XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #include <algorithm> |
| 10 | #include <cfloat> |
| 11 | #include <chrono> |
| 12 | #include <cmath> |
| 13 | #include <functional> |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 14 | #include <limits> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 15 | #include <mutex> |
| 16 | #include <random> |
| 17 | #include <vector> |
| 18 | |
Frank Barchard | fa4daf0 | 2021-10-27 13:31:04 -0700 | [diff] [blame] | 19 | #include <cpuinfo.h> |
| 20 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 21 | #include <benchmark/benchmark.h> |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 22 | #ifdef BENCHMARK_GEMMLOWP |
| 23 | #include "gemmlowp/public/gemmlowp.h" |
| 24 | #endif // BENCHMARK_GEMMLOWP |
| 25 | #ifdef BENCHMARK_RUY |
Benoit Jacob | b038fdc | 2020-03-25 12:14:20 -0700 | [diff] [blame] | 26 | #include "ruy/ruy.h" |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 27 | #endif // BENCHMARK_RUY |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 28 | #include "bench/gemm.h" |
| 29 | #include "bench/utils.h" |
| 30 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 31 | #include <xnnpack/common.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 32 | #include <xnnpack/gemm.h> |
| 33 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 34 | #include <xnnpack/params-init.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 35 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 36 | |
| 37 | |
| 38 | static void GEMMBenchmark(benchmark::State& state, |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 39 | xnn_qu8_gemm_minmax_ukernel_function gemm, |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 40 | xnn_init_qu8_conv_minmax_params_fn init_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 41 | size_t mr, size_t nr, size_t kr, size_t sr, |
Marat Dukhan | 4e89587 | 2020-12-04 15:27:45 -0800 | [diff] [blame] | 42 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 43 | { |
Frank Barchard | fa4daf0 | 2021-10-27 13:31:04 -0700 | [diff] [blame] | 44 | if (!cpuinfo_initialize()) { |
| 45 | state.SkipWithError("cpuinfo initialization failed"); |
| 46 | return; |
| 47 | } |
| 48 | if (isa_check && !isa_check(state)) { |
| 49 | return; |
| 50 | } |
| 51 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 52 | const size_t mc = state.range(0); |
| 53 | const size_t nc = state.range(1); |
| 54 | const size_t kc = state.range(2); |
| 55 | |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 56 | const size_t nc_stride = benchmark::utils::RoundUp(nc, nr); |
Marat Dukhan | fbd67a7 | 2022-01-31 18:03:50 -0800 | [diff] [blame] | 57 | const size_t kc_stride = benchmark::utils::RoundUp(kc, kr * sr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 58 | |
| 59 | std::random_device random_device; |
| 60 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 61 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), std::ref(rng)); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 62 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 63 | |
Marat Dukhan | 91351ef | 2021-08-04 16:32:28 -0700 | [diff] [blame] | 64 | std::vector<uint8_t> a(mc * kc + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 65 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 66 | std::vector<uint8_t> k(nc * kc); |
| 67 | std::generate(k.begin(), k.end(), std::ref(u8rng)); |
| 68 | std::vector<int32_t> b(nc); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 69 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 70 | |
| 71 | const size_t w_elements = kc_stride * nc_stride + nc_stride * sizeof(int32_t) / sizeof(uint8_t); |
| 72 | const size_t c_elements = mc * nc; |
| 73 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 74 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 75 | sizeof(uint8_t) * (w_elements + c_elements)); |
| 76 | |
Marat Dukhan | e13e639 | 2021-07-26 22:22:35 -0700 | [diff] [blame] | 77 | std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> w(w_elements * num_buffers); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 78 | std::fill(w.begin(), w.end(), 0); |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 79 | const xnn_qu8_packing_params packing_params = { 127, 127 }; |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 80 | xnn_pack_qu8_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data(), 0, &packing_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 81 | std::vector<uint8_t> c(c_elements * num_buffers); |
| 82 | std::fill(c.begin(), c.end(), 0xA5); |
| 83 | |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 84 | union xnn_qu8_conv_minmax_params quantization_params; |
Marat Dukhan | 725f47e | 2021-05-22 10:06:19 -0700 | [diff] [blame] | 85 | init_params(&quantization_params, 127, 0.75f, 127, 1, 254); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 86 | |
| 87 | size_t buffer_index = 0; |
| 88 | for (auto _ : state) { |
| 89 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 90 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 91 | // - W is not in cache (for any cache level) |
| 92 | // - C is not in cache (for any cache level) |
| 93 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 94 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 95 | buffer_index = (buffer_index + 1) % num_buffers; |
| 96 | state.ResumeTiming(); |
| 97 | |
| 98 | for (uint32_t m = 0; m < mc; m += mr) { |
| 99 | const uint32_t mb = min(mc - m, mr); |
| 100 | for (uint32_t n = 0; n < nc; n += nr) { |
| 101 | const uint32_t nb = min(nc - n, nr); |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 102 | gemm( |
Marat Dukhan | b186463 | 2019-11-25 16:34:17 -0800 | [diff] [blame] | 103 | mb, nb, kc * sizeof(uint8_t), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 104 | a.data() + m * kc, kc * sizeof(uint8_t), |
| 105 | w.data() + (w_elements * buffer_index + n * (kc_stride + sizeof(int32_t))) / sizeof(uint8_t), |
| 106 | c.data() + (mc * buffer_index + m) * nc + n, nc * sizeof(uint8_t), nr * sizeof(uint8_t), |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 107 | &quantization_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 108 | } |
| 109 | } |
| 110 | } |
| 111 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 112 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 113 | if (cpu_frequency != 0) { |
| 114 | state.counters["cpufreq"] = cpu_frequency; |
| 115 | } |
| 116 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 117 | state.counters["OPS"] = benchmark::Counter( |
| 118 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 119 | } |
| 120 | |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 121 | #ifdef BENCHMARK_GEMMLOWP |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 122 | struct GemmlowpOutputPipeline { |
| 123 | typedef gemmlowp::VectorMap<const int32_t, gemmlowp::VectorShape::Col> ColVectorMap; |
| 124 | typedef std::tuple< |
| 125 | gemmlowp::OutputStageBiasAddition<ColVectorMap>, |
| 126 | gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint, |
| 127 | gemmlowp::OutputStageClamp, |
| 128 | gemmlowp::OutputStageSaturatingCastToUint8> |
| 129 | Pipeline; |
| 130 | |
| 131 | static Pipeline Make( |
| 132 | const int32_t* bias_data, |
| 133 | int output_rows, |
| 134 | int32_t output_offset, |
| 135 | int32_t output_multiplier, |
| 136 | int output_shift, |
| 137 | int32_t output_activation_min, |
| 138 | int32_t output_activation_max) |
| 139 | { |
| 140 | ColVectorMap bias_vector(bias_data, output_rows); |
| 141 | gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage; |
| 142 | bias_addition_stage.bias_vector = bias_vector; |
| 143 | gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint quantize_down_stage; |
| 144 | quantize_down_stage.result_offset_after_shift = output_offset; |
| 145 | quantize_down_stage.result_fixedpoint_multiplier = output_multiplier; |
| 146 | quantize_down_stage.result_shift = output_shift; |
| 147 | gemmlowp::OutputStageClamp clamp_stage; |
| 148 | clamp_stage.min = output_activation_min; |
| 149 | clamp_stage.max = output_activation_max; |
| 150 | gemmlowp::OutputStageSaturatingCastToUint8 saturating_cast_stage; |
| 151 | return std::make_tuple(bias_addition_stage, quantize_down_stage, clamp_stage, saturating_cast_stage); |
| 152 | } |
| 153 | }; |
| 154 | |
| 155 | static void GemmlowpBenchmark(benchmark::State& state, uint32_t threads) |
| 156 | { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 157 | const size_t mc = state.range(0); |
| 158 | const size_t nc = state.range(1); |
| 159 | const size_t kc = state.range(2); |
| 160 | |
| 161 | std::random_device random_device; |
| 162 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 163 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), std::ref(rng)); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 164 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 165 | |
| 166 | std::vector<uint8_t> a(mc * kc); |
| 167 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 168 | |
| 169 | const size_t kElements = nc * kc; |
| 170 | const size_t bElements = nc; |
| 171 | const size_t c_elements = mc * nc; |
| 172 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 173 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 174 | kElements * sizeof(uint8_t) + bElements * sizeof(int32_t) + c_elements * sizeof(uint8_t)); |
| 175 | |
| 176 | std::vector<uint8_t> k(kElements * num_buffers); |
| 177 | std::generate(k.begin(), k.end(), std::ref(u8rng)); |
| 178 | std::vector<int32_t> b(bElements * num_buffers); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 179 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 180 | std::vector<uint8_t> c(c_elements * num_buffers); |
| 181 | std::fill(c.begin(), c.end(), 0xA5); |
| 182 | |
| 183 | gemmlowp::MultiThreadGemmContext threadingContext; |
| 184 | threadingContext.set_max_num_threads(threads); |
| 185 | |
| 186 | size_t buffer_index = 0; |
| 187 | for (auto _ : state) { |
| 188 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 189 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 190 | buffer_index = (buffer_index + 1) % num_buffers; |
| 191 | state.ResumeTiming(); |
| 192 | |
| 193 | gemmlowp::MatrixMap<const uint8_t, gemmlowp::MapOrder::RowMajor> AM(a.data(), mc, kc, kc); |
| 194 | gemmlowp::MatrixMap<const uint8_t, gemmlowp::MapOrder::ColMajor> BM(k.data() + buffer_index * kElements, kc, nc, kc); |
| 195 | gemmlowp::MatrixMap<uint8_t, gemmlowp::MapOrder::RowMajor> CM(c.data() + buffer_index * c_elements, mc, nc, nc); |
