Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 1 | // Copyright 2019 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 <cmath> |
| 9 | #include <functional> |
| 10 | #include <random> |
| 11 | #include <vector> |
| 12 | |
| 13 | #include <cpuinfo.h> |
| 14 | |
| 15 | #include <benchmark/benchmark.h> |
| 16 | #include <fp16/fp16.h> |
| 17 | #include "bench/gemm.h" |
| 18 | #include "bench/utils.h" |
| 19 | #include <xnnpack/AlignedAllocator.h> |
| 20 | #include <xnnpack/common.h> |
| 21 | #include <xnnpack/params-init.h> |
| 22 | #include <xnnpack/params.h> |
| 23 | #include <xnnpack/spmm.h> |
| 24 | |
| 25 | |
| 26 | static void SpMMBenchmark(benchmark::State& state, |
| 27 | xnn_f16_spmm_ukernel_function spmm, uint32_t mr, uint32_t nr, float sparsity) |
| 28 | { |
| 29 | if (!cpuinfo_initialize()) { |
| 30 | state.SkipWithError("cpuinfo initialization failed"); |
| 31 | return; |
| 32 | } |
| 33 | |
| 34 | const size_t mc = state.range(0); |
| 35 | const size_t nc = state.range(1); |
| 36 | const size_t kc = state.range(2); |
| 37 | |
| 38 | std::random_device random_device; |
| 39 | auto rng = std::mt19937(random_device()); |
| 40 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 41 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 42 | |
| 43 | // if using blocks, generate the reduced matrix first and then extrude along |
| 44 | // the block dimension (n), to get the full matrix |
| 45 | size_t ncols = nc / nr + nc % nr; |
| 46 | std::vector<uint16_t> b(ncols * kc); |
| 47 | std::vector<uint16_t> bias(nc); |
| 48 | std::vector<uint16_t> w; |
| 49 | std::vector<uint32_t> nmap; |
| 50 | std::vector<int32_t> dmap; |
| 51 | const size_t sparse_end = std::min(size_t(float(b.size()) * sparsity), b.size()); |
| 52 | const size_t num_nonzeroes = nr * (b.size() - sparse_end); |
| 53 | |
| 54 | const size_t w_elements = num_nonzeroes + nc; |
| 55 | const size_t c_elements = mc * nc; |
| 56 | const size_t dmap_elements = num_nonzeroes / nr; |
| 57 | const size_t nmap_elements = nc; |
| 58 | const size_t num_buffers = 1 + |
| 59 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
| 60 | sizeof(uint16_t) * (w_elements + c_elements) + sizeof(uint32_t) * (dmap_elements + nmap_elements)); |
| 61 | |
| 62 | // Micro-kernel can access one element beyond w and dmap for software pipelining. |
| 63 | w.reserve(num_buffers * w_elements + 1); |
| 64 | dmap.reserve(num_buffers * dmap_elements + 1); |
| 65 | nmap.resize(num_buffers * nmap_elements); |
| 66 | |
| 67 | std::vector<size_t> a_offsets(num_buffers); |
| 68 | |
| 69 | for (size_t buffer_index = 0; buffer_index < num_buffers; buffer_index++) { |
| 70 | // Re-generate weights. Note: each re-generation produces the number of non-zeroes. |
| 71 | std::fill(b.begin(), b.begin() + sparse_end, 0); |
| 72 | std::generate(b.begin() + sparse_end, b.end(), std::ref(f16rng)); |
| 73 | std::shuffle(b.begin(), b.end(), rng); |
| 74 | std::generate(bias.begin(), bias.end(), std::ref(f16rng)); |
| 75 | |
| 76 | uint32_t first_j = 0, last_j = 0; |
| 77 | bool is_first_nonzero = true; |
| 78 | for (uint32_t i = 0; i < nc / nr; i++) { |
| 79 | for (uint32_t n = 0; n < nr; n++) |
| 80 | w.push_back(bias[nr * i + n]); |
| 81 | for (uint32_t j = 0; j < kc; j++) { |
| 82 | if ((b[i * kc + j] & 0x7FFF) != 0) { |
| 83 | for (size_t l = 0; l < nr; l++) |
| 84 | w.push_back(fp16_ieee_from_fp32_value(fp16_ieee_to_fp32_value(b[i * kc + j]) + static_cast<float>(i))); |
| 85 | if (is_first_nonzero) { |
| 86 | first_j = j; |
| 87 | } else { |
| 88 | const ptrdiff_t increment = int32_t(j - last_j) * int32_t(mc) * int32_t(sizeof(uint16_t)); |
| 89 | dmap.push_back(increment); |
| 90 | } |
| 91 | last_j = j; |
| 92 | is_first_nonzero = false; |
| 93 | nmap[buffer_index * nmap_elements + i] += 1; |
| 94 | } |
| 95 | } |
| 96 | } |
| 97 | for (uint32_t i = nc / nr; i < ncols; i++) { |
| 98 | w.push_back(bias[i]); |
| 99 | for (uint32_t j = 0; j < kc; j++) { |
| 100 | if ((b[i * kc + j] & 0x7FFF) != 0) { |
| 101 | w.push_back(b[i * kc + j]); |
| 102 | if (is_first_nonzero) { |
| 103 | first_j = j; |
| 104 | } else { |
| 105 | const ptrdiff_t increment = int32_t(j - last_j) * int32_t(mc) * int32_t(sizeof(uint16_t)); |
| 106 | dmap.push_back(increment); |
| 107 | } |
| 108 | last_j = j; |
| 109 | is_first_nonzero = false; |
| 110 | nmap[buffer_index * nmap_elements + i] += 1; |
| 111 | } |
