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> |
Zhi An Ng | 1425eb5 | 2022-02-03 14:24:44 -0800 | [diff] [blame] | 10 | #include <limits> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 11 | #include <random> |
| 12 | #include <vector> |
| 13 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <benchmark/benchmark.h> |
| 15 | #include "bench/conv.h" |
| 16 | #include "bench/utils.h" |
| 17 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 18 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | #include <xnnpack/igemm.h> |
| 20 | #include <xnnpack/indirection.h> |
| 21 | #include <xnnpack/operator.h> |
| 22 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 23 | #include <xnnpack/params-init.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | |
| 26 | |
| 27 | static void IGEMMBenchmark(benchmark::State& state, |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 28 | xnn_f32_igemm_minmax_ukernel_function f32_igemm, |
Marat Dukhan | c8466f5 | 2019-11-25 18:01:10 -0800 | [diff] [blame] | 29 | uint32_t mr, uint32_t nr, uint32_t kr, uint32_t sr, |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 30 | xnn_init_f32_minmax_params_fn init_params, |
Marat Dukhan | c8466f5 | 2019-11-25 18:01:10 -0800 | [diff] [blame] | 31 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 32 | { |
Marat Dukhan | c8466f5 | 2019-11-25 18:01:10 -0800 | [diff] [blame] | 33 | if (isa_check && !isa_check(state)) { |
| 34 | return; |
| 35 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 36 | |
| 37 | const size_t input_height = state.range(0); |
| 38 | const size_t input_width = state.range(1); |
| 39 | const size_t kernel_height = state.range(2); |
| 40 | const size_t kernel_width = state.range(3); |
| 41 | const size_t kernel_size = kernel_height * kernel_width; |
| 42 | const size_t padding_height = state.range(4); |
| 43 | const size_t padding_width = state.range(5); |
| 44 | const size_t subsampling = state.range(6); |
| 45 | const size_t dilation = state.range(7); |
| 46 | const size_t group_input_channels = state.range(8); |
| 47 | const size_t group_output_channels = state.range(9); |
| 48 | |
| 49 | std::random_device random_device; |
| 50 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 51 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 52 | |
| 53 | const size_t output_pixel_stride = group_output_channels; |
| 54 | const size_t input_pixel_stride = group_input_channels; |
| 55 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 56 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 57 | const size_t padding_left = padding_width / 2; |
| 58 | const size_t padding_top = padding_height / 2; |
| 59 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 60 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 61 | const size_t output_size = output_height * output_width; |
| 62 | |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 63 | const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr); |
| 64 | const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr); |
Marat Dukhan | fbd67a7 | 2022-01-31 18:03:50 -0800 | [diff] [blame] | 65 | const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr * sr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 66 | |
| 67 | std::vector<float> a(input_height * input_width * input_pixel_stride); |
| 68 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 69 | std::vector<float> k(group_output_channels * kernel_height * kernel_width * group_input_channels); |
| 70 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 71 | std::vector<float> b(group_output_channels); |
| 72 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 73 | |
| 74 | std::vector<float> z(group_input_channels); |
| 75 | |
Marat Dukhan | fbd67a7 | 2022-01-31 18:03:50 -0800 | [diff] [blame] | 76 | const size_t w_elements = kernel_size * kc_stride * nc_stride + nc_stride; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 77 | const size_t i_elements = mc_stride * kernel_size; |
| 78 | const size_t c_elements = output_height * output_width * output_pixel_stride; |
| 79 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 80 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 81 | sizeof(float) * (w_elements + c_elements) + sizeof(void*) * i_elements); |
| 82 | |
Marat Dukhan | e13e639 | 2021-07-26 22:22:35 -0700 | [diff] [blame] | 83 | std::vector<float, AlignedAllocator<float, 64>> w(w_elements * num_buffers); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 84 | std::fill(w.begin(), w.end(), 0.0f); |
| 85 | xnn_pack_f32_conv_goki_w( |
| 86 | 1 /* groups */, group_output_channels, kernel_size, group_input_channels, |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 87 | nr, kr, sr, k.data(), b.data(), w.data(), 0 /* extra bytes */, nullptr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 88 | for (size_t n = 1; n < num_buffers; n++) { |
| 89 | std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements); |
| 90 | } |
| 91 | |
| 92 | std::vector<const float*> i(i_elements * num_buffers); |
| 93 | xnn_operator convolution_op = { }; |
| 94 | convolution_op.indirection_buffer = reinterpret_cast<const void**>(i.data()); |
| 95 | convolution_op.input = a.data(); |
| 96 | convolution_op.input_pixel_stride = input_pixel_stride; |
| 97 | convolution_op.zero_buffer = z.data(); |
| 98 | convolution_op.groups = 1; |
| 99 | convolution_op.group_input_channels = group_input_channels; |
| 100 | convolution_op.batch_size = 1; |
| 101 | convolution_op.input_height = input_height; |
| 102 | convolution_op.input_width = input_width; |
| 103 | convolution_op.output_height = output_height; |
| 104 | convolution_op.output_width = output_width; |
| 105 | convolution_op.kernel_height = kernel_height; |
| 106 | convolution_op.kernel_width = kernel_width; |
| 107 | convolution_op.stride_height = subsampling; |
| 108 | convolution_op.stride_width = subsampling; |
| 109 | convolution_op.dilation_height = dilation; |
| 110 | convolution_op.dilation_width = dilation; |
| 111 | convolution_op.padding_top = padding_top; |
| 112 | convolution_op.padding_left = padding_left; |
| 113 | xnn_indirection_init_conv2d(&convolution_op, mr, 2 /* log2(sizeof(float)) */); |
| 114 | for (size_t n = 1; n < num_buffers; n++) { |
| 115 | std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements); |
| 116 | } |
| 117 | |
| 118 | std::vector<float> c(c_elements * num_buffers); |
| 119 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 120 | |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame] | 121 | xnn_f32_minmax_params params; |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 122 | init_params(¶ms, |
| 123 | -std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 124 | |
| 125 | size_t buffer_index = 0; |
| 126 | for (auto _ : state) { |
| 127 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 128 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 129 | buffer_index = (buffer_index + 1) % num_buffers; |
| 130 | state.ResumeTiming(); |
| 131 | |
| 132 | for (uint32_t m = 0; m < output_size; m += mr) { |
| 133 | const uint32_t mb = min(output_size - m, mr); |
Zhi An Ng | 2188833 | 2022-02-03 13:35:48 -0800 | [diff] [blame] | 134 | f32_igemm( |
| 135 | mb, group_output_channels, 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, |
| 138 | c.data() + buffer_index * c_elements + m * group_output_channels, group_output_channels * sizeof(float), nr * sizeof(float), |
| 139 | 0, z.data(), ¶ms); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 140 | } |
| 141 | } |
| 142 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 143 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 144 | if (cpu_frequency != 0) { |
| 145 | state.counters["cpufreq"] = cpu_frequency; |
| 146 | } |
| 147 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 148 | state.counters["FLOPS"] = benchmark::Counter( |
| 149 | uint64_t(state.iterations()) * 2 * |
