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 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 13 | #include <benchmark/benchmark.h> |
| 14 | #include "bench/dwconv.h" |
| 15 | #include "bench/utils.h" |
| 16 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 17 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 18 | #include <xnnpack/dwconv.h> |
| 19 | #include <xnnpack/indirection.h> |
| 20 | #include <xnnpack/operator.h> |
| 21 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 22 | #include <xnnpack/params-init.h> |
| 23 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | |
| 25 | |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 26 | static void DWConv2DBenchmark(benchmark::State& state, |
| 27 | xnn_f32_dwconv2d_chw_ukernel_function dwconv, |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 28 | uint32_t kh, uint32_t kw, uint32_t pw, uint32_t s, |
| 29 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 30 | { |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 31 | if (isa_check && !isa_check(state)) { |
| 32 | return; |
| 33 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 34 | |
| 35 | const size_t input_height = state.range(0); |
| 36 | const size_t input_width = state.range(1); |
| 37 | const size_t kernel_height = state.range(2); |
| 38 | const size_t kernel_width = state.range(3); |
| 39 | const size_t padding_height = state.range(4); |
| 40 | const size_t padding_width = state.range(5); |
| 41 | const size_t subsampling = state.range(6); |
| 42 | const size_t dilation = state.range(7); |
| 43 | const size_t channels = state.range(8); |
| 44 | |
| 45 | if (kernel_height != kh) { |
| 46 | state.SkipWithError("kernel height mismatch"); |
| 47 | return; |
| 48 | } |
| 49 | |
| 50 | if (kernel_width != kw) { |
| 51 | state.SkipWithError("kernel width mismatch"); |
| 52 | return; |
| 53 | } |
| 54 | |
| 55 | if (subsampling != s) { |
| 56 | state.SkipWithError("subsampling mismatch"); |
| 57 | return; |
| 58 | } |
| 59 | |
| 60 | if (padding_width % 2 != 0 || padding_width / 2 != pw) { |
| 61 | state.SkipWithError("padding width mismatch"); |
| 62 | return; |
| 63 | } |
| 64 | |
| 65 | if (dilation != 1) { |
| 66 | state.SkipWithError("unsupported dilation"); |
| 67 | return; |
| 68 | } |
| 69 | |
| 70 | std::random_device random_device; |
| 71 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 72 | 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] | 73 | |
| 74 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 75 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 76 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 77 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 78 | |
| 79 | const size_t inputSize = (input_height + padding_height) * input_width; |
| 80 | const size_t kernel_size = kernel_height * kernel_width; |
| 81 | const size_t output_size = output_height * output_width; |
| 82 | |
Marat Dukhan | ae7e8b2 | 2020-10-20 17:51:51 -0700 | [diff] [blame] | 83 | std::vector<float> input(inputSize * channels + 2 * XNN_EXTRA_BYTES); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 84 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 85 | std::vector<float> bias(channels); |
| 86 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 87 | std::vector<float> kernel(channels * kernel_size); |
| 88 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 89 | std::vector<float> zero(input_width + padding_width); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 90 | |
| 91 | const size_t w_elements = (kernel_size + 1) * channels; |
| 92 | const size_t o_elements = output_size * channels; |
| 93 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 94 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 95 | sizeof(float) * (w_elements + o_elements)); |
| 96 | |
| 97 | std::vector<float, AlignedAllocator<float, 32>> packed_weights(w_elements * num_buffers); |
| 98 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 99 | for (size_t c = 0; c < channels; c++) { |
| 100 | packed_weights[c * kernel_size + c] = bias[c]; |
| 101 | for (size_t i = 0; i < kernel_size; i++) { |
| 102 | packed_weights[c * kernel_size + c + 1 + i] = kernel[c * kernel_size + i]; |
| 103 | } |
| 104 | } |
| 105 | for (size_t n = 1; n < num_buffers; n++) { |
| 106 | std::copy(packed_weights.cbegin(), packed_weights.cbegin() + w_elements, packed_weights.begin() + n * w_elements); |
| 107 | } |
| 108 | |
| 109 | std::vector<float> output(o_elements * num_buffers); |
| 110 | std::fill(output.begin(), output.end(), std::nanf("")); |
| 111 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 112 | xnn_f32_chw_params chw_params = |
| 113 | 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] | 114 | |
| 115 | size_t buffer_index = 0; |
| 116 | for (auto _ : state) { |
| 117 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 118 | benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 119 | buffer_index = (buffer_index + 1) % num_buffers; |
| 120 | state.ResumeTiming(); |
| 121 | |
| 122 | for (uint32_t channel = 0; channel < channels; channel++) { |
| 123 | dwconv( |
Marat Dukhan | 7515777 | 2020-10-21 01:46:28 -0700 | [diff] [blame] | 124 | input_height, input_width * sizeof(float), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 125 | input.data() + channel * inputSize, |
| 126 | packed_weights.data() + channel * (kernel_size + 1) + buffer_index * w_elements, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 127 | zero.data(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 128 | output.data() + channel * output_size + buffer_index * o_elements, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 129 | padding_height / 2, // padding_top |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 130 | &chw_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 131 | } |
| 132 | } |
| 133 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 134 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 135 | if (cpu_frequency != 0) { |
| 136 | state.counters["cpufreq"] = cpu_frequency; |
| 137 | } |
| 138 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 139 | state.counters["FLOPS"] = benchmark::Counter( |
| 140 | uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size, |
| 141 | benchmark::Counter::kIsRate); |
| 142 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 143 | state.counters["bytes"] = benchmark::Counter( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 144 | uint64_t(state.iterations()) * (output_size + inputSize + kernel_size + 1 /* bias */) * channels * sizeof(float), |
| 145 | benchmark::Counter::kIsRate); |
| 146 | } |
| 147 | |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 148 | #if XNN_ARCH_ARM |
| 149 | static void dwconv2d_chw_3x3p1__neon_1x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 150 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 151 | } |
| 152 | static void dwconv2d_chw_3x3p1__neon_2x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 153 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 154 | } |
| 155 | static void dwconv2d_chw_3x3p1__neon_3x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 156 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 157 | } |
| 158 | static void dwconv2d_chw_3x3p1__neon_4x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 159 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 160 | } |
| 161 | static void dwconv2d_chw_3x3p1__neon_5x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 162 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 163 | } |
| 164 | static void dwconv2d_chw_3x3p1__neon_6x4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 165 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 166 | } |
| 167 | static void dwconv2d_chw_3x3p1__neon_1x4_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 168 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 169 | } |
| 170 | static void dwconv2d_chw_3x3p1__neon_1x4_acc3(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 171 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 172 | } |
| 173 | static void dwconv2d_chw_3x3p1__neon_1x4_acc4(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 174 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 175 | } |
| 176 | static void dwconv2d_chw_3x3p1__neon_2x4_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 177 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckNEON); |
| 178 | } |
| 179 | |
| 180 | static void dwconv2d_chw_3x3s2p1__neon_1x4(benchmark::State& state, const char* net) { |
| 181 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 182 | } |
| 183 | static void dwconv2d_chw_3x3s2p1__neon_2x4(benchmark::State& state, const char* net) { |
| 184 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 185 | } |
| 186 | static void dwconv2d_chw_3x3s2p1__neon_3x4(benchmark::State& state, const char* net) { |
| 187 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 188 | } |
| 189 | static void dwconv2d_chw_3x3s2p1__neon_4x4(benchmark::State& state, const char* net) { |
| 190 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 191 | } |
| 192 | static void dwconv2d_chw_3x3s2p1__neon_1x4_acc2(benchmark::State& state, const char* net) { |
| 193 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 194 | } |
| 195 | static void dwconv2d_chw_3x3s2p1__neon_1x4_acc3(benchmark::State& state, const char* net) { |
| 196 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 197 | } |
| 198 | static void dwconv2d_chw_3x3s2p1__neon_1x4_acc4(benchmark::State& state, const char* net) { |
| 199 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 200 | } |
| 201 | static void dwconv2d_chw_3x3s2p1__neon_2x4_acc2(benchmark::State& state, const char* net) { |
| 202 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2, 3, 3, 1, 2, benchmark::utils::CheckNEON); |
| 203 | } |
| 204 | |
| 205 | static void dwconv2d_chw_5x5p2__neon_1x4(benchmark::State& state, const char* net) { |
| 206 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 207 | } |
| 208 | static void dwconv2d_chw_5x5p2__neon_2x4(benchmark::State& state, const char* net) { |
| 209 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 210 | } |
| 211 | static void dwconv2d_chw_5x5p2__neon_3x4(benchmark::State& state, const char* net) { |
| 212 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 213 | } |
| 214 | static void dwconv2d_chw_5x5p2__neon_4x4(benchmark::State& state, const char* net) { |
| 215 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 216 | } |
| 217 | static void dwconv2d_chw_5x5p2__neon_5x4(benchmark::State& state, const char* net) { |
| 218 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 219 | } |
| 220 | static void dwconv2d_chw_5x5p2__neon_1x4_acc2(benchmark::State& state, const char* net) { |
| 221 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 222 | } |
| 223 | static void dwconv2d_chw_5x5p2__neon_1x4_acc3(benchmark::State& state, const char* net) { |
| 224 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 225 | } |
| 226 | static void dwconv2d_chw_5x5p2__neon_1x4_acc4(benchmark::State& state, const char* net) { |
| 227 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 228 | } |
| 229 | static void dwconv2d_chw_5x5p2__neon_1x4_acc5(benchmark::State& state, const char* net) { |
| 230 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 231 | } |
| 232 | static void dwconv2d_chw_5x5p2__neon_2x4_acc2(benchmark::State& state, const char* net) { |
| 233 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 234 | } |
| 235 | static void dwconv2d_chw_5x5p2__neon_2x4_acc3(benchmark::State& state, const char* net) { |
| 236 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 237 | } |
| 238 | static void dwconv2d_chw_5x5p2__neon_3x4_acc2(benchmark::State& state, const char* net) { |
| 239 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 240 | } |
| 241 | static void dwconv2d_chw_5x5p2__neon_4x4_acc2(benchmark::State& state, const char* net) { |
| 242 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2, 5, 5, 2, 1, benchmark::utils::CheckNEON); |
