Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <algorithm> |
| 7 | #include <cfloat> |
| 8 | #include <chrono> |
| 9 | #include <cmath> |
| 10 | #include <functional> |
| 11 | #include <limits> |
| 12 | #include <mutex> |
| 13 | #include <random> |
| 14 | #include <vector> |
| 15 | |
| 16 | #include <cpuinfo.h> |
| 17 | |
| 18 | #include <benchmark/benchmark.h> |
| 19 | #include "bench/gemm.h" |
| 20 | #include "bench/utils.h" |
| 21 | #include <xnnpack/AlignedAllocator.h> |
| 22 | #include <xnnpack/common.h> |
| 23 | #include <xnnpack/gemm.h> |
| 24 | #include <xnnpack/pack.h> |
| 25 | #include <xnnpack/params-init.h> |
| 26 | #include <xnnpack/params.h> |
| 27 | |
| 28 | |
| 29 | static void GEMMBenchmark(benchmark::State& state, |
| 30 | xnn_qs8_gemm_ukernel_function gemm, |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 31 | size_t mr, size_t nr, size_t kr, size_t sr, |
| 32 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 33 | { |
| 34 | if (!cpuinfo_initialize()) { |
| 35 | state.SkipWithError("cpuinfo initialization failed"); |
| 36 | return; |
| 37 | } |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 38 | if (isa_check && !isa_check(state)) { |
| 39 | return; |
| 40 | } |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 41 | |
| 42 | const size_t mc = state.range(0); |
| 43 | const size_t nc = state.range(1); |
| 44 | const size_t kc = state.range(2); |
| 45 | |
| 46 | const size_t nc_stride = benchmark::utils::RoundUp(nc, nr); |
| 47 | const size_t kc_stride = benchmark::utils::RoundUp(kc, kr); |
| 48 | |
| 49 | std::random_device random_device; |
| 50 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 51 | auto s32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), std::ref(rng)); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 52 | auto s8rng = std::bind( |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 53 | std::uniform_int_distribution<uint32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), std::ref(rng)); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 54 | |
| 55 | std::vector<int8_t> a(mc * kc); |
| 56 | std::generate(a.begin(), a.end(), std::ref(s8rng)); |
| 57 | std::vector<int8_t> k(nc * kc); |
| 58 | std::generate(k.begin(), k.end(), std::ref(s8rng)); |
| 59 | std::vector<int32_t> b(nc); |
| 60 | std::generate(b.begin(), b.end(), std::ref(s32rng)); |
| 61 | |
| 62 | const size_t w_elements = kc_stride * nc_stride + nc_stride * sizeof(int32_t) / sizeof(int8_t); |
| 63 | const size_t c_elements = mc * nc; |
| 64 | const size_t num_buffers = 1 + |
| 65 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
| 66 | sizeof(int8_t) * (w_elements + c_elements)); |
| 67 | |
| 68 | std::vector<int8_t, AlignedAllocator<int8_t, 32>> w(w_elements * num_buffers); |
| 69 | std::fill(w.begin(), w.end(), 0); |
| 70 | const xnn_qs8_packing_params packing_params = { 127 }; |
| 71 | xnn_pack_qs8_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data(), &packing_params); |
| 72 | std::vector<int8_t> c(c_elements * num_buffers); |
| 73 | std::fill(c.begin(), c.end(), 0xA5); |
| 74 | |
| 75 | union xnn_qs8_gemm_params quantization_params = xnn_init_qs8_gemm_params(0.75f, 127, -127, 126); |
| 76 | |
| 77 | size_t buffer_index = 0; |
| 78 | for (auto _ : state) { |
| 79 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 80 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 81 | // - W is not in cache (for any cache level) |
| 82 | // - C is not in cache (for any cache level) |
| 83 | state.PauseTiming(); |
| 84 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(int8_t)); |
| 85 | buffer_index = (buffer_index + 1) % num_buffers; |
| 86 | state.ResumeTiming(); |
| 87 | |
| 88 | for (uint32_t m = 0; m < mc; m += mr) { |
| 89 | const uint32_t mb = min(mc - m, mr); |
| 90 | for (uint32_t n = 0; n < nc; n += nr) { |
| 91 | const uint32_t nb = min(nc - n, nr); |
| 92 | gemm( |
| 93 | mb, nb, kc * sizeof(int8_t), |
| 94 | a.data() + m * kc, kc * sizeof(int8_t), |
| 95 | w.data() + (w_elements * buffer_index + n * (kc_stride + sizeof(int32_t))) / sizeof(int8_t), |
| 96 | c.data() + (mc * buffer_index + m) * nc + n, nc * sizeof(int8_t), nr * sizeof(int8_t), |
| 97 | &quantization_params); |
| 98 | } |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 103 | state.counters["OPS"] = benchmark::Counter( |
| 104 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 105 | } |
| 106 | |
| 107 | |
| 108 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 109 | static void qs8_gemm_4x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
