XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [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 | |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 13 | #include <benchmark/benchmark.h> |
Marat Dukhan | 1631e3e | 2020-12-06 19:29:31 -0800 | [diff] [blame] | 14 | #include "bench/spmm.h" |
Frank Barchard | bb4c18b | 2019-09-30 11:05:52 -0700 | [diff] [blame] | 15 | #include "bench/utils.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 16 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 17 | #include <xnnpack/common.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 18 | #include <xnnpack/params-init.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | #include <xnnpack/spmm.h> |
| 21 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | |
| 23 | static void SpMMBenchmark(benchmark::State& state, |
Marat Dukhan | 5b86c43 | 2020-12-06 19:15:03 -0800 | [diff] [blame] | 24 | xnn_f32_spmm_minmax_ukernel_function spmm, uint32_t mr, uint32_t nr, float sparsity, |
| 25 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | { |
Marat Dukhan | 5b86c43 | 2020-12-06 19:15:03 -0800 | [diff] [blame] | 27 | if (isa_check && !isa_check(state)) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 28 | return; |
| 29 | } |
| 30 | |
| 31 | const size_t mc = state.range(0); |
| 32 | const size_t nc = state.range(1); |
| 33 | const size_t kc = state.range(2); |
| 34 | |
| 35 | std::random_device random_device; |
| 36 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 44f0ca7 | 2020-08-02 21:46:58 -0700 | [diff] [blame] | 37 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 38 | |
| 39 | // if using blocks, generate the reduced matrix first and then extrude along |
| 40 | // the block dimension (n), to get the full matrix |
Marat Dukhan | b8ab4cb | 2019-10-03 15:08:04 -0700 | [diff] [blame] | 41 | size_t ncols = nc / nr + nc % nr; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 42 | std::vector<float> b(ncols * kc); |
| 43 | std::vector<float> bias(nc); |
| 44 | std::vector<float> w; |
| 45 | std::vector<uint32_t> nmap; |
| 46 | std::vector<int32_t> dmap; |
| 47 | const size_t sparse_end = std::min(size_t(float(b.size()) * sparsity), b.size()); |
| 48 | const size_t num_nonzeroes = nr * (b.size() - sparse_end); |
| 49 | |
| 50 | const size_t w_elements = num_nonzeroes + nc; |
| 51 | const size_t c_elements = mc * nc; |
| 52 | const size_t dmap_elements = num_nonzeroes / nr; |
| 53 | const size_t nmap_elements = nc; |
| 54 | const size_t num_buffers = 1 + |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 55 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 56 | sizeof(float) * (w_elements + c_elements) + sizeof(uint32_t) * (dmap_elements + nmap_elements)); |
| 57 | |
| 58 | // Micro-kernel can access one element beyond w and dmap for software pipelining. |
| 59 | w.reserve(num_buffers * w_elements + 1); |
| 60 | dmap.reserve(num_buffers * dmap_elements + 1); |
| 61 | nmap.resize(num_buffers * nmap_elements); |
| 62 | |
| 63 | std::vector<size_t> a_offsets(num_buffers); |
| 64 | |
| 65 | for (size_t buffer_index = 0; buffer_index < num_buffers; buffer_index++) { |
| 66 | // Re-generate weights. Note: each re-generation produces the number of non-zeroes. |
| 67 | std::fill(b.begin(), b.begin() + sparse_end, 0.0f); |
| 68 | std::generate(b.begin() + sparse_end, b.end(), std::ref(f32rng)); |
| 69 | std::shuffle(b.begin(), b.end(), rng); |
| 70 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 71 | |
| 72 | uint32_t first_j = 0, last_j = 0; |
| 73 | bool is_first_nonzero = true; |
| 74 | for (uint32_t i = 0; i < nc / nr; i++) { |
| 75 | for (uint32_t n = 0; n < nr; n++) |
| 76 | w.push_back(bias[nr * i + n]); |
| 77 | for (uint32_t j = 0; j < kc; j++) { |
| 78 | if (b[i * kc + j] != 0.0f) { |
| 79 | for (size_t l = 0; l < nr; l++) |
| 80 | w.push_back(b[i * kc + j] + static_cast<float>(i)); |
| 81 | if (is_first_nonzero) { |
| 82 | first_j = j; |
| 83 | } else { |
| 84 | const ptrdiff_t increment = int32_t(j - last_j) * int32_t(mc) * int32_t(sizeof(float)); |
| 85 | dmap.push_back(increment); |
| 86 | } |
| 87 | last_j = j; |
| 88 | is_first_nonzero = false; |
| 89 | nmap[buffer_index * nmap_elements + i] += 1; |
| 90 | } |
| 91 | } |
| 92 | } |
| 93 | for (uint32_t i = nc / nr; i < ncols; i++) { |
| 94 | w.push_back(bias[i]); |
