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/dwconv.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/dwconv.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> |
| 25 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | |
| 27 | |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 28 | static void DWConv2DBenchmark(benchmark::State& state, |
| 29 | xnn_f32_dwconv2d_chw_ukernel_function dwconv, |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 30 | uint32_t kh, uint32_t kw, uint32_t pw, uint32_t s, |
| 31 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 32 | { |
| 33 | if (!cpuinfo_initialize()) { |
| 34 | state.SkipWithError("cpuinfo initialization failed"); |
| 35 | return; |
| 36 | } |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 37 | if (isa_check && !isa_check(state)) { |
| 38 | return; |
| 39 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 40 | |
| 41 | const size_t input_height = state.range(0); |
| 42 | const size_t input_width = state.range(1); |
| 43 | const size_t kernel_height = state.range(2); |
| 44 | const size_t kernel_width = state.range(3); |
| 45 | const size_t padding_height = state.range(4); |
| 46 | const size_t padding_width = state.range(5); |
| 47 | const size_t subsampling = state.range(6); |
| 48 | const size_t dilation = state.range(7); |
| 49 | const size_t channels = state.range(8); |
| 50 | |
| 51 | if (kernel_height != kh) { |
| 52 | state.SkipWithError("kernel height mismatch"); |
| 53 | return; |
| 54 | } |
| 55 | |
| 56 | if (kernel_width != kw) { |
| 57 | state.SkipWithError("kernel width mismatch"); |
| 58 | return; |
| 59 | } |
| 60 | |
| 61 | if (subsampling != s) { |
| 62 | state.SkipWithError("subsampling mismatch"); |
| 63 | return; |
| 64 | } |
| 65 | |
| 66 | if (padding_width % 2 != 0 || padding_width / 2 != pw) { |
| 67 | state.SkipWithError("padding width mismatch"); |
| 68 | return; |
| 69 | } |
| 70 | |
| 71 | if (dilation != 1) { |
| 72 | state.SkipWithError("unsupported dilation"); |
| 73 | return; |
| 74 | } |
| 75 | |
| 76 | std::random_device random_device; |
| 77 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 78 | 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] | 79 | |
| 80 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 81 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 82 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 83 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 84 | |
| 85 | const size_t inputSize = (input_height + padding_height) * input_width; |
| 86 | const size_t kernel_size = kernel_height * kernel_width; |
| 87 | const size_t output_size = output_height * output_width; |
| 88 | |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 89 | std::vector<float> input(inputSize * channels + 2 * XNN_EXTRA_BYTES); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 90 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 91 | std::vector<float> bias(channels); |
| 92 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 93 | std::vector<float> kernel(channels * kernel_size); |
| 94 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 95 | std::vector<float> zero(input_width + padding_width); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 96 | |
| 97 | const size_t w_elements = (kernel_size + 1) * channels; |
| 98 | const size_t o_elements = output_size * channels; |
| 99 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 100 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 101 | sizeof(float) * (w_elements + o_elements)); |
| 102 | |
| 103 | std::vector<float, AlignedAllocator<float, 32>> packed_weights(w_elements * num_buffers); |
| 104 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 105 | for (size_t c = 0; c < channels; c++) { |
| 106 | packed_weights[c * kernel_size + c] = bias[c]; |
| 107 | for (size_t i = 0; i < kernel_size; i++) { |
| 108 | packed_weights[c * kernel_size + c + 1 + i] = kernel[c * kernel_size + i]; |
| 109 | } |
| 110 | } |
| 111 | for (size_t n = 1; n < num_buffers; n++) { |
| 112 | std::copy(packed_weights.cbegin(), packed_weights.cbegin() + w_elements, packed_weights.begin() + n * w_elements); |
| 113 | } |
| 114 | |
| 115 | std::vector<float> output(o_elements * num_buffers); |
| 116 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 117 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 118 | xnn_f32_chw_params chw_params = |
| 119 | xnn_init_f32_chw_params(input_width, -std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 120 | |
| 121 | size_t buffer_index = 0; |
| 122 | for (auto _ : state) { |
| 123 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 124 | benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 125 | buffer_index = (buffer_index + 1) % num_buffers; |
| 126 | state.ResumeTiming(); |
| 127 | |
| 128 | for (uint32_t channel = 0; channel < channels; channel++) { |
| 129 | dwconv( |
Marat Dukhan | 7515777 | 2020-10-21 01:46:28 -0700 | [diff] [blame] | 130 | input_height, input_width * sizeof(float), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 131 | input.data() + channel * inputSize, |
| 132 | packed_weights.data() + channel * (kernel_size + 1) + buffer_index * w_elements, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 133 | zero.data(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 134 | output.data() + channel * output_size + buffer_index * o_elements, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 135 | padding_height / 2, // padding_top |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 136 | &chw_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 137 | } |
| 138 | } |
| 139 | |
| 140 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 141 | state.counters["FLOPS"] = benchmark::Counter( |
| 142 | uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size, |
| 143 | benchmark::Counter::kIsRate); |