| 150 | output_height * output_width * |
| 151 | group_input_channels * group_output_channels * |
| 152 | kernel_height * kernel_width, |
| 153 | benchmark::Counter::kIsRate); |
| 154 | } |
| 155 | |
Zhi An Ng | 1425eb5 | 2022-02-03 14:24:44 -0800 | [diff] [blame] | 156 | #if XNN_PLATFORM_JIT |
Zhi An Ng | 665cb23 | 2022-01-07 15:56:41 -0800 | [diff] [blame] | 157 | static void IGEMMBenchmark(benchmark::State& state, |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 158 | xnn_jit_igemm_code_generator_function generator, |
Zhi An Ng | 665cb23 | 2022-01-07 15:56:41 -0800 | [diff] [blame] | 159 | size_t mr, size_t nr, size_t kr, size_t sr, |
| 160 | xnn_init_f32_minmax_params_fn init_params, |
| 161 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
Zhi An Ng | 1425eb5 | 2022-02-03 14:24:44 -0800 | [diff] [blame] | 162 | { |
| 163 | if (isa_check && !isa_check(state)) { |
| 164 | return; |
Zhi An Ng | 665cb23 | 2022-01-07 15:56:41 -0800 | [diff] [blame] | 165 | } |
| 166 | |
Zhi An Ng | 1425eb5 | 2022-02-03 14:24:44 -0800 | [diff] [blame] | 167 | const size_t input_height = state.range(0); |
| 168 | const size_t input_width = state.range(1); |
| 169 | const size_t kernel_height = state.range(2); |
| 170 | const size_t kernel_width = state.range(3); |
| 171 | const size_t kernel_size = kernel_height * kernel_width; |
| 172 | const size_t padding_height = state.range(4); |
| 173 | const size_t padding_width = state.range(5); |
| 174 | const size_t subsampling = state.range(6); |
| 175 | const size_t dilation = state.range(7); |
| 176 | const size_t group_input_channels = state.range(8); |
| 177 | const size_t group_output_channels = state.range(9); |
| 178 | |
| 179 | std::random_device random_device; |
| 180 | auto rng = std::mt19937(random_device()); |
| 181 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), std::ref(rng)); |
| 182 | |
| 183 | const size_t output_pixel_stride = group_output_channels; |
| 184 | const size_t input_pixel_stride = group_input_channels; |
| 185 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 186 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 187 | const size_t padding_left = padding_width / 2; |
| 188 | const size_t padding_top = padding_height / 2; |
| 189 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 190 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 191 | const size_t output_size = output_height * output_width; |
| 192 | |
| 193 | const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr); |
| 194 | const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr); |
| 195 | const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr * sr); |
| 196 | |
| 197 | std::vector<float> a(input_height * input_width * input_pixel_stride); |
| 198 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 199 | std::vector<float> k(group_output_channels * kernel_height * kernel_width * group_input_channels); |
| 200 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 201 | std::vector<float> b(group_output_channels); |
| 202 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 203 | |
| 204 | std::vector<float> z(group_input_channels); |
| 205 | |
| 206 | const size_t w_elements = kernel_size * kc_stride * nc_stride + nc_stride; |
| 207 | const size_t i_elements = mc_stride * kernel_size; |
| 208 | const size_t c_elements = output_height * output_width * output_pixel_stride; |
| 209 | const size_t num_buffers = 1 + |
| 210 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
| 211 | sizeof(float) * (w_elements + c_elements) + sizeof(void*) * i_elements); |
| 212 | |
| 213 | std::vector<float, AlignedAllocator<float, 64>> w(w_elements * num_buffers); |
| 214 | std::fill(w.begin(), w.end(), 0.0f); |
| 215 | xnn_pack_f32_conv_goki_w( |
| 216 | 1 /* groups */, group_output_channels, kernel_size, group_input_channels, |
| 217 | nr, kr, sr, k.data(), b.data(), w.data(), 0 /* extra bytes */, nullptr); |
| 218 | for (size_t n = 1; n < num_buffers; n++) { |
| 219 | std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements); |
| 220 | } |
| 221 | |
| 222 | std::vector<const float*> i(i_elements * num_buffers); |
| 223 | xnn_operator convolution_op = { }; |
| 224 | convolution_op.indirection_buffer = reinterpret_cast<const void**>(i.data()); |
| 225 | convolution_op.input = a.data(); |
| 226 | convolution_op.input_pixel_stride = input_pixel_stride; |
| 227 | convolution_op.zero_buffer = z.data(); |
| 228 | convolution_op.groups = 1; |
| 229 | convolution_op.group_input_channels = group_input_channels; |
| 230 | convolution_op.batch_size = 1; |
| 231 | convolution_op.input_height = input_height; |
| 232 | convolution_op.input_width = input_width; |
| 233 | convolution_op.output_height = output_height; |
| 234 | convolution_op.output_width = output_width; |
| 235 | convolution_op.kernel_height = kernel_height; |
| 236 | convolution_op.kernel_width = kernel_width; |
| 237 | convolution_op.stride_height = subsampling; |
| 238 | convolution_op.stride_width = subsampling; |
| 239 | convolution_op.dilation_height = dilation; |
| 240 | convolution_op.dilation_width = dilation; |
| 241 | convolution_op.padding_top = padding_top; |
| 242 | convolution_op.padding_left = padding_left; |
| 243 | xnn_indirection_init_conv2d(&convolution_op, mr, 2 /* log2(sizeof(float)) */); |
| 244 | for (size_t n = 1; n < num_buffers; n++) { |
| 245 | std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements); |
| 246 | } |
| 247 | |
| 248 | std::vector<float> c(c_elements * num_buffers); |
| 249 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 250 | |
| 251 | xnn_f32_minmax_params params; |
| 252 | init_params(¶ms, |
| 253 | -std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
| 254 | |
| 255 | jit_gemm_params jit_params = { |
| 256 | .f32_minmax = { |
| 257 | .min = -std::numeric_limits<float>::infinity(), |
| 258 | .max = +std::numeric_limits<float>::infinity() |
| 259 | } |
| 260 | }; |
| 261 | |
| 262 | xnn_code_buffer code_buffer; |
| 263 | xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE); |
| 264 | generator(&code_buffer, group_output_channels, group_input_channels * sizeof(float), kernel_size * mr * sizeof(void *), &jit_params); |
| 265 | auto f32_igemm = reinterpret_cast<xnn_f32_igemm_minmax_ukernel_function>(code_buffer.code); |
| 266 | |
| 267 | size_t buffer_index = 0; |
| 268 | for (auto _ : state) { |
| 269 | state.PauseTiming(); |
| 270 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); |
| 271 | buffer_index = (buffer_index + 1) % num_buffers; |
| 272 | state.ResumeTiming(); |
| 273 | |
| 274 | for (uint32_t m = 0; m < output_size; m += mr) { |
| 275 | const uint32_t mb = min(output_size - m, mr); |
| 276 | f32_igemm( |
| 277 | mb, group_output_channels, group_input_channels * sizeof(float), kernel_size * mr * sizeof(void*), |
| 278 | i.data() + buffer_index * i_elements + m, |
| 279 | w.data() + buffer_index * w_elements, |
| 280 | c.data() + buffer_index * c_elements + m * group_output_channels, group_output_channels * sizeof(float), nr * sizeof(float), |
| 281 | 0, z.data(), ¶ms); |
| 282 | } |
| 283 | } |
| 284 | xnn_release_code_memory(&code_buffer); |
| 285 | |
| 286 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 287 | if (cpu_frequency != 0) { |
| 288 | state.counters["cpufreq"] = cpu_frequency; |
| 289 | } |
| 290 | |
| 291 | state.counters["FLOPS"] = benchmark::Counter( |
| 292 | uint64_t(state.iterations()) * 2 * |
| 293 | output_height * output_width * |
| 294 | group_input_channels * group_output_channels * |
| 295 | kernel_height * kernel_width, |
| 296 | benchmark::Counter::kIsRate); |
| 297 | |
| 298 | } |
| 299 | #endif // XNN_PLATFORM_JIT |
| 300 | |
| 301 | #if XNN_ARCH_ARM64 && XNN_PLATFORM_JIT |
Zhi An Ng | f30a859 | 2022-02-03 16:49:19 -0800 | [diff] [blame] | 302 | static void jit_f32_igemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 303 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, 1, 8, 1, 1, |