| 110 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64, 4, 4, 2, 1); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 111 | } |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 112 | static void qs8_gemm_4x4c2__ssse3_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 113 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64, 4, 4, 2, 1, benchmark::utils::CheckSSSE3); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 114 | } |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 115 | static void qs8_gemm_4x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 116 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld64, 4, 4, 2, 1, benchmark::utils::CheckSSE41); |
| 117 | } |
| 118 | static void qs8_gemm_4x4c2__xop_ld64(benchmark::State& state, const char* net) { |
| 119 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld64, 4, 4, 2, 1, benchmark::utils::CheckXOP); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 120 | } |
| 121 | |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 122 | static void qs8_gemm_4x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
| 123 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128, 4, 4, 2, 1); |
| 124 | } |
| 125 | static void qs8_gemm_4x4c2__ssse3_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 126 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128, 4, 4, 2, 1, benchmark::utils::CheckSSSE3); |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 127 | } |
| 128 | static void qs8_gemm_4x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 129 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld128, 4, 4, 2, 1, benchmark::utils::CheckSSE41); |
| 130 | } |
| 131 | static void qs8_gemm_4x4c2__xop_ld128(benchmark::State& state, const char* net) { |
| 132 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld128, 4, 4, 2, 1, benchmark::utils::CheckXOP); |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 133 | } |
| 134 | |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 135 | static void qs8_gemm_2x4c8__sse2_ld64(benchmark::State& state, const char* net) { |
| 136 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64, 2, 4, 8, 1); |
Marat Dukhan | 733d0be | 2020-07-31 15:55:36 -0700 | [diff] [blame] | 137 | } |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 138 | static void qs8_gemm_2x4c8__ssse3_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 139 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld64, 2, 4, 8, 1, benchmark::utils::CheckSSSE3); |
Marat Dukhan | 733d0be | 2020-07-31 15:55:36 -0700 | [diff] [blame] | 140 | } |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 141 | static void qs8_gemm_2x4c8__sse41_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 142 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld64, 2, 4, 8, 1, benchmark::utils::CheckSSE41); |
| 143 | } |
| 144 | static void qs8_gemm_2x4c8__xop_ld64(benchmark::State& state, const char* net) { |
| 145 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld64, 2, 4, 8, 1, benchmark::utils::CheckXOP); |
Marat Dukhan | 733d0be | 2020-07-31 15:55:36 -0700 | [diff] [blame] | 146 | } |
| 147 | |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 148 | static void qs8_gemm_2x4c8__sse2_ld128(benchmark::State& state, const char* net) { |
| 149 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128, 2, 4, 8, 1); |
| 150 | } |
| 151 | static void qs8_gemm_2x4c8__ssse3_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 152 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld128, 2, 4, 8, 1, benchmark::utils::CheckSSSE3); |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 153 | } |
| 154 | static void qs8_gemm_2x4c8__sse41_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 155 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld128, 2, 4, 8, 1, benchmark::utils::CheckSSE41); |
| 156 | } |
| 157 | static void qs8_gemm_2x4c8__xop_ld128(benchmark::State& state, const char* net) { |
| 158 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld128, 2, 4, 8, 1, benchmark::utils::CheckXOP); |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 159 | } |
| 160 | |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 161 | static void qs8_gemm_3x4c8__sse2_ld64(benchmark::State& state, const char* net) { |
| 162 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64, 3, 4, 8, 1); |
| 163 | } |
| 164 | static void qs8_gemm_3x4c8__ssse3_ld64(benchmark::State& state, const char* net) { |
| 165 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64, 3, 4, 8, 1, benchmark::utils::CheckSSSE3); |
| 166 | } |
| 167 | static void qs8_gemm_3x4c8__sse41_ld64(benchmark::State& state, const char* net) { |
