Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1 | // Copyright 2021 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 <cmath> |
| 8 | #include <functional> |
| 9 | #include <random> |
| 10 | #include <vector> |
| 11 | |
| 12 | #include <xnnpack.h> |
| 13 | |
| 14 | #include <benchmark/benchmark.h> |
| 15 | |
| 16 | #include "bench/end2end.h" |
| 17 | #include "bench/utils.h" |
| 18 | #include "models/models.h" |
| 19 | #include <xnnpack/gemm.h> |
| 20 | #include <xnnpack/igemm.h> |
| 21 | #include <xnnpack/params.h> |
| 22 | #include <xnnpack/params-init.h> |
| 23 | |
| 24 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 25 | static void GEMMEnd2EndBenchmark( |
| 26 | benchmark::State& state, |
| 27 | models::ExecutionPlanFactory model_factory, |
| 28 | xnn_qu8_gemm_minmax_ukernel_function gemm, |
| 29 | xnn_qu8_igemm_minmax_ukernel_function igemm, |
| 30 | xnn_qu8_gemm_minmax_ukernel_function gemm1, |
| 31 | xnn_qu8_igemm_minmax_ukernel_function igemm1, |
| 32 | xnn_init_qu8_conv_minmax_params_fn init_params, |
| 33 | uint8_t mr, uint8_t nr, uint8_t log2_kr = 0, uint8_t log2_sr = 0, |
| 34 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 35 | { |
| 36 | if (isa_check && !isa_check(state)) { |
| 37 | return; |
| 38 | } |
| 39 | if (xnn_initialize(nullptr /* allocator */) != xnn_status_success) { |
| 40 | state.SkipWithError("failed to initialize XNNPACK"); |
| 41 | return; |
| 42 | } |
| 43 | |
| 44 | // Override microkernels chosen in xnn_initialize |
| 45 | // Note: do not directly assign to xnn_params.qu8.gemm because it breaks older gcc. |
| 46 | xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel(xnn_gemm_ukernel_function(gemm)); |
| 47 | xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel(xnn_igemm_ukernel_function(igemm)); |
| 48 | xnn_params.qu8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel(xnn_gemm_ukernel_function(gemm1)); |
| 49 | xnn_params.qu8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel(xnn_igemm_ukernel_function(igemm1)); |
| 50 | xnn_params.qu8.gemm.init.qu8 = init_params; |
| 51 | xnn_params.qu8.gemm.mr = mr; |
| 52 | xnn_params.qu8.gemm.nr = nr; |
| 53 | xnn_params.qu8.gemm.log2_kr = log2_kr; |
| 54 | xnn_params.qu8.gemm.log2_sr = log2_sr; |
| 55 | |
| 56 | auto execution_plan = model_factory(nullptr); |
| 57 | if (execution_plan.empty()) { |
| 58 | state.SkipWithError("failed to create a model"); |
| 59 | return; |
| 60 | } |
| 61 | |
| 62 | for (auto _ : state) { |
| 63 | for (const std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>& op : execution_plan) { |
| 64 | xnn_status status = xnn_run_operator(op.get(), nullptr); |
| 65 | if (status != xnn_status_success) { |
| 66 | state.SkipWithError("failed to run a model"); |
| 67 | return; |
| 68 | } |
| 69 | } |
| 70 | } |
| 71 | |
| 72 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 73 | if (cpu_frequency != 0) { |
| 74 | state.counters["cpufreq"] = cpu_frequency; |
| 75 | } |
| 76 | } |
| 77 | |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 78 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
Frank Barchard | f82410d | 2022-02-02 01:01:38 -0800 | [diff] [blame] | 79 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 80 | GEMMEnd2EndBenchmark(state, model, |
| 81 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, |
| 82 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, |
| 83 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 84 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 85 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 86 | 4 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 87 | benchmark::utils::CheckNEON); |
| 88 | } |
| 89 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 90 | GEMMEnd2EndBenchmark(state, model, |
| 91 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, |
| 92 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, |
| 93 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 94 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 95 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 96 | 4 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 97 | benchmark::utils::CheckNEON); |
| 98 | } |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 99 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 100 | GEMMEnd2EndBenchmark(state, model, |
| 101 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, |
| 102 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, |
| 103 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 104 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 105 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 106 | 4 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 107 | benchmark::utils::CheckNEON); |
| 108 | } |
| 109 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 110 | GEMMEnd2EndBenchmark(state, model, |
| 111 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, |
| 112 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, |
| 113 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 114 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 115 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 116 | 4 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 117 | benchmark::utils::CheckNEON); |
| 118 | } |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 119 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 120 | GEMMEnd2EndBenchmark(state, model, |
| 121 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, |
| 122 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, |
| 123 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 124 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 125 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 126 | 4 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 127 | benchmark::utils::CheckNEON); |
| 128 | } |
| 129 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 130 | GEMMEnd2EndBenchmark(state, model, |
| 131 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, |
| 132 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, |
| 133 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 134 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 135 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 136 | 4 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 137 | benchmark::utils::CheckNEON); |
| 138 | } |
Frank Barchard | f82410d | 2022-02-02 01:01:38 -0800 | [diff] [blame] | 139 | BENCHMARK_QU8_END2END(qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53) |
| 140 | BENCHMARK_QU8_END2END(qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53) |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 141 | BENCHMARK_QU8_END2END(qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7) |
| 142 | BENCHMARK_QU8_END2END(qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7) |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 143 | BENCHMARK_QU8_END2END(qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
| 144 | BENCHMARK_QU8_END2END(qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64) |
| 145 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 146 | |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 147 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | efc3ccf | 2021-08-31 23:20:00 -0700 | [diff] [blame] | 148 | static void qu8_gemm_4x16c4__aarch64_neondot_cortex_a55(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 149 | GEMMEnd2EndBenchmark(state, model, |
| 150 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, |
| 151 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, |
| 152 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 153 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 154 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 155 | 4 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 156 | benchmark::utils::CheckNEONDOT); |
| 157 | } |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 158 | static void qu8_gemm_4x16c4__aarch64_neondot_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 159 | GEMMEnd2EndBenchmark(state, model, |
| 160 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, |
| 161 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, |
| 162 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 163 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 164 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 165 | 4 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 166 | benchmark::utils::CheckNEONDOT); |
| 167 | } |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 168 | static void qu8_gemm_4x8c4__aarch64_neondot_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 169 | GEMMEnd2EndBenchmark(state, model, |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 170 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, |
| 171 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 172 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 173 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 174 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 175 | 4 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 176 | benchmark::utils::CheckNEONDOT); |
| 177 | } |
Frank Barchard | ca4c68e | 2021-08-25 19:06:40 -0700 | [diff] [blame] | 178 | static void qu8_gemm_4x8c4__aarch64_neondot_cortex_a55(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 179 | GEMMEnd2EndBenchmark(state, model, |
| 180 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_cortex_a55, |
| 181 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_cortex_a55, |
| 182 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 183 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 184 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 185 | 4 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 186 | benchmark::utils::CheckNEONDOT); |
| 187 | } |
Frank Barchard | 2366290 | 2021-09-13 15:07:13 -0700 | [diff] [blame] | 188 | |
| 189 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a75(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 190 | GEMMEnd2EndBenchmark(state, model, |
| 191 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, |
| 192 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, |
| 193 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 194 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 195 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 196 | 4 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 197 | benchmark::utils::CheckNEON); |
| 198 | } |
| 199 | |
| 200 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 201 | GEMMEnd2EndBenchmark(state, model, |
| 202 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75, |
| 203 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75, |
| 204 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 205 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 206 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 207 | 4 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 208 | benchmark::utils::CheckNEON); |
| 209 | } |
| 210 | |
| 211 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 212 | GEMMEnd2EndBenchmark(state, model, |
| 213 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, |
| 214 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, |