| 304 | xnn_init_f32_minmax_scalar_params); |
| 305 | } |
| 306 | static void jit_f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 307 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75, 1, 8, 1, 1, |
| 308 | xnn_init_f32_minmax_scalar_params); |
| 309 | } |
Zhi An Ng | 1425eb5 | 2022-02-03 14:24:44 -0800 | [diff] [blame] | 310 | static void jit_f32_igemm_6x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
| 311 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a75, 6, 8, 1, 1, |
| 312 | xnn_init_f32_minmax_scalar_params); |
| 313 | } |
| 314 | static void jit_f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 315 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75, 6, 8, 1, 1, |
| 316 | xnn_init_f32_minmax_scalar_params); |
| 317 | } |
| 318 | |
Zhi An Ng | f30a859 | 2022-02-03 16:49:19 -0800 | [diff] [blame] | 319 | BENCHMARK_CONV(jit_f32_igemm_1x8__aarch64_neonfma_cortex_a75) |
| 320 | BENCHMARK_CONV(jit_f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75) |
Zhi An Ng | 1425eb5 | 2022-02-03 14:24:44 -0800 | [diff] [blame] | 321 | BENCHMARK_CONV(jit_f32_igemm_6x8__aarch64_neonfma_cortex_a75) |
| 322 | BENCHMARK_CONV(jit_f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a75) |
| 323 | #endif // XNN_ARCH_ARM64 && XNN_PLATFORM_JIT |
| 324 | |
| 325 | #if XNN_ARCH_ARM && XNN_PLATFORM_JIT |
Zhi An Ng | 665cb23 | 2022-01-07 15:56:41 -0800 | [diff] [blame] | 326 | static void jit_f32_igemm_4x8__aarch32_neon_ld64(benchmark::State& state, const char* net) { |
| 327 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_4x8__aarch32_neon_ld64, 4, 8, 1, 1, |
| 328 | xnn_init_f32_minmax_scalar_params); |
| 329 | } |
| 330 | static void jit_f32_igemm_4x8__aarch32_neon_cortex_a7(benchmark::State& state, const char* net) { |
| 331 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_4x8__aarch32_neon_cortex_a7, 4, 8, 1, 1, |
| 332 | xnn_init_f32_minmax_scalar_params); |
| 333 | } |
| 334 | static void jit_f32_igemm_4x8__aarch32_neon_cortex_a53(benchmark::State& state, const char* net) { |
| 335 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_4x8__aarch32_neon_cortex_a53, 4, 8, 1, 1, |
| 336 | xnn_init_f32_minmax_scalar_params); |
| 337 | } |
| 338 | static void jit_f32_igemm_4x8__aarch32_neon_cortex_a55(benchmark::State& state, const char* net) { |
| 339 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_4x8__aarch32_neon_cortex_a55, 4, 8, 1, 1, |
| 340 | xnn_init_f32_minmax_scalar_params); |
| 341 | } |
| 342 | static void jit_f32_igemm_4x8__aarch32_neon_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 343 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_4x8__aarch32_neon_prfm_cortex_a75, 4, 8, 1, 1, |
| 344 | xnn_init_f32_minmax_scalar_params); |
| 345 | } |
| 346 | static void jit_f32_igemm_4x8__aarch32_neon_cortex_a75(benchmark::State& state, const char* net) { |
| 347 | IGEMMBenchmark(state, xnn_generate_f32_igemm_ukernel_4x8__aarch32_neon_cortex_a75, 4, 8, 1, 1, |
| 348 | xnn_init_f32_minmax_scalar_params); |
| 349 | } |
| 350 | |
| 351 | BENCHMARK_CONV(jit_f32_igemm_4x8__aarch32_neon_ld64) |
| 352 | BENCHMARK_CONV(jit_f32_igemm_4x8__aarch32_neon_cortex_a7) |
| 353 | BENCHMARK_CONV(jit_f32_igemm_4x8__aarch32_neon_cortex_a53) |
| 354 | BENCHMARK_CONV(jit_f32_igemm_4x8__aarch32_neon_cortex_a55) |
| 355 | BENCHMARK_CONV(jit_f32_igemm_4x8__aarch32_neon_prfm_cortex_a75) |
| 356 | BENCHMARK_CONV(jit_f32_igemm_4x8__aarch32_neon_cortex_a75) |
Zhi An Ng | 1425eb5 | 2022-02-03 14:24:44 -0800 | [diff] [blame] | 357 | #endif // XNN_ARCH_ARM && XNN_PLATFORM_JIT |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 358 | |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 359 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 360 | static void f32_igemm_4x8__aarch32_neon_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 361 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_ld64, 4, 8, 1, 1, |
| 362 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 363 | } |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 364 | static void f32_igemm_4x8__aarch32_neon_cortex_a7(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 365 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7, 4, 8, 1, 1, |
| 366 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 367 | } |
Frank Barchard | b7dd29e | 2020-03-11 12:37:10 -0700 | [diff] [blame] | 368 | static void f32_igemm_4x8__aarch32_neon_cortex_a53(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 369 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53, 4, 8, 1, 1, |
| 370 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | b7dd29e | 2020-03-11 12:37:10 -0700 | [diff] [blame] | 371 | } |
| 372 | static void f32_igemm_4x8__aarch32_neon_cortex_a55(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 373 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55, 4, 8, 1, 1, |
| 374 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | b7dd29e | 2020-03-11 12:37:10 -0700 | [diff] [blame] | 375 | } |
Frank Barchard | 7873586 | 2022-01-04 16:47:44 -0800 | [diff] [blame] | 376 | static void f32_igemm_4x8__aarch32_neon_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 377 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a75, 4, 8, 1, 1, |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 378 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | b7dd29e | 2020-03-11 12:37:10 -0700 | [diff] [blame] | 379 | } |
| 380 | static void f32_igemm_4x8__aarch32_neon_cortex_a75(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 381 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75, 4, 8, 1, 1, |
| 382 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | b7dd29e | 2020-03-11 12:37:10 -0700 | [diff] [blame] | 383 | } |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 384 | |
| 385 | BENCHMARK_CONV(f32_igemm_4x8__aarch32_neon_ld64) |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 386 | BENCHMARK_CONV(f32_igemm_4x8__aarch32_neon_cortex_a7) |
Frank Barchard | b7dd29e | 2020-03-11 12:37:10 -0700 | [diff] [blame] | 387 | BENCHMARK_CONV(f32_igemm_4x8__aarch32_neon_cortex_a53) |
| 388 | BENCHMARK_CONV(f32_igemm_4x8__aarch32_neon_cortex_a55) |
Frank Barchard | 7873586 | 2022-01-04 16:47:44 -0800 | [diff] [blame] | 389 | BENCHMARK_CONV(f32_igemm_4x8__aarch32_neon_prfm_cortex_a75) |
Frank Barchard | b7dd29e | 2020-03-11 12:37:10 -0700 | [diff] [blame] | 390 | BENCHMARK_CONV(f32_igemm_4x8__aarch32_neon_cortex_a75) |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 391 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 392 | |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 393 | |
Frank Barchard | dbafc58 | 2019-10-09 16:30:48 -0700 | [diff] [blame] | 394 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 395 | static void f32_igemm_1x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 396 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53, 1, 12, 1, 1, |
| 397 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 398 | } |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 399 | static void f32_igemm_1x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 400 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53, 1, 8, 1, 1, |
| 401 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 402 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 403 | static void f32_igemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 404 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75, 1, 8, 1, 1, |
| 405 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 406 | } |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 407 | static void f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 408 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75, 1, 8, 1, 1, |
| 409 | xnn_init_f32_minmax_scalar_params); |
