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> |
Frank Barchard | e0601b5 | 2019-10-25 17:43:34 -0700 | [diff] [blame] | 23 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | |
| 25 | |
| 26 | static void DWConvBenchmark(benchmark::State& state, |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 27 | xnn_f32_dwconv_minmax_unipass_ukernel_function dwconv, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 28 | xnn_init_f32_minmax_params_fn init_params, |
| 29 | uint32_t channel_tile, uint32_t primary_tile, |
Marat Dukhan | c8466f5 | 2019-11-25 18:01:10 -0800 | [diff] [blame] | 30 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 31 | { |
Marat Dukhan | c8466f5 | 2019-11-25 18:01:10 -0800 | [diff] [blame] | 32 | if (isa_check && !isa_check(state)) { |
| 33 | return; |
| 34 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 35 | |
| 36 | const size_t input_height = state.range(0); |
| 37 | const size_t input_width = state.range(1); |
| 38 | const size_t kernel_height = state.range(2); |
| 39 | const size_t kernel_width = state.range(3); |
| 40 | const size_t padding_height = state.range(4); |
| 41 | const size_t padding_width = state.range(5); |
| 42 | const size_t subsampling = state.range(6); |
| 43 | const size_t dilation = state.range(7); |
| 44 | const size_t channels = state.range(8); |
| 45 | |
| 46 | const size_t kernel_size = kernel_height * kernel_width; |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 47 | if (kernel_size != primary_tile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 48 | state.SkipWithError("kernel size mismatch"); |
| 49 | return; |
| 50 | } |
| 51 | |
| 52 | std::random_device random_device; |
| 53 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 54 | 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] | 55 | |
| 56 | const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1; |
| 57 | const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1; |
| 58 | const size_t padding_left = padding_width / 2; |
| 59 | const size_t padding_top = padding_height / 2; |
| 60 | const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1; |
| 61 | const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1; |
| 62 | const size_t output_size = output_height * output_width; |
| 63 | const size_t step_width = dilation == 1 ? subsampling : kernel_width; |
Marat Dukhan | 03ff294 | 2019-12-05 09:32:26 -0800 | [diff] [blame] | 64 | const size_t step_height = kernel_size + (output_width - 1) * step_width * kernel_height; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 65 | |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 66 | const size_t c_stride = benchmark::utils::RoundUp<size_t>(channels, channel_tile); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 67 | |
Marat Dukhan | ad74a7b | 2019-12-05 06:18:39 -0800 | [diff] [blame] | 68 | std::vector<float> a(channels * input_height * input_width + XNN_EXTRA_BYTES / sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 69 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 70 | std::vector<float> k(channels * kernel_height * kernel_width); |
| 71 | std::generate(k.begin(), k.end(), std::ref(f32rng)); |
| 72 | std::vector<float> b(channels); |
| 73 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 74 | |
Marat Dukhan | ad74a7b | 2019-12-05 06:18:39 -0800 | [diff] [blame] | 75 | std::vector<float> z(channels + XNN_EXTRA_BYTES / sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 76 | |
| 77 | const size_t w_elements = (kernel_size + 1) * c_stride; |
| 78 | const size_t i_elements = output_height * step_height; |
| 79 | const size_t c_elements = output_size * channels; |
| 80 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 81 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 82 | sizeof(float) * (w_elements + c_elements) + sizeof(void*) * i_elements); |
| 83 | |