| 243 | } |
| 244 | |
| 245 | static void dwconv2d_chw_5x5s2p2__neon_1x4(benchmark::State& state, const char* net) { |
| 246 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 247 | } |
| 248 | static void dwconv2d_chw_5x5s2p2__neon_2x4(benchmark::State& state, const char* net) { |
| 249 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 250 | } |
| 251 | static void dwconv2d_chw_5x5s2p2__neon_3x4(benchmark::State& state, const char* net) { |
| 252 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 253 | } |
| 254 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc2(benchmark::State& state, const char* net) { |
| 255 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 256 | } |
| 257 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc3(benchmark::State& state, const char* net) { |
| 258 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 259 | } |
| 260 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc4(benchmark::State& state, const char* net) { |
| 261 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 262 | } |
| 263 | static void dwconv2d_chw_5x5s2p2__neon_1x4_acc5(benchmark::State& state, const char* net) { |
| 264 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 265 | } |
| 266 | static void dwconv2d_chw_5x5s2p2__neon_2x4_acc2(benchmark::State& state, const char* net) { |
| 267 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 268 | } |
| 269 | static void dwconv2d_chw_5x5s2p2__neon_2x4_acc3(benchmark::State& state, const char* net) { |
| 270 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
| 271 | } |
| 272 | static void dwconv2d_chw_5x5s2p2__neon_3x4_acc2(benchmark::State& state, const char* net) { |
| 273 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2, 5, 5, 2, 2, benchmark::utils::CheckNEON); |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 274 | } |
| 275 | |
| 276 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4) |
| 277 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_2x4) |
| 278 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_3x4) |
| 279 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_4x4) |
| 280 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_5x4) |
| 281 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_6x4) |
| 282 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc2) |
| 283 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc3) |
| 284 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_1x4_acc4) |
| 285 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neon_2x4_acc2) |
Marat Dukhan | c8bbe70 | 2020-12-07 13:01:29 -0800 | [diff] [blame] | 286 | |
| 287 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4) |
| 288 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_2x4) |
| 289 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_3x4) |
| 290 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_4x4) |
| 291 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4_acc2) |
| 292 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4_acc3) |
| 293 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_1x4_acc4) |
| 294 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neon_2x4_acc2) |
| 295 | |
| 296 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4) |
| 297 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_2x4) |
| 298 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_3x4) |
| 299 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_4x4) |
| 300 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_5x4) |
| 301 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc2) |
| 302 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc3) |
| 303 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc4) |
| 304 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_1x4_acc5) |
| 305 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_2x4_acc2) |
| 306 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_2x4_acc3) |
| 307 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_3x4_acc2) |
| 308 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neon_4x4_acc2) |
| 309 | |
| 310 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4) |
| 311 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_2x4) |
| 312 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_3x4) |
| 313 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc2) |
| 314 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc3) |
| 315 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc4) |
| 316 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_1x4_acc5) |
| 317 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_2x4_acc2) |
| 318 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_2x4_acc3) |
| 319 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neon_3x4_acc2) |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 320 | #endif // XNN_ARCH_ARM |
| 321 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 322 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 323 | static void dwconv2d_chw_3x3p1__neonfma_1x4(benchmark::State& state, const char* net) { |
| 324 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4, 3, 3, 1, 1); |
| 325 | } |
| 326 | static void dwconv2d_chw_3x3p1__neonfma_2x4(benchmark::State& state, const char* net) { |
| 327 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4, 3, 3, 1, 1); |
| 328 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 329 | static void dwconv2d_chw_3x3p1__neonfma_3x4(benchmark::State& state, const char* net) { |
| 330 | 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] | 331 | } |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 332 | static void dwconv2d_chw_3x3p1__neonfma_4x4(benchmark::State& state, const char* net) { |
| 333 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4, 3, 3, 1, 1); |
| 334 | } |
| 335 | static void dwconv2d_chw_3x3p1__neonfma_5x4(benchmark::State& state, const char* net) { |
| 336 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4, 3, 3, 1, 1); |
| 337 | } |
| 338 | static void dwconv2d_chw_3x3p1__neonfma_6x4(benchmark::State& state, const char* net) { |
| 339 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4, 3, 3, 1, 1); |
| 340 | } |
| 341 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 342 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2, 3, 3, 1, 1); |
| 343 | } |
| 344 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 345 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3, 3, 3, 1, 1); |
| 346 | } |
| 347 | static void dwconv2d_chw_3x3p1__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 348 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4, 3, 3, 1, 1); |
| 349 | } |
| 350 | static void dwconv2d_chw_3x3p1__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 351 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2, 3, 3, 1, 1); |
| 352 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 353 | |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 354 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4(benchmark::State& state, const char* net) { |
| 355 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4, 3, 3, 1, 2); |
| 356 | } |
| 357 | static void dwconv2d_chw_3x3s2p1__neonfma_2x4(benchmark::State& state, const char* net) { |
| 358 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4, 3, 3, 1, 2); |
| 359 | } |
| 360 | static void dwconv2d_chw_3x3s2p1__neonfma_3x4(benchmark::State& state, const char* net) { |
| 361 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4, 3, 3, 1, 2); |
| 362 | } |
| 363 | static void dwconv2d_chw_3x3s2p1__neonfma_4x4(benchmark::State& state, const char* net) { |
| 364 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4, 3, 3, 1, 2); |
| 365 | } |
| 366 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 367 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2, 3, 3, 1, 2); |
| 368 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 369 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 370 | 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] | 371 | } |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 372 | static void dwconv2d_chw_3x3s2p1__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 373 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4, 3, 3, 1, 2); |
| 374 | } |
| 375 | static void dwconv2d_chw_3x3s2p1__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 376 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2, 3, 3, 1, 2); |
| 377 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 378 | |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 379 | static void dwconv2d_chw_5x5p2__neonfma_1x4(benchmark::State& state, const char* net) { |
| 380 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4, 5, 5, 2, 1); |
| 381 | } |
| 382 | static void dwconv2d_chw_5x5p2__neonfma_2x4(benchmark::State& state, const char* net) { |
| 383 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4, 5, 5, 2, 1); |
| 384 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 385 | static void dwconv2d_chw_5x5p2__neonfma_3x4(benchmark::State& state, const char* net) { |
| 386 | 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] | 387 | } |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 388 | static void dwconv2d_chw_5x5p2__neonfma_4x4(benchmark::State& state, const char* net) { |
| 389 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4, 5, 5, 2, 1); |
| 390 | } |
| 391 | static void dwconv2d_chw_5x5p2__neonfma_5x4(benchmark::State& state, const char* net) { |
| 392 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4, 5, 5, 2, 1); |
| 393 | } |
| 394 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 395 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2, 5, 5, 2, 1); |
| 396 | } |
| 397 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 398 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3, 5, 5, 2, 1); |
| 399 | } |
| 400 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 401 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4, 5, 5, 2, 1); |
| 402 | } |
| 403 | static void dwconv2d_chw_5x5p2__neonfma_1x4_acc5(benchmark::State& state, const char* net) { |
| 404 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5, 5, 5, 2, 1); |
| 405 | } |
| 406 | static void dwconv2d_chw_5x5p2__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 407 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2, 5, 5, 2, 1); |
| 408 | } |
| 409 | static void dwconv2d_chw_5x5p2__neonfma_2x4_acc3(benchmark::State& state, const char* net) { |
| 410 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3, 5, 5, 2, 1); |
| 411 | } |
| 412 | static void dwconv2d_chw_5x5p2__neonfma_3x4_acc2(benchmark::State& state, const char* net) { |
| 413 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2, 5, 5, 2, 1); |
| 414 | } |
| 415 | static void dwconv2d_chw_5x5p2__neonfma_4x4_acc2(benchmark::State& state, const char* net) { |
| 416 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2, 5, 5, 2, 1); |
| 417 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 418 | |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 419 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4(benchmark::State& state, const char* net) { |
| 420 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4, 5, 5, 2, 2); |
| 421 | } |
| 422 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4(benchmark::State& state, const char* net) { |
| 423 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4, 5, 5, 2, 2); |
| 424 | } |
| 425 | static void dwconv2d_chw_5x5s2p2__neonfma_3x4(benchmark::State& state, const char* net) { |
| 426 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4, 5, 5, 2, 2); |
| 427 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 428 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc2(benchmark::State& state, const char* net) { |
| 429 | 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] | 430 | } |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 431 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc3(benchmark::State& state, const char* net) { |