| 215 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 216 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 217 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 218 | 4 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 219 | benchmark::utils::CheckNEON); |
| 220 | } |
| 221 | |
| 222 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 223 | GEMMEnd2EndBenchmark(state, model, |
| 224 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, |
| 225 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, |
| 226 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 227 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 228 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 229 | 4 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 230 | benchmark::utils::CheckNEON); |
| 231 | } |
Frank Barchard | 9cdc10d | 2021-11-22 19:03:54 -0800 | [diff] [blame] | 232 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 233 | GEMMEnd2EndBenchmark(state, model, |
| 234 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64, |
| 235 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64, |
| 236 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 237 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 238 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 239 | 4 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 240 | benchmark::utils::CheckNEON); |
| 241 | } |
Frank Barchard | 2366290 | 2021-09-13 15:07:13 -0700 | [diff] [blame] | 242 | |
Frank Barchard | 9cdc10d | 2021-11-22 19:03:54 -0800 | [diff] [blame] | 243 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 244 | GEMMEnd2EndBenchmark(state, model, |
| 245 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, |
| 246 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, |
| 247 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 248 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 249 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 250 | 4 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 251 | benchmark::utils::CheckNEON); |
| 252 | } |
Frank Barchard | ca4c68e | 2021-08-25 19:06:40 -0700 | [diff] [blame] | 253 | BENCHMARK_QU8_END2END(qu8_gemm_4x8c4__aarch64_neondot_cortex_a55); |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 254 | BENCHMARK_QU8_END2END(qu8_gemm_4x16c4__aarch64_neondot_cortex_a55); |
| 255 | BENCHMARK_QU8_END2END(qu8_gemm_4x8c4__aarch64_neondot_ld128); |
| 256 | BENCHMARK_QU8_END2END(qu8_gemm_4x16c4__aarch64_neondot_ld128); |
Frank Barchard | 2366290 | 2021-09-13 15:07:13 -0700 | [diff] [blame] | 257 | BENCHMARK_QU8_END2END(qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a75); |
| 258 | BENCHMARK_QU8_END2END(qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75); |
| 259 | BENCHMARK_QU8_END2END(qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53); |
| 260 | BENCHMARK_QU8_END2END(qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53); |
Frank Barchard | 9cdc10d | 2021-11-22 19:03:54 -0800 | [diff] [blame] | 261 | BENCHMARK_QU8_END2END(qu8_gemm_4x16__aarch64_neon_mlal_lane_ld64); |
| 262 | BENCHMARK_QU8_END2END(qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 263 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 264 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 265 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 266 | static void qu8_gemm_2x8__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 267 | GEMMEnd2EndBenchmark(state, model, |
| 268 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, |
| 269 | xnn_qu8_igemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, |
| 270 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 271 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 272 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | a30e2df | 2022-01-10 13:44:16 -0800 | [diff] [blame] | 273 | 2 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 274 | benchmark::utils::CheckNEON); |
| 275 | } |
| 276 | |
| 277 | static void qu8_gemm_3x8__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 278 | GEMMEnd2EndBenchmark(state, model, |
| 279 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, |
| 280 | xnn_qu8_igemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, |
| 281 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 282 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 283 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | a30e2df | 2022-01-10 13:44:16 -0800 | [diff] [blame] | 284 | 3 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 285 | benchmark::utils::CheckNEON); |
| 286 | } |
| 287 | |
Frank Barchard | 2df7542 | 2021-09-08 19:02:40 -0700 | [diff] [blame] | 288 | static void qu8_gemm_4x8__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 289 | GEMMEnd2EndBenchmark(state, model, |
| 290 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane, |
| 291 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane, |
| 292 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 293 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 294 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 295 | 4 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 296 | benchmark::utils::CheckNEON); |
| 297 | } |
| 298 | |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 299 | static void qu8_gemm_6x8__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 300 | GEMMEnd2EndBenchmark(state, model, |
| 301 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, |
| 302 | xnn_qu8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, |
| 303 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 304 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 305 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | a30e2df | 2022-01-10 13:44:16 -0800 | [diff] [blame] | 306 | 6 /* mr */, 8 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 307 | benchmark::utils::CheckNEON); |
| 308 | } |
| 309 | |
| 310 | static void qu8_gemm_2x16__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 311 | GEMMEnd2EndBenchmark(state, model, |
| 312 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, |
| 313 | xnn_qu8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, |
| 314 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 315 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 316 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | a30e2df | 2022-01-10 13:44:16 -0800 | [diff] [blame] | 317 | 2 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 318 | benchmark::utils::CheckNEON); |
| 319 | } |
| 320 | |
| 321 | static void qu8_gemm_3x16__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 322 | GEMMEnd2EndBenchmark(state, model, |
| 323 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, |
| 324 | xnn_qu8_igemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, |
| 325 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 326 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 327 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | a30e2df | 2022-01-10 13:44:16 -0800 | [diff] [blame] | 328 | 3 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 329 | benchmark::utils::CheckNEON); |
| 330 | } |
| 331 | |
Frank Barchard | a29b57e | 2021-08-20 19:06:28 -0700 | [diff] [blame] | 332 | static void qu8_gemm_4x16__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 333 | GEMMEnd2EndBenchmark(state, model, |
| 334 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, |
| 335 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, |
| 336 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 337 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 338 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 339 | 4 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
| 340 | benchmark::utils::CheckNEON); |
| 341 | } |
| 342 | |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 343 | static void qu8_gemm_6x16__neon_mlal_lane(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 344 | GEMMEnd2EndBenchmark(state, model, |
| 345 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, |
| 346 | xnn_qu8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, |
| 347 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 348 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 349 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | a30e2df | 2022-01-10 13:44:16 -0800 | [diff] [blame] | 350 | 6 /* mr */, 16 /* nr */, 0 /* log2_kr */, 0 /* log2_sr */, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 351 | benchmark::utils::CheckNEON); |
| 352 | } |
| 353 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 354 | static void qu8_gemm_1x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 355 | GEMMEnd2EndBenchmark(state, model, |
| 356 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 357 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 358 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 359 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 360 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 361 | 1 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 362 | benchmark::utils::CheckNEONDOT); |
| 363 | } |
| 364 | static void qu8_gemm_2x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 365 | GEMMEnd2EndBenchmark(state, model, |
| 366 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, |
| 367 | xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, |
| 368 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 369 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 370 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 371 | 2 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 372 | benchmark::utils::CheckNEONDOT); |
| 373 | } |
| 374 | static void qu8_gemm_3x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 375 | GEMMEnd2EndBenchmark(state, model, |
| 376 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, |
| 377 | xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, |
| 378 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 379 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 380 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 381 | 3 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 382 | benchmark::utils::CheckNEONDOT); |
| 383 | } |
| 384 | static void qu8_gemm_4x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 385 | GEMMEnd2EndBenchmark(state, model, |
| 386 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, |
| 387 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot, |
| 388 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 389 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 390 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 391 | 4 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 392 | benchmark::utils::CheckNEONDOT); |
| 393 | } |
| 394 | static void qu8_gemm_5x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 395 | GEMMEnd2EndBenchmark(state, model, |
| 396 | xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, |
| 397 | xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, |
| 398 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 399 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 400 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 401 | 5 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 402 | benchmark::utils::CheckNEONDOT); |
| 403 | } |