Marat Dukhan | e13e639 | 2021-07-26 22:22:35 -0700 | [diff] [blame] | 84 | std::vector<float, AlignedAllocator<float, 64>> w(w_elements * num_buffers); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 85 | std::fill(w.begin(), w.end(), 0.0f); |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 86 | xnn_pack_f32_dwconv_ghw_w(kernel_height, kernel_width, channels, channel_tile, |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 87 | k.data(), b.data(), w.data(), 0 /* extra bytes */, nullptr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 88 | for (size_t n = 1; n < num_buffers; n++) { |
| 89 | std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements); |
| 90 | } |
| 91 | |
| 92 | std::vector<const float*> i(i_elements * num_buffers); |
| 93 | xnn_operator convolution_op = { }; |
| 94 | convolution_op.indirection_buffer = reinterpret_cast<const void**>(i.data()); |
| 95 | convolution_op.input = a.data(); |
| 96 | convolution_op.input_pixel_stride = channels; |
| 97 | convolution_op.zero_buffer = z.data(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 98 | convolution_op.input_height = input_height; |
| 99 | convolution_op.input_width = input_width; |
| 100 | convolution_op.output_height = output_height; |
| 101 | convolution_op.output_width = output_width; |
| 102 | convolution_op.kernel_height = kernel_height; |
| 103 | convolution_op.kernel_width = kernel_width; |
| 104 | convolution_op.stride_height = subsampling; |
| 105 | convolution_op.stride_width = subsampling; |
| 106 | convolution_op.dilation_height = dilation; |
| 107 | convolution_op.dilation_width = dilation; |
| 108 | convolution_op.padding_top = padding_top; |
| 109 | convolution_op.padding_left = padding_left; |
| 110 | |
Marat Dukhan | c79427c | 2020-10-15 09:04:21 -0700 | [diff] [blame] | 111 | xnn_indirection_init_dwconv2d(&convolution_op, step_height, step_width, 2 /* log2(sizeof(float)) */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 112 | for (size_t n = 1; n < num_buffers; n++) { |
| 113 | std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements); |
| 114 | } |
| 115 | |
| 116 | std::vector<float> c(c_elements * num_buffers); |
| 117 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 118 | |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame] | 119 | xnn_f32_minmax_params params; |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 120 | init_params(¶ms, -std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 121 | |
| 122 | size_t buffer_index = 0; |
| 123 | for (auto _ : state) { |
| 124 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 125 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 126 | buffer_index = (buffer_index + 1) % num_buffers; |
| 127 | state.ResumeTiming(); |
| 128 | |
Marat Dukhan | c79427c | 2020-10-15 09:04:21 -0700 | [diff] [blame] | 129 | for (size_t y = 0; y < output_height; y++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 130 | dwconv(channels, output_width, |
| 131 | i.data() + buffer_index * i_elements + step_height * y, |
| 132 | w.data() + buffer_index * w_elements, |
| 133 | c.data() + buffer_index * c_elements + y * output_width * channels, |
| 134 | kernel_height * step_width * sizeof(void*), 0, |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 135 | 0, z.data(), ¶ms); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 136 | } |
| 137 | } |
| 138 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 139 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 140 | if (cpu_frequency != 0) { |
| 141 | state.counters["cpufreq"] = cpu_frequency; |
| 142 | } |
| 143 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 144 | state.counters["FLOPS"] = benchmark::Counter( |
| 145 | uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size, |
| 146 | benchmark::Counter::kIsRate); |
| 147 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 148 | state.counters["bytes"] = benchmark::Counter( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 149 | uint64_t(state.iterations()) * (output_size + input_height * input_width + kernel_size + 1 /* bias */) * channels * sizeof(float), |
| 150 | benchmark::Counter::kIsRate); |
| 151 | } |
| 152 | |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 153 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 154 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 155 | static void f32_dwconv_4x9__aarch64_neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 156 | DWConvBenchmark(state, |
