | } |
| 221 | |
| 222 | static void f32_igemm_6x8s4__neonfma(benchmark::State& state, const char* net) { |
| 223 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8s4__neonfma, 6, 8, 1, 4); |
| 224 | } |
| 225 | |
| 226 | static void f32_igemm_8x8s4__neon(benchmark::State& state, const char* net) { |
| 227 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_8x8s4__neon, 8, 8, 1, 4); |
| 228 | } |
| 229 | |
| 230 | static void f32_igemm_8x8s4__neonfma(benchmark::State& state, const char* net) { |
| 231 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_8x8s4__neonfma, 8, 8, 1, 4); |
| 232 | } |
| 233 | |
Frank Barchard | db45b6a | 2019-10-09 16:42:45 -0700 | [diff] [blame] | 234 | BENCHMARK_CONV(f32_igemm_1x8__neon_ld64) |
| 235 | BENCHMARK_CONV(f32_igemm_1x8__neonfma_ld64) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 236 | BENCHMARK_CONV(f32_igemm_4x2__neon_ld64) |
| 237 | BENCHMARK_CONV(f32_igemm_4x2__neonfma_ld64) |
| 238 | BENCHMARK_CONV(f32_igemm_4x4__neon_ld64) |
| 239 | BENCHMARK_CONV(f32_igemm_4x4__neonfma_ld64) |
| 240 | BENCHMARK_CONV(f32_igemm_4x8__neon_ld128) |
| 241 | BENCHMARK_CONV(f32_igemm_4x8__neon_ld64) |
| 242 | BENCHMARK_CONV(f32_igemm_4x8__neonfma_ld128) |
| 243 | BENCHMARK_CONV(f32_igemm_4x8__neonfma_ld64) |
Frank Barchard | df06d80 | 2019-11-20 15:53:46 -0800 | [diff] [blame] | 244 | BENCHMARK_CONV(f32_igemm_1x8s4__neon) |
| 245 | BENCHMARK_CONV(f32_igemm_1x8s4__neonfma) |
| 246 | BENCHMARK_CONV(f32_igemm_4x8s4__neon) |
| 247 | BENCHMARK_CONV(f32_igemm_4x8s4__neonfma) |
| 248 | BENCHMARK_CONV(f32_igemm_6x8s4__neon) |
| 249 | BENCHMARK_CONV(f32_igemm_6x8s4__neonfma) |
| 250 | BENCHMARK_CONV(f32_igemm_8x8s4__neon) |
| 251 | BENCHMARK_CONV(f32_igemm_8x8s4__neonfma) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 252 | BENCHMARK_CONV(f32_igemm_6x8__neon_ld64) |
| 253 | BENCHMARK_CONV(f32_igemm_6x8__neonfma_ld64) |
| 254 | #endif |
| 255 | |
Frank Barchard | dbafc58 | 2019-10-09 16:30:48 -0700 | [diff] [blame] | 256 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 257 | static void f32_igemm_1x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
| 258 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x12__aarch64_neonfma_cortex_a53, 1, 12, 1, 1); |
| 259 | } |
| 260 | |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 261 | static void f32_igemm_1x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
| 262 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a53, 1, 8, 1, 1); |
| 263 | } |
| 264 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 265 | static void f32_igemm_1x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { |
| 266 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a57, 1, 8, 1, 1); |
| 267 | } |
| 268 | |
| 269 | static void f32_igemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 270 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, 1, 8, 1, 1); |
| 271 | } |
| 272 | |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 273 | static void f32_igemm_4x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
| 274 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__aarch64_neonfma_cortex_a53, 4, 8, 1, 1); |
| 275 | } |
| 276 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 277 | static void f32_igemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 278 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__aarch64_neonfma_cortex_a75, 4, 8, 1, 1); |
| 279 | } |
| 280 | |
| 281 | static void f32_igemm_5x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 282 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_5x8__aarch64_neonfma_cortex_a75, 5, 8, 1, 1); |
| 283 | } |
| 284 | |
| 285 | static void f32_igemm_4x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
| 286 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x12__aarch64_neonfma_cortex_a53, 4, 12, 1, 1); |
| 287 | } |
| 288 | |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 289 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
| 290 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a53, 6, 8, 1, 1); |
| 291 | } |
| 292 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 293 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { |
| 294 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a57, 6, 8, 1, 1); |
| 295 | } |
| 296 | |
| 297 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a73(benchmark::State& state, const char* net) { |
| 298 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a73, 6, 8, 1, 1); |
| 299 | } |
| 300 | |
| 301 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 302 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a75, 6, 8, 1, 1); |
| 303 | } |
| 304 | |
