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 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | #include <benchmark/benchmark.h> |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 20 | #ifdef BENCHMARK_GEMMLOWP |
| 21 | #include "gemmlowp/public/gemmlowp.h" |
| 22 | #endif // BENCHMARK_GEMMLOWP |
| 23 | #ifdef BENCHMARK_RUY |
Benoit Jacob | b038fdc | 2020-03-25 12:14:20 -0700 | [diff] [blame] | 24 | #include "ruy/ruy.h" |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 25 | #endif // BENCHMARK_RUY |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 26 | #include "bench/gemm.h" |
| 27 | #include "bench/utils.h" |
| 28 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 29 | #include <xnnpack/common.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 30 | #include <xnnpack/gemm.h> |
| 31 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 32 | #include <xnnpack/params-init.h> |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 33 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 34 | |
| 35 | |
| 36 | static void GEMMBenchmark(benchmark::State& state, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 37 | xnn_qu8_gemm_ukernel_function gemm, |
Marat Dukhan | 4e89587 | 2020-12-04 15:27:45 -0800 | [diff] [blame^] | 38 | size_t mr, size_t nr, size_t kr, size_t sr, |
| 39 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 40 | { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 41 | const size_t mc = state.range(0); |
| 42 | const size_t nc = state.range(1); |
| 43 | const size_t kc = state.range(2); |
| 44 | |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 45 | const size_t nc_stride = benchmark::utils::RoundUp(nc, nr); |
| 46 | const size_t kc_stride = benchmark::utils::RoundUp(kc, kr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | |
| 48 | std::random_device random_device; |
| 49 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 50 | 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] | 51 | 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] | 52 | |
| 53 | std::vector<uint8_t> a(mc * kc); |
| 54 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 55 | std::vector<uint8_t> k(nc * kc); |
| 56 | std::generate(k.begin(), k.end(), std::ref(u8rng)); |
| 57 | std::vector<int32_t> b(nc); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 58 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 59 | |
| 60 | const size_t w_elements = kc_stride * nc_stride + nc_stride * sizeof(int32_t) / sizeof(uint8_t); |
| 61 | const size_t c_elements = mc * nc; |
| 62 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 63 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 64 | sizeof(uint8_t) * (w_elements + c_elements)); |
| 65 | |
| 66 | std::vector<uint8_t, AlignedAllocator<uint8_t, 32>> w(w_elements * num_buffers); |
| 67 | std::fill(w.begin(), w.end(), 0); |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 68 | const xnn_qu8_packing_params packing_params = { 127, 127 }; |
| 69 | xnn_pack_qu8_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data(), &packing_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 70 | std::vector<uint8_t> c(c_elements * num_buffers); |
| 71 | std::fill(c.begin(), c.end(), 0xA5); |
| 72 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 73 | union xnn_qu8_gemm_params quantization_params = |
Marat Dukhan | 9199246 | 2020-07-30 00:06:34 -0700 | [diff] [blame] | 74 | xnn_init_qu8_gemm_params(127, 0.75f, 127, 1, 254); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 75 | |
| 76 | size_t buffer_index = 0; |
| 77 | for (auto _ : state) { |
| 78 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 79 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 80 | // - W is not in cache (for any cache level) |
| 81 | // - C is not in cache (for any cache level) |
| 82 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 83 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 84 | buffer_index = (buffer_index + 1) % num_buffers; |
| 85 | state.ResumeTiming(); |
| 86 | |
| 87 | for (uint32_t m = 0; m < mc; m += mr) { |
| 88 | const uint32_t mb = min(mc - m, mr); |
| 89 | for (uint32_t n = 0; n < nc; n += nr) { |
| 90 | const uint32_t nb = min(nc - n, nr); |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 91 | gemm( |
Marat Dukhan | b186463 | 2019-11-25 16:34:17 -0800 | [diff] [blame] | 92 | mb, nb, kc * sizeof(uint8_t), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 93 | a.data() + m * kc, kc * sizeof(uint8_t), |