| 157 | xnn_f32_dwconv_minmax_ukernel_up4x9__neon, |
| 158 | xnn_init_f32_minmax_scalar_params, |
| 159 | 4 /* channel tile */, 9 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 160 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 161 | static void f32_dwconv_4x9__aarch64_neonfma_cortex_a55(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 162 | DWConvBenchmark(state, |
| 163 | xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma, |
| 164 | xnn_init_f32_minmax_scalar_params, |
| 165 | 4 /* channel tile */, 9 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 166 | } |
| 167 | |
| 168 | BENCHMARK_DWCONV(f32_dwconv_4x9__aarch64_neonfma) |
| 169 | BENCHMARK_DWCONV(f32_dwconv_4x9__aarch64_neonfma_cortex_a55) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 170 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 171 | |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 172 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 173 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 174 | static void f32_dwconv_4x4__neon_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 175 | DWConvBenchmark(state, |
| 176 | xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2, |
| 177 | xnn_init_f32_minmax_scalar_params, |
| 178 | 4 /* channel tile */, 4 /* primary tile */, |
| 179 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 180 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 181 | static void f32_dwconv_4x4__neon(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 182 | DWConvBenchmark(state, |
| 183 | xnn_f32_dwconv_minmax_ukernel_up4x4__neon, |
| 184 | xnn_init_f32_minmax_scalar_params, |
| 185 | 4 /* channel tile */, 4 /* primary tile */, |
| 186 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 187 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 188 | static void f32_dwconv_4x4__neonfma_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 189 | DWConvBenchmark(state, |
| 190 | xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2, |
| 191 | xnn_init_f32_minmax_scalar_params, |
| 192 | 4 /* channel tile */, 4 /* primary tile */, |
| 193 | benchmark::utils::CheckNEONFMA); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 194 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 195 | static void f32_dwconv_4x4__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 196 | DWConvBenchmark(state, |
| 197 | xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma, |
| 198 | xnn_init_f32_minmax_scalar_params, |
| 199 | 4 /* channel tile */, 4 /* primary tile */, |
| 200 | benchmark::utils::CheckNEONFMA); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 201 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 202 | static void f32_dwconv_4x9__neon_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 203 | DWConvBenchmark(state, |
| 204 | xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2, |
| 205 | xnn_init_f32_minmax_scalar_params, |
| 206 | 4 /* channel tile */, 9 /* primary tile */, |
| 207 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 208 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 209 | static void f32_dwconv_4x9__neon(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 210 | DWConvBenchmark(state, |
| 211 | xnn_f32_dwconv_minmax_ukernel_up4x9__neon, |
| 212 | xnn_init_f32_minmax_scalar_params, |
| 213 | 4 /* channel tile */, 9 /* primary tile */, |
| 214 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 215 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 216 | static void f32_dwconv_4x9__neonfma_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 217 | DWConvBenchmark(state, |
| 218 | xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2, |
| 219 | xnn_init_f32_minmax_scalar_params, |
| 220 | 4 /* channel tile */, 9 /* primary tile */, |
| 221 | benchmark::utils::CheckNEONFMA); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 222 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 223 | static void f32_dwconv_4x9__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 224 | DWConvBenchmark(state, |
| 225 | xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma, |