| 305 | BENCHMARK_CONV(f32_igemm_1x12__aarch64_neonfma_cortex_a53) |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 306 | BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 307 | BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_cortex_a57) |
| 308 | BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_cortex_a75) |
| 309 | BENCHMARK_CONV(f32_igemm_4x12__aarch64_neonfma_cortex_a53) |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 310 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 311 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_cortex_a75) |
| 312 | BENCHMARK_CONV(f32_igemm_5x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 313 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 314 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a57) |
| 315 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a73) |
| 316 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a75) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 317 | #endif /* XNN_ARCH_ARM64 */ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 318 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 319 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 320 | static void f32_igemm_1x8__sse_load1(benchmark::State& state, const char* net) { |
| 321 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__sse_load1, 1, 8, 1, 1); |
| 322 | } |
| 323 | |
| 324 | static void f32_igemm_4x8__sse_load1(benchmark::State& state, const char* net) { |
| 325 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__sse_load1, 4, 8, 1, 1); |
| 326 | } |
| 327 | |
| 328 | static void f32_igemm_1x8__sse_dup(benchmark::State& state, const char* net) { |
| 329 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__sse_dup, 1, 8, 1, 1); |
| 330 | } |
| 331 | |
| 332 | static void f32_igemm_4x8__sse_dup(benchmark::State& state, const char* net) { |
| 333 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__sse_dup, 4, 8, 1, 1); |
| 334 | } |
| 335 | |
| 336 | static void f32_igemm_1x8s4__sse(benchmark::State& state, const char* net) { |
| 337 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8s4__sse, 1, 8, 1, 4); |
| 338 | } |
| 339 | |
| 340 | static void f32_igemm_4x8s4__sse(benchmark::State& state, const char* net) { |
| 341 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8s4__sse, 4, 8, 1, 4); |
| 342 | } |
| 343 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame^] | 344 | static void f32_igemm_1x8__avx_broadcast(benchmark::State& state, const char* net) { |
| 345 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__avx_broadcast, 1, 8, 1, 1); |
| 346 | } |
| 347 | |
| 348 | static void f32_igemm_4x8__avx_broadcast(benchmark::State& state, const char* net) { |
| 349 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__avx_broadcast, 4, 8, 1, 1); |
| 350 | } |
| 351 | |
| 352 | static void f32_igemm_5x8__avx_broadcast(benchmark::State& state, const char* net) { |
| 353 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_5x8__avx_broadcast, 5, 8, 1, 1); |
| 354 | } |
| 355 | |
| 356 | static void f32_igemm_6x8__avx_broadcast(benchmark::State& state, const char* net) { |
| 357 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__avx_broadcast, 6, 8, 1, 1); |
| 358 | } |
| 359 | |
| 360 | static void f32_igemm_7x8__avx_broadcast(benchmark::State& state, const char* net) { |
| 361 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_7x8__avx_broadcast, 7, 8, 1, 1); |
| 362 | } |
| 363 | |
| 364 | static void f32_igemm_1x8__fma3_broadcast(benchmark::State& state, const char* net) { |
| 365 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__fma3_broadcast, 1, 8, 1, 1); |
| 366 | } |
| 367 | |
| 368 | static void f32_igemm_4x8__fma3_broadcast(benchmark::State& state, const char* net) { |
| 369 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__fma3_broadcast, 4, 8, 1, 1); |
| 370 | } |
| 371 | |
| 372 | static void f32_igemm_5x8__fma3_broadcast(benchmark::State& state, const char* net) { |
| 373 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_5x8__fma3_broadcast, 5, 8, 1, 1); |
| 374 | } |
| 375 | |
| 376 | static void f32_igemm_6x8__fma3_broadcast(benchmark::State& state, const char* net) { |
| 377 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__fma3_broadcast, 6, 8, 1, 1); |
| 378 | } |
| 379 | |
| 380 | static void f32_igemm_7x8__fma3_broadcast(benchmark::State& state, const char* net) { |
| 381 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_7x8__fma3_broadcast, 7, 8, 1, 1); |
| 382 | } |
| 383 | |
| 384 | static void f32_igemm_8x8__fma3_broadcast(benchmark::State& state, const char* net) { |
| 385 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_8x8__fma3_broadcast, 8, 8, 1, 1); |
| 386 | } |