| 404 | static void qu8_gemm_6x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 405 | GEMMEnd2EndBenchmark(state, model, |
| 406 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot, |
| 407 | xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, |
| 408 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 409 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 410 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 411 | 6 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 412 | benchmark::utils::CheckNEONDOT); |
| 413 | } |
| 414 | static void qu8_gemm_8x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 415 | GEMMEnd2EndBenchmark(state, model, |
| 416 | xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot, |
| 417 | xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, |
| 418 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 419 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 420 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 421 | 8 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 422 | benchmark::utils::CheckNEONDOT); |
| 423 | } |
| 424 | static void qu8_gemm_1x16c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 425 | GEMMEnd2EndBenchmark(state, model, |
| 426 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 427 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 428 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 429 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 430 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 431 | 1 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 432 | benchmark::utils::CheckNEONDOT); |
| 433 | } |
| 434 | static void qu8_gemm_2x16c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 435 | GEMMEnd2EndBenchmark(state, model, |
| 436 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, |
| 437 | xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, |
| 438 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 439 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 440 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 441 | 2 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 442 | benchmark::utils::CheckNEONDOT); |
| 443 | } |
| 444 | static void qu8_gemm_3x16c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 445 | GEMMEnd2EndBenchmark(state, model, |
| 446 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, |
| 447 | xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, |
| 448 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 449 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 450 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 451 | 3 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 452 | benchmark::utils::CheckNEONDOT); |
| 453 | } |
| 454 | static void qu8_gemm_4x16c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 455 | GEMMEnd2EndBenchmark(state, model, |
| 456 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, |
| 457 | xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, |
| 458 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 459 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 460 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 461 | 4 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 462 | benchmark::utils::CheckNEONDOT); |
| 463 | } |
| 464 | static void qu8_gemm_5x16c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 465 | GEMMEnd2EndBenchmark(state, model, |
| 466 | xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, |
| 467 | xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, |
| 468 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 469 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 470 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 471 | 5 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 472 | benchmark::utils::CheckNEONDOT); |
| 473 | } |
| 474 | static void qu8_gemm_6x16c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 475 | GEMMEnd2EndBenchmark(state, model, |
| 476 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, |
| 477 | xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, |
| 478 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 479 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 480 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 481 | 6 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 482 | benchmark::utils::CheckNEONDOT); |
| 483 | } |
| 484 | static void qu8_gemm_8x16c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 485 | GEMMEnd2EndBenchmark(state, model, |
| 486 | xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, |
| 487 | xnn_qu8_igemm_minmax_rndnu_ukernel_8x16c4__neondot, |
| 488 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 489 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 490 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 491 | 8 /* mr */, 16 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 492 | benchmark::utils::CheckNEONDOT); |
| 493 | } |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 494 | static void qu8_gemm_2x32c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 495 | GEMMEnd2EndBenchmark(state, model, |
| 496 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot, |
| 497 | xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot, |
| 498 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, |
| 499 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot, |
| 500 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 501 | 2 /* mr */, 32 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 502 | benchmark::utils::CheckNEONDOT); |
| 503 | } |
| 504 | static void qu8_gemm_3x32c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 505 | GEMMEnd2EndBenchmark(state, model, |
| 506 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot, |
| 507 | xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot, |
| 508 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, |
| 509 | xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot, |
| 510 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 511 | 3 /* mr */, 32 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */, |
| 512 | benchmark::utils::CheckNEONDOT); |
| 513 | } |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 514 | BENCHMARK_QU8_END2END(qu8_gemm_1x8c4__neondot); |
| 515 | BENCHMARK_QU8_END2END(qu8_gemm_2x8c4__neondot); |
| 516 | BENCHMARK_QU8_END2END(qu8_gemm_3x8c4__neondot); |
| 517 | BENCHMARK_QU8_END2END(qu8_gemm_4x8c4__neondot); |
| 518 | BENCHMARK_QU8_END2END(qu8_gemm_5x8c4__neondot); |
| 519 | BENCHMARK_QU8_END2END(qu8_gemm_6x8c4__neondot); |
| 520 | BENCHMARK_QU8_END2END(qu8_gemm_8x8c4__neondot); |
| 521 | BENCHMARK_QU8_END2END(qu8_gemm_1x16c4__neondot); |
| 522 | BENCHMARK_QU8_END2END(qu8_gemm_2x16c4__neondot); |
| 523 | BENCHMARK_QU8_END2END(qu8_gemm_3x16c4__neondot); |
| 524 | BENCHMARK_QU8_END2END(qu8_gemm_4x16c4__neondot); |
| 525 | BENCHMARK_QU8_END2END(qu8_gemm_5x16c4__neondot); |
| 526 | BENCHMARK_QU8_END2END(qu8_gemm_6x16c4__neondot); |
| 527 | BENCHMARK_QU8_END2END(qu8_gemm_8x16c4__neondot); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 528 | BENCHMARK_QU8_END2END(qu8_gemm_2x32c4__neondot); |
| 529 | BENCHMARK_QU8_END2END(qu8_gemm_3x32c4__neondot); |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 530 | |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 531 | BENCHMARK_QU8_END2END(qu8_gemm_2x8__neon_mlal_lane); |
| 532 | BENCHMARK_QU8_END2END(qu8_gemm_3x8__neon_mlal_lane); |
Frank Barchard | 2df7542 | 2021-09-08 19:02:40 -0700 | [diff] [blame] | 533 | BENCHMARK_QU8_END2END(qu8_gemm_4x8__neon_mlal_lane); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 534 | BENCHMARK_QU8_END2END(qu8_gemm_6x8__neon_mlal_lane); |
| 535 | BENCHMARK_QU8_END2END(qu8_gemm_2x16__neon_mlal_lane); |
| 536 | BENCHMARK_QU8_END2END(qu8_gemm_3x16__neon_mlal_lane); |
Frank Barchard | a29b57e | 2021-08-20 19:06:28 -0700 | [diff] [blame] | 537 | BENCHMARK_QU8_END2END(qu8_gemm_4x16__neon_mlal_lane); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 538 | BENCHMARK_QU8_END2END(qu8_gemm_6x16__neon_mlal_lane); |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 539 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 540 | |
| 541 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 542 | static void qu8_gemm_2x16c8__avx512skx(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 543 | GEMMEnd2EndBenchmark(state, model, |
| 544 | xnn_qu8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx, |
| 545 | xnn_qu8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx, |
| 546 | xnn_qu8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx, |
| 547 | xnn_qu8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx, |
| 548 | xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 549 | 2 /* mr */, 16 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 550 | benchmark::utils::CheckAVX512F); |
| 551 | } |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 552 | |
| 553 | static void qu8_gemm_3x16c8__avx512skx(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 554 | GEMMEnd2EndBenchmark(state, model, |
| 555 | xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx, |
| 556 | xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx, |
| 557 | xnn_qu8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx, |
| 558 | xnn_qu8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx, |
| 559 | xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 560 | 3 /* mr */, 16 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 561 | benchmark::utils::CheckAVX512F); |
| 562 | } |
Frank Barchard | 07228a3 | 2021-11-10 14:23:59 -0800 | [diff] [blame] | 563 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 564 | static void qu8_gemm_4x16c8__avx512skx(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 565 | GEMMEnd2EndBenchmark(state, model, |
| 566 | xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx, |
| 567 | xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx, |
| 568 | xnn_qu8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx, |
| 569 | xnn_qu8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx, |
| 570 | xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 571 | 4 /* mr */, 16 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 572 | benchmark::utils::CheckAVX512F); |
| 573 | } |
| 574 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 575 | static void qu8_gemm_2x8c8__avx2(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 576 | GEMMEnd2EndBenchmark(state, model, |
| 577 | xnn_qu8_gemm_minmax_fp32_ukernel_2x8c8__avx2, |
| 578 | xnn_qu8_igemm_minmax_fp32_ukernel_2x8c8__avx2, |
| 579 | xnn_qu8_gemm_minmax_fp32_ukernel_1x8c8__avx2, |
| 580 | xnn_qu8_igemm_minmax_fp32_ukernel_1x8c8__avx2, |
| 581 | xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 582 | 2 /* mr */, 8 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 583 | benchmark::utils::CheckAVX2); |
| 584 | } |
| 585 | static void qu8_gemm_3x8c8__avx2(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 586 | GEMMEnd2EndBenchmark(state, model, |
| 587 | xnn_qu8_gemm_minmax_fp32_ukernel_3x8c8__avx2, |
| 588 | xnn_qu8_igemm_minmax_fp32_ukernel_3x8c8__avx2, |
| 589 | xnn_qu8_gemm_minmax_fp32_ukernel_1x8c8__avx2, |
| 590 | xnn_qu8_igemm_minmax_fp32_ukernel_1x8c8__avx2, |