| 226 | xnn_init_f32_minmax_scalar_params, |
| 227 | 4 /* channel tile */, 9 /* primary tile */, |
| 228 | benchmark::utils::CheckNEONFMA); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 229 | } |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 230 | static void f32_dwconv_4x25__neon_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 231 | DWConvBenchmark(state, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 232 | xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 233 | xnn_init_f32_minmax_scalar_params, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 234 | 4 /* channel tile */, 25 /* primary tile */, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 235 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 236 | } |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 237 | static void f32_dwconv_4x25__neon(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 238 | DWConvBenchmark(state, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 239 | xnn_f32_dwconv_minmax_ukernel_up4x25__neon, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 240 | xnn_init_f32_minmax_scalar_params, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 241 | 4 /* channel tile */, 25 /* primary tile */, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 242 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 243 | } |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 244 | static void f32_dwconv_4x25__neonfma_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 245 | DWConvBenchmark(state, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 246 | xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 247 | xnn_init_f32_minmax_scalar_params, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 248 | 4 /* channel tile */, 25 /* primary tile */, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 249 | benchmark::utils::CheckNEONFMA); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 250 | } |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 251 | static void f32_dwconv_4x25__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 252 | DWConvBenchmark(state, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 253 | xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 254 | xnn_init_f32_minmax_scalar_params, |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 255 | 4 /* channel tile */, 25 /* primary tile */, |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 256 | benchmark::utils::CheckNEONFMA); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 257 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 258 | static void f32_dwconv_8x4__neon_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 259 | DWConvBenchmark(state, |
| 260 | xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2, |
| 261 | xnn_init_f32_minmax_scalar_params, |
| 262 | 8 /* channel tile */, 4 /* primary tile */, |
| 263 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 264 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 265 | static void f32_dwconv_8x4__neon(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 266 | DWConvBenchmark(state, |
| 267 | xnn_f32_dwconv_minmax_ukernel_up8x4__neon, |
| 268 | xnn_init_f32_minmax_scalar_params, |
| 269 | 8 /* channel tile */, 4 /* primary tile */, |
| 270 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 271 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 272 | static void f32_dwconv_8x4__neonfma_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 273 | DWConvBenchmark(state, |
| 274 | xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2, |
| 275 | xnn_init_f32_minmax_scalar_params, |
| 276 | 8 /* channel tile */, 4 /* primary tile */, |
| 277 | benchmark::utils::CheckNEONFMA); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 278 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 279 | static void f32_dwconv_8x4__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 280 | DWConvBenchmark(state, |
| 281 | xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma, |
| 282 | xnn_init_f32_minmax_scalar_params, |
| 283 | 8 /* channel tile */, 4 /* primary tile */, |