| 410 | } |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 411 | static void f32_igemm_4x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 412 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53, 4, 8, 1, 1, |
| 413 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 414 | } |
Frank Barchard | 8fb9055 | 2020-03-16 11:36:09 -0700 | [diff] [blame] | 415 | static void f32_igemm_4x8__aarch64_neonfma_cortex_a55(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 416 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55, 4, 8, 1, 1, |
| 417 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 8fb9055 | 2020-03-16 11:36:09 -0700 | [diff] [blame] | 418 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 419 | static void f32_igemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 420 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75, 4, 8, 1, 1, |
| 421 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 422 | } |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 423 | static void f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 424 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75, 4, 8, 1, 1, |
| 425 | xnn_init_f32_minmax_scalar_params); |
| 426 | } |
Frank Barchard | e349124 | 2021-06-11 14:04:57 -0700 | [diff] [blame] | 427 | static void f32_igemm_4x8__aarch64_neonfma_ld64(benchmark::State& state, const char* net) { |
| 428 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_ld64, 4, 8, 1, 1, |
| 429 | xnn_init_f32_minmax_scalar_params); |
| 430 | } |
Frank Barchard | 79cd5f9 | 2021-06-21 17:34:59 -0700 | [diff] [blame] | 431 | static void f32_igemm_4x8__aarch64_neonfma_ld128(benchmark::State& state, const char* net) { |
| 432 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_ld128, 4, 8, 1, 1, |
| 433 | xnn_init_f32_minmax_scalar_params); |
| 434 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 435 | static void f32_igemm_5x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 436 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75, 5, 8, 1, 1, |
| 437 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 438 | } |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 439 | static void f32_igemm_5x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 440 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__aarch64_neonfma_prfm_cortex_a75, 5, 8, 1, 1, |
| 441 | xnn_init_f32_minmax_scalar_params); |
| 442 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 443 | static void f32_igemm_4x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 444 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53, 4, 12, 1, 1, |
| 445 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 446 | } |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 447 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 448 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53, 6, 8, 1, 1, |
| 449 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 450 | } |
Frank Barchard | 91e1999 | 2020-03-09 18:46:14 -0700 | [diff] [blame] | 451 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a55(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 452 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55, 6, 8, 1, 1, |
| 453 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 91e1999 | 2020-03-09 18:46:14 -0700 | [diff] [blame] | 454 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 455 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a73(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 456 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73, 6, 8, 1, 1, |
| 457 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 458 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 459 | static void f32_igemm_6x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 460 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75, 6, 8, 1, 1, |
| 461 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 462 | } |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 463 | static void f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 464 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75, 6, 8, 1, 1, |
| 465 | xnn_init_f32_minmax_scalar_params); |
| 466 | } |
Frank Barchard | e349124 | 2021-06-11 14:04:57 -0700 | [diff] [blame] | 467 | static void f32_igemm_6x8__aarch64_neonfma_ld64(benchmark::State& state, const char* net) { |
| 468 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_ld64, 6, 8, 1, 1, |
| 469 | xnn_init_f32_minmax_scalar_params); |
| 470 | } |
Frank Barchard | 79cd5f9 | 2021-06-21 17:34:59 -0700 | [diff] [blame] | 471 | static void f32_igemm_6x8__aarch64_neonfma_ld128(benchmark::State& state, const char* net) { |
| 472 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_ld128, 6, 8, 1, 1, |
| 473 | xnn_init_f32_minmax_scalar_params); |
| 474 | } |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 475 | static void f32_igemm_1x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 476 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neonfma_lane_ld64, 1, 8, 1, 1, |
| 477 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 478 | } |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 479 | static void f32_igemm_4x2__neonfma_lane_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 480 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64, 4, 2, 1, 1, |
| 481 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 482 | } |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 483 | static void f32_igemm_4x4__neonfma_lane_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 484 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64, 4, 4, 1, 1, |
| 485 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 486 | } |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 487 | static void f32_igemm_4x8__neonfma_lane_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 488 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld128, 4, 8, 1, 1, |
| 489 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 490 | } |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 491 | static void f32_igemm_4x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 492 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld64, 4, 8, 1, 1, |
| 493 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 494 | } |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 495 | static void f32_igemm_6x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 496 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64, 6, 8, 1, 1, |
| 497 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 498 | } |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 499 | static void f32_igemm_6x8__neonfma_lane_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 500 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld128, 6, 8, 1, 1, |
| 501 | xnn_init_f32_minmax_scalar_params); |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 502 | } |
| 503 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 504 | BENCHMARK_CONV(f32_igemm_1x12__aarch64_neonfma_cortex_a53) |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 505 | BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 506 | BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 507 | BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 508 | BENCHMARK_CONV(f32_igemm_4x12__aarch64_neonfma_cortex_a53) |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 509 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_cortex_a53) |
Frank Barchard | 8fb9055 | 2020-03-16 11:36:09 -0700 | [diff] [blame] | 510 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_cortex_a55) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 511 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 512 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a75) |
Frank Barchard | e349124 | 2021-06-11 14:04:57 -0700 | [diff] [blame] | 513 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_ld64) |