| 387 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 388 | BENCHMARK_CONV(f32_igemm_1x8__sse_load1) |
| 389 | BENCHMARK_CONV(f32_igemm_4x8__sse_load1) |
| 390 | BENCHMARK_CONV(f32_igemm_1x8__sse_dup) |
| 391 | BENCHMARK_CONV(f32_igemm_4x8__sse_dup) |
| 392 | BENCHMARK_CONV(f32_igemm_1x8s4__sse) |
| 393 | BENCHMARK_CONV(f32_igemm_4x8s4__sse) |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame^] | 394 | BENCHMARK_CONV(f32_igemm_1x8__avx_broadcast) |
| 395 | BENCHMARK_CONV(f32_igemm_4x8__avx_broadcast) |
| 396 | BENCHMARK_CONV(f32_igemm_5x8__avx_broadcast) |
| 397 | BENCHMARK_CONV(f32_igemm_6x8__avx_broadcast) |
| 398 | BENCHMARK_CONV(f32_igemm_7x8__avx_broadcast) |
| 399 | BENCHMARK_CONV(f32_igemm_1x8__fma3_broadcast) |
| 400 | BENCHMARK_CONV(f32_igemm_4x8__fma3_broadcast) |
| 401 | BENCHMARK_CONV(f32_igemm_5x8__fma3_broadcast) |
| 402 | BENCHMARK_CONV(f32_igemm_6x8__fma3_broadcast) |
| 403 | BENCHMARK_CONV(f32_igemm_7x8__fma3_broadcast) |
| 404 | BENCHMARK_CONV(f32_igemm_8x8__fma3_broadcast) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 405 | #endif /* XNN_ARCH_X86 || XNN_ARCH_X86_64 */ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 406 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 407 | #if !XNN_ARCH_WASM && !XNN_ARCH_ASMJS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 408 | static void f32_igemm_1x8__psimd_loadsplat(benchmark::State& state, const char* net) { |
| 409 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__psimd_loadsplat, 1, 8, 1, 1); |
| 410 | } |
| 411 | |
| 412 | static void f32_igemm_4x8__psimd_loadsplat(benchmark::State& state, const char* net) { |
| 413 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__psimd_loadsplat, 4, 8, 1, 1); |
| 414 | } |
| 415 | |
| 416 | static void f32_igemm_6x8__psimd_loadsplat(benchmark::State& state, const char* net) { |
| 417 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__psimd_loadsplat, 6, 8, 1, 1); |
| 418 | } |
| 419 | |
| 420 | static void f32_igemm_1x8__psimd_splat(benchmark::State& state, const char* net) { |
| 421 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__psimd_splat, 1, 8, 1, 1); |
| 422 | } |
| 423 | |
| 424 | static void f32_igemm_4x8__psimd_splat(benchmark::State& state, const char* net) { |
| 425 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__psimd_splat, 4, 8, 1, 1); |
| 426 | } |
| 427 | |
| 428 | static void f32_igemm_6x8__psimd_splat(benchmark::State& state, const char* net) { |
| 429 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__psimd_splat, 6, 8, 1, 1); |
| 430 | } |
| 431 | |
| 432 | static void f32_igemm_1x8s4__psimd(benchmark::State& state, const char* net) { |
| 433 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8s4__psimd, 1, 8, 1, 4); |
| 434 | } |
| 435 | |
| 436 | static void f32_igemm_4x8s4__psimd(benchmark::State& state, const char* net) { |
| 437 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8s4__psimd, 4, 8, 1, 4); |
| 438 | } |
| 439 | |
| 440 | static void f32_igemm_6x8s4__psimd(benchmark::State& state, const char* net) { |
| 441 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8s4__psimd, 6, 8, 1, 4); |
| 442 | } |
| 443 | |
| 444 | BENCHMARK_CONV(f32_igemm_1x8__psimd_loadsplat) |
| 445 | BENCHMARK_CONV(f32_igemm_4x8__psimd_loadsplat) |
| 446 | BENCHMARK_CONV(f32_igemm_6x8__psimd_loadsplat) |
| 447 | |
| 448 | BENCHMARK_CONV(f32_igemm_1x8__psimd_splat) |
| 449 | BENCHMARK_CONV(f32_igemm_4x8__psimd_splat) |
| 450 | BENCHMARK_CONV(f32_igemm_6x8__psimd_splat) |
| 451 | |
| 452 | BENCHMARK_CONV(f32_igemm_1x8s4__psimd) |
| 453 | BENCHMARK_CONV(f32_igemm_4x8s4__psimd) |
| 454 | BENCHMARK_CONV(f32_igemm_6x8s4__psimd) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 455 | #endif /* !XNN_ARCH_WASM && !XNN_ARCH_ASMJS */ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 456 | |
| 457 | static void f32_igemm_1x4__scalar(benchmark::State& state, const char* net) { |
| 458 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x4__scalar, 1, 4, 1, 1); |
| 459 | } |
| 460 | |
| 461 | static void f32_igemm_2x4__scalar(benchmark::State& state, const char* net) { |
| 462 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_2x4__scalar, 2, 4, 1, 1); |
| 463 | } |
| 464 | |
| 465 | static void f32_igemm_4x4__scalar(benchmark::State& state, const char* net) { |
| 466 | IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x4__scalar, 4, 4, 1, 1); |
| 467 | } |
| 468 | |
| 469 | BENCHMARK_CONV(f32_igemm_1x4__scalar) |
| 470 | BENCHMARK_CONV(f32_igemm_2x4__scalar) |
| 471 | BENCHMARK_CONV(f32_igemm_4x4__scalar) |
| 472 | |
| 473 | |
| 474 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 475 | BENCHMARK_MAIN(); |
| 476 | #endif |