| 284 | benchmark::utils::CheckNEONFMA); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 285 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 286 | static void f32_dwconv_8x9__neon_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 287 | DWConvBenchmark(state, |
| 288 | xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2, |
| 289 | xnn_init_f32_minmax_scalar_params, |
| 290 | 8 /* channel tile */, 9 /* primary tile */, |
| 291 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 292 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 293 | static void f32_dwconv_8x9__neon(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 294 | DWConvBenchmark(state, |
| 295 | xnn_f32_dwconv_minmax_ukernel_up8x9__neon, |
| 296 | xnn_init_f32_minmax_scalar_params, |
| 297 | 8 /* channel tile */, 9 /* primary tile */, |
| 298 | benchmark::utils::CheckNEON); |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 299 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 300 | static void f32_dwconv_8x9__neonfma_acc2(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 301 | DWConvBenchmark(state, |
| 302 | xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2, |
| 303 | xnn_init_f32_minmax_scalar_params, |
| 304 | 8 /* channel tile */, 9 /* primary tile */, |
| 305 | benchmark::utils::CheckNEONFMA); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 306 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 307 | static void f32_dwconv_8x9__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 308 | DWConvBenchmark(state, |
| 309 | xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma, |
| 310 | xnn_init_f32_minmax_scalar_params, |
| 311 | 8 /* channel tile */, 9 /* primary tile */, |
| 312 | benchmark::utils::CheckNEONFMA); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 313 | } |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 314 | static void f32_dwconv_8x25__neon_acc2(benchmark::State& state, const char* net) { |
| 315 | DWConvBenchmark(state, |
| 316 | xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2, |
| 317 | xnn_init_f32_minmax_scalar_params, |
| 318 | 8 /* channel tile */, 25 /* primary tile */, |
| 319 | benchmark::utils::CheckNEON); |
| 320 | } |
| 321 | static void f32_dwconv_8x25__neon(benchmark::State& state, const char* net) { |
| 322 | DWConvBenchmark(state, |
| 323 | xnn_f32_dwconv_minmax_ukernel_up8x25__neon, |
| 324 | xnn_init_f32_minmax_scalar_params, |
| 325 | 8 /* channel tile */, 25 /* primary tile */, |
| 326 | benchmark::utils::CheckNEON); |
| 327 | } |
| 328 | static void f32_dwconv_8x25__neonfma_acc2(benchmark::State& state, const char* net) { |
| 329 | DWConvBenchmark(state, |
| 330 | xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2, |
| 331 | xnn_init_f32_minmax_scalar_params, |
| 332 | 8 /* channel tile */, 25 /* primary tile */, |
| 333 | benchmark::utils::CheckNEONFMA); |
| 334 | } |
| 335 | static void f32_dwconv_8x25__neonfma(benchmark::State& state, const char* net) { |
| 336 | DWConvBenchmark(state, |
| 337 | xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma, |
| 338 | xnn_init_f32_minmax_scalar_params, |
| 339 | 8 /* channel tile */, 25 /* primary tile */, |
| 340 | benchmark::utils::CheckNEONFMA); |
| 341 | } |
| 342 | static void f32_dwconv_16x4__neon_acc2(benchmark::State& state, const char* net) { |
| 343 | DWConvBenchmark(state, |
| 344 | xnn_f32_dwconv_minmax_ukernel_up16x4__neon_acc2, |
| 345 | xnn_init_f32_minmax_scalar_params, |
| 346 | 16 /* channel tile */, 4 /* primary tile */, |
| 347 | benchmark::utils::CheckNEON); |
| 348 | } |
| 349 | static void f32_dwconv_16x4__neon(benchmark::State& state, const char* net) { |
| 350 | DWConvBenchmark(state, |
| 351 | xnn_f32_dwconv_minmax_ukernel_up16x4__neon, |
| 352 | xnn_init_f32_minmax_scalar_params, |
| 353 | 16 /* channel tile */, 4 /* primary tile */, |
| 354 | benchmark::utils::CheckNEON); |
| 355 | } |
| 356 | static void f32_dwconv_16x4__neonfma_acc2(benchmark::State& state, const char* net) { |
| 357 | DWConvBenchmark(state, |
| 358 | xnn_f32_dwconv_minmax_ukernel_up16x4__neonfma_acc2, |
| 359 | xnn_init_f32_minmax_scalar_params, |
| 360 | 16 /* channel tile */, 4 /* primary tile */, |