| 95 | for (uint32_t j = 0; j < kc; j++) { |
| 96 | if (b[i * kc + j] != 0.0f) { |
| 97 | w.push_back(b[i * kc + j]); |
| 98 | if (is_first_nonzero) { |
| 99 | first_j = j; |
| 100 | } else { |
| 101 | const ptrdiff_t increment = int32_t(j - last_j) * int32_t(mc) * int32_t(sizeof(float)); |
| 102 | dmap.push_back(increment); |
| 103 | } |
| 104 | last_j = j; |
| 105 | is_first_nonzero = false; |
| 106 | nmap[buffer_index * nmap_elements + i] += 1; |
| 107 | } |
| 108 | } |
| 109 | } |
| 110 | { |
| 111 | const ptrdiff_t increment = int32_t(first_j - last_j) * int32_t(mc) * int32_t(sizeof(float)); |
| 112 | dmap.push_back(increment); |
| 113 | } |
| 114 | |
| 115 | a_offsets[buffer_index] = first_j * mc; |
| 116 | } |
| 117 | |
| 118 | // Micro-kernel can access one element beyond w and dmap for software pipelining. |
| 119 | w.resize(w.size() + 1); |
| 120 | dmap.resize(dmap.size() + 1); |
| 121 | |
| 122 | std::vector<float, AlignedAllocator<float, 64>> a(kc * mc); |
| 123 | std::vector<float, AlignedAllocator<float, 64>> c(num_buffers * c_elements); |
| 124 | |
| 125 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 126 | std::fill(c.begin(), c.end(), nanf("")); |
| 127 | |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 128 | xnn_f32_minmax_params params = |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 129 | xnn_init_f32_minmax_params(-std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 130 | |
| 131 | size_t buffer_index = 0; |
| 132 | for (auto _ : state) { |
| 133 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 134 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 135 | // - W, Kmap, and Nmap is not in cache (for any cache level) |
| 136 | // - C is not in cache (for any cache level) |
| 137 | state.PauseTiming(); |
Marat Dukhan | 4232323 | 2019-10-23 02:09:02 -0700 | [diff] [blame] | 138 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 139 | buffer_index = (buffer_index + 1) % num_buffers; |
| 140 | state.ResumeTiming(); |
| 141 | |
Marat Dukhan | e8bfcc8 | 2020-11-16 12:28:13 -0800 | [diff] [blame] | 142 | spmm(mc * sizeof(float), nc, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 143 | a.data() + a_offsets[buffer_index], |
| 144 | w.data() + buffer_index * w_elements, |
| 145 | dmap.data() + buffer_index * dmap_elements, |
| 146 | nmap.data() + buffer_index * nmap_elements, |
Marat Dukhan | e8bfcc8 | 2020-11-16 12:28:13 -0800 | [diff] [blame] | 147 | c.data() + buffer_index * c_elements, mc * sizeof(float), |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 148 | ¶ms); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 149 | } |
| 150 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 151 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 152 | if (cpu_frequency != 0) { |
| 153 | state.counters["cpufreq"] = cpu_frequency; |
| 154 | } |
| 155 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 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 | |
Marat Dukhan | 5b86c43 | 2020-12-06 19:15:03 -0800 | [diff] [blame] | 164 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 165 | static void spmm80_4x1__neon(benchmark::State& state, const char* net) { |
| 166 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__neon, 4, 1, 0.8f, benchmark::utils::CheckNEON); |
| 167 | } |
| 168 | |
| 169 | static void spmm80_8x1__neon(benchmark::State& state, const char* net) { |
| 170 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__neon, 8, 1, 0.8f, benchmark::utils::CheckNEON); |
| 171 | } |
| 172 | |
| 173 | static void spmm80_12x1__neon(benchmark::State& state, const char* net) { |
| 174 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_12x1__neon, 12, 1, 0.8f, benchmark::utils::CheckNEON); |
| 175 | } |
| 176 | |
| 177 | static void spmm80_16x1__neon(benchmark::State& state, const char* net) { |
| 178 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__neon, 16, 1, 0.8f, benchmark::utils::CheckNEON); |
| 179 | } |
| 180 | |
| 181 | static void spmm80_32x1__neon(benchmark::State& state, const char* net) { |
| 182 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__neon, 32, 1, 0.8f, benchmark::utils::CheckNEON); |
| 183 | } |
| 184 | |
| 185 | static void spmm80_4x1__neon_x2(benchmark::State& state, const char* net) { |
| 186 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__neon_x2, 4, 1, 0.8f, benchmark::utils::CheckNEON); |