| 196 | const auto& outputPipeline = GemmlowpOutputPipeline::Make(b.data() + buffer_index * bElements, nc, 127, 127, 127, 0, 255); |
| 197 | gemmlowp::GemmWithOutputPipeline<uint8_t, uint8_t, gemmlowp::L8R8WithLhsNonzeroBitDepthParams>( |
| 198 | &threadingContext, AM, BM, &CM, 127, 127, outputPipeline); |
| 199 | } |
| 200 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 201 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 202 | if (cpu_frequency != 0) { |
| 203 | state.counters["cpufreq"] = cpu_frequency; |
| 204 | } |
| 205 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 206 | state.counters["OPS"] = benchmark::Counter( |
| 207 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 208 | } |
| 209 | |
| 210 | static void gemmlowp_st(benchmark::State& state, const char* net) |
| 211 | { |
| 212 | GemmlowpBenchmark(state, 1); |
| 213 | } |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 214 | #endif // BENCHMARK_GEMMLOWP |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 215 | |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 216 | |
| 217 | #ifdef BENCHMARK_RUY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 218 | static void RuyBenchmark(benchmark::State& state, size_t threads) |
| 219 | { |
| 220 | const size_t mc = state.range(0); |
| 221 | const size_t nc = state.range(1); |
| 222 | const size_t kc = state.range(2); |
| 223 | |
| 224 | std::random_device random_device; |
| 225 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 226 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), std::ref(rng)); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 227 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 228 | |
| 229 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 230 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 231 | nc * (sizeof(uint8_t) * (mc + kc) + sizeof(int32_t))); |
| 232 | |
| 233 | std::vector<uint8_t> a(mc * kc); |
| 234 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 235 | std::vector<uint8_t> k(num_buffers * nc * kc); |
| 236 | std::generate(k.begin(), k.end(), std::ref(u8rng)); |
| 237 | std::vector<int32_t> b(num_buffers * nc); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 238 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 239 | std::vector<uint8_t> c(num_buffers * nc * mc); |
| 240 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 241 | |
| 242 | // Note: context must be static to avoid the cost of re-creating it for each benchmark. |
| 243 | static ruy::Context context; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 244 | context.set_max_num_threads(threads); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 245 | |
| 246 | ruy::Matrix<uint8_t> ruy_a; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 247 | ruy::MakeSimpleLayout(nc, kc, ruy::Order::kRowMajor, ruy_a.mutable_layout()); |
| 248 | ruy_a.set_zero_point(127); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 249 | ruy::Matrix<uint8_t> ruy_b; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 250 | ruy::MakeSimpleLayout(kc, mc, ruy::Order::kColMajor, ruy_b.mutable_layout()); |
| 251 | ruy_b.set_data(a.data()); |
| 252 | ruy_b.set_zero_point(127); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 253 | ruy::Matrix<uint8_t> ruy_c; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 254 | ruy::MakeSimpleLayout(nc, mc, ruy::Order::kColMajor, ruy_c.mutable_layout()); |
| 255 | ruy_c.set_zero_point(127); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 256 | |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 257 | ruy::MulParams<int32_t, uint8_t> mul_params; |
| 258 | mul_params.set_multiplier_fixedpoint(0x40000000); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 259 | |
| 260 | // ruy::Context uses deferred initialization, which affects percieved GEMM performance. Initialization happens during |
| 261 | // the first GEMM calls, and per Benoit Jacob it takes up to ~250 milliseconds for performance to stabilize. |
| 262 | // Thus, on the first benchmark, we compute GEMM for 500 milliseconds (to be safe) without recording performance, and |
| 263 | // keep the ruy::Context object initialized (by being static) between subsequent benchmarks. |
| 264 | static std::once_flag warmup; |
| 265 | std::call_once(warmup, [&](){ |
| 266 | auto start = std::chrono::steady_clock::now(); |
| 267 | do { |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 268 | ruy_a.set_data(k.data()); |
| 269 | ruy_c.set_data(c.data()); |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 270 | mul_params.set_bias(b.data()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 271 | |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 272 | ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 273 | } while (std::chrono::duration<double>(std::chrono::steady_clock::now() - start).count() < 0.5); |
| 274 | }); |
| 275 | |
| 276 | size_t buffer_index = 0; |
| 277 | for (auto _ : state) { |
| 278 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 279 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 280 | // - K is not in cache (for any cache level) |
| 281 | // - B is not in cache (for any cache level) |
| 282 | // - C is not in cache (for any cache level) |
| 283 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 284 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 285 | buffer_index = (buffer_index + 1) % num_buffers; |
| 286 | state.ResumeTiming(); |
| 287 | |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 288 | ruy_a.set_data(k.data() + buffer_index * nc * kc); |
| 289 | ruy_c.set_data(c.data() + buffer_index * mc * nc); |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 290 | mul_params.set_bias(b.data() + buffer_index * nc); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 291 | |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 292 | ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 293 | } |
| 294 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 295 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 296 | if (cpu_frequency != 0) { |
| 297 | state.counters["cpufreq"] = cpu_frequency; |
| 298 | } |
| 299 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 300 | state.counters["OPS"] = benchmark::Counter( |
| 301 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 302 | } |
| 303 | |
| 304 | static void ruy_st(benchmark::State& state, const char* net) |
| 305 | { |
| 306 | RuyBenchmark(state, 1); |
| 307 | } |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 308 | #endif // BENCHMARK_RUY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 309 | |
| 310 | |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 311 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
Frank Barchard | f82410d | 2022-02-02 01:01:38 -0800 | [diff] [blame] | 312 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53(benchmark::State& state, const char* net) { |
| 313 | GEMMBenchmark(state, |
| 314 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, |
| 315 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 316 | 4, 8, 1, 1, benchmark::utils::CheckNEON); |
| 317 | } |
| 318 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, const char* net) { |
| 319 | GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, |
| 320 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 321 | 4, 8, 1, 1, benchmark::utils::CheckNEON); |
| 322 | } |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 323 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7(benchmark::State& state, const char* net) { |
| 324 | GEMMBenchmark(state, |
| 325 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, |
| 326 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 327 | 4, 8, 1, 1, benchmark::utils::CheckNEON); |
| 328 | } |
| 329 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7(benchmark::State& state, const char* net) { |
| 330 | GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, |
| 331 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 332 | 4, 8, 1, 1, benchmark::utils::CheckNEON); |
| 333 | } |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 334 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { |
| 335 | GEMMBenchmark(state, |
| 336 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, |
| 337 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 338 | 4, 8, 1, 1, benchmark::utils::CheckNEON); |
| 339 | } |
| 340 | static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* net) { |
| 341 | GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, |
| 342 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 343 | 4, 8, 1, 1, benchmark::utils::CheckNEON); |
| 344 | } |
Frank Barchard | f82410d | 2022-02-02 01:01:38 -0800 | [diff] [blame] | 345 | BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53) |
| 346 | BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53) |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 347 | BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7) |
| 348 | BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7) |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 349 | BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64) |
| 350 | BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
| 351 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 352 | |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 353 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | efc3ccf | 2021-08-31 23:20:00 -0700 | [diff] [blame] | 354 | static void qu8_gemm_4x16c4__aarch64_neondot_cortex_a55(benchmark::State& state, const char* net) { |
Frank Barchard | ca4c68e | 2021-08-25 19:06:40 -0700 | [diff] [blame] | 355 | GEMMBenchmark(state, |
Frank Barchard | efc3ccf | 2021-08-31 23:20:00 -0700 | [diff] [blame] | 356 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, |
Frank Barchard | ca4c68e | 2021-08-25 19:06:40 -0700 | [diff] [blame] | 357 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | efc3ccf | 2021-08-31 23:20:00 -0700 | [diff] [blame] | 358 | 4, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | ca4c68e | 2021-08-25 19:06:40 -0700 | [diff] [blame] | 359 | } |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 360 | static void qu8_gemm_4x16c4__aarch64_neondot_ld128(benchmark::State& state, const char* net) { |
| 361 | GEMMBenchmark(state, |
| 362 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, |
| 363 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 364 | 4, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