| 144 | |
| 145 | state.counters["BYTES"] = benchmark::Counter( |
| 146 | uint64_t(state.iterations()) * (output_size + inputSize + kernel_size + 1 /* bias */) * channels * sizeof(float), |
| 147 | benchmark::Counter::kIsRate); |
| 148 | } |
| 149 | |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 150 | #if XNN_ARCH_ARM |
| 151 | static void dwconv2d_chw_3x3p1__neon_1x4(benchmark::State& state, const char* net) { |
| 152 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4, 3, 3, 1, 1); |
| 153 | } |
| 154 | static void dwconv2d_chw_3x3p1__neon_2x4(benchmark::State& state, const char* net) { |
| 155 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4, 3, 3, 1, 1); |
| 156 | } |
| 157 | static void dwconv2d_chw_3x3p1__neon_3x4(benchmark::State& state, const char* net) { |
| 158 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4, 3, 3, 1, 1); |
| 159 | } |
| 160 | static void dwconv2d_chw_3x3p1__neon_4x4(benchmark::State& state, const char* net) { |
| 161 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4, 3, 3, 1, 1); |
| 162 | } |
| 163 | static void dwconv2d_chw_3x3p1__neon_5x4(benchmark::State& state, const char* net) { |
| 164 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4, 3, 3, 1, 1); |
| 165 | } |
| 166 | static void dwconv2d_chw_3x3p1__neon_6x4(benchmark::State& state, const char* net) { |
| 167 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4, 3, 3, 1, 1); |
| 168 | } |
| 169 | static void dwconv2d_chw_3x3p1__neon_1x4_acc2(benchmark::State& state, const char* net) { |
| 170 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2, 3, 3, 1, 1); |
| 171 | } |
| 172 | static void dwconv2d_chw_3x3p1__neon_1x4_acc3(benchmark::State& state, const char* net) { |
| 173 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3, 3, 3, 1, 1); |
| 174 | } |
| 175 | static void dwconv2d_chw_3x3p1__neon_1x4_acc4(benchmark::State& state, const char* net) { |
| 176 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4, 3, 3, 1, 1); |
| 177 | } |
| 178 | static void dwconv2d_chw_3x3p1__neon_2x4_acc2(benchmark::State& state, const char* net) { |
| 179 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2, 3, 3, 1, 1); |
| 180 | } |
| 181 | |
| 182 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4) |
| 183 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_2x4) |
| 184 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_3x4) |
| 185 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_4x4) |
| 186 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_5x4) |
| 187 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_6x4) |
| 188 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc2) |
| 189 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc3) |
| 190 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc4) |
| 191 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_2x4_acc2) |
| 192 | #endif // XNN_ARCH_ARM |
| 193 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 194 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 195 | static void dwconv2d_chw_3x3p1__neonfma_1x4(benchmark::State& state, const char* net) { |
| 196 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4, 3, 3, 1, 1); |
| 197 | } |
| 198 | static void dwconv2d_chw_3x3p1__neonfma_2x4(benchmark::State& state, const char* net) { |
| 199 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4, 3, 3, 1, 1); |
| 200 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 201 | static void dwconv2d_chw_3x3p1__neonfma_3x4(benchmark::State& state, const char* net) { |
| 202 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4, 3, 3, 1, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 203 | } |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 204 | static void dwconv2d_chw_3x3p1__neonfma_4x4(benchmark::State& state, const char* net) { |
| 205 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4, 3, 3, 1, 1); |
| 206 | } |
| 207 | static void dwconv2d_chw_3x3p1__neonfma_5x4(benchmark::State& state, const char* net) { |
| 208 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4, 3, 3, 1, 1); |
| 209 | } |
| 210 | static void dwconv2d_chw_3x3p1__neonfma_6x4(benchmark::State& state, const char* net) { |
| 211 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4, 3, 3, 1, 1); |
| 212 | } |
| 213 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 214 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2, 3, 3, 1, 1); |
| 215 | } |
| 216 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 217 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3, 3, 3, 1, 1); |
| 218 | } |
| 219 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 220 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4, 3, 3, 1, 1); |
| 221 | } |
| 222 | static void dwconv2d_chw_3x3p1__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 223 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2, 3, 3, 1, 1); |
| 224 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 225 | |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 226 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4(benchmark::State& state, const char* net) { |
| 227 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4, 3, 3, 1, 2); |
| 228 | } |
| 229 | static void dwconv2d_chw_3x3s2p1__neonfma_2x4(benchmark::State& state, const char* net) { |
| 230 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4, 3, 3, 1, 2); |
| 231 | } |
| 232 | static void dwconv2d_chw_3x3s2p1__neonfma_3x4(benchmark::State& state, const char* net) { |
| 233 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4, 3, 3, 1, 2); |
| 234 | } |
| 235 | static void dwconv2d_chw_3x3s2p1__neonfma_4x4(benchmark::State& state, const char* net) { |
| 236 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4, 3, 3, 1, 2); |
| 237 | } |