| 432 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3, 5, 5, 2, 2); |
| 433 | } |
| 434 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc4(benchmark::State& state, const char* net) { |
| 435 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4, 5, 5, 2, 2); |
| 436 | } |
| 437 | static void dwconv2d_chw_5x5s2p2__neonfma_1x4_acc5(benchmark::State& state, const char* net) { |
| 438 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5, 5, 5, 2, 2); |
| 439 | } |
| 440 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4_acc2(benchmark::State& state, const char* net) { |
| 441 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2, 5, 5, 2, 2); |
| 442 | } |
| 443 | static void dwconv2d_chw_5x5s2p2__neonfma_2x4_acc3(benchmark::State& state, const char* net) { |
| 444 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3, 5, 5, 2, 2); |
| 445 | } |
| 446 | static void dwconv2d_chw_5x5s2p2__neonfma_3x4_acc2(benchmark::State& state, const char* net) { |
| 447 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2, 5, 5, 2, 2); |
| 448 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 449 | |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 450 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4) |
| 451 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_2x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 452 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_3x4) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 453 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_4x4) |
| 454 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_5x4) |
| 455 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_6x4) |
| 456 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc2) |
| 457 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc3) |
| 458 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_1x4_acc4) |
| 459 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfma_2x4_acc2) |
| 460 | |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 461 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4) |
| 462 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_2x4) |
| 463 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_3x4) |
| 464 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_4x4) |
| 465 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 466 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc3) |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 467 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_1x4_acc4) |
| 468 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfma_2x4_acc2) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 469 | |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 470 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4) |
| 471 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 472 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_3x4) |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 473 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_4x4) |
| 474 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_5x4) |
| 475 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc2) |
| 476 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc3) |
| 477 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc4) |
| 478 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_1x4_acc5) |
| 479 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4_acc2) |
| 480 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_2x4_acc3) |
| 481 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_3x4_acc2) |
| 482 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfma_4x4_acc2) |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 483 | |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 484 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4) |
| 485 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4) |
| 486 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_3x4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 487 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc2) |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 488 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc3) |
| 489 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc4) |
| 490 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_1x4_acc5) |
| 491 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4_acc2) |
| 492 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_2x4_acc3) |
| 493 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfma_3x4_acc2) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 494 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 495 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 496 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 497 | static void dwconv2d_chw_3x3p1__sse_1x4(benchmark::State& state, const char* net) { |
| 498 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4, 3, 3, 1, 1); |
| 499 | } |
| 500 | static void dwconv2d_chw_3x3p1__sse_2x4(benchmark::State& state, const char* net) { |
| 501 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4, 3, 3, 1, 1); |
| 502 | } |
| 503 | static void dwconv2d_chw_3x3p1__sse_3x4(benchmark::State& state, const char* net) { |
| 504 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4, 3, 3, 1, 1); |
| 505 | } |
| 506 | static void dwconv2d_chw_3x3p1__sse_4x4(benchmark::State& state, const char* net) { |
| 507 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4, 3, 3, 1, 1); |
| 508 | } |
| 509 | static void dwconv2d_chw_3x3p1__sse_5x4(benchmark::State& state, const char* net) { |
| 510 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4, 3, 3, 1, 1); |
| 511 | } |
| 512 | static void dwconv2d_chw_3x3p1__sse_6x4(benchmark::State& state, const char* net) { |
| 513 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4, 3, 3, 1, 1); |
| 514 | } |
| 515 | static void dwconv2d_chw_3x3p1__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 516 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2, 3, 3, 1, 1); |
| 517 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 518 | static void dwconv2d_chw_3x3p1__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 519 | 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] | 520 | } |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 521 | static void dwconv2d_chw_3x3p1__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 522 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4, 3, 3, 1, 1); |
| 523 | } |
| 524 | static void dwconv2d_chw_3x3p1__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 525 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2, 3, 3, 1, 1); |
| 526 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 527 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 528 | static void dwconv2d_chw_3x3p1__ssse3_1x4(benchmark::State& state, const char* net) { |
| 529 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 530 | } |
| 531 | static void dwconv2d_chw_3x3p1__ssse3_2x4(benchmark::State& state, const char* net) { |
| 532 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 533 | } |
| 534 | static void dwconv2d_chw_3x3p1__ssse3_3x4(benchmark::State& state, const char* net) { |
| 535 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 536 | } |
| 537 | static void dwconv2d_chw_3x3p1__ssse3_4x4(benchmark::State& state, const char* net) { |
| 538 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 539 | } |
| 540 | static void dwconv2d_chw_3x3p1__ssse3_5x4(benchmark::State& state, const char* net) { |
| 541 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 542 | } |
| 543 | static void dwconv2d_chw_3x3p1__ssse3_6x4(benchmark::State& state, const char* net) { |
| 544 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 545 | } |
| 546 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc2(benchmark::State& state, const char* net) { |
| 547 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 548 | } |
| 549 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc3(benchmark::State& state, const char* net) { |
| 550 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 551 | } |
| 552 | static void dwconv2d_chw_3x3p1__ssse3_1x4_acc4(benchmark::State& state, const char* net) { |
| 553 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 554 | } |
| 555 | static void dwconv2d_chw_3x3p1__ssse3_2x4_acc2(benchmark::State& state, const char* net) { |
| 556 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2, 3, 3, 1, 1, benchmark::utils::CheckSSSE3); |
| 557 | } |
| 558 | |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 559 | static void dwconv2d_chw_3x3s2p1__sse_1x4(benchmark::State& state, const char* net) { |
| 560 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4, 3, 3, 1, 2); |
| 561 | } |
| 562 | static void dwconv2d_chw_3x3s2p1__sse_2x4(benchmark::State& state, const char* net) { |
| 563 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4, 3, 3, 1, 2); |
| 564 | } |
| 565 | static void dwconv2d_chw_3x3s2p1__sse_3x4(benchmark::State& state, const char* net) { |
| 566 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4, 3, 3, 1, 2); |
| 567 | } |
| 568 | static void dwconv2d_chw_3x3s2p1__sse_4x4(benchmark::State& state, const char* net) { |
| 569 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4, 3, 3, 1, 2); |
| 570 | } |
| 571 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 572 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2, 3, 3, 1, 2); |
| 573 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 574 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 575 | 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] | 576 | } |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 577 | static void dwconv2d_chw_3x3s2p1__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 578 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4, 3, 3, 1, 2); |
| 579 | } |
| 580 | static void dwconv2d_chw_3x3s2p1__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 581 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2, 3, 3, 1, 2); |
| 582 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 583 | |
Marat Dukhan | d050389 | 2020-10-30 08:22:04 -0700 | [diff] [blame] | 584 | static void dwconv2d_chw_5x5p2__sse_1x4(benchmark::State& state, const char* net) { |
| 585 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4, 5, 5, 2, 1); |
| 586 | } |
| 587 | static void dwconv2d_chw_5x5p2__sse_2x4(benchmark::State& state, const char* net) { |
| 588 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4, 5, 5, 2, 1); |
| 589 | } |
| 590 | static void dwconv2d_chw_5x5p2__sse_3x4(benchmark::State& state, const char* net) { |
| 591 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4, 5, 5, 2, 1); |
| 592 | } |
| 593 | static void dwconv2d_chw_5x5p2__sse_4x4(benchmark::State& state, const char* net) { |
| 594 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4, 5, 5, 2, 1); |
| 595 | } |
| 596 | static void dwconv2d_chw_5x5p2__sse_5x4(benchmark::State& state, const char* net) { |
| 597 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4, 5, 5, 2, 1); |
| 598 | } |
| 599 | static void dwconv2d_chw_5x5p2__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 600 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2, 5, 5, 2, 1); |
| 601 | } |
| 602 | static void dwconv2d_chw_5x5p2__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 603 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3, 5, 5, 2, 1); |
| 604 | } |
| 605 | static void dwconv2d_chw_5x5p2__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 606 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4, 5, 5, 2, 1); |
| 607 | } |
| 608 | static void dwconv2d_chw_5x5p2__sse_1x4_acc5(benchmark::State& state, const char* net) { |
| 609 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5, 5, 5, 2, 1); |
| 610 | } |
| 611 | static void dwconv2d_chw_5x5p2__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 612 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2, 5, 5, 2, 1); |
| 613 | } |