| 591 | xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 592 | 3 /* mr */, 8 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 593 | benchmark::utils::CheckAVX2); |
| 594 | } |
| 595 | |
| 596 | static void qu8_gemm_2x4c2__xop_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 597 | GEMMEnd2EndBenchmark(state, model, |
| 598 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64, |
| 599 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld64, |
| 600 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64, |
| 601 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld64, |
| 602 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 603 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 604 | benchmark::utils::CheckXOP); |
| 605 | } |
| 606 | static void qu8_gemm_2x4c2__xop_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 607 | GEMMEnd2EndBenchmark(state, model, |
| 608 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128, |
| 609 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld128, |
| 610 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128, |
| 611 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128, |
| 612 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 613 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 614 | benchmark::utils::CheckXOP); |
| 615 | } |
| 616 | static void qu8_gemm_3x4c2__xop_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 617 | GEMMEnd2EndBenchmark(state, model, |
| 618 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64, |
| 619 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld64, |
| 620 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64, |
| 621 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld64, |
| 622 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 623 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 624 | benchmark::utils::CheckXOP); |
| 625 | } |
| 626 | static void qu8_gemm_3x4c2__xop_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 627 | GEMMEnd2EndBenchmark(state, model, |
| 628 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128, |
| 629 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld128, |
| 630 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128, |
| 631 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128, |
| 632 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 633 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 634 | benchmark::utils::CheckXOP); |
| 635 | } |
| 636 | static void qu8_gemm_4x4c2__xop_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 637 | GEMMEnd2EndBenchmark(state, model, |
| 638 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64, |
| 639 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__xop_ld64, |
| 640 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64, |
| 641 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld64, |
| 642 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 643 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 644 | benchmark::utils::CheckXOP); |
| 645 | } |
| 646 | static void qu8_gemm_4x4c2__xop_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 647 | GEMMEnd2EndBenchmark(state, model, |
| 648 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128, |
| 649 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__xop_ld128, |
| 650 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128, |
| 651 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128, |
| 652 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 653 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 654 | benchmark::utils::CheckXOP); |
| 655 | } |
| 656 | |
| 657 | static void qu8_gemm_2x4c8__xop_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 658 | GEMMEnd2EndBenchmark(state, model, |
| 659 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64, |
| 660 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld64, |
| 661 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64, |
| 662 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld64, |
| 663 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 664 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 665 | benchmark::utils::CheckXOP); |
| 666 | } |
| 667 | static void qu8_gemm_3x4c8__xop_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 668 | GEMMEnd2EndBenchmark(state, model, |
| 669 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64, |
| 670 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__xop_ld64, |
| 671 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64, |
| 672 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld64, |
| 673 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 674 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 675 | benchmark::utils::CheckXOP); |
| 676 | } |
| 677 | |
| 678 | static void qu8_gemm_2x4c8__xop_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 679 | GEMMEnd2EndBenchmark(state, model, |
| 680 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128, |
| 681 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld128, |
| 682 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128, |
| 683 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld128, |
| 684 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 685 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 686 | benchmark::utils::CheckXOP); |
| 687 | } |
| 688 | static void qu8_gemm_3x4c8__xop_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 689 | GEMMEnd2EndBenchmark(state, model, |
| 690 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128, |
| 691 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__xop_ld128, |
| 692 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128, |
| 693 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld128, |
| 694 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 695 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 696 | benchmark::utils::CheckXOP); |
| 697 | } |
| 698 | |
| 699 | static void qu8_gemm_2x4c2__avx_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 700 | GEMMEnd2EndBenchmark(state, model, |
| 701 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64, |
| 702 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__avx_ld64, |
| 703 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64, |
| 704 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld64, |
| 705 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 706 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 707 | benchmark::utils::CheckAVX); |
| 708 | } |
| 709 | static void qu8_gemm_2x4c2__avx_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 710 | GEMMEnd2EndBenchmark(state, model, |
| 711 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128, |
| 712 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__avx_ld128, |
| 713 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128, |
| 714 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld128, |
| 715 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 716 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 717 | benchmark::utils::CheckAVX); |
| 718 | } |
| 719 | static void qu8_gemm_3x4c2__avx_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 720 | GEMMEnd2EndBenchmark(state, model, |
| 721 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64, |
| 722 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__avx_ld64, |
| 723 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64, |
| 724 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld64, |
| 725 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 726 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 727 | benchmark::utils::CheckAVX); |
| 728 | } |
| 729 | static void qu8_gemm_3x4c2__avx_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 730 | GEMMEnd2EndBenchmark(state, model, |
| 731 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128, |
| 732 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__avx_ld128, |
| 733 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128, |
| 734 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld128, |
| 735 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 736 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 737 | benchmark::utils::CheckAVX); |
| 738 | } |
| 739 | static void qu8_gemm_4x4c2__avx_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 740 | GEMMEnd2EndBenchmark(state, model, |
| 741 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64, |
| 742 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__avx_ld64, |
| 743 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64, |
| 744 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld64, |
| 745 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 746 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 747 | benchmark::utils::CheckAVX); |
| 748 | } |
| 749 | static void qu8_gemm_4x4c2__avx_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 750 | GEMMEnd2EndBenchmark(state, model, |
| 751 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128, |
| 752 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__avx_ld128, |
| 753 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128, |
| 754 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld128, |
| 755 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 756 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 757 | benchmark::utils::CheckAVX); |
| 758 | } |
| 759 | |
| 760 | |
| 761 | static void qu8_gemm_2x4c8__avx_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 762 | GEMMEnd2EndBenchmark(state, model, |
| 763 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64, |
| 764 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__avx_ld64, |
| 765 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64, |
| 766 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld64, |
| 767 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 768 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 769 | benchmark::utils::CheckAVX); |
| 770 | } |
| 771 | static void qu8_gemm_2x4c8__avx_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 772 | GEMMEnd2EndBenchmark(state, model, |
| 773 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128, |
| 774 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__avx_ld128, |
| 775 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128, |
| 776 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld128, |
| 777 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 778 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 779 | benchmark::utils::CheckAVX); |
| 780 | } |
| 781 | static void qu8_gemm_3x4c8__avx_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 782 | GEMMEnd2EndBenchmark(state, model, |
| 783 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64, |
| 784 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__avx_ld64, |
| 785 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64, |
| 786 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld64, |
| 787 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 788 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 789 | benchmark::utils::CheckAVX); |
| 790 | } |
| 791 | static void qu8_gemm_3x4c8__avx_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 792 | GEMMEnd2EndBenchmark(state, model, |