| 94 | w.data() + (w_elements * buffer_index + n * (kc_stride + sizeof(int32_t))) / sizeof(uint8_t), |
| 95 | 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] | 96 | &quantization_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 97 | } |
| 98 | } |
| 99 | } |
| 100 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 101 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 102 | if (cpu_frequency != 0) { |
| 103 | state.counters["cpufreq"] = cpu_frequency; |
| 104 | } |
| 105 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 106 | state.counters["OPS"] = benchmark::Counter( |
| 107 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 108 | } |
| 109 | |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 110 | #ifdef BENCHMARK_GEMMLOWP |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 111 | struct GemmlowpOutputPipeline { |
| 112 | typedef gemmlowp::VectorMap<const int32_t, gemmlowp::VectorShape::Col> ColVectorMap; |
| 113 | typedef std::tuple< |
| 114 | gemmlowp::OutputStageBiasAddition<ColVectorMap>, |
| 115 | gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint, |
| 116 | gemmlowp::OutputStageClamp, |
| 117 | gemmlowp::OutputStageSaturatingCastToUint8> |
| 118 | Pipeline; |
| 119 | |
| 120 | static Pipeline Make( |
| 121 | const int32_t* bias_data, |
| 122 | int output_rows, |
| 123 | int32_t output_offset, |
| 124 | int32_t output_multiplier, |
| 125 | int output_shift, |
| 126 | int32_t output_activation_min, |
| 127 | int32_t output_activation_max) |
| 128 | { |
| 129 | ColVectorMap bias_vector(bias_data, output_rows); |
| 130 | gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage; |
| 131 | bias_addition_stage.bias_vector = bias_vector; |
| 132 | gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint quantize_down_stage; |
| 133 | quantize_down_stage.result_offset_after_shift = output_offset; |
| 134 | quantize_down_stage.result_fixedpoint_multiplier = output_multiplier; |
| 135 | quantize_down_stage.result_shift = output_shift; |
| 136 | gemmlowp::OutputStageClamp clamp_stage; |
| 137 | clamp_stage.min = output_activation_min; |
| 138 | clamp_stage.max = output_activation_max; |
| 139 | gemmlowp::OutputStageSaturatingCastToUint8 saturating_cast_stage; |
| 140 | return std::make_tuple(bias_addition_stage, quantize_down_stage, clamp_stage, saturating_cast_stage); |
| 141 | } |
| 142 | }; |
| 143 | |
| 144 | static void GemmlowpBenchmark(benchmark::State& state, uint32_t threads) |
| 145 | { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 146 | const size_t mc = state.range(0); |
| 147 | const size_t nc = state.range(1); |
| 148 | const size_t kc = state.range(2); |
| 149 | |
| 150 | std::random_device random_device; |
| 151 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 152 | 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] | 153 | 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] | 154 | |
| 155 | std::vector<uint8_t> a(mc * kc); |
| 156 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 157 | |
| 158 | const size_t kElements = nc * kc; |
| 159 | const size_t bElements = nc; |
| 160 | const size_t c_elements = mc * nc; |
| 161 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 162 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 163 | kElements * sizeof(uint8_t) + bElements * sizeof(int32_t) + c_elements * sizeof(uint8_t)); |
| 164 | |
| 165 | std::vector<uint8_t> k(kElements * num_buffers); |
| 166 | std::generate(k.begin(), k.end(), std::ref(u8rng)); |
| 167 | std::vector<int32_t> b(bElements * num_buffers); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 168 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 169 | std::vector<uint8_t> c(c_elements * num_buffers); |
| 170 | std::fill(c.begin(), c.end(), 0xA5); |
| 171 | |
| 172 | gemmlowp::MultiThreadGemmContext threadingContext; |
| 173 | threadingContext.set_max_num_threads(threads); |
| 174 | |
| 175 | size_t buffer_index = 0; |
| 176 | for (auto _ : state) { |
| 177 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 178 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 179 | buffer_index = (buffer_index + 1) % num_buffers; |
| 180 | state.ResumeTiming(); |
| 181 | |
| 182 | gemmlowp::MatrixMap<const uint8_t, gemmlowp::MapOrder::RowMajor> AM(a.data(), mc, kc, kc); |
| 183 | gemmlowp::MatrixMap<const uint8_t, gemmlowp::MapOrder::ColMajor> BM(k.data() + buffer_index * kElements, kc, nc, kc); |
| 184 | gemmlowp::MatrixMap<uint8_t, gemmlowp::MapOrder::RowMajor> CM(c.data() + buffer_index * c_elements, mc, nc, nc); |