| 361 | benchmark::utils::CheckNEONFMA); |
| 362 | } |
| 363 | static void f32_dwconv_16x4__neonfma(benchmark::State& state, const char* net) { |
| 364 | DWConvBenchmark(state, |
| 365 | xnn_f32_dwconv_minmax_ukernel_up16x4__neonfma, |
| 366 | xnn_init_f32_minmax_scalar_params, |
| 367 | 16 /* channel tile */, 4 /* primary tile */, |
| 368 | benchmark::utils::CheckNEONFMA); |
| 369 | } |
| 370 | static void f32_dwconv_16x9__neon_acc2(benchmark::State& state, const char* net) { |
| 371 | DWConvBenchmark(state, |
| 372 | xnn_f32_dwconv_minmax_ukernel_up16x9__neon_acc2, |
| 373 | xnn_init_f32_minmax_scalar_params, |
| 374 | 16 /* channel tile */, 9 /* primary tile */, |
| 375 | benchmark::utils::CheckNEON); |
| 376 | } |
| 377 | static void f32_dwconv_16x9__neon(benchmark::State& state, const char* net) { |
| 378 | DWConvBenchmark(state, |
| 379 | xnn_f32_dwconv_minmax_ukernel_up16x9__neon, |
| 380 | xnn_init_f32_minmax_scalar_params, |
| 381 | 16 /* channel tile */, 9 /* primary tile */, |
| 382 | benchmark::utils::CheckNEON); |
| 383 | } |
| 384 | static void f32_dwconv_16x9__neonfma_acc2(benchmark::State& state, const char* net) { |
| 385 | DWConvBenchmark(state, |
| 386 | xnn_f32_dwconv_minmax_ukernel_up16x9__neonfma_acc2, |
| 387 | xnn_init_f32_minmax_scalar_params, |
| 388 | 16 /* channel tile */, 9 /* primary tile */, |
| 389 | benchmark::utils::CheckNEONFMA); |
| 390 | } |
| 391 | static void f32_dwconv_16x9__neonfma(benchmark::State& state, const char* net) { |
| 392 | DWConvBenchmark(state, |
| 393 | xnn_f32_dwconv_minmax_ukernel_up16x9__neonfma, |
| 394 | xnn_init_f32_minmax_scalar_params, |
| 395 | 16 /* channel tile */, 9 /* primary tile */, |
| 396 | benchmark::utils::CheckNEONFMA); |
| 397 | } |
| 398 | static void f32_dwconv_16x25__neon_acc2(benchmark::State& state, const char* net) { |
| 399 | DWConvBenchmark(state, |
| 400 | xnn_f32_dwconv_minmax_ukernel_up16x25__neon_acc2, |
| 401 | xnn_init_f32_minmax_scalar_params, |
| 402 | 16 /* channel tile */, 25 /* primary tile */, |
| 403 | benchmark::utils::CheckNEON); |
| 404 | } |
| 405 | static void f32_dwconv_16x25__neon(benchmark::State& state, const char* net) { |
| 406 | DWConvBenchmark(state, |
| 407 | xnn_f32_dwconv_minmax_ukernel_up16x25__neon, |
| 408 | xnn_init_f32_minmax_scalar_params, |
| 409 | 16 /* channel tile */, 25 /* primary tile */, |
| 410 | benchmark::utils::CheckNEON); |
| 411 | } |
| 412 | static void f32_dwconv_16x25__neonfma_acc2(benchmark::State& state, const char* net) { |
| 413 | DWConvBenchmark(state, |
| 414 | xnn_f32_dwconv_minmax_ukernel_up16x25__neonfma_acc2, |
| 415 | xnn_init_f32_minmax_scalar_params, |
| 416 | 16 /* channel tile */, 25 /* primary tile */, |
| 417 | benchmark::utils::CheckNEONFMA); |
| 418 | } |
| 419 | static void f32_dwconv_16x25__neonfma(benchmark::State& state, const char* net) { |
| 420 | DWConvBenchmark(state, |
| 421 | xnn_f32_dwconv_minmax_ukernel_up16x25__neonfma, |
| 422 | xnn_init_f32_minmax_scalar_params, |
| 423 | 16 /* channel tile */, 25 /* primary tile */, |
| 424 | benchmark::utils::CheckNEONFMA); |
| 425 | } |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 426 | BENCHMARK_DWCONV(f32_dwconv_4x4__neonfma) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 427 | BENCHMARK_DWCONV(f32_dwconv_4x4__neonfma_acc2) |
Frank Barchard | 2b95ea1 | 2020-06-01 11:58:43 -0700 | [diff] [blame] | 428 | BENCHMARK_DWCONV(f32_dwconv_8x4__neonfma) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 429 | BENCHMARK_DWCONV(f32_dwconv_8x4__neonfma_acc2) |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 430 | BENCHMARK_DWCONV(f32_dwconv_16x4__neonfma) |
| 431 | BENCHMARK_DWCONV(f32_dwconv_16x4__neonfma_acc2) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 432 | |
| 433 | BENCHMARK_DWCONV(f32_dwconv_4x9__neonfma) |
| 434 | BENCHMARK_DWCONV(f32_dwconv_4x9__neonfma_acc2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 435 | BENCHMARK_DWCONV(f32_dwconv_8x9__neonfma) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 436 | BENCHMARK_DWCONV(f32_dwconv_8x9__neonfma_acc2) |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 437 | BENCHMARK_DWCONV(f32_dwconv_16x9__neonfma) |