| 793 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128, |
| 794 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__avx_ld128, |
| 795 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128, |
| 796 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld128, |
| 797 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 798 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 799 | benchmark::utils::CheckAVX); |
| 800 | } |
| 801 | |
| 802 | static void qu8_gemm_2x4c2__sse41_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 803 | GEMMEnd2EndBenchmark(state, model, |
| 804 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64, |
| 805 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__sse41_ld64, |
| 806 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64, |
| 807 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld64, |
| 808 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 809 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 810 | benchmark::utils::CheckSSE41); |
| 811 | } |
| 812 | static void qu8_gemm_2x4c2__sse41_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 813 | GEMMEnd2EndBenchmark(state, model, |
| 814 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128, |
| 815 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__sse41_ld128, |
| 816 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128, |
| 817 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld128, |
| 818 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 819 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 820 | benchmark::utils::CheckSSE41); |
| 821 | } |
| 822 | static void qu8_gemm_3x4c2__sse41_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 823 | GEMMEnd2EndBenchmark(state, model, |
| 824 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64, |
| 825 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld64, |
| 826 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64, |
| 827 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld64, |
| 828 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 829 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 830 | benchmark::utils::CheckSSE41); |
| 831 | } |
| 832 | static void qu8_gemm_3x4c2__sse41_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 833 | GEMMEnd2EndBenchmark(state, model, |
| 834 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128, |
| 835 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld128, |
| 836 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128, |
| 837 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld128, |
| 838 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 839 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 840 | benchmark::utils::CheckSSE41); |
| 841 | } |
| 842 | static void qu8_gemm_4x4c2__sse41_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 843 | GEMMEnd2EndBenchmark(state, model, |
| 844 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64, |
| 845 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse41_ld64, |
| 846 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64, |
| 847 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld64, |
| 848 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 849 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 850 | benchmark::utils::CheckSSE41); |
| 851 | } |
| 852 | static void qu8_gemm_4x4c2__sse41_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 853 | GEMMEnd2EndBenchmark(state, model, |
| 854 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128, |
| 855 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse41_ld128, |
| 856 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128, |
| 857 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld128, |
| 858 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 859 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */, |
| 860 | benchmark::utils::CheckSSE41); |
| 861 | } |
| 862 | |
| 863 | static void qu8_gemm_2x4c8__sse41_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 864 | GEMMEnd2EndBenchmark(state, model, |
| 865 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64, |
| 866 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse41_ld64, |
| 867 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64, |
| 868 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld64, |
| 869 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 870 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 871 | benchmark::utils::CheckSSE41); |
| 872 | } |
| 873 | static void qu8_gemm_2x4c8__sse41_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 874 | GEMMEnd2EndBenchmark(state, model, |
| 875 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128, |
| 876 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse41_ld128, |
| 877 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128, |
| 878 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld128, |
| 879 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 880 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 881 | benchmark::utils::CheckSSE41); |
| 882 | } |
| 883 | static void qu8_gemm_3x4c8__sse41_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 884 | GEMMEnd2EndBenchmark(state, model, |
| 885 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64, |
| 886 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld64, |
| 887 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64, |
| 888 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld64, |
| 889 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 890 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 891 | benchmark::utils::CheckSSE41); |
| 892 | } |
| 893 | static void qu8_gemm_3x4c8__sse41_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 894 | GEMMEnd2EndBenchmark(state, model, |
| 895 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128, |
| 896 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld128, |
| 897 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128, |
| 898 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld128, |
| 899 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 900 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */, |
| 901 | benchmark::utils::CheckSSE41); |
| 902 | } |
| 903 | |
| 904 | static void qu8_gemm_2x4c2__sse2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 905 | GEMMEnd2EndBenchmark(state, model, |
| 906 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64, |
| 907 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__sse2_ld64, |
| 908 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64, |
| 909 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld64, |
| 910 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 911 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */); |
| 912 | } |
| 913 | static void qu8_gemm_2x4c2__sse2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 914 | GEMMEnd2EndBenchmark(state, model, |
| 915 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128, |
| 916 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__sse2_ld128, |
| 917 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128, |
| 918 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld128, |
| 919 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 920 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */); |
| 921 | } |
| 922 | static void qu8_gemm_3x4c2__sse2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 923 | GEMMEnd2EndBenchmark(state, model, |
| 924 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64, |
| 925 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__sse2_ld64, |
| 926 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64, |
| 927 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld64, |
| 928 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 929 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */); |
| 930 | } |
| 931 | static void qu8_gemm_3x4c2__sse2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 932 | GEMMEnd2EndBenchmark(state, model, |
| 933 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128, |
| 934 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__sse2_ld128, |
| 935 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128, |
| 936 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld128, |
| 937 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 938 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */); |
| 939 | } |
| 940 | static void qu8_gemm_4x4c2__sse2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 941 | GEMMEnd2EndBenchmark(state, model, |
| 942 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64, |
| 943 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse2_ld64, |
| 944 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64, |
| 945 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld64, |
| 946 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 947 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */); |
| 948 | } |
| 949 | static void qu8_gemm_4x4c2__sse2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 950 | GEMMEnd2EndBenchmark(state, model, |
| 951 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128, |
| 952 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse2_ld128, |
| 953 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128, |
| 954 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld128, |
| 955 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 956 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 0 /* log2_sr */); |
| 957 | } |
| 958 | |
| 959 | static void qu8_gemm_2x4c8__sse2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 960 | GEMMEnd2EndBenchmark(state, model, |
| 961 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64, |
| 962 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse2_ld64, |
| 963 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64, |
| 964 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64, |
| 965 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 966 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */); |
| 967 | } |
| 968 | static void qu8_gemm_2x4c8__sse2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 969 | GEMMEnd2EndBenchmark(state, model, |
| 970 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128, |
| 971 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse2_ld128, |
| 972 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128, |
| 973 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld128, |
| 974 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 975 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */); |
| 976 | } |
| 977 | static void qu8_gemm_3x4c8__sse2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 978 | GEMMEnd2EndBenchmark(state, model, |
| 979 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64, |
| 980 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64, |
| 981 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64, |
| 982 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64, |
| 983 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 984 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */); |
| 985 | } |
| 986 | static void qu8_gemm_3x4c8__sse2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 987 | GEMMEnd2EndBenchmark(state, model, |