| 614 | static void dwconv2d_chw_5x5p2__sse_2x4_acc3(benchmark::State& state, const char* net) { |
| 615 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3, 5, 5, 2, 1); |
| 616 | } |
| 617 | static void dwconv2d_chw_5x5p2__sse_3x4_acc2(benchmark::State& state, const char* net) { |
| 618 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2, 5, 5, 2, 1); |
| 619 | } |
| 620 | static void dwconv2d_chw_5x5p2__sse_4x4_acc2(benchmark::State& state, const char* net) { |
| 621 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2, 5, 5, 2, 1); |
| 622 | } |
| 623 | |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 624 | static void dwconv2d_chw_5x5s2p2__sse_1x4(benchmark::State& state, const char* net) { |
| 625 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4, 5, 5, 2, 2); |
| 626 | } |
| 627 | static void dwconv2d_chw_5x5s2p2__sse_2x4(benchmark::State& state, const char* net) { |
| 628 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4, 5, 5, 2, 2); |
| 629 | } |
| 630 | static void dwconv2d_chw_5x5s2p2__sse_3x4(benchmark::State& state, const char* net) { |
| 631 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4, 5, 5, 2, 2); |
| 632 | } |
| 633 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc2(benchmark::State& state, const char* net) { |
| 634 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2, 5, 5, 2, 2); |
| 635 | } |
| 636 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc3(benchmark::State& state, const char* net) { |
| 637 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3, 5, 5, 2, 2); |
| 638 | } |
| 639 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc4(benchmark::State& state, const char* net) { |
| 640 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4, 5, 5, 2, 2); |
| 641 | } |
| 642 | static void dwconv2d_chw_5x5s2p2__sse_1x4_acc5(benchmark::State& state, const char* net) { |
| 643 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5, 5, 5, 2, 2); |
| 644 | } |
| 645 | static void dwconv2d_chw_5x5s2p2__sse_2x4_acc2(benchmark::State& state, const char* net) { |
| 646 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc2, 5, 5, 2, 2); |
| 647 | } |
| 648 | static void dwconv2d_chw_5x5s2p2__sse_2x4_acc3(benchmark::State& state, const char* net) { |
| 649 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4_acc3, 5, 5, 2, 2); |
| 650 | } |
| 651 | static void dwconv2d_chw_5x5s2p2__sse_3x4_acc2(benchmark::State& state, const char* net) { |
| 652 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2, 5, 5, 2, 2); |
| 653 | } |
| 654 | |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 655 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4) |
| 656 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_2x4) |
| 657 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_3x4) |
| 658 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_4x4) |
| 659 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_5x4) |
| 660 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_6x4) |
| 661 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 662 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc3) |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 663 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_1x4_acc4) |
| 664 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__sse_2x4_acc2) |
| 665 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 666 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4) |
| 667 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_2x4) |
| 668 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_3x4) |
| 669 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_4x4) |
| 670 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_5x4) |
| 671 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_6x4) |
| 672 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc2) |
| 673 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc3) |
| 674 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_1x4_acc4) |
| 675 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__ssse3_2x4_acc2) |
| 676 | |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 677 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4) |
| 678 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_2x4) |
| 679 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_3x4) |
| 680 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_4x4) |
| 681 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 682 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc3) |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 683 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_1x4_acc4) |
| 684 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__sse_2x4_acc2) |
Marat Dukhan | d050389 | 2020-10-30 08:22:04 -0700 | [diff] [blame] | 685 | |
| 686 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4) |
| 687 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4) |
| 688 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_3x4) |
| 689 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_4x4) |
| 690 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_5x4) |
| 691 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc2) |
| 692 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc3) |
| 693 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc4) |
| 694 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_1x4_acc5) |
| 695 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4_acc2) |
| 696 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_2x4_acc3) |
| 697 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_3x4_acc2) |
| 698 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__sse_4x4_acc2) |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 699 | |
| 700 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4) |
| 701 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4) |
| 702 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_3x4) |
| 703 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc2) |
| 704 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc3) |
| 705 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc4) |
| 706 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_1x4_acc5) |
| 707 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4_acc2) |
| 708 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_2x4_acc3) |
| 709 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__sse_3x4_acc2) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 710 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 711 | |
Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame] | 712 | #if XNN_ARCH_WASMSIMD |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 713 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 714 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 715 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 716 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 717 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 718 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 719 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 720 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 721 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 722 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 723 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_4x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 724 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 725 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 726 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 727 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 728 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_6x4(benchmark::State& state, const char* net) { |
| 729 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 730 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 731 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 732 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 733 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 734 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 735 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3, 3, 3, 1, 1); |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 736 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 737 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 738 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4, 3, 3, 1, 1); |
Erich Elsen | fd7a6e3 | 2020-06-11 12:04:44 -0700 | [diff] [blame] | 739 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 740 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 741 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2, 3, 3, 1, 1); |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 742 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 743 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 744 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4, 3, 3, 1, 1); |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 745 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 746 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 747 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 748 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 749 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 750 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 751 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 752 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 753 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_4x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 754 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 755 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 756 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 757 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 758 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_6x4(benchmark::State& state, const char* net) { |
| 759 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 760 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 761 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 762 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc2, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 763 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 764 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 765 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc3, 3, 3, 1, 1); |
Frank Barchard | db5c32d | 2020-11-16 23:58:42 -0800 | [diff] [blame] | 766 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 767 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 768 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 769 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 770 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 771 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2, 3, 3, 1, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 772 | } |
| 773 | |
Frank Barchard | 02bb429 | 2020-12-15 18:25:32 -0800 | [diff] [blame] | 774 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 775 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4, 3, 3, 1, 1); |
| 776 | } |
| 777 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 778 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4, 3, 3, 1, 1); |
| 779 | } |
| 780 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 781 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4, 3, 3, 1, 1); |
| 782 | } |
| 783 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_4x4(benchmark::State& state, const char* net) { |
| 784 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4, 3, 3, 1, 1); |
| 785 | } |
| 786 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_5x4(benchmark::State& state, const char* net) { |
| 787 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4, 3, 3, 1, 1); |
| 788 | } |
| 789 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_6x4(benchmark::State& state, const char* net) { |
| 790 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4, 3, 3, 1, 1); |
| 791 | } |
| 792 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 793 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2, 3, 3, 1, 1); |