| 438 | BENCHMARK_DWCONV(f32_dwconv_16x9__neonfma_acc2) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 439 | |
| 440 | BENCHMARK_DWCONV(f32_dwconv_4x25__neonfma) |
| 441 | BENCHMARK_DWCONV(f32_dwconv_4x25__neonfma_acc2) |
| 442 | BENCHMARK_DWCONV(f32_dwconv_8x25__neonfma) |
| 443 | BENCHMARK_DWCONV(f32_dwconv_8x25__neonfma_acc2) |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 444 | BENCHMARK_DWCONV(f32_dwconv_16x25__neonfma) |
| 445 | BENCHMARK_DWCONV(f32_dwconv_16x25__neonfma_acc2) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 446 | |
| 447 | BENCHMARK_DWCONV(f32_dwconv_4x4__neon) |
| 448 | BENCHMARK_DWCONV(f32_dwconv_4x4__neon_acc2) |
| 449 | BENCHMARK_DWCONV(f32_dwconv_8x4__neon) |
| 450 | BENCHMARK_DWCONV(f32_dwconv_8x4__neon_acc2) |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 451 | BENCHMARK_DWCONV(f32_dwconv_16x4__neon) |
| 452 | BENCHMARK_DWCONV(f32_dwconv_16x4__neon_acc2) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 453 | |
| 454 | BENCHMARK_DWCONV(f32_dwconv_4x9__neon) |
| 455 | BENCHMARK_DWCONV(f32_dwconv_4x9__neon_acc2) |
| 456 | BENCHMARK_DWCONV(f32_dwconv_8x9__neon) |
| 457 | BENCHMARK_DWCONV(f32_dwconv_8x9__neon_acc2) |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 458 | BENCHMARK_DWCONV(f32_dwconv_16x9__neon) |
| 459 | BENCHMARK_DWCONV(f32_dwconv_16x9__neon_acc2) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 460 | |
| 461 | BENCHMARK_DWCONV(f32_dwconv_4x25__neon) |
| 462 | BENCHMARK_DWCONV(f32_dwconv_4x25__neon_acc2) |
| 463 | BENCHMARK_DWCONV(f32_dwconv_8x25__neon) |
| 464 | BENCHMARK_DWCONV(f32_dwconv_8x25__neon_acc2) |
Frank Barchard | c9f9d67 | 2021-10-18 12:51:59 -0700 | [diff] [blame] | 465 | BENCHMARK_DWCONV(f32_dwconv_16x25__neon) |
| 466 | BENCHMARK_DWCONV(f32_dwconv_16x25__neon_acc2) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 467 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 468 | |
| 469 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 470 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 471 | static void f32_dwconv_4x4__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 472 | DWConvBenchmark(state, |
| 473 | xnn_f32_dwconv_minmax_ukernel_up4x4__sse, |
| 474 | xnn_init_f32_minmax_sse_params, |
| 475 | 4 /* channel tile */, 4 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 476 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 477 | static void f32_dwconv_4x9__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 478 | DWConvBenchmark(state, |
| 479 | xnn_f32_dwconv_minmax_ukernel_up4x9__sse, |
| 480 | xnn_init_f32_minmax_sse_params, |
| 481 | 4 /* channel tile */, 9 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 482 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 483 | static void f32_dwconv_4x25__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 484 | DWConvBenchmark(state, |
| 485 | xnn_f32_dwconv_minmax_ukernel_up4x25__sse, |
| 486 | xnn_init_f32_minmax_sse_params, |
| 487 | 4 /* channel tile */, 25 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 488 | } |
| 489 | |
| 490 | BENCHMARK_DWCONV(f32_dwconv_4x4__sse) |
| 491 | BENCHMARK_DWCONV(f32_dwconv_4x9__sse) |
| 492 | BENCHMARK_DWCONV(f32_dwconv_4x25__sse) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 493 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 494 | |
| 495 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 496 | static void f32_dwconv_1x4__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 497 | DWConvBenchmark(state, |
| 498 | xnn_f32_dwconv_minmax_ukernel_up1x4__scalar, |
| 499 | xnn_init_f32_minmax_scalar_params, |
| 500 | 1 /* channel tile */, 4 /* primary tile */); |
| 501 | } |
| 502 | static void f32_dwconv_1x4__scalar_acc2(benchmark::State& state, const char* net) { |
| 503 | DWConvBenchmark(state, |
| 504 | xnn_f32_dwconv_minmax_ukernel_up1x4__scalar_acc2, |
| 505 | xnn_init_f32_minmax_scalar_params, |
| 506 | 1 /* channel tile */, 4 /* primary tile */); |
| 507 | } |
| 508 | static void f32_dwconv_2x4__scalar(benchmark::State& state, const char* net) { |
| 509 | DWConvBenchmark(state, |
| 510 | xnn_f32_dwconv_minmax_ukernel_up2x4__scalar, |
| 511 | xnn_init_f32_minmax_scalar_params, |