Frank Barchard | 79cd5f9 | 2021-06-21 17:34:59 -0700 | [diff] [blame] | 514 | BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_ld128) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 515 | BENCHMARK_CONV(f32_igemm_5x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 516 | BENCHMARK_CONV(f32_igemm_5x8__aarch64_neonfma_prfm_cortex_a75) |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 517 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a53) |
Frank Barchard | 91e1999 | 2020-03-09 18:46:14 -0700 | [diff] [blame] | 518 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a55) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 519 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a73) |
| 520 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 521 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a75) |
Frank Barchard | e349124 | 2021-06-11 14:04:57 -0700 | [diff] [blame] | 522 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_ld64) |
Frank Barchard | 79cd5f9 | 2021-06-21 17:34:59 -0700 | [diff] [blame] | 523 | BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_ld128) |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 524 | BENCHMARK_CONV(f32_igemm_1x8__neonfma_lane_ld64) |
| 525 | BENCHMARK_CONV(f32_igemm_4x2__neonfma_lane_ld64) |
| 526 | BENCHMARK_CONV(f32_igemm_4x4__neonfma_lane_ld64) |
| 527 | BENCHMARK_CONV(f32_igemm_4x8__neonfma_lane_ld128) |
| 528 | BENCHMARK_CONV(f32_igemm_4x8__neonfma_lane_ld64) |
| 529 | BENCHMARK_CONV(f32_igemm_6x8__neonfma_lane_ld64) |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 530 | BENCHMARK_CONV(f32_igemm_6x8__neonfma_lane_ld128) |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 531 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 532 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 533 | |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 534 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 535 | static void f32_igemm_1x8__neon_lane_ld64(benchmark::State& state, const char* net) { |
| 536 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64, 1, 8, 1, 1, |
| 537 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 538 | } |
| 539 | static void f32_igemm_4x2__neon_lane_ld64(benchmark::State& state, const char* net) { |
| 540 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x2__neon_lane_ld64, 4, 2, 1, 1, |
| 541 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 542 | } |
| 543 | static void f32_igemm_4x4__neon_lane_ld64(benchmark::State& state, const char* net) { |
| 544 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64, 4, 4, 1, 1, |
| 545 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 546 | } |
| 547 | static void f32_igemm_4x8__neon_lane_ld64(benchmark::State& state, const char* net) { |
| 548 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld64, 4, 8, 1, 1, |
| 549 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 550 | } |
| 551 | static void f32_igemm_4x8__neon_lane_ld128(benchmark::State& state, const char* net) { |
| 552 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld128, 4, 8, 1, 1, |
| 553 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 554 | } |
| 555 | static void f32_igemm_6x8__neon_lane_ld64(benchmark::State& state, const char* net) { |
| 556 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64, 6, 8, 1, 1, |
| 557 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 558 | } |
| 559 | static void f32_igemm_6x8__neon_lane_ld128(benchmark::State& state, const char* net) { |
| 560 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld128, 6, 8, 1, 1, |
| 561 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 562 | } |
| 563 | static void f32_igemm_1x8__neon_dup_ld64(benchmark::State& state, const char* net) { |
| 564 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neon_dup_ld64, 1, 8, 1, 1, |
| 565 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 566 | } |
| 567 | static void f32_igemm_4x8__neon_dup_ld128(benchmark::State& state, const char* net) { |
| 568 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld128, 4, 8, 1, 1, |
| 569 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 570 | } |
| 571 | static void f32_igemm_4x8__neon_dup_ld64(benchmark::State& state, const char* net) { |
| 572 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64, 4, 8, 1, 1, |
| 573 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 574 | } |
| 575 | static void f32_igemm_6x8__neon_dup_ld64(benchmark::State& state, const char* net) { |
| 576 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64, 6, 8, 1, 1, |
| 577 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 578 | } |
| 579 | static void f32_igemm_6x8__neon_dup_ld128(benchmark::State& state, const char* net) { |
| 580 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128, 6, 8, 1, 1, |
| 581 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 582 | } |
| 583 | static void f32_igemm_1x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { |
| 584 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neonfma_dup_ld64, 1, 8, 1, 1, |
| 585 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 586 | } |
| 587 | static void f32_igemm_4x8__neonfma_dup_ld128(benchmark::State& state, const char* net) { |
| 588 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld128, 4, 8, 1, 1, |
| 589 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 590 | } |
| 591 | static void f32_igemm_4x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { |
| 592 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64, 4, 8, 1, 1, |
| 593 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 594 | } |
| 595 | static void f32_igemm_6x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { |
| 596 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64, 6, 8, 1, 1, |
| 597 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 598 | } |
| 599 | static void f32_igemm_6x8__neonfma_dup_ld128(benchmark::State& state, const char* net) { |
| 600 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128, 6, 8, 1, 1, |
| 601 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 602 | } |
| 603 | static void f32_igemm_1x8s4__neon(benchmark::State& state, const char* net) { |
| 604 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8s4__neon, 1, 8, 1, 4, |
| 605 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 606 | } |
| 607 | static void f32_igemm_4x8s4__neon(benchmark::State& state, const char* net) { |
| 608 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__neon, 4, 8, 1, 4, |
| 609 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 610 | } |
| 611 | static void f32_igemm_6x8s4__neon(benchmark::State& state, const char* net) { |
| 612 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__neon, 6, 8, 1, 4, |
| 613 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 614 | } |
| 615 | static void f32_igemm_8x8s4__neon(benchmark::State& state, const char* net) { |
| 616 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x8s4__neon, 8, 8, 1, 4, |
| 617 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); |
| 618 | } |
| 619 | static void f32_igemm_1x8s4__neonfma(benchmark::State& state, const char* net) { |
| 620 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8s4__neonfma, 1, 8, 1, 4, |
| 621 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 622 | } |
| 623 | static void f32_igemm_4x8s4__neonfma(benchmark::State& state, const char* net) { |
| 624 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__neonfma, 4, 8, 1, 4, |
| 625 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 626 | } |
| 627 | static void f32_igemm_6x8s4__neonfma(benchmark::State& state, const char* net) { |
| 628 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__neonfma, 6, 8, 1, 4, |
| 629 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 630 | } |
| 631 | static void f32_igemm_8x8s4__neonfma(benchmark::State& state, const char* net) { |
| 632 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x8s4__neonfma, 8, 8, 1, 4, |