| 238 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 239 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2, 3, 3, 1, 2); |
| 240 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 241 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 242 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3, 3, 3, 1, 2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 243 | } |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 244 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 245 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4, 3, 3, 1, 2); |
| 246 | } |
| 247 | static void dwconv2d_chw_3x3s2p1__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 248 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2, 3, 3, 1, 2); |
| 249 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 250 | |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 251 | static void dwconv2d_chw_5x5p2__neonfma_1x4(benchmark::State& state, const char* net) { |
| 252 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4, 5, 5, 2, 1); |
| 253 | } |
| 254 | static void dwconv2d_chw_5x5p2__neonfma_2x4(benchmark::State& state, const char* net) { |
| 255 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4, 5, 5, 2, 1); |
| 256 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 257 | static void dwconv2d_chw_5x5p2__neonfma_3x4(benchmark::State& state, const char* net) { |
| 258 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4, 5, 5, 2, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 259 | } |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 260 | static void dwconv2d_chw_5x5p2__neonfma_4x4(benchmark::State& state, const char* net) { |
| 261 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4, 5, 5, 2, 1); |
| 262 | } |
| 263 | static void dwconv2d_chw_5x5p2__neonfma_5x4(benchmark::State& state, const char* net) { |
| 264 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4, 5, 5, 2, 1); |
| 265 | } |
| 266 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 267 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2, 5, 5, 2, 1); |
| 268 | } |
| 269 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 270 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3, 5, 5, 2, 1); |
| 271 | } |
| 272 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 273 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4, 5, 5, 2, 1); |
| 274 | } |
| 275 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc5(benchmark::State& state, const char* net) { |
| 276 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5, 5, 5, 2, 1); |
| 277 | } |
| 278 | static void dwconv2d_chw_5x5p2__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 279 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2, 5, 5, 2, 1); |
| 280 | } |
| 281 | static void dwconv2d_chw_5x5p2__neonfma_2x4_acc3(benchmark::State& state, const char* net) { |
| 282 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3, 5, 5, 2, 1); |
| 283 | } |
| 284 | static void dwconv2d_chw_5x5p2__neonfma_3x4_acc2(benchmark::State& state, const char* net) { |
| 285 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2, 5, 5, 2, 1); |
| 286 | } |
| 287 | static void dwconv2d_chw_5x5p2__neonfma_4x4_acc2(benchmark::State& state, const char* net) { |
| 288 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2, 5, 5, 2, 1); |
| 289 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 290 | |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 291 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4(benchmark::State& state, const char* net) { |
| 292 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4, 5, 5, 2, 2); |
| 293 | } |
| 294 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4(benchmark::State& state, const char* net) { |
| 295 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4, 5, 5, 2, 2); |
| 296 | } |
| 297 | static void dwconv2d_chw_5x5s2p2__neonfma_3x4(benchmark::State& state, const char* net) { |
| 298 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4, 5, 5, 2, 2); |
| 299 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 300 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 301 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2, 5, 5, 2, 2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 302 | } |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 303 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 304 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3, 5, 5, 2, 2); |
| 305 | } |
| 306 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 307 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4, 5, 5, 2, 2); |
| 308 | } |
| 309 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc5(benchmark::State& state, const char* net) { |
| 310 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5, 5, 5, 2, 2); |
| 311 | } |
| 312 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 313 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2, 5, 5, 2, 2); |
| 314 | } |
| 315 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4_acc3(benchmark::State& state, const char* net) { |
| 316 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3, 5, 5, 2, 2); |
| 317 | } |
| 318 | static void dwconv2d_chw_5x5s2p2__neonfma_3x4_acc2(benchmark::State& state, const char* net) { |
| 319 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2, 5, 5, 2, 2); |
| 320 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 321 | |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 322 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4) |
| 323 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_2x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 324 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_3x4) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 325 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_4x4) |
| 326 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_5x4) |