| 187 | } |
| 188 | |
| 189 | static void spmm80_8x1__neon_x2(benchmark::State& state, const char* net) { |
| 190 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__neon_x2, 8, 1, 0.8f, benchmark::utils::CheckNEON); |
| 191 | } |
| 192 | |
| 193 | static void spmm80_16x1__neon_x2(benchmark::State& state, const char* net) { |
| 194 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__neon_x2, 16, 1, 0.8f, benchmark::utils::CheckNEON); |
| 195 | } |
| 196 | |
| 197 | static void spmm80_32x1__neon_x2(benchmark::State& state, const char* net) { |
| 198 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__neon_x2, 32, 1, 0.8f, benchmark::utils::CheckNEON); |
| 199 | } |
| 200 | |
| 201 | static void spmm80_4x1__neon_pipelined(benchmark::State& state, const char* net) { |
| 202 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__neon_pipelined, 4, 1, 0.8f, benchmark::utils::CheckNEON); |
| 203 | } |
| 204 | |
| 205 | static void spmm80_8x1__neon_pipelined(benchmark::State& state, const char* net) { |
| 206 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__neon_pipelined, 8, 1, 0.8f, benchmark::utils::CheckNEON); |
| 207 | } |
| 208 | |
| 209 | static void spmm80_16x1__neon_pipelined(benchmark::State& state, const char* net) { |
| 210 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined, 16, 1, 0.8f, benchmark::utils::CheckNEON); |
| 211 | } |
| 212 | |
| 213 | static void spmm80_32x1__neon_pipelined(benchmark::State& state, const char* net) { |
| 214 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined, 32, 1, 0.8f, benchmark::utils::CheckNEON); |
| 215 | } |
| 216 | |
Marat Dukhan | 1631e3e | 2020-12-06 19:29:31 -0800 | [diff] [blame] | 217 | BENCHMARK_SPMM(spmm80_4x1__neon) |
| 218 | BENCHMARK_SPMM(spmm80_4x1__neon_pipelined) |
| 219 | BENCHMARK_SPMM(spmm80_4x1__neon_x2) |
| 220 | BENCHMARK_SPMM(spmm80_8x1__neon) |
| 221 | BENCHMARK_SPMM(spmm80_8x1__neon_pipelined) |
| 222 | BENCHMARK_SPMM(spmm80_8x1__neon_x2) |
| 223 | BENCHMARK_SPMM(spmm80_12x1__neon) |
| 224 | BENCHMARK_SPMM(spmm80_16x1__neon) |
| 225 | BENCHMARK_SPMM(spmm80_16x1__neon_pipelined) |
| 226 | BENCHMARK_SPMM(spmm80_16x1__neon_x2) |
| 227 | BENCHMARK_SPMM(spmm80_32x1__neon) |
| 228 | BENCHMARK_SPMM(spmm80_32x1__neon_pipelined) |
| 229 | BENCHMARK_SPMM(spmm80_32x1__neon_x2) |
Marat Dukhan | 5b86c43 | 2020-12-06 19:15:03 -0800 | [diff] [blame] | 230 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 231 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 232 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 233 | static void spmm80_4x1__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 234 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__neonfma, 4, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 235 | } |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 236 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 237 | static void spmm80_4x2__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 238 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x2__neonfma, 4, 2, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 239 | } |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 240 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 241 | static void spmm80_4x4__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 242 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x4__neonfma, 4, 4, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 243 | } |
| 244 | |
| 245 | static void spmm80_8x1__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 246 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__neonfma, 8, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 247 | } |
| 248 | |
| 249 | static void spmm80_8x2__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 250 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x2__neonfma, 8, 2, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 251 | } |
| 252 | |
| 253 | static void spmm80_8x4__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 254 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x4__neonfma, 8, 4, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 255 | } |
| 256 | |
| 257 | static void spmm80_12x1__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 258 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_12x1__neonfma, 12, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 259 | } |