| 633 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); |
| 634 | } |
| 635 | |
| 636 | BENCHMARK_CONV(f32_igemm_1x8__neon_lane_ld64) |
| 637 | BENCHMARK_CONV(f32_igemm_4x2__neon_lane_ld64) |
| 638 | BENCHMARK_CONV(f32_igemm_4x4__neon_lane_ld64) |
| 639 | BENCHMARK_CONV(f32_igemm_4x8__neon_lane_ld128) |
| 640 | BENCHMARK_CONV(f32_igemm_4x8__neon_lane_ld64) |
| 641 | BENCHMARK_CONV(f32_igemm_6x8__neon_lane_ld64) |
| 642 | BENCHMARK_CONV(f32_igemm_6x8__neon_lane_ld128) |
| 643 | BENCHMARK_CONV(f32_igemm_1x8__neon_dup_ld64) |
| 644 | BENCHMARK_CONV(f32_igemm_4x8__neon_dup_ld128) |
| 645 | BENCHMARK_CONV(f32_igemm_4x8__neon_dup_ld64) |
| 646 | BENCHMARK_CONV(f32_igemm_6x8__neon_dup_ld64) |
| 647 | BENCHMARK_CONV(f32_igemm_6x8__neon_dup_ld128) |
| 648 | BENCHMARK_CONV(f32_igemm_1x8__neonfma_dup_ld64) |
| 649 | BENCHMARK_CONV(f32_igemm_4x8__neonfma_dup_ld128) |
| 650 | BENCHMARK_CONV(f32_igemm_4x8__neonfma_dup_ld64) |
| 651 | BENCHMARK_CONV(f32_igemm_6x8__neonfma_dup_ld64) |
| 652 | BENCHMARK_CONV(f32_igemm_6x8__neonfma_dup_ld128) |
| 653 | BENCHMARK_CONV(f32_igemm_1x8s4__neon) |
| 654 | BENCHMARK_CONV(f32_igemm_4x8s4__neon) |
| 655 | BENCHMARK_CONV(f32_igemm_6x8s4__neon) |
| 656 | BENCHMARK_CONV(f32_igemm_8x8s4__neon) |
| 657 | BENCHMARK_CONV(f32_igemm_1x8s4__neonfma) |
| 658 | BENCHMARK_CONV(f32_igemm_4x8s4__neonfma) |
| 659 | BENCHMARK_CONV(f32_igemm_6x8s4__neonfma) |
| 660 | BENCHMARK_CONV(f32_igemm_8x8s4__neonfma) |
| 661 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 662 | |
| 663 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 664 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 665 | static void f32_igemm_1x8__sse_load1(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 666 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__sse_load1, 1, 8, 1, 1, |
| 667 | xnn_init_f32_minmax_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 668 | } |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 669 | static void f32_igemm_3x8__sse_load1(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 670 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__sse_load1, 3, 8, 1, 1, |
| 671 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 672 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 673 | static void f32_igemm_4x8__sse_load1(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 674 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__sse_load1, 4, 8, 1, 1, |
| 675 | xnn_init_f32_minmax_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 676 | } |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 677 | static void f32_igemm_5x8__sse_load1(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 678 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__sse_load1, 5, 8, 1, 1, |
| 679 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 680 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 681 | |
| 682 | static void f32_igemm_1x8__sse_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 683 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__sse_dup, 1, 8, 1, 1, |
| 684 | xnn_init_f32_minmax_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 685 | } |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 686 | static void f32_igemm_3x8__sse_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 687 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__sse_dup, 3, 8, 1, 1, |
| 688 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 689 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 690 | static void f32_igemm_4x8__sse_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 691 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__sse_dup, 4, 8, 1, 1, |
| 692 | xnn_init_f32_minmax_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 693 | } |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 694 | static void f32_igemm_5x8__sse_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 695 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__sse_dup, 5, 8, 1, 1, |
| 696 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 697 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 698 | |
| 699 | static void f32_igemm_1x8s4__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 700 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8s4__sse, 1, 8, 1, 4, |
| 701 | xnn_init_f32_minmax_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 702 | } |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 703 | static void f32_igemm_3x8s4__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 704 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8s4__sse, 3, 8, 1, 4, |
| 705 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 706 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 707 | static void f32_igemm_4x8s4__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 708 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__sse, 4, 8, 1, 4, |
| 709 | xnn_init_f32_minmax_sse_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 710 | } |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 711 | static void f32_igemm_5x8s4__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 712 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8s4__sse, 5, 8, 1, 4, |
| 713 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 714 | } |
| 715 | |
| 716 | static void f32_igemm_1x8__sse2_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 717 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__sse2_dup, 1, 8, 1, 1, |
| 718 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 719 | } |
| 720 | static void f32_igemm_3x8__sse2_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 721 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup, 3, 8, 1, 1, |
| 722 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 723 | } |
| 724 | static void f32_igemm_4x8__sse2_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 725 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup, 4, 8, 1, 1, |
| 726 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 727 | } |
| 728 | static void f32_igemm_5x8__sse2_dup(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 729 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup, 5, 8, 1, 1, |
| 730 | xnn_init_f32_minmax_sse_params); |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 731 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 732 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 733 | static void f32_igemm_1x8__avx_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 734 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__avx_broadcast, 1, 8, 1, 1, |
| 735 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 736 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 737 | static void f32_igemm_4x8__avx_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 738 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__avx_broadcast, 4, 8, 1, 1, |
| 739 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 740 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 741 | static void f32_igemm_5x8__avx_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 742 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__avx_broadcast, 5, 8, 1, 1, |
| 743 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 744 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 745 | static void f32_igemm_6x8__avx_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 746 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__avx_broadcast, 6, 8, 1, 1, |
| 747 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 748 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 749 | static void f32_igemm_7x8__avx_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 750 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_7x8__avx_broadcast, 7, 8, 1, 1, |
| 751 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 752 | } |
| 753 | |