| 327 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_6x4) |
| 328 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc2) |
| 329 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc3) |
| 330 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc4) |
| 331 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_2x4_acc2) |
| 332 | |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 333 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4) |
| 334 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_2x4) |
| 335 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_3x4) |
| 336 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_4x4) |
| 337 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 338 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc3) |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 339 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc4) |
| 340 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_2x4_acc2) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 341 | |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 342 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4) |
| 343 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 344 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_3x4) |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 345 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_4x4) |
| 346 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_5x4) |
| 347 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc2) |
| 348 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc3) |
| 349 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc4) |
| 350 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc5) |
| 351 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4_acc2) |
| 352 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4_acc3) |
| 353 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_3x4_acc2) |
| 354 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_4x4_acc2) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 355 | |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 356 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4) |
| 357 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4) |
| 358 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_3x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 359 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc2) |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 360 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc3) |
| 361 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc4) |
| 362 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc5) |
| 363 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4_acc2) |
| 364 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4_acc3) |
| 365 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_3x4_acc2) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 366 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 367 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 368 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 369 | static void dwconv2d_chw_3x3p1__sse_1x4(benchmark::State& state, const char* net) { |
| 370 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4, 3, 3, 1, 1); |
| 371 | } |
| 372 | static void dwconv2d_chw_3x3p1__sse_2x4(benchmark::State& state, const char* net) { |
| 373 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4, 3, 3, 1, 1); |
| 374 | } |
| 375 | static void dwconv2d_chw_3x3p1__sse_3x4(benchmark::State& state, const char* net) { |
| 376 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4, 3, 3, 1, 1); |
| 377 | } |
| 378 | static void dwconv2d_chw_3x3p1__sse_4x4(benchmark::State& state, const char* net) { |
| 379 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4, 3, 3, 1, 1); |
| 380 | } |
| 381 | static void dwconv2d_chw_3x3p1__sse_5x4(benchmark::State& state, const char* net) { |
| 382 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4, 3, 3, 1, 1); |
| 383 | } |
| 384 | static void dwconv2d_chw_3x3p1__sse_6x4(benchmark::State& state, const char* net) { |
| 385 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4, 3, 3, 1, 1); |
| 386 | } |
| 387 | static void dwconv2d_chw_3x3p1__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 388 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2, 3, 3, 1, 1); |
| 389 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 390 | static void dwconv2d_chw_3x3p1__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 391 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3, 3, 3, 1, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 392 | } |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 393 | static void dwconv2d_chw_3x3p1__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 394 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4, 3, 3, 1, 1); |
| 395 | } |
| 396 | static void dwconv2d_chw_3x3p1__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 397 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2, 3, 3, 1, 1); |
| 398 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 399 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 400 | static void dwconv2d_chw_3x3p1__ssse3_1x4(benchmark::State& state, const char* net) { |
| 401 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 402 | } |
| 403 | static void dwconv2d_chw_3x3p1__ssse3_2x4(benchmark::State& state, const char* net) { |
| 404 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 405 | } |
| 406 | static void dwconv2d_chw_3x3p1__ssse3_3x4(benchmark::State& state, const char* net) { |
| 407 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 408 | } |
| 409 | static void dwconv2d_chw_3x3p1__ssse3_4x4(benchmark::State& state, const char* net) { |
| 410 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 411 | } |