| 988 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128, |
| 989 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld128, |
| 990 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128, |
| 991 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld128, |
| 992 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 993 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */, 0 /* log2_sr */); |
| 994 | } |
| 995 | |
| 996 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 997 | BENCHMARK_QU8_END2END(qu8_gemm_2x16c8__avx512skx); |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 998 | BENCHMARK_QU8_END2END(qu8_gemm_3x16c8__avx512skx); |
| 999 | BENCHMARK_QU8_END2END(qu8_gemm_4x16c8__avx512skx); |
| 1000 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1001 | BENCHMARK_QU8_END2END(qu8_gemm_2x8c8__avx2); |
| 1002 | BENCHMARK_QU8_END2END(qu8_gemm_3x8c8__avx2); |
| 1003 | |
| 1004 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__xop_ld64); |
| 1005 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__xop_ld128); |
| 1006 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__xop_ld64); |
| 1007 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__xop_ld128); |
| 1008 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__xop_ld64); |
| 1009 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__xop_ld128); |
| 1010 | |
| 1011 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__xop_ld64); |
| 1012 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__xop_ld128); |
| 1013 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__xop_ld64); |
| 1014 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__xop_ld128); |
| 1015 | |
| 1016 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__avx_ld64); |
| 1017 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__avx_ld128); |
| 1018 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__avx_ld64); |
| 1019 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__avx_ld128); |
| 1020 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__avx_ld64); |
| 1021 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__avx_ld128); |
| 1022 | |
| 1023 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__avx_ld64); |
| 1024 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__avx_ld128); |
| 1025 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__avx_ld64); |
| 1026 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__avx_ld128); |
| 1027 | |
| 1028 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__sse41_ld64); |
| 1029 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__sse41_ld128); |
| 1030 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__sse41_ld64); |
| 1031 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__sse41_ld128); |
| 1032 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__sse41_ld64); |
| 1033 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__sse41_ld128); |
| 1034 | |
| 1035 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__sse41_ld64); |
| 1036 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__sse41_ld128); |
| 1037 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__sse41_ld64); |
| 1038 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__sse41_ld128); |
| 1039 | |
| 1040 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__sse2_ld64); |
| 1041 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__sse2_ld128); |
| 1042 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__sse2_ld64); |
| 1043 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__sse2_ld128); |
| 1044 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__sse2_ld64); |
| 1045 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__sse2_ld128); |
| 1046 | |
| 1047 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__sse2_ld64); |
| 1048 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__sse2_ld128); |
| 1049 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__sse2_ld64); |
| 1050 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__sse2_ld128); |
| 1051 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1052 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1053 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 58cdcf2 | 2022-02-01 02:05:00 -0800 | [diff] [blame] | 1054 | static void qu8_gemm_2x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1055 | GEMMEnd2EndBenchmark(state, model, |
| 1056 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64, |
| 1057 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64, |
| 1058 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64, |
| 1059 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64, |
| 1060 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1061 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */); |
| 1062 | } |
| 1063 | static void qu8_gemm_2x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1064 | GEMMEnd2EndBenchmark(state, model, |
| 1065 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128, |
| 1066 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128, |
| 1067 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128, |
| 1068 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128, |
| 1069 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1070 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */); |
| 1071 | } |
| 1072 | static void qu8_gemm_3x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1073 | GEMMEnd2EndBenchmark(state, model, |
| 1074 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64, |
| 1075 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64, |
| 1076 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64, |
| 1077 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64, |
| 1078 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1079 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */); |
| 1080 | } |
| 1081 | static void qu8_gemm_3x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1082 | GEMMEnd2EndBenchmark(state, model, |
| 1083 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128, |
| 1084 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128, |
| 1085 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128, |
| 1086 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128, |
| 1087 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1088 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */); |
| 1089 | } |
| 1090 | static void qu8_gemm_4x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1091 | GEMMEnd2EndBenchmark(state, model, |
| 1092 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64, |
| 1093 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64, |
| 1094 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64, |
| 1095 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64, |
| 1096 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1097 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */); |
| 1098 | } |
| 1099 | static void qu8_gemm_4x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1100 | GEMMEnd2EndBenchmark(state, model, |
| 1101 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128, |
| 1102 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128, |
| 1103 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128, |
| 1104 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128, |
| 1105 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1106 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */); |
| 1107 | } |
| 1108 | |
| 1109 | static void qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1110 | GEMMEnd2EndBenchmark(state, model, |
| 1111 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64, |
| 1112 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64, |
| 1113 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64, |
| 1114 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64, |
| 1115 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1116 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 2 /* log2_sr */); |
| 1117 | } |
| 1118 | static void qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1119 | GEMMEnd2EndBenchmark(state, model, |
| 1120 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128, |
| 1121 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128, |
| 1122 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128, |
| 1123 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128, |
| 1124 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1125 | 2 /* mr */, 4 /* nr */, 1 /* log2_kr */, 2 /* log2_sr */); |
| 1126 | } |
| 1127 | static void qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1128 | GEMMEnd2EndBenchmark(state, model, |
| 1129 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64, |
| 1130 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64, |
| 1131 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64, |
| 1132 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64, |
| 1133 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1134 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 2 /* log2_sr */); |
| 1135 | } |
| 1136 | static void qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1137 | GEMMEnd2EndBenchmark(state, model, |
| 1138 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128, |
| 1139 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128, |
| 1140 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128, |
| 1141 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128, |
| 1142 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1143 | 3 /* mr */, 4 /* nr */, 1 /* log2_kr */, 2 /* log2_sr */); |
| 1144 | } |
| 1145 | static void qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1146 | GEMMEnd2EndBenchmark(state, model, |
| 1147 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64, |
| 1148 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64, |
| 1149 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64, |
| 1150 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64, |
| 1151 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1152 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 2 /* log2_sr */); |
| 1153 | } |
| 1154 | static void qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1155 | GEMMEnd2EndBenchmark(state, model, |
| 1156 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128, |
| 1157 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128, |
| 1158 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128, |
| 1159 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128, |
| 1160 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1161 | 4 /* mr */, 4 /* nr */, 1 /* log2_kr */, 2 /* log2_sr */); |
| 1162 | } |
| 1163 | |
| 1164 | static void qu8_gemm_2x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1165 | GEMMEnd2EndBenchmark(state, model, |
| 1166 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64, |
| 1167 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64, |
| 1168 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64, |
| 1169 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64, |
| 1170 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1171 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1172 | } |