| 754 | static void f32_igemm_1x8__fma3_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 755 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__fma3_broadcast, 1, 8, 1, 1, |
| 756 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 757 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 758 | static void f32_igemm_4x8__fma3_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 759 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__fma3_broadcast, 4, 8, 1, 1, |
| 760 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 761 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 762 | static void f32_igemm_5x8__fma3_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 763 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__fma3_broadcast, 5, 8, 1, 1, |
| 764 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 765 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 766 | static void f32_igemm_6x8__fma3_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 767 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__fma3_broadcast, 6, 8, 1, 1, |
| 768 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 769 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 770 | static void f32_igemm_7x8__fma3_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 771 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_7x8__fma3_broadcast, 7, 8, 1, 1, |
| 772 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 773 | } |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 774 | static void f32_igemm_8x8__fma3_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 775 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x8__fma3_broadcast, 8, 8, 1, 1, |
| 776 | xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 777 | } |
| 778 | |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 779 | static void f32_igemm_1x16__avx512f_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 780 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast, 1, 16, 1, 1, |
| 781 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 782 | } |
| 783 | static void f32_igemm_4x16__avx512f_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 784 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast, 4, 16, 1, 1, |
| 785 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 786 | } |
| 787 | static void f32_igemm_5x16__avx512f_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 788 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast, 5, 16, 1, 1, |
| 789 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 790 | } |
| 791 | static void f32_igemm_6x16__avx512f_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 792 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast, 6, 16, 1, 1, |
| 793 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 794 | } |
| 795 | static void f32_igemm_7x16__avx512f_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 796 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast, 7, 16, 1, 1, |
| 797 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 798 | } |
| 799 | static void f32_igemm_8x16__avx512f_broadcast(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 800 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast, 8, 16, 1, 1, |
| 801 | xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 802 | } |
| 803 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 804 | BENCHMARK_CONV(f32_igemm_1x8__sse_load1) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 805 | BENCHMARK_CONV(f32_igemm_3x8__sse_load1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 806 | BENCHMARK_CONV(f32_igemm_4x8__sse_load1) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 807 | BENCHMARK_CONV(f32_igemm_5x8__sse_load1) |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 808 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 809 | BENCHMARK_CONV(f32_igemm_1x8__sse_dup) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 810 | BENCHMARK_CONV(f32_igemm_3x8__sse_dup) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 811 | BENCHMARK_CONV(f32_igemm_4x8__sse_dup) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 812 | BENCHMARK_CONV(f32_igemm_5x8__sse_dup) |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 813 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 814 | BENCHMARK_CONV(f32_igemm_1x8s4__sse) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 815 | BENCHMARK_CONV(f32_igemm_3x8s4__sse) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 816 | BENCHMARK_CONV(f32_igemm_4x8s4__sse) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 817 | BENCHMARK_CONV(f32_igemm_5x8s4__sse) |
| 818 | |
| 819 | BENCHMARK_CONV(f32_igemm_1x8__sse2_dup) |
| 820 | BENCHMARK_CONV(f32_igemm_3x8__sse2_dup) |
| 821 | BENCHMARK_CONV(f32_igemm_4x8__sse2_dup) |
| 822 | BENCHMARK_CONV(f32_igemm_5x8__sse2_dup) |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 823 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 824 | BENCHMARK_CONV(f32_igemm_1x8__avx_broadcast) |
| 825 | BENCHMARK_CONV(f32_igemm_4x8__avx_broadcast) |
| 826 | BENCHMARK_CONV(f32_igemm_5x8__avx_broadcast) |
| 827 | BENCHMARK_CONV(f32_igemm_6x8__avx_broadcast) |
| 828 | BENCHMARK_CONV(f32_igemm_7x8__avx_broadcast) |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 829 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 830 | BENCHMARK_CONV(f32_igemm_1x8__fma3_broadcast) |
| 831 | BENCHMARK_CONV(f32_igemm_4x8__fma3_broadcast) |
| 832 | BENCHMARK_CONV(f32_igemm_5x8__fma3_broadcast) |
| 833 | BENCHMARK_CONV(f32_igemm_6x8__fma3_broadcast) |
| 834 | BENCHMARK_CONV(f32_igemm_7x8__fma3_broadcast) |
| 835 | BENCHMARK_CONV(f32_igemm_8x8__fma3_broadcast) |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 836 | |
| 837 | BENCHMARK_CONV(f32_igemm_1x16__avx512f_broadcast) |
| 838 | BENCHMARK_CONV(f32_igemm_4x16__avx512f_broadcast) |
| 839 | BENCHMARK_CONV(f32_igemm_5x16__avx512f_broadcast) |
| 840 | BENCHMARK_CONV(f32_igemm_6x16__avx512f_broadcast) |
| 841 | BENCHMARK_CONV(f32_igemm_7x16__avx512f_broadcast) |
| 842 | BENCHMARK_CONV(f32_igemm_8x16__avx512f_broadcast) |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 843 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 844 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 845 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 846 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 847 | static void f32_igemm_3x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 848 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_arm_loadsplat, 3, 8, 1, 1, |
| 849 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 850 | } |
| 851 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 852 | static void f32_igemm_4x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 853 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_arm_loadsplat, 4, 8, 1, 1, |
| 854 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 855 | } |
| 856 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 857 | static void f32_igemm_5x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 858 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_loadsplat, 5, 8, 1, 1, |
| 859 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 860 | } |
| 861 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 862 | static void f32_igemm_6x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 863 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_loadsplat, 6, 8, 1, 1, |
| 864 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 865 | } |
| 866 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 867 | static void f32_igemm_3x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 868 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_x86_loadsplat, 3, 8, 1, 1, |
| 869 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 870 | } |
| 871 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 872 | static void f32_igemm_4x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 873 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_x86_loadsplat, 4, 8, 1, 1, |