| 412 | static void dwconv2d_chw_3x3p1__ssse3_5x4(benchmark::State& state, const char* net) { |
| 413 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 414 | } |
| 415 | static void dwconv2d_chw_3x3p1__ssse3_6x4(benchmark::State& state, const char* net) { |
| 416 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 417 | } |
| 418 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc2(benchmark::State& state, const char* net) { |
| 419 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 420 | } |
| 421 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc3(benchmark::State& state, const char* net) { |
| 422 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 423 | } |
| 424 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc4(benchmark::State& state, const char* net) { |
| 425 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 426 | } |
| 427 | static void dwconv2d_chw_3x3p1__ssse3_2x4_acc2(benchmark::State& state, const char* net) { |
| 428 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 429 | } |
| 430 | |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 431 | static void dwconv2d_chw_3x3s2p1__sse_1x4(benchmark::State& state, const char* net) { |
| 432 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4, 3, 3, 1, 2); |
| 433 | } |
| 434 | static void dwconv2d_chw_3x3s2p1__sse_2x4(benchmark::State& state, const char* net) { |
| 435 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4, 3, 3, 1, 2); |
| 436 | } |
| 437 | static void dwconv2d_chw_3x3s2p1__sse_3x4(benchmark::State& state, const char* net) { |
| 438 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4, 3, 3, 1, 2); |
| 439 | } |
| 440 | static void dwconv2d_chw_3x3s2p1__sse_4x4(benchmark::State& state, const char* net) { |
| 441 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4, 3, 3, 1, 2); |
| 442 | } |
| 443 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 444 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2, 3, 3, 1, 2); |
| 445 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 446 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 447 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3, 3, 3, 1, 2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 448 | } |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 449 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 450 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4, 3, 3, 1, 2); |
| 451 | } |
| 452 | static void dwconv2d_chw_3x3s2p1__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 453 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2, 3, 3, 1, 2); |
| 454 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 455 | |
Marat Dukhan | d050389 | 2020-10-30 08:22:04 -0700 | [diff] [blame] | 456 | static void dwconv2d_chw_5x5p2__sse_1x4(benchmark::State& state, const char* net) { |
| 457 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4, 5, 5, 2, 1); |
| 458 | } |
| 459 | static void dwconv2d_chw_5x5p2__sse_2x4(benchmark::State& state, const char* net) { |
| 460 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4, 5, 5, 2, 1); |
| 461 | } |
| 462 | static void dwconv2d_chw_5x5p2__sse_3x4(benchmark::State& state, const char* net) { |
| 463 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4, 5, 5, 2, 1); |
| 464 | } |
| 465 | static void dwconv2d_chw_5x5p2__sse_4x4(benchmark::State& state, const char* net) { |
| 466 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4, 5, 5, 2, 1); |
| 467 | } |
| 468 | static void dwconv2d_chw_5x5p2__sse_5x4(benchmark::State& state, const char* net) { |
| 469 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4, 5, 5, 2, 1); |
| 470 | } |
| 471 | static void dwconv2d_chw_5x5p2__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 472 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2, 5, 5, 2, 1); |
| 473 | } |
| 474 | static void dwconv2d_chw_5x5p2__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 475 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3, 5, 5, 2, 1); |
| 476 | } |
| 477 | static void dwconv2d_chw_5x5p2__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 478 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4, 5, 5, 2, 1); |
| 479 | } |
| 480 | static void dwconv2d_chw_5x5p2__sse_1x4_acc5(benchmark::State& state, const char* net) { |
| 481 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5, 5, 5, 2, 1); |
| 482 | } |
| 483 | static void dwconv2d_chw_5x5p2__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 484 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2, 5, 5, 2, 1); |
| 485 | } |
| 486 | static void dwconv2d_chw_5x5p2__sse_2x4_acc3(benchmark::State& state, const char* net) { |
| 487 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3, 5, 5, 2, 1); |
| 488 | } |
| 489 | static void dwconv2d_chw_5x5p2__sse_3x4_acc2(benchmark::State& state, const char* net) { |
| 490 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2, 5, 5, 2, 1); |
| 491 | } |
| 492 | static void dwconv2d_chw_5x5p2__sse_4x4_acc2(benchmark::State& state, const char* net) { |
| 493 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2, 5, 5, 2, 1); |
| 494 | } |
| 495 | |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 496 | static void dwconv2d_chw_5x5s2p2__sse_1x4(benchmark::State& state, const char* net) { |
| 497 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4, 5, 5, 2, 2); |
| 498 | } |
| 499 | static void dwconv2d_chw_5x5s2p2__sse_2x4(benchmark::State& state, const char* net) { |
| 500 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4, 5, 5, 2, 2); |
| 501 | } |
| 502 | static void dwconv2d_chw_5x5s2p2__sse_3x4(benchmark::State& state, const char* net) { |
| 503 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4, 5, 5, 2, 2); |
| 504 | } |
| 505 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 506 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2, 5, 5, 2, 2); |