| 365 | } |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 366 | static void qu8_gemm_4x8c4__aarch64_neondot_ld128(benchmark::State& state, const char* net) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 367 | GEMMBenchmark(state, |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 368 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 369 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 370 | 4, 8, 4, 1, |
| 371 | benchmark::utils::CheckNEONDOT); |
| 372 | } |
Frank Barchard | efc3ccf | 2021-08-31 23:20:00 -0700 | [diff] [blame] | 373 | static void qu8_gemm_4x8c4__aarch64_neondot_cortex_a55(benchmark::State& state, const char* net) { |
| 374 | GEMMBenchmark(state, |
| 375 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_cortex_a55, |
| 376 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 377 | 4, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
| 378 | } |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 379 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53(benchmark::State& state, const char* net) { |
| 380 | GEMMBenchmark(state, |
| 381 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, |
| 382 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 383 | 4, 16, 1, 1, |
| 384 | benchmark::utils::CheckNEON); |
| 385 | } |
| 386 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, const char* net) { |
| 387 | GEMMBenchmark(state, |
| 388 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, |
| 389 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 390 | 4, 16, 1, 1, |
| 391 | benchmark::utils::CheckNEON); |
| 392 | } |
Frank Barchard | 9cdc10d | 2021-11-22 19:03:54 -0800 | [diff] [blame] | 393 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { |
| 394 | GEMMBenchmark(state, |
| 395 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64, |
| 396 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 397 | 4, 16, 1, 1, |
| 398 | benchmark::utils::CheckNEON); |
| 399 | } |
| 400 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* net) { |
| 401 | GEMMBenchmark(state, |
| 402 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, |
| 403 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 404 | 4, 16, 1, 1, |
| 405 | benchmark::utils::CheckNEON); |
| 406 | } |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 407 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a75(benchmark::State& state, const char* net) { |
| 408 | GEMMBenchmark(state, |
| 409 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, |
| 410 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 411 | 4, 16, 1, 1, |
| 412 | benchmark::utils::CheckNEON); |
| 413 | } |
| 414 | static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75(benchmark::State& state, const char* net) { |
| 415 | GEMMBenchmark(state, |
| 416 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75, |
| 417 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 418 | 4, 16, 1, 1, |
| 419 | benchmark::utils::CheckNEON); |
| 420 | } |
Frank Barchard | efc3ccf | 2021-08-31 23:20:00 -0700 | [diff] [blame] | 421 | BENCHMARK_GEMM(qu8_gemm_4x8c4__aarch64_neondot_cortex_a55) |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 422 | BENCHMARK_GEMM(qu8_gemm_4x16c4__aarch64_neondot_cortex_a55) |
| 423 | BENCHMARK_GEMM(qu8_gemm_4x8c4__aarch64_neondot_ld128) |
| 424 | BENCHMARK_GEMM(qu8_gemm_4x16c4__aarch64_neondot_ld128) |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 425 | BENCHMARK_GEMM(qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53) |
| 426 | BENCHMARK_GEMM(qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53) |
Frank Barchard | 9cdc10d | 2021-11-22 19:03:54 -0800 | [diff] [blame] | 427 | BENCHMARK_GEMM(qu8_gemm_4x16__aarch64_neon_mlal_lane_ld64) |
| 428 | BENCHMARK_GEMM(qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64) |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 429 | BENCHMARK_GEMM(qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a75) |
| 430 | BENCHMARK_GEMM(qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75) |
| 431 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 432 | |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 433 | |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 434 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 435 | static void qu8_gemm_1x8c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 436 | GEMMBenchmark(state, |
| 437 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, |
| 438 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 439 | 1, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
| 440 | } |
| 441 | static void qu8_gemm_2x8c4__neondot(benchmark::State& state, const char* net) { |
| 442 | GEMMBenchmark(state, |
| 443 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, |
| 444 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 445 | 2, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
| 446 | } |
| 447 | static void qu8_gemm_3x8c4__neondot(benchmark::State& state, const char* net) { |
| 448 | GEMMBenchmark(state, |
| 449 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, |
| 450 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 451 | 3, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 452 | } |
| 453 | static void qu8_gemm_4x8c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 454 | GEMMBenchmark(state, |
| 455 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, |
| 456 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 457 | 4, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
| 458 | } |
| 459 | static void qu8_gemm_5x8c4__neondot(benchmark::State& state, const char* net) { |
| 460 | GEMMBenchmark(state, |
| 461 | xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, |
| 462 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 463 | 5, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 464 | } |
| 465 | static void qu8_gemm_6x8c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 466 | GEMMBenchmark(state, |
| 467 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot, |
| 468 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 469 | 6, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 470 | } |
| 471 | static void qu8_gemm_8x8c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 472 | GEMMBenchmark(state, |
| 473 | xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot, |
| 474 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 475 | 8, 8, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 476 | } |
| 477 | static void qu8_gemm_1x16c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 478 | GEMMBenchmark(state, |
| 479 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, |
| 480 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 481 | 1, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
| 482 | } |
| 483 | static void qu8_gemm_2x16c4__neondot(benchmark::State& state, const char* net) { |
| 484 | GEMMBenchmark(state, |
| 485 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, |
| 486 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 487 | 2, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
| 488 | } |
| 489 | static void qu8_gemm_3x16c4__neondot(benchmark::State& state, const char* net) { |
| 490 | GEMMBenchmark(state, |
| 491 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, |
| 492 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 493 | 3, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 494 | } |
| 495 | static void qu8_gemm_4x16c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 496 | GEMMBenchmark(state, |
| 497 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, |
| 498 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 499 | 4, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
| 500 | } |
| 501 | static void qu8_gemm_5x16c4__neondot(benchmark::State& state, const char* net) { |
| 502 | GEMMBenchmark(state, |
| 503 | xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, |
| 504 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 505 | 5, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 506 | } |
| 507 | static void qu8_gemm_6x16c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 508 | GEMMBenchmark(state, |
| 509 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, |
| 510 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 511 | 6, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 512 | } |
| 513 | static void qu8_gemm_8x16c4__neondot(benchmark::State& state, const char* net) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 514 | GEMMBenchmark(state, |
| 515 | xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, |
| 516 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 517 | 8, 16, 4, 1, benchmark::utils::CheckNEONDOT); |
| 518 | } |
| 519 | static void qu8_gemm_1x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 520 | GEMMBenchmark(state, |
| 521 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, |
| 522 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 523 | 1, 8, 1, 1, benchmark::utils::CheckNEON); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 524 | } |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 525 | static void qu8_gemm_2x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 526 | GEMMBenchmark(state, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 527 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, |
Marat Dukhan | afd2ed9 | 2021-07-27 00:21:05 -0700 | [diff] [blame] | 528 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 529 | 2, 8, 1, 1, benchmark::utils::CheckNEON); |
| 530 | } |
| 531 | static void qu8_gemm_3x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 532 | GEMMBenchmark(state, |
| 533 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, |
| 534 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 535 | 3, 8, 1, 1, benchmark::utils::CheckNEON); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 536 | } |
| 537 | static void qu8_gemm_4x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 538 | GEMMBenchmark(state, |
| 539 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane, |
| 540 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 541 | 4, 8, 1, 1, benchmark::utils::CheckNEON); |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 542 | } |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 543 | static void qu8_gemm_6x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 544 | GEMMBenchmark(state, |
| 545 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, |
| 546 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 547 | 6, 8, 1, 1, benchmark::utils::CheckNEON); |