| 794 | } |
| 795 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 796 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3, 3, 3, 1, 1); |
| 797 | } |
| 798 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 799 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4, 3, 3, 1, 1); |
| 800 | } |
| 801 | static void dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 802 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2, 3, 3, 1, 1); |
| 803 | } |
| 804 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 805 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4, 3, 3, 1, 1); |
| 806 | } |
| 807 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 808 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4, 3, 3, 1, 1); |
| 809 | } |
| 810 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 811 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4, 3, 3, 1, 1); |
| 812 | } |
| 813 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_4x4(benchmark::State& state, const char* net) { |
| 814 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4, 3, 3, 1, 1); |
| 815 | } |
| 816 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_5x4(benchmark::State& state, const char* net) { |
| 817 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4, 3, 3, 1, 1); |
| 818 | } |
| 819 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_6x4(benchmark::State& state, const char* net) { |
| 820 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4, 3, 3, 1, 1); |
| 821 | } |
| 822 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 823 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2, 3, 3, 1, 1); |
| 824 | } |
| 825 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 826 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3, 3, 3, 1, 1); |
| 827 | } |
| 828 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 829 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4, 3, 3, 1, 1); |
| 830 | } |
| 831 | static void dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 832 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2, 3, 3, 1, 1); |
| 833 | } |
| 834 | |
Frank Barchard | c5704bf | 2020-12-21 23:09:00 -0800 | [diff] [blame] | 835 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 836 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4, 3, 3, 1, 2); |
| 837 | } |
| 838 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 839 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4, 3, 3, 1, 2); |
| 840 | } |
| 841 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 842 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4, 3, 3, 1, 2); |
| 843 | } |
| 844 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 845 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4, 3, 3, 1, 2); |
| 846 | } |
| 847 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 848 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2, 3, 3, 1, 2); |
| 849 | } |
| 850 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 851 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3, 3, 3, 1, 2); |
| 852 | } |
| 853 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 854 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4, 3, 3, 1, 2); |
| 855 | } |
| 856 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 857 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2, 3, 3, 1, 2); |
| 858 | } |
| 859 | |
| 860 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 861 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4, 3, 3, 1, 2); |
| 862 | } |
| 863 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 864 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4, 3, 3, 1, 2); |
| 865 | } |
| 866 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 867 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4, 3, 3, 1, 2); |
| 868 | } |
| 869 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 870 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4, 3, 3, 1, 2); |
| 871 | } |
| 872 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 873 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2, 3, 3, 1, 2); |
| 874 | } |
| 875 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 876 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3, 3, 3, 1, 2); |
| 877 | } |
| 878 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 879 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4, 3, 3, 1, 2); |
| 880 | } |
| 881 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 882 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2, 3, 3, 1, 2); |
| 883 | } |
| 884 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 885 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 886 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 887 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 888 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 889 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 890 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 891 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 892 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 893 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 894 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_4x4(benchmark::State& state, const char* net) { |
| 895 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 896 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 897 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 898 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 899 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 900 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 901 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 902 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 903 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 904 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 905 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 906 | static void dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 907 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 908 | } |
| 909 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 910 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 911 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 912 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 913 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 914 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 915 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 916 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 917 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 918 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 919 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_4x4(benchmark::State& state, const char* net) { |
| 920 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 921 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 922 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 923 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 924 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 925 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 926 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 927 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 928 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 929 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 930 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 931 | static void dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 932 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2, 3, 3, 1, 2); |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 933 | } |
| 934 | |
Frank Barchard | b20dcd6 | 2020-12-15 16:46:14 -0800 | [diff] [blame] | 935 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 936 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4, 5, 5, 2, 1); |
| 937 | } |
| 938 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 939 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4, 5, 5, 2, 1); |
| 940 | } |
| 941 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 942 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4, 5, 5, 2, 1); |
| 943 | } |
| 944 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 945 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4, 5, 5, 2, 1); |
| 946 | } |
| 947 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 948 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4, 5, 5, 2, 1); |
| 949 | } |
| 950 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 951 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2, 5, 5, 2, 1); |
| 952 | } |
| 953 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 954 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3, 5, 5, 2, 1); |
| 955 | } |
| 956 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 957 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4, 5, 5, 2, 1); |
| 958 | } |
| 959 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 960 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5, 5, 5, 2, 1); |
| 961 | } |
| 962 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 963 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2, 5, 5, 2, 1); |
| 964 | } |
| 965 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 966 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3, 5, 5, 2, 1); |
| 967 | } |
| 968 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 969 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2, 5, 5, 2, 1); |
| 970 | } |
| 971 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2(benchmark::State& state, const char* net) { |
| 972 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2, 5, 5, 2, 1); |
| 973 | } |
| 974 | |
| 975 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 976 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4, 5, 5, 2, 1); |
| 977 | } |
| 978 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 979 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4, 5, 5, 2, 1); |
| 980 | } |
| 981 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 982 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4, 5, 5, 2, 1); |
| 983 | } |
| 984 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4(benchmark::State& state, const char* net) { |
| 985 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4, 5, 5, 2, 1); |
| 986 | } |
| 987 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_5x4(benchmark::State& state, const char* net) { |
| 988 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4, 5, 5, 2, 1); |
| 989 | } |
| 990 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 991 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2, 5, 5, 2, 1); |
| 992 | } |
| 993 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 994 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3, 5, 5, 2, 1); |
| 995 | } |
| 996 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 997 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4, 5, 5, 2, 1); |
| 998 | } |
| 999 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1000 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5, 5, 5, 2, 1); |
| 1001 | } |
| 1002 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1003 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2, 5, 5, 2, 1); |
| 1004 | } |
| 1005 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1006 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc3, 5, 5, 2, 1); |
| 1007 | } |
| 1008 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1009 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2, 5, 5, 2, 1); |
| 1010 | } |
| 1011 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2(benchmark::State& state, const char* net) { |
| 1012 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2, 5, 5, 2, 1); |
| 1013 | } |