| 1173 | static void qu8_gemm_2x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1174 | GEMMEnd2EndBenchmark(state, model, |
| 1175 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128, |
| 1176 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128, |
| 1177 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128, |
| 1178 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128, |
| 1179 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1180 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1181 | } |
| 1182 | static void qu8_gemm_3x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1183 | GEMMEnd2EndBenchmark(state, model, |
| 1184 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64, |
| 1185 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64, |
| 1186 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64, |
| 1187 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64, |
| 1188 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1189 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1190 | } |
| 1191 | static void qu8_gemm_3x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1192 | GEMMEnd2EndBenchmark(state, model, |
| 1193 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128, |
| 1194 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128, |
| 1195 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128, |
| 1196 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128, |
| 1197 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1198 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1199 | } |
| 1200 | static void qu8_gemm_4x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1201 | GEMMEnd2EndBenchmark(state, model, |
| 1202 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64, |
| 1203 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64, |
| 1204 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64, |
| 1205 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64, |
| 1206 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1207 | 4 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1208 | } |
| 1209 | static void qu8_gemm_4x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1210 | GEMMEnd2EndBenchmark(state, model, |
| 1211 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128, |
| 1212 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128, |
| 1213 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128, |
| 1214 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128, |
| 1215 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1216 | 4 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1217 | } |
| 1218 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1219 | static void qu8_gemm_2x4c8__wasmsimd_mul32_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1220 | GEMMEnd2EndBenchmark(state, model, |
| 1221 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld64, |
| 1222 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld64, |
| 1223 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld64, |
| 1224 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld64, |
| 1225 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1226 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1227 | } |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1228 | static void qu8_gemm_2x4c8__wasmsimd_mul32_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1229 | GEMMEnd2EndBenchmark(state, model, |
| 1230 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld128, |
| 1231 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld128, |
| 1232 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld128, |
| 1233 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld128, |
| 1234 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1235 | 2 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1236 | } |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1237 | static void qu8_gemm_3x4c8__wasmsimd_mul32_ld64(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1238 | GEMMEnd2EndBenchmark(state, model, |
| 1239 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld64, |
| 1240 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld64, |
| 1241 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld64, |
| 1242 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld64, |
| 1243 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1244 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1245 | } |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1246 | static void qu8_gemm_3x4c8__wasmsimd_mul32_ld128(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1247 | GEMMEnd2EndBenchmark(state, model, |
| 1248 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128, |
| 1249 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128, |
| 1250 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld128, |
| 1251 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld128, |
| 1252 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1253 | 3 /* mr */, 4 /* nr */, 3 /* log2_kr */); |
| 1254 | } |
| 1255 | |
Marat Dukhan | 58cdcf2 | 2022-02-01 02:05:00 -0800 | [diff] [blame] | 1256 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__wasmsimd_dot16x2_ld64) |
| 1257 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2__wasmsimd_dot16x2_ld128) |
| 1258 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__wasmsimd_dot16x2_ld64) |
| 1259 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2__wasmsimd_dot16x2_ld128) |
| 1260 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__wasmsimd_dot16x2_ld64) |
| 1261 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2__wasmsimd_dot16x2_ld128) |
| 1262 | |
| 1263 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64) |
| 1264 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128) |
| 1265 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64) |
| 1266 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128) |
| 1267 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64) |
| 1268 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128) |
| 1269 | |
| 1270 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__wasmsimd_dot16x2_ld64) |
| 1271 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__wasmsimd_dot16x2_ld128) |
| 1272 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__wasmsimd_dot16x2_ld64) |
| 1273 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__wasmsimd_dot16x2_ld128) |
| 1274 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c8__wasmsimd_dot16x2_ld64) |
| 1275 | BENCHMARK_QU8_END2END(qu8_gemm_4x4c8__wasmsimd_dot16x2_ld128) |
| 1276 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1277 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__wasmsimd_mul32_ld64) |
| 1278 | BENCHMARK_QU8_END2END(qu8_gemm_2x4c8__wasmsimd_mul32_ld128) |
| 1279 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__wasmsimd_mul32_ld64) |
| 1280 | BENCHMARK_QU8_END2END(qu8_gemm_3x4c8__wasmsimd_mul32_ld128) |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1281 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1282 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1283 | |
| 1284 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| 1285 | static void qu8_gemm_2x2__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1286 | GEMMEnd2EndBenchmark(state, model, |
| 1287 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__wasm_fmagic, |
| 1288 | xnn_qu8_igemm_minmax_fp32_ukernel_2x2__wasm_fmagic, |
| 1289 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic, |
| 1290 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__wasm_fmagic, |
| 1291 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1292 | 2 /* mr */, 2 /* nr */); |
| 1293 | } |
| 1294 | static void qu8_gemm_3x2__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1295 | GEMMEnd2EndBenchmark(state, model, |
| 1296 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic, |
| 1297 | xnn_qu8_igemm_minmax_fp32_ukernel_3x2__wasm_fmagic, |
| 1298 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic, |
| 1299 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__wasm_fmagic, |
| 1300 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1301 | 3 /* mr */, 2 /* nr */); |
| 1302 | } |
| 1303 | static void qu8_gemm_4x2__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1304 | GEMMEnd2EndBenchmark(state, model, |
| 1305 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic, |
| 1306 | xnn_qu8_igemm_minmax_fp32_ukernel_4x2__wasm_fmagic, |
| 1307 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic, |
| 1308 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__wasm_fmagic, |
| 1309 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1310 | 4 /* mr */, 2 /* nr */); |
| 1311 | } |
| 1312 | static void qu8_gemm_2x4__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1313 | GEMMEnd2EndBenchmark(state, model, |
| 1314 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic, |
| 1315 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4__wasm_fmagic, |
| 1316 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic, |
| 1317 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__wasm_fmagic, |
| 1318 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1319 | 2 /* mr */, 4 /* nr */); |
| 1320 | } |
| 1321 | static void qu8_gemm_3x4__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1322 | GEMMEnd2EndBenchmark(state, model, |
| 1323 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__wasm_fmagic, |
| 1324 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4__wasm_fmagic, |
| 1325 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic, |
| 1326 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__wasm_fmagic, |
| 1327 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1328 | 3 /* mr */, 4 /* nr */); |
| 1329 | } |
| 1330 | static void qu8_gemm_4x4__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| 1331 | GEMMEnd2EndBenchmark(state, model, |
| 1332 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic, |
| 1333 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4__wasm_fmagic, |
| 1334 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic, |
| 1335 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__wasm_fmagic, |
| 1336 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1337 | 4 /* mr */, 4 /* nr */); |
| 1338 | } |
| 1339 | |
| 1340 | BENCHMARK_QU8_END2END(qu8_gemm_2x2__wasm_fmagic) |
| 1341 | BENCHMARK_QU8_END2END(qu8_gemm_3x2__wasm_fmagic) |
| 1342 | BENCHMARK_QU8_END2END(qu8_gemm_4x2__wasm_fmagic) |
| 1343 | BENCHMARK_QU8_END2END(qu8_gemm_2x4__wasm_fmagic) |
| 1344 | BENCHMARK_QU8_END2END(qu8_gemm_3x4__wasm_fmagic) |
| 1345 | BENCHMARK_QU8_END2END(qu8_gemm_4x4__wasm_fmagic) |
| 1346 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| 1347 | |
| 1348 | |
| 1349 | static void qu8_gemm_2x2__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1350 | GEMMEnd2EndBenchmark(state, model, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1351 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_fmagic, |
| 1352 | xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_fmagic, |
| 1353 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic, |
| 1354 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_fmagic, |