| 507 | } |
| 508 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 509 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3, 5, 5, 2, 2); |
| 510 | } |
| 511 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 512 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4, 5, 5, 2, 2); |
| 513 | } |
| 514 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc5(benchmark::State& state, const char* net) { |
| 515 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5, 5, 5, 2, 2); |
| 516 | } |
| 517 | static void dwconv2d_chw_5x5s2p2__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 518 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc2, 5, 5, 2, 2); |
| 519 | } |
| 520 | static void dwconv2d_chw_5x5s2p2__sse_2x4_acc3(benchmark::State& state, const char* net) { |
| 521 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc3, 5, 5, 2, 2); |
| 522 | } |
| 523 | static void dwconv2d_chw_5x5s2p2__sse_3x4_acc2(benchmark::State& state, const char* net) { |
| 524 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2, 5, 5, 2, 2); |
| 525 | } |
| 526 | |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 527 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4) |
| 528 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_2x4) |
| 529 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_3x4) |
| 530 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_4x4) |
| 531 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_5x4) |
| 532 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_6x4) |
| 533 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 534 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc3) |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 535 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc4) |
| 536 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_2x4_acc2) |
| 537 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 538 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4) |
| 539 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_2x4) |
| 540 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_3x4) |
| 541 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_4x4) |
| 542 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_5x4) |
| 543 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_6x4) |
| 544 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc2) |
| 545 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc3) |
| 546 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc4) |
| 547 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_2x4_acc2) |
| 548 | |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 549 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4) |
| 550 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_2x4) |
| 551 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_3x4) |
| 552 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_4x4) |
| 553 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 554 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc3) |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 555 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc4) |
| 556 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_2x4_acc2) |
Marat Dukhan | d050389 | 2020-10-30 08:22:04 -0700 | [diff] [blame] | 557 | |
| 558 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4) |
| 559 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4) |
| 560 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_3x4) |
| 561 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_4x4) |
| 562 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_5x4) |
| 563 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc2) |
| 564 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc3) |
| 565 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc4) |
| 566 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc5) |
| 567 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4_acc2) |
| 568 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4_acc3) |
| 569 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_3x4_acc2) |
| 570 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_4x4_acc2) |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 571 | |
| 572 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4) |
| 573 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4) |
| 574 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_3x4) |
| 575 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc2) |
| 576 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc3) |
| 577 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc4) |
| 578 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc5) |
| 579 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4_acc2) |
| 580 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4_acc3) |
| 581 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_3x4_acc2) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 582 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 583 | |
Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame^] | 584 | #if XNN_ARCH_WASMSIMD |
| 585 | static void dwconv2d_chw_3x3p1__wasmsimd_1x4_acc3(benchmark::State& state, const char* net) { |
| 586 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_1x4_acc3, 3, 3, 1, 1); |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 587 | } |
Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame^] | 588 | static void dwconv2d_chw_3x3s2p1__wasmsimd_1x4_acc3(benchmark::State& state, const char* net) { |
| 589 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_1x4_acc3, 3, 3, 1, 2); |
Erich Elsen | fd7a6e3 | 2020-06-11 12:04:44 -0700 | [diff] [blame] | 590 | } |
Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame^] | 591 | static void dwconv2d_chw_5x5p2__wasmsimd_3x4(benchmark::State& state, const char* net) { |