| 1014 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1015 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 1016 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1017 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1018 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 1019 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4, 5, 5, 2, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1020 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1021 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 1022 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4, 5, 5, 2, 1); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1023 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1024 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4(benchmark::State& state, const char* net) { |
| 1025 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1026 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1027 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_5x4(benchmark::State& state, const char* net) { |
| 1028 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1029 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1030 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1031 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1032 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1033 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1034 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1035 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1036 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1037 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1038 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1039 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1040 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1041 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1042 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1043 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1044 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1045 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1046 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1047 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1048 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1049 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1050 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1051 | static void dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4_acc2(benchmark::State& state, const char* net) { |
| 1052 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1053 | } |
| 1054 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1055 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 1056 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1057 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1058 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 1059 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1060 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1061 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 1062 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1063 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1064 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4(benchmark::State& state, const char* net) { |
| 1065 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1066 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1067 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_5x4(benchmark::State& state, const char* net) { |
| 1068 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1069 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1070 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1071 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1072 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1073 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1074 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1075 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1076 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1077 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1078 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1079 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1080 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1081 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1082 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1083 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1084 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1085 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1086 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1087 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1088 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1089 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1090 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1091 | static void dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4_acc2(benchmark::State& state, const char* net) { |
| 1092 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2, 5, 5, 2, 1); |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1093 | } |
| 1094 | |
Frank Barchard | c6889b3 | 2020-12-21 11:27:22 -0800 | [diff] [blame] | 1095 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 1096 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4, 5, 5, 2, 2); |
| 1097 | } |
| 1098 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 1099 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4, 5, 5, 2, 2); |
| 1100 | } |
| 1101 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 1102 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4, 5, 5, 2, 2); |
| 1103 | } |
| 1104 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1105 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2, 5, 5, 2, 2); |
| 1106 | } |
| 1107 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1108 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3, 5, 5, 2, 2); |
| 1109 | } |
| 1110 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1111 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4, 5, 5, 2, 2); |
| 1112 | } |
| 1113 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1114 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5, 5, 5, 2, 2); |
| 1115 | } |
| 1116 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1117 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2, 5, 5, 2, 2); |
| 1118 | } |
| 1119 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1120 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3, 5, 5, 2, 2); |
| 1121 | } |
| 1122 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1123 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2, 5, 5, 2, 2); |
| 1124 | } |
| 1125 | |
| 1126 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4(benchmark::State& state, const char* net) { |
| 1127 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4, 5, 5, 2, 2); |
| 1128 | } |
| 1129 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4(benchmark::State& state, const char* net) { |
| 1130 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4, 5, 5, 2, 2); |
| 1131 | } |
| 1132 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4(benchmark::State& state, const char* net) { |
| 1133 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4, 5, 5, 2, 2); |
| 1134 | } |
| 1135 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1136 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2, 5, 5, 2, 2); |
| 1137 | } |
| 1138 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1139 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3, 5, 5, 2, 2); |
| 1140 | } |
| 1141 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1142 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4, 5, 5, 2, 2); |
| 1143 | } |
| 1144 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1145 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5, 5, 5, 2, 2); |
| 1146 | } |
| 1147 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1148 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2, 5, 5, 2, 2); |
| 1149 | } |
| 1150 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1151 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3, 5, 5, 2, 2); |
| 1152 | } |
| 1153 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1154 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2, 5, 5, 2, 2); |
| 1155 | } |
| 1156 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1157 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4(benchmark::State& state, const char* net) { |
| 1158 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1159 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1160 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4(benchmark::State& state, const char* net) { |
| 1161 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_2x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1162 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1163 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4(benchmark::State& state, const char* net) { |
| 1164 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1165 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1166 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1167 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2, 5, 5, 2, 2); |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1168 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1169 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1170 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1171 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1172 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1173 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1174 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1175 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1176 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1177 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1178 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1179 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_2x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1180 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1181 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1182 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_2x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1183 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1184 | static void dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1185 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1186 | } |
| 1187 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1188 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4(benchmark::State& state, const char* net) { |
| 1189 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1190 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1191 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4(benchmark::State& state, const char* net) { |
| 1192 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1193 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1194 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4(benchmark::State& state, const char* net) { |
| 1195 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1196 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1197 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc2(benchmark::State& state, const char* net) { |
| 1198 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2, 5, 5, 2, 2); |
Frank Barchard | db5c32d | 2020-11-16 23:58:42 -0800 | [diff] [blame] | 1199 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1200 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc3(benchmark::State& state, const char* net) { |
| 1201 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1202 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1203 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc4(benchmark::State& state, const char* net) { |
| 1204 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1205 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1206 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc5(benchmark::State& state, const char* net) { |
| 1207 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1208 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1209 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc2(benchmark::State& state, const char* net) { |