| 168 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64, 3, 4, 8, 1, benchmark::utils::CheckSSE41); |
| 169 | } |
| 170 | static void qs8_gemm_3x4c8__xop_ld64(benchmark::State& state, const char* net) { |
| 171 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64, 3, 4, 8, 1, benchmark::utils::CheckXOP); |
| 172 | } |
| 173 | |
| 174 | static void qs8_gemm_3x4c8__sse2_ld128(benchmark::State& state, const char* net) { |
| 175 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128, 3, 4, 8, 1); |
| 176 | } |
| 177 | static void qs8_gemm_3x4c8__ssse3_ld128(benchmark::State& state, const char* net) { |
| 178 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128, 3, 4, 8, 1, benchmark::utils::CheckSSSE3); |
| 179 | } |
| 180 | static void qs8_gemm_3x4c8__sse41_ld128(benchmark::State& state, const char* net) { |
| 181 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128, 3, 4, 8, 1, benchmark::utils::CheckSSE41); |
| 182 | } |
| 183 | static void qs8_gemm_3x4c8__xop_ld128(benchmark::State& state, const char* net) { |
| 184 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128, 3, 4, 8, 1, benchmark::utils::CheckXOP); |
| 185 | } |
| 186 | |
Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 187 | static void qs8_gemm_2x8c8__avx2(benchmark::State& state, const char* net) { |
| 188 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2, 2, 8, 8, 1, benchmark::utils::CheckAVX2); |
| 189 | } |
| 190 | static void qs8_gemm_3x8c8__avx2(benchmark::State& state, const char* net) { |
| 191 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2, 3, 8, 8, 1, benchmark::utils::CheckAVX2); |
| 192 | } |
| 193 | |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 194 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse2_ld64) |
| 195 | BENCHMARK_GEMM(qs8_gemm_4x4c2__ssse3_ld64) |
| 196 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse41_ld64) |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 197 | BENCHMARK_GEMM(qs8_gemm_4x4c2__xop_ld64) |
Marat Dukhan | 733d0be | 2020-07-31 15:55:36 -0700 | [diff] [blame] | 198 | |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 199 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse2_ld128) |
| 200 | BENCHMARK_GEMM(qs8_gemm_4x4c2__ssse3_ld128) |
| 201 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse41_ld128) |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 202 | BENCHMARK_GEMM(qs8_gemm_4x4c2__xop_ld128) |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 203 | |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 204 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse2_ld64) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 205 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse2_ld64) |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 206 | BENCHMARK_GEMM(qs8_gemm_2x4c8__ssse3_ld64) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 207 | BENCHMARK_GEMM(qs8_gemm_3x4c8__ssse3_ld64) |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 208 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse41_ld64) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 209 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse41_ld64) |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 210 | BENCHMARK_GEMM(qs8_gemm_2x4c8__xop_ld64) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 211 | BENCHMARK_GEMM(qs8_gemm_3x4c8__xop_ld64) |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 212 | |
| 213 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse2_ld128) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 214 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse2_ld128) |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 215 | BENCHMARK_GEMM(qs8_gemm_2x4c8__ssse3_ld128) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 216 | BENCHMARK_GEMM(qs8_gemm_3x4c8__ssse3_ld128) |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 217 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse41_ld128) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 218 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse41_ld128) |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 219 | BENCHMARK_GEMM(qs8_gemm_2x4c8__xop_ld128) |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame^] | 220 | BENCHMARK_GEMM(qs8_gemm_3x4c8__xop_ld128) |
Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 221 | |
| 222 | BENCHMARK_GEMM(qs8_gemm_2x8c8__avx2) |
| 223 | BENCHMARK_GEMM(qs8_gemm_3x8c8__avx2) |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 224 | #endif |
| 225 | |
| 226 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 227 | BENCHMARK_MAIN(); |
| 228 | #endif |