| 548 | } |
| 549 | static void qu8_gemm_1x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 550 | GEMMBenchmark(state, |
| 551 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, |
| 552 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 553 | 1, 16, 1, 1, benchmark::utils::CheckNEON); |
| 554 | } |
| 555 | static void qu8_gemm_2x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 556 | GEMMBenchmark(state, |
| 557 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, |
| 558 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 559 | 2, 16, 1, 1, benchmark::utils::CheckNEON); |
| 560 | } |
| 561 | static void qu8_gemm_3x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 562 | GEMMBenchmark(state, |
| 563 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, |
| 564 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 565 | 3, 16, 1, 1, benchmark::utils::CheckNEON); |
| 566 | } |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 567 | static void qu8_gemm_4x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 568 | GEMMBenchmark(state, |
Marat Dukhan | afd2ed9 | 2021-07-27 00:21:05 -0700 | [diff] [blame] | 569 | xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, |
| 570 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 571 | 4, 16, 1, 1, benchmark::utils::CheckNEON); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 572 | } |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 573 | static void qu8_gemm_6x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 574 | GEMMBenchmark(state, |
| 575 | xnn_qu8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, |
| 576 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 577 | 6, 16, 1, 1, benchmark::utils::CheckNEON); |
| 578 | } |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 579 | static void qu8_gemm_1x32c4__neondot(benchmark::State& state, const char* net) { |
| 580 | GEMMBenchmark(state, |
| 581 | xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, |
| 582 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 583 | 1, 32, 4, 1, benchmark::utils::CheckNEONDOT); |
| 584 | } |
| 585 | static void qu8_gemm_2x32c4__neondot(benchmark::State& state, const char* net) { |
| 586 | GEMMBenchmark(state, |
| 587 | xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot, |
| 588 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 589 | 2, 32, 4, 1, benchmark::utils::CheckNEONDOT); |
| 590 | } |
| 591 | static void qu8_gemm_3x32c4__neondot(benchmark::State& state, const char* net) { |
| 592 | GEMMBenchmark(state, |
| 593 | xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot, |
| 594 | xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 595 | 3, 32, 4, 1, benchmark::utils::CheckNEONDOT); |
| 596 | } |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 597 | BENCHMARK_GEMM(qu8_gemm_1x8c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 598 | BENCHMARK_GEMM(qu8_gemm_2x8c4__neondot) |
| 599 | BENCHMARK_GEMM(qu8_gemm_3x8c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 600 | BENCHMARK_GEMM(qu8_gemm_4x8c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 601 | BENCHMARK_GEMM(qu8_gemm_5x8c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 602 | BENCHMARK_GEMM(qu8_gemm_6x8c4__neondot) |
| 603 | BENCHMARK_GEMM(qu8_gemm_8x8c4__neondot) |
| 604 | BENCHMARK_GEMM(qu8_gemm_1x16c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 605 | BENCHMARK_GEMM(qu8_gemm_2x16c4__neondot) |
| 606 | BENCHMARK_GEMM(qu8_gemm_3x16c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 607 | BENCHMARK_GEMM(qu8_gemm_4x16c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 608 | BENCHMARK_GEMM(qu8_gemm_5x16c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 609 | BENCHMARK_GEMM(qu8_gemm_6x16c4__neondot) |
| 610 | BENCHMARK_GEMM(qu8_gemm_8x16c4__neondot) |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 611 | BENCHMARK_GEMM(qu8_gemm_1x32c4__neondot) |
| 612 | BENCHMARK_GEMM(qu8_gemm_2x32c4__neondot) |
| 613 | BENCHMARK_GEMM(qu8_gemm_3x32c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 614 | BENCHMARK_GEMM(qu8_gemm_1x8__neon_mlal_lane) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 615 | BENCHMARK_GEMM(qu8_gemm_2x8__neon_mlal_lane) |
| 616 | BENCHMARK_GEMM(qu8_gemm_3x8__neon_mlal_lane) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 617 | BENCHMARK_GEMM(qu8_gemm_4x8__neon_mlal_lane) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 618 | BENCHMARK_GEMM(qu8_gemm_6x8__neon_mlal_lane) |
| 619 | BENCHMARK_GEMM(qu8_gemm_1x16__neon_mlal_lane) |
| 620 | BENCHMARK_GEMM(qu8_gemm_2x16__neon_mlal_lane) |
| 621 | BENCHMARK_GEMM(qu8_gemm_3x16__neon_mlal_lane) |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 622 | BENCHMARK_GEMM(qu8_gemm_4x16__neon_mlal_lane) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 623 | BENCHMARK_GEMM(qu8_gemm_6x16__neon_mlal_lane) |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 624 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 625 | |
Frank Barchard | ca4c68e | 2021-08-25 19:06:40 -0700 | [diff] [blame] | 626 | |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 627 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 628 | static void qu8_gemm_1x16c8__avx512skx(benchmark::State& state, const char* net) { |
| 629 | GEMMBenchmark(state, |
| 630 | xnn_qu8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx, |
| 631 | xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 632 | 1, 16, 8, 1, |
| 633 | benchmark::utils::CheckAVX512SKX); |
| 634 | } |
| 635 | static void qu8_gemm_2x16c8__avx512skx(benchmark::State& state, const char* net) { |
| 636 | GEMMBenchmark(state, |
| 637 | xnn_qu8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx, |
| 638 | xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 639 | 2, 16, 8, 1, |
| 640 | benchmark::utils::CheckAVX512SKX); |
| 641 | } |
| 642 | static void qu8_gemm_3x16c8__avx512skx(benchmark::State& state, const char* net) { |
| 643 | GEMMBenchmark(state, |
| 644 | xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx, |
| 645 | xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 646 | 3, 16, 8, 1, |
| 647 | benchmark::utils::CheckAVX512SKX); |
| 648 | } |
| 649 | static void qu8_gemm_4x16c8__avx512skx(benchmark::State& state, const char* net) { |
| 650 | GEMMBenchmark(state, |
| 651 | xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx, |
| 652 | xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 653 | 4, 16, 8, 1, |
| 654 | benchmark::utils::CheckAVX512SKX); |
| 655 | } |
| 656 | static void qu8_gemm_1x8c8__avx2(benchmark::State& state, const char* net) { |
| 657 | GEMMBenchmark(state, |
| 658 | xnn_qu8_gemm_minmax_fp32_ukernel_1x8c8__avx2, |
| 659 | xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 660 | 1, 8, 8, 1, |
| 661 | benchmark::utils::CheckAVX2); |
| 662 | } |
| 663 | static void qu8_gemm_2x8c8__avx2(benchmark::State& state, const char* net) { |
| 664 | GEMMBenchmark(state, |
| 665 | xnn_qu8_gemm_minmax_fp32_ukernel_2x8c8__avx2, |
| 666 | xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 667 | 2, 8, 8, 1, |
| 668 | benchmark::utils::CheckAVX2); |
| 669 | } |
| 670 | static void qu8_gemm_3x8c8__avx2(benchmark::State& state, const char* net) { |
| 671 | GEMMBenchmark(state, |
| 672 | xnn_qu8_gemm_minmax_fp32_ukernel_3x8c8__avx2, |
| 673 | xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 674 | 3, 8, 8, 1, |
| 675 | benchmark::utils::CheckAVX2); |
| 676 | } |
| 677 | static void qu8_gemm_1x4c2__xop_ld64(benchmark::State& state, const char* net) { |
| 678 | GEMMBenchmark(state, |
| 679 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64, |
| 680 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 681 | 1, 4, 2, 1, |
| 682 | benchmark::utils::CheckXOP); |
| 683 | } |
| 684 | static void qu8_gemm_2x4c2__xop_ld64(benchmark::State& state, const char* net) { |
| 685 | GEMMBenchmark(state, |
| 686 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64, |
| 687 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 688 | 2, 4, 2, 1, |
| 689 | benchmark::utils::CheckXOP); |
| 690 | } |
| 691 | static void qu8_gemm_3x4c2__xop_ld64(benchmark::State& state, const char* net) { |
| 692 | GEMMBenchmark(state, |
| 693 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64, |
| 694 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 695 | 3, 4, 2, 1, |
| 696 | benchmark::utils::CheckXOP); |
| 697 | } |
| 698 | static void qu8_gemm_4x4c2__xop_ld64(benchmark::State& state, const char* net) { |
| 699 | GEMMBenchmark(state, |
| 700 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64, |
| 701 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 702 | 4, 4, 2, 1, |
| 703 | benchmark::utils::CheckXOP); |
| 704 | } |
| 705 | static void qu8_gemm_1x4c2__xop_ld128(benchmark::State& state, const char* net) { |
| 706 | GEMMBenchmark(state, |
| 707 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128, |
| 708 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 709 | 1, 4, 2, 1, |
| 710 | benchmark::utils::CheckXOP); |
| 711 | } |
| 712 | static void qu8_gemm_2x4c2__xop_ld128(benchmark::State& state, const char* net) { |
| 713 | GEMMBenchmark(state, |
| 714 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128, |
| 715 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 716 | 2, 4, 2, 1, |
| 717 | benchmark::utils::CheckXOP); |
| 718 | } |
| 719 | static void qu8_gemm_3x4c2__xop_ld128(benchmark::State& state, const char* net) { |
| 720 | GEMMBenchmark(state, |
| 721 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128, |
| 722 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 723 | 3, 4, 2, 1, |
| 724 | benchmark::utils::CheckXOP); |
| 725 | } |
| 726 | static void qu8_gemm_4x4c2__xop_ld128(benchmark::State& state, const char* net) { |
| 727 | GEMMBenchmark(state, |
| 728 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128, |
| 729 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 730 | 4, 4, 2, 1, |
| 731 | benchmark::utils::CheckXOP); |
| 732 | } |
| 733 | static void qu8_gemm_1x4c8__xop_ld64(benchmark::State& state, const char* net) { |
| 734 | GEMMBenchmark(state, |
| 735 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64, |
| 736 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 737 | 1, 4, 8, 1, |
| 738 | benchmark::utils::CheckXOP); |
| 739 | } |
| 740 | static void qu8_gemm_2x4c8__xop_ld64(benchmark::State& state, const char* net) { |
| 741 | GEMMBenchmark(state, |
| 742 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64, |