| 592 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_3x4, 5, 5, 2, 1); |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 593 | } |
Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame^] | 594 | static void dwconv2d_chw_5x5s2p2__wasmsimd_1x4_acc2(benchmark::State& state, const char* net) { |
| 595 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_1x4_acc2, 5, 5, 2, 2); |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 596 | } |
| 597 | |
Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame^] | 598 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_1x4_acc3) |
| 599 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_1x4_acc3) |
| 600 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_3x4) |
| 601 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_1x4_acc2) |
| 602 | #endif // XNN_ARCH_WASMSIMD |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 603 | |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 604 | static void dwconv2d_chw_3x3p1__scalar_1x1(benchmark::State& state, const char* net) { |
| 605 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1, 3, 3, 1, 1); |
| 606 | } |
| 607 | static void dwconv2d_chw_3x3p1__scalar_2x1(benchmark::State& state, const char* net) { |
| 608 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1, 3, 3, 1, 1); |
| 609 | } |
| 610 | static void dwconv2d_chw_3x3p1__scalar_3x1(benchmark::State& state, const char* net) { |
| 611 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1, 3, 3, 1, 1); |
| 612 | } |
| 613 | static void dwconv2d_chw_3x3p1__scalar_4x1(benchmark::State& state, const char* net) { |
| 614 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1, 3, 3, 1, 1); |
| 615 | } |
| 616 | static void dwconv2d_chw_3x3p1__scalar_5x1(benchmark::State& state, const char* net) { |
| 617 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1, 3, 3, 1, 1); |
| 618 | } |
| 619 | static void dwconv2d_chw_3x3p1__scalar_6x1(benchmark::State& state, const char* net) { |
| 620 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1, 3, 3, 1, 1); |
| 621 | } |
| 622 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 623 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2, 3, 3, 1, 1); |
| 624 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 625 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 626 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3, 3, 3, 1, 1); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 627 | } |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 628 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 629 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4, 3, 3, 1, 1); |
| 630 | } |
| 631 | static void dwconv2d_chw_3x3p1__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 632 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2, 3, 3, 1, 1); |
| 633 | } |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 634 | |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 635 | static void dwconv2d_chw_3x3s2p1__scalar_1x1(benchmark::State& state, const char* net) { |
| 636 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1, 3, 3, 1, 2); |
| 637 | } |
| 638 | static void dwconv2d_chw_3x3s2p1__scalar_2x1(benchmark::State& state, const char* net) { |
| 639 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1, 3, 3, 1, 2); |
| 640 | } |
| 641 | static void dwconv2d_chw_3x3s2p1__scalar_3x1(benchmark::State& state, const char* net) { |
| 642 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1, 3, 3, 1, 2); |
| 643 | } |
| 644 | static void dwconv2d_chw_3x3s2p1__scalar_4x1(benchmark::State& state, const char* net) { |
| 645 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1, 3, 3, 1, 2); |
| 646 | } |
| 647 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 648 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2, 3, 3, 1, 2); |
| 649 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 650 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 651 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3, 3, 3, 1, 2); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 652 | } |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 653 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 654 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4, 3, 3, 1, 2); |
| 655 | } |
| 656 | static void dwconv2d_chw_3x3s2p1__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 657 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2, 3, 3, 1, 2); |
| 658 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 659 | |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 660 | static void dwconv2d_chw_5x5p2__scalar_1x1(benchmark::State& state, const char* net) { |
| 661 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1, 5, 5, 2, 1); |
| 662 | } |
| 663 | static void dwconv2d_chw_5x5p2__scalar_2x1(benchmark::State& state, const char* net) { |
| 664 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1, 5, 5, 2, 1); |
| 665 | } |
| 666 | static void dwconv2d_chw_5x5p2__scalar_3x1(benchmark::State& state, const char* net) { |
| 667 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1, 5, 5, 2, 1); |
| 668 | } |
| 669 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 670 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2, 5, 5, 2, 1); |
| 671 | } |
| 672 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 673 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3, 5, 5, 2, 1); |
| 674 | } |
| 675 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 676 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4, 5, 5, 2, 1); |
| 677 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 678 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc5(benchmark::State& state, const char* net) { |