| 1210 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1211 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1212 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc3(benchmark::State& state, const char* net) { |
| 1213 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc3, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1214 | } |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1215 | static void dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4_acc2(benchmark::State& state, const char* net) { |
| 1216 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2, 5, 5, 2, 2); |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1217 | } |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 1218 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1219 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4) |
| 1220 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4) |
| 1221 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_3x4) |
| 1222 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_4x4) |
| 1223 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_5x4) |
| 1224 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_6x4) |
| 1225 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1226 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1227 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1228 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2) |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1229 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1230 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4) |
| 1231 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4) |
| 1232 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_3x4) |
| 1233 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_4x4) |
| 1234 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_5x4) |
| 1235 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_6x4) |
| 1236 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1237 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1238 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1239 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2) |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 1240 | |
Frank Barchard | 02bb429 | 2020-12-15 18:25:32 -0800 | [diff] [blame] | 1241 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4) |
| 1242 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4) |
| 1243 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_3x4) |
| 1244 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_4x4) |
| 1245 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_5x4) |
| 1246 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_6x4) |
| 1247 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc2) |
| 1248 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc3) |
| 1249 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_1x4_acc4) |
| 1250 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_arm_splat_2x4_acc2) |
| 1251 | |
| 1252 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4) |
| 1253 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4) |
| 1254 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_3x4) |
| 1255 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_4x4) |
| 1256 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_5x4) |
| 1257 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_6x4) |
| 1258 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc2) |
| 1259 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc3) |
| 1260 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_1x4_acc4) |
| 1261 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__wasmsimd_x86_splat_2x4_acc2) |
| 1262 | |
Frank Barchard | c5704bf | 2020-12-21 23:09:00 -0800 | [diff] [blame] | 1263 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4) |
| 1264 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4) |
| 1265 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_3x4) |
| 1266 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_4x4) |
| 1267 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1268 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1269 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1270 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2) |
| 1271 | |
| 1272 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4) |
| 1273 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4) |
| 1274 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_3x4) |
| 1275 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_4x4) |
| 1276 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1277 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1278 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1279 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2) |
| 1280 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1281 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4) |
| 1282 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4) |
| 1283 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_3x4) |
| 1284 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_4x4) |
| 1285 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc2) |
| 1286 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc3) |
| 1287 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_1x4_acc4) |
| 1288 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_arm_splat_2x4_acc2) |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 1289 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1290 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4) |
| 1291 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4) |
| 1292 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_3x4) |
| 1293 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_4x4) |
| 1294 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc2) |
| 1295 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc3) |
| 1296 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_1x4_acc4) |
| 1297 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__wasmsimd_x86_splat_2x4_acc2) |
Frank Barchard | ff0624e | 2020-12-04 11:55:48 -0800 | [diff] [blame] | 1298 | |
Frank Barchard | b20dcd6 | 2020-12-15 16:46:14 -0800 | [diff] [blame] | 1299 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4) |
| 1300 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4) |
| 1301 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4) |
| 1302 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4) |
| 1303 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_5x4) |
| 1304 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1305 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1306 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1307 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5) |
| 1308 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2) |
| 1309 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3) |
| 1310 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2) |
| 1311 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2) |
| 1312 | |
| 1313 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4) |
| 1314 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4) |
| 1315 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4) |
| 1316 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4) |
| 1317 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_5x4) |
| 1318 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1319 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1320 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1321 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5) |
| 1322 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2) |
| 1323 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_2x4_acc3) |
| 1324 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2) |
| 1325 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2) |
| 1326 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1327 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4) |
| 1328 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4) |
| 1329 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4) |
| 1330 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4) |
| 1331 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_5x4) |
| 1332 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc2) |
| 1333 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc3) |
| 1334 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc4) |
| 1335 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_1x4_acc5) |
| 1336 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc2) |
| 1337 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_2x4_acc3) |
| 1338 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_3x4_acc2) |
| 1339 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_arm_splat_4x4_acc2) |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1340 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1341 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4) |
| 1342 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4) |
| 1343 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4) |
| 1344 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4) |
| 1345 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_5x4) |
| 1346 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc2) |
| 1347 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc3) |
| 1348 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc4) |
| 1349 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_1x4_acc5) |
| 1350 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc2) |
| 1351 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_2x4_acc3) |
| 1352 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_3x4_acc2) |
| 1353 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__wasmsimd_x86_splat_4x4_acc2) |
Frank Barchard | 20a0741 | 2020-11-30 23:30:00 -0800 | [diff] [blame] | 1354 | |
Frank Barchard | c6889b3 | 2020-12-21 11:27:22 -0800 | [diff] [blame] | 1355 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4) |
| 1356 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4) |
| 1357 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4) |
| 1358 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2) |
| 1359 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3) |
| 1360 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4) |
| 1361 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5) |
| 1362 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2) |
| 1363 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3) |
| 1364 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2) |
| 1365 | |
| 1366 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4) |
| 1367 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4) |
| 1368 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4) |
| 1369 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2) |
| 1370 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3) |
| 1371 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4) |
| 1372 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5) |
| 1373 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2) |
| 1374 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3) |
| 1375 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2) |
| 1376 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1377 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4) |
| 1378 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4) |
| 1379 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4) |
| 1380 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc2) |
| 1381 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc3) |
| 1382 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc4) |
| 1383 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_1x4_acc5) |
| 1384 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc2) |
| 1385 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_2x4_acc3) |
| 1386 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_arm_splat_3x4_acc2) |