| 743 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 744 | 2, 4, 8, 1, |
| 745 | benchmark::utils::CheckXOP); |
| 746 | } |
| 747 | static void qu8_gemm_3x4c8__xop_ld64(benchmark::State& state, const char* net) { |
| 748 | GEMMBenchmark(state, |
| 749 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64, |
| 750 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 751 | 3, 4, 8, 1, |
| 752 | benchmark::utils::CheckXOP); |
| 753 | } |
| 754 | static void qu8_gemm_1x4c8__xop_ld128(benchmark::State& state, const char* net) { |
| 755 | GEMMBenchmark(state, |
| 756 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128, |
| 757 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 758 | 1, 4, 8, 1, |
| 759 | benchmark::utils::CheckXOP); |
| 760 | } |
| 761 | static void qu8_gemm_2x4c8__xop_ld128(benchmark::State& state, const char* net) { |
| 762 | GEMMBenchmark(state, |
| 763 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128, |
| 764 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 765 | 2, 4, 8, 1, |
| 766 | benchmark::utils::CheckXOP); |
| 767 | } |
| 768 | static void qu8_gemm_3x4c8__xop_ld128(benchmark::State& state, const char* net) { |
| 769 | GEMMBenchmark(state, |
| 770 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128, |
| 771 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 772 | 3, 4, 8, 1, |
| 773 | benchmark::utils::CheckXOP); |
| 774 | } |
| 775 | static void qu8_gemm_1x4c2__avx_ld64(benchmark::State& state, const char* net) { |
| 776 | GEMMBenchmark(state, |
| 777 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64, |
| 778 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 779 | 1, 4, 2, 1, |
| 780 | benchmark::utils::CheckAVX); |
| 781 | } |
| 782 | static void qu8_gemm_2x4c2__avx_ld64(benchmark::State& state, const char* net) { |
| 783 | GEMMBenchmark(state, |
| 784 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64, |
| 785 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 786 | 2, 4, 2, 1, |
| 787 | benchmark::utils::CheckAVX); |
| 788 | } |
| 789 | static void qu8_gemm_3x4c2__avx_ld64(benchmark::State& state, const char* net) { |
| 790 | GEMMBenchmark(state, |
| 791 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64, |
| 792 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 793 | 3, 4, 2, 1, |
| 794 | benchmark::utils::CheckAVX); |
| 795 | } |
| 796 | static void qu8_gemm_4x4c2__avx_ld64(benchmark::State& state, const char* net) { |
| 797 | GEMMBenchmark(state, |
| 798 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64, |
| 799 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 800 | 4, 4, 2, 1, |
| 801 | benchmark::utils::CheckAVX); |
| 802 | } |
| 803 | static void qu8_gemm_1x4c2__avx_ld128(benchmark::State& state, const char* net) { |
| 804 | GEMMBenchmark(state, |
| 805 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128, |
| 806 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 807 | 1, 4, 2, 1, |
| 808 | benchmark::utils::CheckAVX); |
| 809 | } |
| 810 | static void qu8_gemm_2x4c2__avx_ld128(benchmark::State& state, const char* net) { |
| 811 | GEMMBenchmark(state, |
| 812 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128, |
| 813 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 814 | 2, 4, 2, 1, |
| 815 | benchmark::utils::CheckAVX); |
| 816 | } |
| 817 | static void qu8_gemm_3x4c2__avx_ld128(benchmark::State& state, const char* net) { |
| 818 | GEMMBenchmark(state, |
| 819 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128, |
| 820 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 821 | 3, 4, 2, 1, |
| 822 | benchmark::utils::CheckAVX); |
| 823 | } |
| 824 | static void qu8_gemm_4x4c2__avx_ld128(benchmark::State& state, const char* net) { |
| 825 | GEMMBenchmark(state, |
| 826 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128, |
| 827 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 828 | 4, 4, 2, 1, |
| 829 | benchmark::utils::CheckAVX); |
| 830 | } |
| 831 | static void qu8_gemm_1x4c8__avx_ld64(benchmark::State& state, const char* net) { |
| 832 | GEMMBenchmark(state, |
| 833 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64, |
| 834 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 835 | 1, 4, 8, 1, |
| 836 | benchmark::utils::CheckAVX); |
| 837 | } |
| 838 | static void qu8_gemm_2x4c8__avx_ld64(benchmark::State& state, const char* net) { |
| 839 | GEMMBenchmark(state, |
| 840 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64, |
| 841 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 842 | 2, 4, 8, 1, |
| 843 | benchmark::utils::CheckAVX); |
| 844 | } |
| 845 | static void qu8_gemm_3x4c8__avx_ld64(benchmark::State& state, const char* net) { |
| 846 | GEMMBenchmark(state, |
| 847 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64, |
| 848 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 849 | 3, 4, 8, 1, |
| 850 | benchmark::utils::CheckAVX); |
| 851 | } |
| 852 | static void qu8_gemm_1x4c8__avx_ld128(benchmark::State& state, const char* net) { |
| 853 | GEMMBenchmark(state, |
| 854 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128, |
| 855 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 856 | 1, 4, 8, 1, |
| 857 | benchmark::utils::CheckAVX); |
| 858 | } |
| 859 | static void qu8_gemm_2x4c8__avx_ld128(benchmark::State& state, const char* net) { |
| 860 | GEMMBenchmark(state, |
| 861 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128, |
| 862 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 863 | 2, 4, 8, 1, |
| 864 | benchmark::utils::CheckAVX); |
| 865 | } |
| 866 | static void qu8_gemm_3x4c8__avx_ld128(benchmark::State& state, const char* net) { |
| 867 | GEMMBenchmark(state, |
| 868 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128, |
| 869 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 870 | 3, 4, 8, 1, |
| 871 | benchmark::utils::CheckAVX); |
| 872 | } |
| 873 | static void qu8_gemm_1x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
| 874 | GEMMBenchmark(state, |
| 875 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64, |
| 876 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 877 | 1, 4, 2, 1, |
| 878 | benchmark::utils::CheckSSE41); |
| 879 | } |
| 880 | static void qu8_gemm_2x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
| 881 | GEMMBenchmark(state, |
| 882 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64, |
| 883 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 884 | 2, 4, 2, 1, |
| 885 | benchmark::utils::CheckSSE41); |
| 886 | } |
| 887 | static void qu8_gemm_3x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
| 888 | GEMMBenchmark(state, |
| 889 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64, |
| 890 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 891 | 3, 4, 2, 1, |
| 892 | benchmark::utils::CheckSSE41); |
| 893 | } |
| 894 | static void qu8_gemm_4x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
| 895 | GEMMBenchmark(state, |
| 896 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64, |
| 897 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 898 | 4, 4, 2, 1, |
| 899 | benchmark::utils::CheckSSE41); |
| 900 | } |
| 901 | static void qu8_gemm_1x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
| 902 | GEMMBenchmark(state, |
| 903 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128, |
| 904 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 905 | 1, 4, 2, 1, |
| 906 | benchmark::utils::CheckSSE41); |
| 907 | } |
| 908 | static void qu8_gemm_2x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
| 909 | GEMMBenchmark(state, |
| 910 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128, |
| 911 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 912 | 2, 4, 2, 1, |
| 913 | benchmark::utils::CheckSSE41); |
| 914 | } |
| 915 | static void qu8_gemm_3x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
| 916 | GEMMBenchmark(state, |
| 917 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128, |
| 918 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 919 | 3, 4, 2, 1, |
| 920 | benchmark::utils::CheckSSE41); |
| 921 | } |
| 922 | static void qu8_gemm_4x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
| 923 | GEMMBenchmark(state, |
| 924 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128, |
| 925 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 926 | 4, 4, 2, 1, |
| 927 | benchmark::utils::CheckSSE41); |
| 928 | } |
| 929 | static void qu8_gemm_1x4c8__sse41_ld64(benchmark::State& state, const char* net) { |
| 930 | GEMMBenchmark(state, |
| 931 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64, |
| 932 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 933 | 1, 4, 8, 1, |
| 934 | benchmark::utils::CheckSSE41); |
| 935 | } |
| 936 | static void qu8_gemm_2x4c8__sse41_ld64(benchmark::State& state, const char* net) { |
| 937 | GEMMBenchmark(state, |
| 938 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64, |
| 939 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 940 | 2, 4, 8, 1, |
| 941 | benchmark::utils::CheckSSE41); |
| 942 | } |
| 943 | static void qu8_gemm_3x4c8__sse41_ld64(benchmark::State& state, const char* net) { |
| 944 | GEMMBenchmark(state, |
| 945 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64, |
| 946 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 947 | 3, 4, 8, 1, |
| 948 | benchmark::utils::CheckSSE41); |
| 949 | } |
| 950 | static void qu8_gemm_1x4c8__sse41_ld128(benchmark::State& state, const char* net) { |
| 951 | GEMMBenchmark(state, |
| 952 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128, |
| 953 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 954 | 1, 4, 8, 1, |
| 955 | benchmark::utils::CheckSSE41); |
| 956 | } |
| 957 | static void qu8_gemm_2x4c8__sse41_ld128(benchmark::State& state, const char* net) { |
| 958 | GEMMBenchmark(state, |
| 959 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128, |
| 960 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 961 | 2, 4, 8, 1, |
| 962 | benchmark::utils::CheckSSE41); |
| 963 | } |
| 964 | static void qu8_gemm_3x4c8__sse41_ld128(benchmark::State& state, const char* net) { |
| 965 | GEMMBenchmark(state, |
| 966 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128, |
| 967 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 968 | 3, 4, 8, 1, |
| 969 | benchmark::utils::CheckSSE41); |
| 970 | } |
| 971 | static void qu8_gemm_1x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
| 972 | GEMMBenchmark(state, |
| 973 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64, |
| 974 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 975 | 1, 4, 2, 1); |