| 185 | const auto& outputPipeline = GemmlowpOutputPipeline::Make(b.data() + buffer_index * bElements, nc, 127, 127, 127, 0, 255); |
| 186 | gemmlowp::GemmWithOutputPipeline<uint8_t, uint8_t, gemmlowp::L8R8WithLhsNonzeroBitDepthParams>( |
| 187 | &threadingContext, AM, BM, &CM, 127, 127, outputPipeline); |
| 188 | } |
| 189 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 190 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 191 | if (cpu_frequency != 0) { |
| 192 | state.counters["cpufreq"] = cpu_frequency; |
| 193 | } |
| 194 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 195 | state.counters["OPS"] = benchmark::Counter( |
| 196 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 197 | } |
| 198 | |
| 199 | static void gemmlowp_st(benchmark::State& state, const char* net) |
| 200 | { |
| 201 | GemmlowpBenchmark(state, 1); |
| 202 | } |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 203 | #endif // BENCHMARK_GEMMLOWP |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 204 | |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 205 | |
| 206 | #ifdef BENCHMARK_RUY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 207 | static void RuyBenchmark(benchmark::State& state, size_t threads) |
| 208 | { |
| 209 | const size_t mc = state.range(0); |
| 210 | const size_t nc = state.range(1); |
| 211 | const size_t kc = state.range(2); |
| 212 | |
| 213 | std::random_device random_device; |
| 214 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 215 | 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] | 216 | 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] | 217 | |
| 218 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 219 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 220 | nc * (sizeof(uint8_t) * (mc + kc) + sizeof(int32_t))); |
| 221 | |
| 222 | std::vector<uint8_t> a(mc * kc); |
| 223 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 224 | std::vector<uint8_t> k(num_buffers * nc * kc); |
| 225 | std::generate(k.begin(), k.end(), std::ref(u8rng)); |
| 226 | std::vector<int32_t> b(num_buffers * nc); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 227 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 228 | std::vector<uint8_t> c(num_buffers * nc * mc); |
| 229 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 230 | |
| 231 | // Note: context must be static to avoid the cost of re-creating it for each benchmark. |
| 232 | static ruy::Context context; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 233 | context.set_max_num_threads(threads); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 234 | |
| 235 | ruy::Matrix<uint8_t> ruy_a; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 236 | ruy::MakeSimpleLayout(nc, kc, ruy::Order::kRowMajor, ruy_a.mutable_layout()); |
| 237 | ruy_a.set_zero_point(127); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 238 | ruy::Matrix<uint8_t> ruy_b; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 239 | ruy::MakeSimpleLayout(kc, mc, ruy::Order::kColMajor, ruy_b.mutable_layout()); |
| 240 | ruy_b.set_data(a.data()); |
| 241 | ruy_b.set_zero_point(127); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 242 | ruy::Matrix<uint8_t> ruy_c; |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 243 | ruy::MakeSimpleLayout(nc, mc, ruy::Order::kColMajor, ruy_c.mutable_layout()); |
| 244 | ruy_c.set_zero_point(127); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 245 | |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 246 | ruy::MulParams<int32_t, uint8_t> mul_params; |
| 247 | mul_params.set_multiplier_fixedpoint(0x40000000); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 248 | |
| 249 | // ruy::Context uses deferred initialization, which affects percieved GEMM performance. Initialization happens during |
| 250 | // the first GEMM calls, and per Benoit Jacob it takes up to ~250 milliseconds for performance to stabilize. |
| 251 | // Thus, on the first benchmark, we compute GEMM for 500 milliseconds (to be safe) without recording performance, and |
| 252 | // keep the ruy::Context object initialized (by being static) between subsequent benchmarks. |
| 253 | static std::once_flag warmup; |
| 254 | std::call_once(warmup, [&](){ |
| 255 | auto start = std::chrono::steady_clock::now(); |
| 256 | do { |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 257 | ruy_a.set_data(k.data()); |
