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 <cmath> |
| 12 | #include <functional> |
| 13 | #include <random> |
| 14 | #include <vector> |
| 15 | |
| 16 | #include <cpuinfo.h> |
| 17 | |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 18 | #include <benchmark/benchmark.h> |
Marat Dukhan | 629a33e | 2019-10-01 10:39:14 -0700 | [diff] [blame] | 19 | #include <fp16/fp16.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | #include "bench/gemm.h" |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 21 | #include "bench/utils.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 23 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | #include <xnnpack/gemm.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 26 | #include <xnnpack/params-init.h> |
Frank Barchard | e0601b5 | 2019-10-25 17:43:34 -0700 | [diff] [blame] | 27 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 28 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | |
| 30 | static void GEMMBenchmark(benchmark::State& state, |
| 31 | xnn_f16_gemm_ukernel_function hgemm, |
| 32 | size_t mr, size_t nr, size_t kr) |
| 33 | { |
| 34 | if (!cpuinfo_initialize()) { |
| 35 | state.SkipWithError("cpuinfo initialization failed"); |
| 36 | return; |
| 37 | } |
| 38 | |
| 39 | const size_t mc = state.range(0); |
| 40 | const size_t nc = state.range(1); |
| 41 | const size_t kc = state.range(2); |
| 42 | |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 43 | const size_t nc_stride = benchmark::utils::RoundUp(nc, nr); |
| 44 | const size_t kc_stride = benchmark::utils::RoundUp(kc, kr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 45 | |
| 46 | std::random_device random_device; |
| 47 | auto rng = std::mt19937(random_device()); |
| 48 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 49 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 50 | |
| 51 | std::vector<uint16_t> a(mc * kc); |
| 52 | std::generate(a.begin(), a.end(), std::ref(f16rng)); |
| 53 | std::vector<uint16_t> k(nc * kc); |
| 54 | std::generate(k.begin(), k.end(), std::ref(f16rng)); |
| 55 | std::vector<uint16_t> b(nc); |
| 56 | std::generate(b.begin(), b.end(), std::ref(f16rng)); |
| 57 | |
| 58 | const size_t w_elements = nc_stride * kc_stride + nc_stride; |
| 59 | const size_t c_elements = mc * nc; |
| 60 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 61 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 62 | sizeof(uint16_t) * (w_elements + c_elements)); |
| 63 | |
| 64 | std::vector<uint16_t, AlignedAllocator<uint16_t, 32>> w(w_elements * num_buffers); |
| 65 | std::fill(w.begin(), w.end(), 0); |
| 66 | xnn_pack_f16_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, k.data(), b.data(), w.data()); |
| 67 | std::vector<uint16_t> c(c_elements * num_buffers); |
| 68 | std::fill(c.begin(), c.end(), UINT16_C(0x7E00) /* NaN */); |
| 69 | |
| 70 | xnn_f16_output_params output_params{ |
| 71 | 0x3C00 /* 1.0 */, 0x7C00 /* inf */, 0xFC00 /* -inf */}; |
| 72 | |
| 73 | size_t buffer_index = 0; |
| 74 | for (auto _ : state) { |
| 75 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 76 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 77 | // - W is not in cache (for any cache level) |
| 78 | // - C is not in cache (for any cache level) |
| 79 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 80 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint16_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 81 | buffer_index = (buffer_index + 1) % num_buffers; |
| 82 | state.ResumeTiming(); |
| 83 | |
| 84 | for (uint32_t m = 0; m < mc; m += mr) { |
| 85 | const uint32_t mb = min(mc - m, mr); |
| 86 | for (uint32_t n = 0; n < nc; n += nr) { |
| 87 | const uint32_t nb = min(nc - n, nr); |
| 88 | hgemm( |
Marat Dukhan | b186463 | 2019-11-25 16:34:17 -0800 | [diff] [blame] | 89 | mb, nb, kc * sizeof(uint16_t), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 90 | a.data() + m * kc, kc * sizeof(uint16_t), |
| 91 | w.data() + (nc_stride * buffer_index + n) * (kc_stride + 1), |
| 92 | c.data() + (mc * buffer_index + m) * nc + n, nc * sizeof(uint16_t), nr * sizeof(uint16_t), |
| 93 | &output_params); |
| 94 | } |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 99 | state.counters["FLOPS"] = benchmark::Counter( |
| 100 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 101 | } |
| 102 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 103 | #if XNN_ARCH_ARM64 |
Frank Barchard | 54a9d9d | 2019-11-25 16:46:38 -0800 | [diff] [blame] | 104 | static void f16_gemm_4x8__neonfp16arith_ld64(benchmark::State& state, const char* net) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 105 | GEMMBenchmark(state, xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64, 4, 8, 1); |
| 106 | } |
| 107 | |
Frank Barchard | 54a9d9d | 2019-11-25 16:46:38 -0800 | [diff] [blame] | 108 | static void f16_gemm_6x8__neonfp16arith_ld64(benchmark::State& state, const char* net) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 109 | GEMMBenchmark(state, xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64, 6, 8, 1); |
| 110 | } |
| 111 | |
Frank Barchard | 54a9d9d | 2019-11-25 16:46:38 -0800 | [diff] [blame] | 112 | static void f16_gemm_8x8__neonfp16arith_ld64(benchmark::State& state, const char* net) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 113 | GEMMBenchmark(state, xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64, 8, 8, 1); |
| 114 | } |
| 115 | |
Frank Barchard | 54a9d9d | 2019-11-25 16:46:38 -0800 | [diff] [blame] | 116 | BENCHMARK_GEMM(f16_gemm_4x8__neonfp16arith_ld64) |
| 117 | BENCHMARK_GEMM(f16_gemm_6x8__neonfp16arith_ld64) |
| 118 | BENCHMARK_GEMM(f16_gemm_8x8__neonfp16arith_ld64) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 119 | #endif |
| 120 | |
| 121 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 122 | BENCHMARK_MAIN(); |
| 123 | #endif |