| 976 | } |
| 977 | static void qu8_gemm_2x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
| 978 | GEMMBenchmark(state, |
| 979 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64, |
| 980 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 981 | 2, 4, 2, 1); |
| 982 | } |
| 983 | static void qu8_gemm_3x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
| 984 | GEMMBenchmark(state, |
| 985 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64, |
| 986 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 987 | 3, 4, 2, 1); |
| 988 | } |
| 989 | static void qu8_gemm_4x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
| 990 | GEMMBenchmark(state, |
| 991 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64, |
| 992 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 993 | 4, 4, 2, 1); |
| 994 | } |
| 995 | static void qu8_gemm_1x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
| 996 | GEMMBenchmark(state, |
| 997 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128, |
| 998 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 999 | 1, 4, 2, 1); |
| 1000 | } |
| 1001 | static void qu8_gemm_2x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
| 1002 | GEMMBenchmark(state, |
| 1003 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128, |
| 1004 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1005 | 2, 4, 2, 1); |
| 1006 | } |
| 1007 | static void qu8_gemm_3x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
| 1008 | GEMMBenchmark(state, |
| 1009 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128, |
| 1010 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1011 | 3, 4, 2, 1); |
| 1012 | } |
| 1013 | static void qu8_gemm_4x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
| 1014 | GEMMBenchmark(state, |
| 1015 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128, |
| 1016 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1017 | 4, 4, 2, 1); |
| 1018 | } |
| 1019 | static void qu8_gemm_1x4c8__sse2_ld64(benchmark::State& state, const char* net) { |
| 1020 | GEMMBenchmark(state, |
| 1021 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64, |
| 1022 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1023 | 1, 4, 8, 1); |
| 1024 | } |
| 1025 | static void qu8_gemm_2x4c8__sse2_ld64(benchmark::State& state, const char* net) { |
| 1026 | GEMMBenchmark(state, |
| 1027 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64, |
| 1028 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1029 | 2, 4, 8, 1); |
| 1030 | } |
| 1031 | static void qu8_gemm_3x4c8__sse2_ld64(benchmark::State& state, const char* net) { |
| 1032 | GEMMBenchmark(state, |
| 1033 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64, |
| 1034 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1035 | 3, 4, 8, 1); |
| 1036 | } |
| 1037 | static void qu8_gemm_1x4c8__sse2_ld128(benchmark::State& state, const char* net) { |
| 1038 | GEMMBenchmark(state, |
| 1039 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128, |
| 1040 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1041 | 1, 4, 8, 1); |
| 1042 | } |
| 1043 | static void qu8_gemm_2x4c8__sse2_ld128(benchmark::State& state, const char* net) { |
| 1044 | GEMMBenchmark(state, |
| 1045 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128, |
| 1046 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1047 | 2, 4, 8, 1); |
| 1048 | } |
| 1049 | static void qu8_gemm_3x4c8__sse2_ld128(benchmark::State& state, const char* net) { |
| 1050 | GEMMBenchmark(state, |
| 1051 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128, |
| 1052 | xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 1053 | 3, 4, 8, 1); |
| 1054 | } |
| 1055 | |
| 1056 | BENCHMARK_GEMM(qu8_gemm_1x16c8__avx512skx) |
| 1057 | BENCHMARK_GEMM(qu8_gemm_2x16c8__avx512skx) |
| 1058 | BENCHMARK_GEMM(qu8_gemm_3x16c8__avx512skx) |
| 1059 | BENCHMARK_GEMM(qu8_gemm_4x16c8__avx512skx) |
| 1060 | |
| 1061 | BENCHMARK_GEMM(qu8_gemm_1x8c8__avx2) |
| 1062 | BENCHMARK_GEMM(qu8_gemm_2x8c8__avx2) |
| 1063 | BENCHMARK_GEMM(qu8_gemm_3x8c8__avx2) |
| 1064 | |
| 1065 | BENCHMARK_GEMM(qu8_gemm_1x4c2__xop_ld64) |
| 1066 | BENCHMARK_GEMM(qu8_gemm_2x4c2__xop_ld64) |
| 1067 | BENCHMARK_GEMM(qu8_gemm_3x4c2__xop_ld64) |
| 1068 | BENCHMARK_GEMM(qu8_gemm_4x4c2__xop_ld64) |
| 1069 | BENCHMARK_GEMM(qu8_gemm_1x4c2__xop_ld128) |
| 1070 | BENCHMARK_GEMM(qu8_gemm_2x4c2__xop_ld128) |
| 1071 | BENCHMARK_GEMM(qu8_gemm_3x4c2__xop_ld128) |
| 1072 | BENCHMARK_GEMM(qu8_gemm_4x4c2__xop_ld128) |
| 1073 | BENCHMARK_GEMM(qu8_gemm_1x4c8__xop_ld64) |
| 1074 | BENCHMARK_GEMM(qu8_gemm_2x4c8__xop_ld64) |
| 1075 | BENCHMARK_GEMM(qu8_gemm_3x4c8__xop_ld64) |
| 1076 | BENCHMARK_GEMM(qu8_gemm_1x4c8__xop_ld128) |
| 1077 | BENCHMARK_GEMM(qu8_gemm_2x4c8__xop_ld128) |
| 1078 | BENCHMARK_GEMM(qu8_gemm_3x4c8__xop_ld128) |
| 1079 | |
| 1080 | BENCHMARK_GEMM(qu8_gemm_1x4c2__avx_ld64) |
| 1081 | BENCHMARK_GEMM(qu8_gemm_2x4c2__avx_ld64) |
| 1082 | BENCHMARK_GEMM(qu8_gemm_3x4c2__avx_ld64) |
| 1083 | BENCHMARK_GEMM(qu8_gemm_4x4c2__avx_ld64) |
| 1084 | BENCHMARK_GEMM(qu8_gemm_1x4c2__avx_ld128) |
| 1085 | BENCHMARK_GEMM(qu8_gemm_2x4c2__avx_ld128) |
| 1086 | BENCHMARK_GEMM(qu8_gemm_3x4c2__avx_ld128) |
| 1087 | BENCHMARK_GEMM(qu8_gemm_4x4c2__avx_ld128) |
| 1088 | BENCHMARK_GEMM(qu8_gemm_1x4c8__avx_ld64) |
| 1089 | BENCHMARK_GEMM(qu8_gemm_2x4c8__avx_ld64) |
| 1090 | BENCHMARK_GEMM(qu8_gemm_3x4c8__avx_ld64) |
| 1091 | BENCHMARK_GEMM(qu8_gemm_1x4c8__avx_ld128) |
| 1092 | BENCHMARK_GEMM(qu8_gemm_2x4c8__avx_ld128) |
| 1093 | BENCHMARK_GEMM(qu8_gemm_3x4c8__avx_ld128) |
| 1094 | |
| 1095 | BENCHMARK_GEMM(qu8_gemm_1x4c2__sse41_ld64) |
| 1096 | BENCHMARK_GEMM(qu8_gemm_2x4c2__sse41_ld64) |
| 1097 | BENCHMARK_GEMM(qu8_gemm_3x4c2__sse41_ld64) |
| 1098 | BENCHMARK_GEMM(qu8_gemm_4x4c2__sse41_ld64) |
| 1099 | BENCHMARK_GEMM(qu8_gemm_1x4c2__sse41_ld128) |
| 1100 | BENCHMARK_GEMM(qu8_gemm_2x4c2__sse41_ld128) |
| 1101 | BENCHMARK_GEMM(qu8_gemm_3x4c2__sse41_ld128) |
| 1102 | BENCHMARK_GEMM(qu8_gemm_4x4c2__sse41_ld128) |
| 1103 | BENCHMARK_GEMM(qu8_gemm_1x4c8__sse41_ld64) |
| 1104 | BENCHMARK_GEMM(qu8_gemm_2x4c8__sse41_ld64) |
| 1105 | BENCHMARK_GEMM(qu8_gemm_3x4c8__sse41_ld64) |
| 1106 | BENCHMARK_GEMM(qu8_gemm_1x4c8__sse41_ld128) |
| 1107 | BENCHMARK_GEMM(qu8_gemm_2x4c8__sse41_ld128) |
| 1108 | BENCHMARK_GEMM(qu8_gemm_3x4c8__sse41_ld128) |
| 1109 | |
| 1110 | BENCHMARK_GEMM(qu8_gemm_1x4c2__sse2_ld64) |
| 1111 | BENCHMARK_GEMM(qu8_gemm_2x4c2__sse2_ld64) |
| 1112 | BENCHMARK_GEMM(qu8_gemm_3x4c2__sse2_ld64) |
| 1113 | BENCHMARK_GEMM(qu8_gemm_4x4c2__sse2_ld64) |
| 1114 | BENCHMARK_GEMM(qu8_gemm_1x4c2__sse2_ld128) |
| 1115 | BENCHMARK_GEMM(qu8_gemm_2x4c2__sse2_ld128) |
| 1116 | BENCHMARK_GEMM(qu8_gemm_3x4c2__sse2_ld128) |
| 1117 | BENCHMARK_GEMM(qu8_gemm_4x4c2__sse2_ld128) |
| 1118 | BENCHMARK_GEMM(qu8_gemm_1x4c8__sse2_ld64) |
| 1119 | BENCHMARK_GEMM(qu8_gemm_2x4c8__sse2_ld64) |
| 1120 | BENCHMARK_GEMM(qu8_gemm_3x4c8__sse2_ld64) |
| 1121 | BENCHMARK_GEMM(qu8_gemm_1x4c8__sse2_ld128) |
| 1122 | BENCHMARK_GEMM(qu8_gemm_2x4c8__sse2_ld128) |
| 1123 | BENCHMARK_GEMM(qu8_gemm_3x4c8__sse2_ld128) |
Marat Dukhan | 725f47e | 2021-05-22 10:06:19 -0700 | [diff] [blame] | 1124 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1125 | |
| 1126 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1127 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 348c377 | 2022-02-01 00:36:50 -0800 | [diff] [blame] | 1128 | static void qu8_gemm_1x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1129 | GEMMBenchmark(state, |
| 1130 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64, |
| 1131 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1132 | 1, 4, 2, 1); |
| 1133 | } |
| 1134 | static void qu8_gemm_2x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1135 | GEMMBenchmark(state, |
| 1136 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64, |
| 1137 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1138 | 2, 4, 2, 1); |
| 1139 | } |
| 1140 | static void qu8_gemm_3x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1141 | GEMMBenchmark(state, |
| 1142 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64, |
| 1143 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1144 | 3, 4, 2, 1); |
| 1145 | } |
| 1146 | static void qu8_gemm_4x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1147 | GEMMBenchmark(state, |
| 1148 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64, |
| 1149 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1150 | 4, 4, 2, 1); |
| 1151 | } |
| 1152 | |
| 1153 | static void qu8_gemm_1x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1154 | GEMMBenchmark(state, |
| 1155 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128, |
| 1156 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1157 | 1, 4, 2, 1); |
| 1158 | } |
| 1159 | static void qu8_gemm_2x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1160 | GEMMBenchmark(state, |
| 1161 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128, |
| 1162 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1163 | 2, 4, 2, 1); |
| 1164 | } |
| 1165 | static void qu8_gemm_3x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1166 | GEMMBenchmark(state, |
| 1167 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128, |
| 1168 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1169 | 3, 4, 2, 1); |
| 1170 | } |
| 1171 | static void qu8_gemm_4x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1172 | GEMMBenchmark(state, |
| 1173 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128, |
| 1174 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1175 | 4, 4, 2, 1); |
| 1176 | } |
| 1177 | |
| 1178 | static void qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1179 | GEMMBenchmark(state, |
| 1180 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64, |
| 1181 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1182 | 1, 4, 2, 4); |
| 1183 | } |
| 1184 | static void qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1185 | GEMMBenchmark(state, |
| 1186 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64, |
| 1187 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1188 | 2, 4, 2, 4); |
| 1189 | } |
| 1190 | static void qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1191 | GEMMBenchmark(state, |
| 1192 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64, |
| 1193 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1194 | 3, 4, 2, 4); |
| 1195 | } |
| 1196 | static void qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1197 | GEMMBenchmark(state, |