| 112 | } |
| 113 | } |
| 114 | { |
| 115 | const ptrdiff_t increment = int32_t(first_j - last_j) * int32_t(mc) * int32_t(sizeof(uint16_t)); |
| 116 | dmap.push_back(increment); |
| 117 | } |
| 118 | |
| 119 | a_offsets[buffer_index] = first_j * mc; |
| 120 | } |
| 121 | |
| 122 | // Micro-kernel can access one element beyond w and dmap for software pipelining. |
| 123 | w.resize(w.size() + 1); |
| 124 | dmap.resize(dmap.size() + 1); |
| 125 | |
| 126 | std::vector<float, AlignedAllocator<float, 64>> a(kc * mc); |
| 127 | std::vector<float, AlignedAllocator<float, 64>> c(num_buffers * c_elements); |
| 128 | |
| 129 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 130 | std::fill(c.begin(), c.end(), nanf("")); |
| 131 | |
| 132 | xnn_f16_output_params output_params{ |
| 133 | 0x3C00 /* 1.0 */, 0x7C00 /* inf */, 0xFC00 /* -inf */}; |
| 134 | |
| 135 | size_t buffer_index = 0; |
| 136 | for (auto _ : state) { |
| 137 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 138 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 139 | // - W, Kmap, and Nmap is not in cache (for any cache level) |
| 140 | // - C is not in cache (for any cache level) |
| 141 | state.PauseTiming(); |
| 142 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint16_t)); |
| 143 | buffer_index = (buffer_index + 1) % num_buffers; |
| 144 | state.ResumeTiming(); |
| 145 | |
| 146 | spmm(mc, nc, |
| 147 | a.data() + a_offsets[buffer_index], |
| 148 | w.data() + buffer_index * w_elements, |
| 149 | dmap.data() + buffer_index * dmap_elements, |
| 150 | nmap.data() + buffer_index * nmap_elements, |
| 151 | c.data() + buffer_index * c_elements, |
| 152 | &output_params); |
| 153 | } |
| 154 | |
| 155 | state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency(); |
| 156 | state.counters["FLOPS"] = benchmark::Counter( |
| 157 | uint64_t(state.iterations()) * 2 * mc * num_nonzeroes, benchmark::Counter::kIsRate); |
| 158 | |
| 159 | state.counters["EffFLOPS"] = benchmark::Counter( |
| 160 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 161 | } |
| 162 | |
| 163 | |
| 164 | #if XNN_ARCH_ARM64 |
| 165 | static void spmm80_8x1__neonfp16arith(benchmark::State& state, const char* net) { |
| 166 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_8x1__neonfp16arith, 8, 1, 0.8f); |
| 167 | } |
| 168 | static void spmm80_8x1__neonfp16arith_unroll2(benchmark::State& state, const char* net) { |
| 169 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_8x1__neonfp16arith_unroll2, 8, 1, 0.8f); |
| 170 | } |
| 171 | static void spmm80_16x1__neonfp16arith(benchmark::State& state, const char* net) { |
| 172 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_16x1__neonfp16arith, 16, 1, 0.8f); |
| 173 | } |
| 174 | static void spmm80_16x1__neonfp16arith_unroll2(benchmark::State& state, const char* net) { |
| 175 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_16x1__neonfp16arith_unroll2, 16, 1, 0.8f); |
| 176 | } |
| 177 | static void spmm80_24x1__neonfp16arith(benchmark::State& state, const char* net) { |
| 178 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_24x1__neonfp16arith, 24, 1, 0.8f); |
| 179 | } |
| 180 | static void spmm80_24x1__neonfp16arith_unroll2(benchmark::State& state, const char* net) { |
| 181 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_24x1__neonfp16arith_unroll2, 24, 1, 0.8f); |
| 182 | } |
| 183 | static void spmm80_32x1__neonfp16arith(benchmark::State& state, const char* net) { |
| 184 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_32x1__neonfp16arith, 32, 1, 0.8f); |
| 185 | } |
| 186 | static void spmm80_32x1__neonfp16arith_unroll2(benchmark::State& state, const char* net) { |
| 187 | SpMMBenchmark(state, xnn_f16_spmm_ukernel_32x1__neonfp16arith_unroll2, 32, 1, 0.8f); |
| 188 | } |
| 189 | |
| 190 | BENCHMARK_GEMM(spmm80_8x1__neonfp16arith) |
| 191 | BENCHMARK_GEMM(spmm80_8x1__neonfp16arith_unroll2) |
| 192 | BENCHMARK_GEMM(spmm80_16x1__neonfp16arith) |
| 193 | BENCHMARK_GEMM(spmm80_16x1__neonfp16arith_unroll2) |
| 194 | BENCHMARK_GEMM(spmm80_24x1__neonfp16arith) |
| 195 | BENCHMARK_GEMM(spmm80_24x1__neonfp16arith_unroll2) |
| 196 | BENCHMARK_GEMM(spmm80_32x1__neonfp16arith) |
| 197 | BENCHMARK_GEMM(spmm80_32x1__neonfp16arith_unroll2) |
| 198 | #endif // XNN_ARCH_ARM64 |
| 199 | |
| 200 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 201 | BENCHMARK_MAIN(); |
| 202 | #endif |