| 874 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 875 | } |
| 876 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 877 | static void f32_igemm_5x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 878 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat, 5, 8, 1, 1, |
| 879 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 880 | } |
| 881 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 882 | static void f32_igemm_6x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 883 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat, 6, 8, 1, 1, |
| 884 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 885 | } |
| 886 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 887 | static void f32_igemm_3x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 888 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_arm_splat, 3, 8, 1, 1, |
| 889 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 890 | } |
| 891 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 892 | static void f32_igemm_4x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 893 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_arm_splat, 4, 8, 1, 1, |
| 894 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 895 | } |
| 896 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 897 | static void f32_igemm_5x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 898 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_splat, 5, 8, 1, 1, |
| 899 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 900 | } |
| 901 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 902 | static void f32_igemm_6x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 903 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat, 6, 8, 1, 1, |
| 904 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 905 | } |
| 906 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 907 | static void f32_igemm_3x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 908 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_x86_splat, 3, 8, 1, 1, |
| 909 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 910 | } |
| 911 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 912 | static void f32_igemm_4x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 913 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_x86_splat, 4, 8, 1, 1, |
| 914 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 915 | } |
| 916 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 917 | static void f32_igemm_5x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 918 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_splat, 5, 8, 1, 1, |
| 919 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 920 | } |
| 921 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 922 | static void f32_igemm_6x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 923 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat, 6, 8, 1, 1, |
| 924 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 925 | } |
| 926 | |
| 927 | static void f32_igemm_3x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 928 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm, 3, 8, 1, 4, |
| 929 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 930 | } |
| 931 | |
| 932 | static void f32_igemm_4x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 933 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__wasmsimd_arm, 4, 8, 1, 4, |
| 934 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 935 | } |
| 936 | |
| 937 | static void f32_igemm_5x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 938 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm, 5, 8, 1, 4, |
| 939 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 940 | } |
| 941 | |
| 942 | static void f32_igemm_6x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 943 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__wasmsimd_arm, 6, 8, 1, 4, |
| 944 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 945 | } |
| 946 | |
| 947 | static void f32_igemm_3x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 948 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_x86, 3, 8, 1, 4, |
| 949 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 950 | } |
| 951 | |
| 952 | static void f32_igemm_4x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 953 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__wasmsimd_x86, 4, 8, 1, 4, |
| 954 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 955 | } |
| 956 | |
| 957 | static void f32_igemm_5x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 958 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_x86, 5, 8, 1, 4, |
| 959 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 960 | } |
| 961 | |
| 962 | static void f32_igemm_6x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 963 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__wasmsimd_x86, 6, 8, 1, 4, |
| 964 | xnn_init_f32_minmax_scalar_params); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 965 | } |
| 966 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 967 | BENCHMARK_CONV(f32_igemm_3x8__wasmsimd_arm_loadsplat) |
| 968 | BENCHMARK_CONV(f32_igemm_4x8__wasmsimd_arm_loadsplat) |
| 969 | BENCHMARK_CONV(f32_igemm_5x8__wasmsimd_arm_loadsplat) |
| 970 | BENCHMARK_CONV(f32_igemm_6x8__wasmsimd_arm_loadsplat) |
| 971 | BENCHMARK_CONV(f32_igemm_3x8__wasmsimd_x86_loadsplat) |
| 972 | BENCHMARK_CONV(f32_igemm_4x8__wasmsimd_x86_loadsplat) |
| 973 | BENCHMARK_CONV(f32_igemm_5x8__wasmsimd_x86_loadsplat) |
| 974 | BENCHMARK_CONV(f32_igemm_6x8__wasmsimd_x86_loadsplat) |
| 975 | BENCHMARK_CONV(f32_igemm_3x8__wasmsimd_arm_splat) |
| 976 | BENCHMARK_CONV(f32_igemm_4x8__wasmsimd_arm_splat) |
| 977 | BENCHMARK_CONV(f32_igemm_5x8__wasmsimd_arm_splat) |
| 978 | BENCHMARK_CONV(f32_igemm_6x8__wasmsimd_arm_splat) |
| 979 | BENCHMARK_CONV(f32_igemm_3x8__wasmsimd_x86_splat) |
| 980 | BENCHMARK_CONV(f32_igemm_4x8__wasmsimd_x86_splat) |
| 981 | BENCHMARK_CONV(f32_igemm_5x8__wasmsimd_x86_splat) |
| 982 | BENCHMARK_CONV(f32_igemm_6x8__wasmsimd_x86_splat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 983 | BENCHMARK_CONV(f32_igemm_3x8s4__wasmsimd_arm) |
| 984 | BENCHMARK_CONV(f32_igemm_4x8s4__wasmsimd_arm) |
| 985 | BENCHMARK_CONV(f32_igemm_5x8s4__wasmsimd_arm) |
| 986 | BENCHMARK_CONV(f32_igemm_6x8s4__wasmsimd_arm) |
| 987 | BENCHMARK_CONV(f32_igemm_3x8s4__wasmsimd_x86) |
| 988 | BENCHMARK_CONV(f32_igemm_4x8s4__wasmsimd_x86) |
| 989 | BENCHMARK_CONV(f32_igemm_5x8s4__wasmsimd_x86) |
| 990 | BENCHMARK_CONV(f32_igemm_6x8s4__wasmsimd_x86) |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 991 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 992 | |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 993 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 994 | static void f32_igemm_1x4__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 995 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x4__scalar, 1, 4, 1, 1, |
| 996 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 997 | } |
| 998 | |
| 999 | static void f32_igemm_2x4__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 1000 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_2x4__scalar, 2, 4, 1, 1, |
| 1001 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1002 | } |
| 1003 | |
| 1004 | static void f32_igemm_4x4__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 1005 | IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x4__scalar, 4, 4, 1, 1, |
| 1006 | xnn_init_f32_minmax_scalar_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1007 | } |
| 1008 | |
| 1009 | BENCHMARK_CONV(f32_igemm_1x4__scalar) |
| 1010 | BENCHMARK_CONV(f32_igemm_2x4__scalar) |
| 1011 | BENCHMARK_CONV(f32_igemm_4x4__scalar) |
| 1012 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1013 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 1014 | BENCHMARK_MAIN(); |
| 1015 | #endif |