| 1198 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64, |
| 1199 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1200 | 4, 4, 2, 4); |
| 1201 | } |
| 1202 | |
| 1203 | static void qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1204 | GEMMBenchmark(state, |
| 1205 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128, |
| 1206 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1207 | 1, 4, 2, 4); |
| 1208 | } |
| 1209 | static void qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1210 | GEMMBenchmark(state, |
| 1211 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128, |
| 1212 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1213 | 2, 4, 2, 4); |
| 1214 | } |
| 1215 | static void qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1216 | GEMMBenchmark(state, |
| 1217 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128, |
| 1218 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1219 | 3, 4, 2, 4); |
| 1220 | } |
| 1221 | static void qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1222 | GEMMBenchmark(state, |
| 1223 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128, |
| 1224 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1225 | 4, 4, 2, 4); |
| 1226 | } |
| 1227 | |
| 1228 | static void qu8_gemm_1x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1229 | GEMMBenchmark(state, |
| 1230 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64, |
| 1231 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1232 | 1, 4, 8, 1); |
| 1233 | } |
| 1234 | static void qu8_gemm_2x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1235 | GEMMBenchmark(state, |
| 1236 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64, |
| 1237 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1238 | 2, 4, 8, 1); |
| 1239 | } |
| 1240 | static void qu8_gemm_3x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1241 | GEMMBenchmark(state, |
| 1242 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64, |
| 1243 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1244 | 3, 4, 8, 1); |
| 1245 | } |
| 1246 | static void qu8_gemm_4x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1247 | GEMMBenchmark(state, |
| 1248 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64, |
| 1249 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1250 | 4, 4, 8, 1); |
| 1251 | } |
| 1252 | |
| 1253 | static void qu8_gemm_1x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1254 | GEMMBenchmark(state, |
| 1255 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128, |
| 1256 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1257 | 1, 4, 8, 1); |
| 1258 | } |
| 1259 | static void qu8_gemm_2x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1260 | GEMMBenchmark(state, |
| 1261 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128, |
| 1262 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1263 | 2, 4, 8, 1); |
| 1264 | } |
| 1265 | static void qu8_gemm_3x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1266 | GEMMBenchmark(state, |
| 1267 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128, |
| 1268 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1269 | 3, 4, 8, 1); |
| 1270 | } |
| 1271 | static void qu8_gemm_4x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1272 | GEMMBenchmark(state, |
| 1273 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128, |
| 1274 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1275 | 4, 4, 8, 1); |
| 1276 | } |
| 1277 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1278 | static void qu8_gemm_1x4c8__wasmsimd_mul32_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1279 | GEMMBenchmark(state, |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1280 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld64, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1281 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1282 | 1, 4, 8, 1); |
| 1283 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1284 | static void qu8_gemm_2x4c8__wasmsimd_mul32_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1285 | GEMMBenchmark(state, |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1286 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld64, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1287 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1288 | 2, 4, 8, 1); |
| 1289 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1290 | static void qu8_gemm_3x4c8__wasmsimd_mul32_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1291 | GEMMBenchmark(state, |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1292 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld64, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1293 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1294 | 3, 4, 8, 1); |
| 1295 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1296 | static void qu8_gemm_1x4c8__wasmsimd_mul32_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1297 | GEMMBenchmark(state, |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1298 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld128, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1299 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1300 | 1, 4, 8, 1); |
| 1301 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1302 | static void qu8_gemm_2x4c8__wasmsimd_mul32_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1303 | GEMMBenchmark(state, |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1304 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld128, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1305 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1306 | 2, 4, 8, 1); |
| 1307 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1308 | static void qu8_gemm_3x4c8__wasmsimd_mul32_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1309 | GEMMBenchmark(state, |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1310 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1311 | xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 1312 | 3, 4, 8, 1); |
| 1313 | } |
| 1314 | |
Marat Dukhan | 348c377 | 2022-02-01 00:36:50 -0800 | [diff] [blame] | 1315 | BENCHMARK_GEMM(qu8_gemm_1x4c2__wasmsimd_dot16x2_ld64) |
| 1316 | BENCHMARK_GEMM(qu8_gemm_2x4c2__wasmsimd_dot16x2_ld64) |
| 1317 | BENCHMARK_GEMM(qu8_gemm_3x4c2__wasmsimd_dot16x2_ld64) |
| 1318 | BENCHMARK_GEMM(qu8_gemm_4x4c2__wasmsimd_dot16x2_ld64) |
| 1319 | BENCHMARK_GEMM(qu8_gemm_1x4c2__wasmsimd_dot16x2_ld128) |
| 1320 | BENCHMARK_GEMM(qu8_gemm_2x4c2__wasmsimd_dot16x2_ld128) |
| 1321 | BENCHMARK_GEMM(qu8_gemm_3x4c2__wasmsimd_dot16x2_ld128) |
| 1322 | BENCHMARK_GEMM(qu8_gemm_4x4c2__wasmsimd_dot16x2_ld128) |
| 1323 | |
| 1324 | BENCHMARK_GEMM(qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld64) |
| 1325 | BENCHMARK_GEMM(qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64) |
| 1326 | BENCHMARK_GEMM(qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64) |
| 1327 | BENCHMARK_GEMM(qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64) |
| 1328 | BENCHMARK_GEMM(qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld128) |
| 1329 | BENCHMARK_GEMM(qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128) |
| 1330 | BENCHMARK_GEMM(qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128) |
| 1331 | BENCHMARK_GEMM(qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128) |
| 1332 | |
| 1333 | BENCHMARK_GEMM(qu8_gemm_1x4c8__wasmsimd_dot16x2_ld64) |
| 1334 | BENCHMARK_GEMM(qu8_gemm_2x4c8__wasmsimd_dot16x2_ld64) |
| 1335 | BENCHMARK_GEMM(qu8_gemm_3x4c8__wasmsimd_dot16x2_ld64) |
| 1336 | BENCHMARK_GEMM(qu8_gemm_4x4c8__wasmsimd_dot16x2_ld64) |
| 1337 | BENCHMARK_GEMM(qu8_gemm_1x4c8__wasmsimd_dot16x2_ld128) |
| 1338 | BENCHMARK_GEMM(qu8_gemm_2x4c8__wasmsimd_dot16x2_ld128) |
| 1339 | BENCHMARK_GEMM(qu8_gemm_3x4c8__wasmsimd_dot16x2_ld128) |
| 1340 | BENCHMARK_GEMM(qu8_gemm_4x4c8__wasmsimd_dot16x2_ld128) |
| 1341 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1342 | BENCHMARK_GEMM(qu8_gemm_1x4c8__wasmsimd_mul32_ld64) |
| 1343 | BENCHMARK_GEMM(qu8_gemm_2x4c8__wasmsimd_mul32_ld64) |
| 1344 | BENCHMARK_GEMM(qu8_gemm_3x4c8__wasmsimd_mul32_ld64) |
| 1345 | BENCHMARK_GEMM(qu8_gemm_1x4c8__wasmsimd_mul32_ld128) |
| 1346 | BENCHMARK_GEMM(qu8_gemm_2x4c8__wasmsimd_mul32_ld128) |
| 1347 | BENCHMARK_GEMM(qu8_gemm_3x4c8__wasmsimd_mul32_ld128) |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1348 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1349 | |
| 1350 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1351 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| 1352 | static void qu8_gemm_1x2__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1353 | GEMMBenchmark(state, |
| 1354 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic, |
| 1355 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1356 | 1, 2, 1, 1); |
| 1357 | } |
| 1358 | static void qu8_gemm_2x2__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1359 | GEMMBenchmark(state, |
| 1360 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__wasm_fmagic, |
| 1361 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1362 | 2, 2, 1, 1); |
| 1363 | } |
| 1364 | static void qu8_gemm_3x2__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1365 | GEMMBenchmark(state, |
| 1366 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic, |
| 1367 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1368 | 3, 2, 1, 1); |
| 1369 | } |
| 1370 | static void qu8_gemm_4x2__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1371 | GEMMBenchmark(state, |
| 1372 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic, |
| 1373 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1374 | 4, 2, 1, 1); |
| 1375 | } |
| 1376 | static void qu8_gemm_1x4__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1377 | GEMMBenchmark(state, |
| 1378 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic, |
| 1379 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1380 | 1, 4, 1, 1); |
| 1381 | } |
| 1382 | static void qu8_gemm_2x4__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1383 | GEMMBenchmark(state, |
| 1384 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic, |
| 1385 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1386 | 2, 4, 1, 1); |
| 1387 | } |
| 1388 | static void qu8_gemm_3x4__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1389 | GEMMBenchmark(state, |
| 1390 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__wasm_fmagic, |