| 512 | 2 /* channel tile */, 4 /* primary tile */); |
| 513 | } |
| 514 | static void f32_dwconv_2x4__scalar_acc2(benchmark::State& state, const char* net) { |
| 515 | DWConvBenchmark(state, |
| 516 | xnn_f32_dwconv_minmax_ukernel_up2x4__scalar_acc2, |
| 517 | xnn_init_f32_minmax_scalar_params, |
| 518 | 2 /* channel tile */, 4 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 519 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 520 | static void f32_dwconv_1x9__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 521 | DWConvBenchmark(state, |
| 522 | xnn_f32_dwconv_minmax_ukernel_up1x9__scalar, |
| 523 | xnn_init_f32_minmax_scalar_params, |
| 524 | 1 /* channel tile */, 9 /* primary tile */); |
| 525 | } |
| 526 | static void f32_dwconv_1x9__scalar_acc2(benchmark::State& state, const char* net) { |
| 527 | DWConvBenchmark(state, |
| 528 | xnn_f32_dwconv_minmax_ukernel_up1x9__scalar_acc2, |
| 529 | xnn_init_f32_minmax_scalar_params, |
| 530 | 1 /* channel tile */, 9 /* primary tile */); |
| 531 | } |
| 532 | static void f32_dwconv_2x9__scalar(benchmark::State& state, const char* net) { |
| 533 | DWConvBenchmark(state, |
| 534 | xnn_f32_dwconv_minmax_ukernel_up2x9__scalar, |
| 535 | xnn_init_f32_minmax_scalar_params, |
| 536 | 2 /* channel tile */, 9 /* primary tile */); |
| 537 | } |
| 538 | static void f32_dwconv_2x9__scalar_acc2(benchmark::State& state, const char* net) { |
| 539 | DWConvBenchmark(state, |
| 540 | xnn_f32_dwconv_minmax_ukernel_up2x9__scalar_acc2, |
| 541 | xnn_init_f32_minmax_scalar_params, |
| 542 | 2 /* channel tile */, 9 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 543 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 544 | static void f32_dwconv_1x25__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 545 | DWConvBenchmark(state, |
| 546 | xnn_f32_dwconv_minmax_ukernel_up1x25__scalar, |
| 547 | xnn_init_f32_minmax_scalar_params, |
| 548 | 1 /* channel tile */, 25 /* primary tile */); |
| 549 | } |
| 550 | static void f32_dwconv_1x25__scalar_acc2(benchmark::State& state, const char* net) { |
| 551 | DWConvBenchmark(state, |
| 552 | xnn_f32_dwconv_minmax_ukernel_up1x25__scalar_acc2, |
| 553 | xnn_init_f32_minmax_scalar_params, |
| 554 | 1 /* channel tile */, 25 /* primary tile */); |
| 555 | } |
| 556 | static void f32_dwconv_2x25__scalar(benchmark::State& state, const char* net) { |
| 557 | DWConvBenchmark(state, |
| 558 | xnn_f32_dwconv_minmax_ukernel_up1x25__scalar, |
| 559 | xnn_init_f32_minmax_scalar_params, |
| 560 | 2 /* channel tile */, 25 /* primary tile */); |
| 561 | } |
| 562 | static void f32_dwconv_2x25__scalar_acc2(benchmark::State& state, const char* net) { |
| 563 | DWConvBenchmark(state, |
| 564 | xnn_f32_dwconv_minmax_ukernel_up1x25__scalar_acc2, |
| 565 | xnn_init_f32_minmax_scalar_params, |
| 566 | 2 /* channel tile */, 25 /* primary tile */); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 567 | } |
| 568 | |
| 569 | BENCHMARK_DWCONV(f32_dwconv_1x4__scalar) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 570 | BENCHMARK_DWCONV(f32_dwconv_1x4__scalar_acc2) |
| 571 | BENCHMARK_DWCONV(f32_dwconv_2x4__scalar) |
| 572 | BENCHMARK_DWCONV(f32_dwconv_2x4__scalar_acc2) |
| 573 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 574 | BENCHMARK_DWCONV(f32_dwconv_1x9__scalar) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 575 | BENCHMARK_DWCONV(f32_dwconv_1x9__scalar_acc2) |
| 576 | BENCHMARK_DWCONV(f32_dwconv_2x9__scalar) |
| 577 | BENCHMARK_DWCONV(f32_dwconv_2x9__scalar_acc2) |
| 578 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 579 | BENCHMARK_DWCONV(f32_dwconv_1x25__scalar) |
Marat Dukhan | 42b441b | 2021-07-26 17:40:52 -0700 | [diff] [blame] | 580 | BENCHMARK_DWCONV(f32_dwconv_1x25__scalar_acc2) |
| 581 | BENCHMARK_DWCONV(f32_dwconv_2x25__scalar) |
| 582 | BENCHMARK_DWCONV(f32_dwconv_2x25__scalar_acc2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 583 | |
Marat Dukhan | 104ae5e | 2021-05-24 13:41:57 -0700 | [diff] [blame] | 584 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 585 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 586 | BENCHMARK_MAIN(); |
| 587 | #endif |