Frank Barchard | e7223ee | 2020-12-04 19:04:01 -0800 | [diff] [blame] | 1387 | |
Frank Barchard | 412e2f4 | 2020-12-11 11:40:50 -0800 | [diff] [blame] | 1388 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4) |
| 1389 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4) |
| 1390 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4) |
| 1391 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc2) |
| 1392 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc3) |
| 1393 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc4) |
| 1394 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_1x4_acc5) |
| 1395 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc2) |
| 1396 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_2x4_acc3) |
| 1397 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__wasmsimd_x86_splat_3x4_acc2) |
Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame] | 1398 | #endif // XNN_ARCH_WASMSIMD |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 1399 | |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1400 | static void dwconv2d_chw_3x3p1__scalar_1x1(benchmark::State& state, const char* net) { |
| 1401 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1, 3, 3, 1, 1); |
| 1402 | } |
| 1403 | static void dwconv2d_chw_3x3p1__scalar_2x1(benchmark::State& state, const char* net) { |
| 1404 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1, 3, 3, 1, 1); |
| 1405 | } |
| 1406 | static void dwconv2d_chw_3x3p1__scalar_3x1(benchmark::State& state, const char* net) { |
| 1407 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1, 3, 3, 1, 1); |
| 1408 | } |
| 1409 | static void dwconv2d_chw_3x3p1__scalar_4x1(benchmark::State& state, const char* net) { |
| 1410 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1, 3, 3, 1, 1); |
| 1411 | } |
| 1412 | static void dwconv2d_chw_3x3p1__scalar_5x1(benchmark::State& state, const char* net) { |
| 1413 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1, 3, 3, 1, 1); |
| 1414 | } |
| 1415 | static void dwconv2d_chw_3x3p1__scalar_6x1(benchmark::State& state, const char* net) { |
| 1416 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1, 3, 3, 1, 1); |
| 1417 | } |
| 1418 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1419 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2, 3, 3, 1, 1); |
| 1420 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1421 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1422 | 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] | 1423 | } |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1424 | static void dwconv2d_chw_3x3p1__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1425 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4, 3, 3, 1, 1); |
| 1426 | } |
| 1427 | static void dwconv2d_chw_3x3p1__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1428 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2, 3, 3, 1, 1); |
| 1429 | } |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1430 | |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1431 | static void dwconv2d_chw_3x3s2p1__scalar_1x1(benchmark::State& state, const char* net) { |
| 1432 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1, 3, 3, 1, 2); |
| 1433 | } |
| 1434 | static void dwconv2d_chw_3x3s2p1__scalar_2x1(benchmark::State& state, const char* net) { |
| 1435 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1, 3, 3, 1, 2); |
| 1436 | } |
| 1437 | static void dwconv2d_chw_3x3s2p1__scalar_3x1(benchmark::State& state, const char* net) { |
| 1438 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1, 3, 3, 1, 2); |
| 1439 | } |
| 1440 | static void dwconv2d_chw_3x3s2p1__scalar_4x1(benchmark::State& state, const char* net) { |
| 1441 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1, 3, 3, 1, 2); |
| 1442 | } |
| 1443 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1444 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2, 3, 3, 1, 2); |
| 1445 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1446 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1447 | 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] | 1448 | } |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1449 | static void dwconv2d_chw_3x3s2p1__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1450 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4, 3, 3, 1, 2); |
| 1451 | } |
| 1452 | static void dwconv2d_chw_3x3s2p1__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1453 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2, 3, 3, 1, 2); |
| 1454 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 1455 | |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1456 | static void dwconv2d_chw_5x5p2__scalar_1x1(benchmark::State& state, const char* net) { |
| 1457 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1, 5, 5, 2, 1); |
| 1458 | } |
| 1459 | static void dwconv2d_chw_5x5p2__scalar_2x1(benchmark::State& state, const char* net) { |
| 1460 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1, 5, 5, 2, 1); |
| 1461 | } |
| 1462 | static void dwconv2d_chw_5x5p2__scalar_3x1(benchmark::State& state, const char* net) { |
| 1463 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1, 5, 5, 2, 1); |
| 1464 | } |
| 1465 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1466 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2, 5, 5, 2, 1); |
| 1467 | } |
| 1468 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1469 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3, 5, 5, 2, 1); |
| 1470 | } |
| 1471 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1472 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4, 5, 5, 2, 1); |
| 1473 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1474 | static void dwconv2d_chw_5x5p2__scalar_1x1_acc5(benchmark::State& state, const char* net) { |
| 1475 | 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] | 1476 | } |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1477 | static void dwconv2d_chw_5x5p2__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1478 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2, 5, 5, 2, 1); |
| 1479 | } |
| 1480 | static void dwconv2d_chw_5x5p2__scalar_2x1_acc3(benchmark::State& state, const char* net) { |
| 1481 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3, 5, 5, 2, 1); |
| 1482 | } |
| 1483 | static void dwconv2d_chw_5x5p2__scalar_3x1_acc2(benchmark::State& state, const char* net) { |
| 1484 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2, 5, 5, 2, 1); |
| 1485 | } |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1486 | |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1487 | static void dwconv2d_chw_5x5s2p2__scalar_1x1(benchmark::State& state, const char* net) { |
| 1488 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1, 5, 5, 2, 2); |
| 1489 | } |
| 1490 | static void dwconv2d_chw_5x5s2p2__scalar_2x1(benchmark::State& state, const char* net) { |
| 1491 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1, 5, 5, 2, 2); |
| 1492 | } |
| 1493 | static void dwconv2d_chw_5x5s2p2__scalar_3x1(benchmark::State& state, const char* net) { |
| 1494 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1, 5, 5, 2, 2); |
| 1495 | } |
| 1496 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc2(benchmark::State& state, const char* net) { |
| 1497 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2, 5, 5, 2, 2); |
| 1498 | } |
| 1499 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc3(benchmark::State& state, const char* net) { |
| 1500 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3, 5, 5, 2, 2); |
| 1501 | } |
| 1502 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc4(benchmark::State& state, const char* net) { |
| 1503 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4, 5, 5, 2, 2); |
| 1504 | } |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1505 | static void dwconv2d_chw_5x5s2p2__scalar_1x1_acc5(benchmark::State& state, const char* net) { |
| 1506 | 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] | 1507 | } |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1508 | static void dwconv2d_chw_5x5s2p2__scalar_2x1_acc2(benchmark::State& state, const char* net) { |
| 1509 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2, 5, 5, 2, 2); |
| 1510 | } |
| 1511 | static void dwconv2d_chw_5x5s2p2__scalar_2x1_acc3(benchmark::State& state, const char* net) { |
| 1512 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3, 5, 5, 2, 2); |
| 1513 | } |
| 1514 | static void dwconv2d_chw_5x5s2p2__scalar_3x1_acc2(benchmark::State& state, const char* net) { |
| 1515 | DWConv2DBenchmark(state, xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2, 5, 5, 2, 2); |
| 1516 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 1517 | |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1518 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1) |
| 1519 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_2x1) |
| 1520 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_3x1) |
| 1521 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_4x1) |
| 1522 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_5x1) |
| 1523 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_6x1) |
| 1524 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1525 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc3) |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 1526 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_1x1_acc4) |
| 1527 | BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__scalar_2x1_acc2) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1528 | |
| 1529 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1) |
| 1530 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_2x1) |
| 1531 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_3x1) |
| 1532 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_4x1) |
| 1533 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1534 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc3) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1535 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_1x1_acc4) |
| 1536 | BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__scalar_2x1_acc2) |
| 1537 | |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1538 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1) |
| 1539 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1) |
| 1540 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_3x1) |
| 1541 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc2) |
| 1542 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc3) |
| 1543 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1544 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_1x1_acc5) |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 1545 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1_acc2) |
| 1546 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_2x1_acc3) |
| 1547 | BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__scalar_3x1_acc2) |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 1548 | |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1549 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1) |
| 1550 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1) |
| 1551 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_3x1) |
| 1552 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc2) |
| 1553 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc3) |
| 1554 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc4) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1555 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_1x1_acc5) |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 1556 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1_acc2) |
| 1557 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_2x1_acc3) |
| 1558 | BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__scalar_3x1_acc2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1559 | |
| 1560 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 1561 | BENCHMARK_MAIN(); |
| 1562 | #endif |