| 1391 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1392 | 3, 4, 1, 1); |
| 1393 | } |
| 1394 | static void qu8_gemm_4x4__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1395 | GEMMBenchmark(state, |
| 1396 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic, |
| 1397 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1398 | 4, 4, 1, 1); |
| 1399 | } |
| 1400 | |
| 1401 | BENCHMARK_GEMM(qu8_gemm_1x2__wasm_fmagic) |
| 1402 | BENCHMARK_GEMM(qu8_gemm_2x2__wasm_fmagic) |
| 1403 | BENCHMARK_GEMM(qu8_gemm_3x2__wasm_fmagic) |
| 1404 | BENCHMARK_GEMM(qu8_gemm_4x2__wasm_fmagic) |
| 1405 | BENCHMARK_GEMM(qu8_gemm_1x4__wasm_fmagic) |
| 1406 | BENCHMARK_GEMM(qu8_gemm_2x4__wasm_fmagic) |
| 1407 | BENCHMARK_GEMM(qu8_gemm_3x4__wasm_fmagic) |
| 1408 | BENCHMARK_GEMM(qu8_gemm_4x4__wasm_fmagic) |
| 1409 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| 1410 | |
| 1411 | |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1412 | static void qu8_gemm_1x2__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1413 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1414 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic, |
| 1415 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1416 | 1, 2, 1, 1); |
| 1417 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1418 | static void qu8_gemm_2x2__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1419 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1420 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_fmagic, |
| 1421 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1422 | 2, 2, 1, 1); |
| 1423 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1424 | static void qu8_gemm_3x2__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1425 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1426 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic, |
| 1427 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1428 | 3, 2, 1, 1); |
| 1429 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1430 | static void qu8_gemm_4x2__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1431 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1432 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic, |
| 1433 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1434 | 4, 2, 1, 1); |
| 1435 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1436 | static void qu8_gemm_1x4__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1437 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1438 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic, |
| 1439 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1440 | 1, 4, 1, 1); |
| 1441 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1442 | static void qu8_gemm_2x4__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1443 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1444 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic, |
| 1445 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1446 | 2, 4, 1, 1); |
| 1447 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1448 | static void qu8_gemm_3x4__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1449 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1450 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic, |
| 1451 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1452 | 3, 4, 1, 1); |
| 1453 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1454 | static void qu8_gemm_4x4__scalar_fmagic(benchmark::State& state, const char* net) { |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1455 | GEMMBenchmark(state, |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1456 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic, |
| 1457 | xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1458 | 4, 4, 1, 1); |
| 1459 | } |
| 1460 | |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1461 | static void qu8_gemm_1x2__scalar_imagic(benchmark::State& state, const char* net) { |
| 1462 | GEMMBenchmark(state, |
| 1463 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, |
| 1464 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1465 | 1, 2, 1, 1); |
| 1466 | } |
| 1467 | static void qu8_gemm_2x2__scalar_imagic(benchmark::State& state, const char* net) { |
| 1468 | GEMMBenchmark(state, |
| 1469 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, |
| 1470 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1471 | 2, 2, 1, 1); |
| 1472 | } |
| 1473 | static void qu8_gemm_3x2__scalar_imagic(benchmark::State& state, const char* net) { |
| 1474 | GEMMBenchmark(state, |
| 1475 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, |
| 1476 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1477 | 3, 2, 1, 1); |
| 1478 | } |
| 1479 | static void qu8_gemm_4x2__scalar_imagic(benchmark::State& state, const char* net) { |
| 1480 | GEMMBenchmark(state, |
| 1481 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, |
| 1482 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1483 | 4, 2, 1, 1); |
| 1484 | } |
| 1485 | static void qu8_gemm_1x4__scalar_imagic(benchmark::State& state, const char* net) { |
| 1486 | GEMMBenchmark(state, |
| 1487 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, |
| 1488 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1489 | 1, 4, 1, 1); |
| 1490 | } |
| 1491 | static void qu8_gemm_2x4__scalar_imagic(benchmark::State& state, const char* net) { |
| 1492 | GEMMBenchmark(state, |
| 1493 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, |
| 1494 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1495 | 2, 4, 1, 1); |
| 1496 | } |
| 1497 | static void qu8_gemm_3x4__scalar_imagic(benchmark::State& state, const char* net) { |
| 1498 | GEMMBenchmark(state, |
| 1499 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, |
| 1500 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1501 | 3, 4, 1, 1); |
| 1502 | } |
| 1503 | static void qu8_gemm_4x4__scalar_imagic(benchmark::State& state, const char* net) { |
| 1504 | GEMMBenchmark(state, |
| 1505 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, |
| 1506 | xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1507 | 4, 4, 1, 1); |
| 1508 | } |
| 1509 | |
| 1510 | static void qu8_gemm_1x2__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1511 | GEMMBenchmark(state, |
| 1512 | xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, |
| 1513 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1514 | 1, 2, 1, 1); |
| 1515 | } |
| 1516 | static void qu8_gemm_2x2__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1517 | GEMMBenchmark(state, |
| 1518 | xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, |
| 1519 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1520 | 2, 2, 1, 1); |
| 1521 | } |
| 1522 | static void qu8_gemm_3x2__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1523 | GEMMBenchmark(state, |
| 1524 | xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, |
| 1525 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1526 | 3, 2, 1, 1); |
| 1527 | } |
| 1528 | static void qu8_gemm_4x2__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1529 | GEMMBenchmark(state, |
| 1530 | xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, |
| 1531 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1532 | 4, 2, 1, 1); |
| 1533 | } |
| 1534 | static void qu8_gemm_1x4__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1535 | GEMMBenchmark(state, |
| 1536 | xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, |
| 1537 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1538 | 1, 4, 1, 1); |
| 1539 | } |
| 1540 | static void qu8_gemm_2x4__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1541 | GEMMBenchmark(state, |
| 1542 | xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, |
| 1543 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1544 | 2, 4, 1, 1); |
| 1545 | } |
| 1546 | static void qu8_gemm_3x4__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1547 | GEMMBenchmark(state, |
| 1548 | xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, |
| 1549 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1550 | 3, 4, 1, 1); |
| 1551 | } |
| 1552 | static void qu8_gemm_4x4__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1553 | GEMMBenchmark(state, |
| 1554 | xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, |
| 1555 | xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1556 | 4, 4, 1, 1); |
| 1557 | } |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1558 | |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1559 | BENCHMARK_GEMM(qu8_gemm_1x2__scalar_fmagic) |
| 1560 | BENCHMARK_GEMM(qu8_gemm_2x2__scalar_fmagic) |
| 1561 | BENCHMARK_GEMM(qu8_gemm_3x2__scalar_fmagic) |
| 1562 | BENCHMARK_GEMM(qu8_gemm_4x2__scalar_fmagic) |
| 1563 | BENCHMARK_GEMM(qu8_gemm_1x4__scalar_fmagic) |
| 1564 | BENCHMARK_GEMM(qu8_gemm_2x4__scalar_fmagic) |
| 1565 | BENCHMARK_GEMM(qu8_gemm_3x4__scalar_fmagic) |
| 1566 | BENCHMARK_GEMM(qu8_gemm_4x4__scalar_fmagic) |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1567 | |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1568 | BENCHMARK_GEMM(qu8_gemm_1x2__scalar_imagic) |
| 1569 | BENCHMARK_GEMM(qu8_gemm_2x2__scalar_imagic) |
| 1570 | BENCHMARK_GEMM(qu8_gemm_3x2__scalar_imagic) |
| 1571 | BENCHMARK_GEMM(qu8_gemm_4x2__scalar_imagic) |
| 1572 | BENCHMARK_GEMM(qu8_gemm_1x4__scalar_imagic) |
| 1573 | BENCHMARK_GEMM(qu8_gemm_2x4__scalar_imagic) |
| 1574 | BENCHMARK_GEMM(qu8_gemm_3x4__scalar_imagic) |
| 1575 | BENCHMARK_GEMM(qu8_gemm_4x4__scalar_imagic) |
| 1576 | |
| 1577 | BENCHMARK_GEMM(qu8_gemm_1x2__scalar_lrintf) |
| 1578 | BENCHMARK_GEMM(qu8_gemm_2x2__scalar_lrintf) |
| 1579 | BENCHMARK_GEMM(qu8_gemm_3x2__scalar_lrintf) |
| 1580 | BENCHMARK_GEMM(qu8_gemm_4x2__scalar_lrintf) |
| 1581 | BENCHMARK_GEMM(qu8_gemm_1x4__scalar_lrintf) |
| 1582 | BENCHMARK_GEMM(qu8_gemm_2x4__scalar_lrintf) |
| 1583 | BENCHMARK_GEMM(qu8_gemm_3x4__scalar_lrintf) |
| 1584 | BENCHMARK_GEMM(qu8_gemm_4x4__scalar_lrintf) |
| 1585 | |
Marat Dukhan | d8e2d71 | 2021-07-26 23:35:50 -0700 | [diff] [blame] | 1586 | |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 1587 | #ifdef BENCHMARK_RUY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1588 | BENCHMARK_GEMM(ruy_st) |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 1589 | #endif // BENCHMARK_RUY |
| 1590 | #ifdef BENCHMARK_GEMMLOWP |
| 1591 | BENCHMARK_GEMM(gemmlowp_st) |
| 1592 | #endif // BENCHMARK_GEMMLOWP |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1593 | |
| 1594 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 1595 | BENCHMARK_MAIN(); |
| 1596 | #endif |