| 679 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5, 5, 5, 2, 1); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 680 | } |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 681 | static void dwconv2d_chw_5x5p2__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 682 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2, 5, 5, 2, 1); |
| 683 | } |
| 684 | static void dwconv2d_chw_5x5p2__scalar_2x1_acc3(benchmark::State& state, const char* net) { |
| 685 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3, 5, 5, 2, 1); |
| 686 | } |
| 687 | static void dwconv2d_chw_5x5p2__scalar_3x1_acc2(benchmark::State& state, const char* net) { |
| 688 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2, 5, 5, 2, 1); |
| 689 | } |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 690 | |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 691 | static void dwconv2d_chw_5x5s2p2__scalar_1x1(benchmark::State& state, const char* net) { |
| 692 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1, 5, 5, 2, 2); |
| 693 | } |
| 694 | static void dwconv2d_chw_5x5s2p2__scalar_2x1(benchmark::State& state, const char* net) { |
| 695 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1, 5, 5, 2, 2); |
| 696 | } |
| 697 | static void dwconv2d_chw_5x5s2p2__scalar_3x1(benchmark::State& state, const char* net) { |
| 698 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1, 5, 5, 2, 2); |
| 699 | } |
| 700 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 701 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2, 5, 5, 2, 2); |
| 702 | } |
| 703 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 704 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3, 5, 5, 2, 2); |
| 705 | } |
| 706 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 707 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4, 5, 5, 2, 2); |
| 708 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 709 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc5(benchmark::State& state, const char* net) { |
| 710 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5, 5, 5, 2, 2); |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 711 | } |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 712 | static void dwconv2d_chw_5x5s2p2__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 713 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2, 5, 5, 2, 2); |
| 714 | } |
| 715 | static void dwconv2d_chw_5x5s2p2__scalar_2x1_acc3(benchmark::State& state, const char* net) { |
| 716 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3, 5, 5, 2, 2); |
| 717 | } |
| 718 | static void dwconv2d_chw_5x5s2p2__scalar_3x1_acc2(benchmark::State& state, const char* net) { |
| 719 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2, 5, 5, 2, 2); |
| 720 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 721 | |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 722 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1) |
| 723 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_2x1) |
| 724 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_3x1) |
| 725 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_4x1) |
| 726 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_5x1) |
| 727 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_6x1) |
| 728 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 729 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc3) |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 730 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc4) |
| 731 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_2x1_acc2) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 732 | |
| 733 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1) |
| 734 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_2x1) |
| 735 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_3x1) |
| 736 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_4x1) |
| 737 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 738 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc3) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 739 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc4) |
| 740 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_2x1_acc2) |
| 741 | |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 742 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1) |
| 743 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1) |
| 744 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_3x1) |
| 745 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc2) |
| 746 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc3) |
| 747 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 748 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc5) |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 749 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1_acc2) |
| 750 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1_acc3) |
| 751 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_3x1_acc2) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 752 | |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 753 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1) |
| 754 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1) |
| 755 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_3x1) |
| 756 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc2) |
| 757 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc3) |
| 758 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 759 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc5) |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 760 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1_acc2) |
| 761 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1_acc3) |
| 762 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_3x1_acc2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 763 | |
| 764 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 765 | BENCHMARK_MAIN(); |
| 766 | #endif |