| 258 | ruy_c.set_data(c.data()); |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 259 | mul_params.set_bias(b.data()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 260 | |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 261 | ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 262 | } while (std::chrono::duration<double>(std::chrono::steady_clock::now() - start).count() < 0.5); |
| 263 | }); |
| 264 | |
| 265 | size_t buffer_index = 0; |
| 266 | for (auto _ : state) { |
| 267 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 268 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 269 | // - K is not in cache (for any cache level) |
| 270 | // - B is not in cache (for any cache level) |
| 271 | // - C is not in cache (for any cache level) |
| 272 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 273 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 274 | buffer_index = (buffer_index + 1) % num_buffers; |
| 275 | state.ResumeTiming(); |
| 276 | |
Benoit Jacob | 349701a | 2020-04-15 19:35:24 -0700 | [diff] [blame] | 277 | ruy_a.set_data(k.data() + buffer_index * nc * kc); |
| 278 | ruy_c.set_data(c.data() + buffer_index * mc * nc); |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 279 | mul_params.set_bias(b.data() + buffer_index * nc); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 280 | |
Benoit Jacob | b026e22 | 2020-04-16 12:30:03 -0700 | [diff] [blame] | 281 | ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 282 | } |
| 283 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 284 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 285 | if (cpu_frequency != 0) { |
| 286 | state.counters["cpufreq"] = cpu_frequency; |
| 287 | } |
| 288 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 289 | state.counters["OPS"] = benchmark::Counter( |
| 290 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 291 | } |
| 292 | |
| 293 | static void ruy_st(benchmark::State& state, const char* net) |
| 294 | { |
| 295 | RuyBenchmark(state, 1); |
| 296 | } |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 297 | #endif // BENCHMARK_RUY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 298 | |
| 299 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 300 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 301 | static void qu8_gemm_4x8__neon(benchmark::State& state, const char* net) { |
Marat Dukhan | 4e89587 | 2020-12-04 15:27:45 -0800 | [diff] [blame^] | 302 | GEMMBenchmark(state, xnn_qu8_gemm_minmax_ukernel_4x8__neon, 4, 8, 1, 1, benchmark::utils::CheckNEON); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 303 | } |
| 304 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 305 | static void qu8_gemm_8x8__neon(benchmark::State& state, const char* net) { |
Marat Dukhan | 4e89587 | 2020-12-04 15:27:45 -0800 | [diff] [blame^] | 306 | GEMMBenchmark(state, xnn_qu8_gemm_minmax_ukernel_8x8__neon, 8, 8, 1, 1, benchmark::utils::CheckNEON); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 307 | } |
| 308 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 309 | BENCHMARK_GEMM(qu8_gemm_4x8__neon) |
| 310 | BENCHMARK_GEMM(qu8_gemm_8x8__neon) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 311 | #endif |
| 312 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 313 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 314 | static void qu8_gemm_4x4c2__sse2(benchmark::State& state, const char* net) { |
| 315 | GEMMBenchmark(state, xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2, 4, 4, 2, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 316 | } |
| 317 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 318 | static void qu8_gemm_2x4c8__sse2(benchmark::State& state, const char* net) { |
| 319 | GEMMBenchmark(state, xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2, 2, 4, 8, 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 320 | } |
| 321 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 322 | BENCHMARK_GEMM(qu8_gemm_4x4c2__sse2) |
| 323 | BENCHMARK_GEMM(qu8_gemm_2x4c8__sse2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 324 | #endif |
| 325 | |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 326 | #ifdef BENCHMARK_RUY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 327 | BENCHMARK_GEMM(ruy_st) |
Marat Dukhan | 33f0c7a | 2019-10-01 13:33:08 -0700 | [diff] [blame] | 328 | #endif // BENCHMARK_RUY |
| 329 | #ifdef BENCHMARK_GEMMLOWP |
| 330 | BENCHMARK_GEMM(gemmlowp_st) |
| 331 | #endif // BENCHMARK_GEMMLOWP |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 332 | |
| 333 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 334 | BENCHMARK_MAIN(); |
| 335 | #endif |