| 1355 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1356 | 2 /* mr */, 2 /* nr */); |
| 1357 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1358 | static void qu8_gemm_3x2__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1359 | GEMMEnd2EndBenchmark(state, model, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1360 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic, |
| 1361 | xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_fmagic, |
| 1362 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic, |
| 1363 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_fmagic, |
| 1364 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1365 | 3 /* mr */, 2 /* nr */); |
| 1366 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1367 | static void qu8_gemm_4x2__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1368 | GEMMEnd2EndBenchmark(state, model, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1369 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic, |
| 1370 | xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_fmagic, |
| 1371 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic, |
| 1372 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_fmagic, |
| 1373 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1374 | 4 /* mr */, 2 /* nr */); |
| 1375 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1376 | static void qu8_gemm_2x4__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1377 | GEMMEnd2EndBenchmark(state, model, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1378 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic, |
| 1379 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_fmagic, |
| 1380 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic, |
| 1381 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_fmagic, |
| 1382 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1383 | 2 /* mr */, 4 /* nr */); |
| 1384 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1385 | static void qu8_gemm_3x4__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1386 | GEMMEnd2EndBenchmark(state, model, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1387 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic, |
| 1388 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_fmagic, |
| 1389 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic, |
| 1390 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_fmagic, |
| 1391 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1392 | 3 /* mr */, 4 /* nr */); |
| 1393 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1394 | static void qu8_gemm_4x4__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1395 | GEMMEnd2EndBenchmark(state, model, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1396 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic, |
| 1397 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic, |
| 1398 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic, |
| 1399 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_fmagic, |
| 1400 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1401 | 4 /* mr */, 4 /* nr */); |
| 1402 | } |
| 1403 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1404 | static void qu8_gemm_2x2__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1405 | GEMMEnd2EndBenchmark(state, model, |
| 1406 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, |
| 1407 | xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, |
| 1408 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, |
| 1409 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, |
| 1410 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1411 | 2 /* mr */, 2 /* nr */); |
| 1412 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1413 | static void qu8_gemm_3x2__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1414 | GEMMEnd2EndBenchmark(state, model, |
| 1415 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, |
| 1416 | xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, |
| 1417 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, |
| 1418 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, |
| 1419 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1420 | 3 /* mr */, 2 /* nr */); |
| 1421 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1422 | static void qu8_gemm_4x2__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1423 | GEMMEnd2EndBenchmark(state, model, |
| 1424 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, |
| 1425 | xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, |
| 1426 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, |
| 1427 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, |
| 1428 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1429 | 4 /* mr */, 2 /* nr */); |
| 1430 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1431 | static void qu8_gemm_2x4__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1432 | GEMMEnd2EndBenchmark(state, model, |
| 1433 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, |
| 1434 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, |
| 1435 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, |
| 1436 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, |
| 1437 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1438 | 2 /* mr */, 4 /* nr */); |
| 1439 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1440 | static void qu8_gemm_3x4__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1441 | GEMMEnd2EndBenchmark(state, model, |
| 1442 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, |
| 1443 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, |
| 1444 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, |
| 1445 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, |
| 1446 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1447 | 3 /* mr */, 4 /* nr */); |
| 1448 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1449 | static void qu8_gemm_4x4__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1450 | GEMMEnd2EndBenchmark(state, model, |
| 1451 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, |
| 1452 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, |
| 1453 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, |
| 1454 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, |
| 1455 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1456 | 4 /* mr */, 4 /* nr */); |
| 1457 | } |
| 1458 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1459 | static void qu8_gemm_2x2__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1460 | GEMMEnd2EndBenchmark(state, model, |
| 1461 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, |
| 1462 | xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, |
| 1463 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, |
| 1464 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, |
| 1465 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1466 | 2 /* mr */, 2 /* nr */); |
| 1467 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1468 | static void qu8_gemm_3x2__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1469 | GEMMEnd2EndBenchmark(state, model, |
| 1470 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, |
| 1471 | xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, |
| 1472 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, |
| 1473 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, |
| 1474 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1475 | 3 /* mr */, 2 /* nr */); |
| 1476 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1477 | static void qu8_gemm_4x2__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1478 | GEMMEnd2EndBenchmark(state, model, |
| 1479 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, |
| 1480 | xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, |
| 1481 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, |
| 1482 | xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, |
| 1483 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1484 | 4 /* mr */, 2 /* nr */); |
| 1485 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1486 | static void qu8_gemm_2x4__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1487 | GEMMEnd2EndBenchmark(state, model, |
| 1488 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, |
| 1489 | xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, |
| 1490 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, |
| 1491 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, |
| 1492 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1493 | 2 /* mr */, 4 /* nr */); |
| 1494 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1495 | static void qu8_gemm_3x4__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1496 | GEMMEnd2EndBenchmark(state, model, |
| 1497 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, |
| 1498 | xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, |
| 1499 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, |
| 1500 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, |
| 1501 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1502 | 3 /* mr */, 4 /* nr */); |
| 1503 | } |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1504 | static void qu8_gemm_4x4__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1505 | GEMMEnd2EndBenchmark(state, model, |
| 1506 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, |
| 1507 | xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, |
| 1508 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, |
| 1509 | xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, |
| 1510 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1511 | 4 /* mr */, 4 /* nr */); |
| 1512 | } |
| 1513 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1514 | BENCHMARK_QU8_END2END(qu8_gemm_2x2__scalar_fmagic) |
| 1515 | BENCHMARK_QU8_END2END(qu8_gemm_3x2__scalar_fmagic) |
| 1516 | BENCHMARK_QU8_END2END(qu8_gemm_4x2__scalar_fmagic) |
| 1517 | BENCHMARK_QU8_END2END(qu8_gemm_2x4__scalar_fmagic) |
| 1518 | BENCHMARK_QU8_END2END(qu8_gemm_3x4__scalar_fmagic) |
| 1519 | BENCHMARK_QU8_END2END(qu8_gemm_4x4__scalar_fmagic) |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1520 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1521 | BENCHMARK_QU8_END2END(qu8_gemm_2x2__scalar_imagic) |
| 1522 | BENCHMARK_QU8_END2END(qu8_gemm_3x2__scalar_imagic) |
| 1523 | BENCHMARK_QU8_END2END(qu8_gemm_4x2__scalar_imagic) |
| 1524 | BENCHMARK_QU8_END2END(qu8_gemm_2x4__scalar_imagic) |
| 1525 | BENCHMARK_QU8_END2END(qu8_gemm_3x4__scalar_imagic) |
| 1526 | BENCHMARK_QU8_END2END(qu8_gemm_4x4__scalar_imagic) |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1527 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1528 | BENCHMARK_QU8_END2END(qu8_gemm_2x2__scalar_lrintf) |
| 1529 | BENCHMARK_QU8_END2END(qu8_gemm_3x2__scalar_lrintf) |
| 1530 | BENCHMARK_QU8_END2END(qu8_gemm_4x2__scalar_lrintf) |
| 1531 | BENCHMARK_QU8_END2END(qu8_gemm_2x4__scalar_lrintf) |
| 1532 | BENCHMARK_QU8_END2END(qu8_gemm_3x4__scalar_lrintf) |
| 1533 | BENCHMARK_QU8_END2END(qu8_gemm_4x4__scalar_lrintf) |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1534 | |
Frank Barchard | 9098aba | 2021-08-12 12:20:03 -0700 | [diff] [blame] | 1535 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 1536 | BENCHMARK_MAIN(); |
| 1537 | #endif |