| 260 | |
| 261 | static void spmm80_12x2__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 262 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_12x2__neonfma, 12, 2, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 263 | } |
| 264 | |
| 265 | static void spmm80_12x4__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 266 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_12x4__neonfma, 12, 4, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 267 | } |
| 268 | |
| 269 | static void spmm80_16x1__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 270 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__neonfma, 16, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 271 | } |
| 272 | |
| 273 | static void spmm80_16x2__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 274 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x2__neonfma, 16, 2, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 275 | } |
| 276 | |
| 277 | static void spmm80_16x4__neonfma(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 278 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x4__neonfma, 16, 4, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 279 | } |
| 280 | |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 281 | static void spmm80_32x1__neonfma(benchmark::State& state, const char* net) { |
| 282 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__neonfma, 32, 1, 0.8f); |
| 283 | } |
| 284 | |
| 285 | static void spmm80_32x2__neonfma(benchmark::State& state, const char* net) { |
| 286 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x2__neonfma, 32, 2, 0.8f); |
| 287 | } |
| 288 | |
| 289 | static void spmm80_32x4__neonfma(benchmark::State& state, const char* net) { |
| 290 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x4__neonfma, 32, 4, 0.8f); |
| 291 | } |
| 292 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 293 | static void spmm80_4x1__neonfma_x2(benchmark::State& state, const char* net) { |
| 294 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__neonfma_x2, 4, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 295 | } |
| 296 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 297 | static void spmm80_8x1__neonfma_x2(benchmark::State& state, const char* net) { |
| 298 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__neonfma_x2, 8, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 299 | } |
| 300 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 301 | static void spmm80_16x1__neonfma_x2(benchmark::State& state, const char* net) { |
| 302 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2, 16, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 303 | } |
| 304 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 305 | static void spmm80_32x1__neonfma_x2(benchmark::State& state, const char* net) { |
| 306 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__neonfma_x2, 32, 1, 0.8f); |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 307 | } |
| 308 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 309 | static void spmm80_4x1__neonfma_pipelined(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 310 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__neonfma_pipelined, 4, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 311 | } |
| 312 | |
| 313 | static void spmm80_8x1__neonfma_pipelined(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 314 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__neonfma_pipelined, 8, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 315 | } |
| 316 | |
| 317 | static void spmm80_16x1__neonfma_pipelined(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 318 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined, 16, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 319 | } |
| 320 | |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 321 | static void spmm80_32x1__neonfma_pipelined(benchmark::State& state, const char* net) { |
| 322 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined, 32, 1, 0.8f); |
| 323 | } |
| 324 | |
Marat Dukhan | 1631e3e | 2020-12-06 19:29:31 -0800 | [diff] [blame] | 325 | BENCHMARK_SPMM(spmm80_4x1__neonfma) |
| 326 | BENCHMARK_SPMM(spmm80_4x1__neonfma_pipelined) |
| 327 | BENCHMARK_SPMM(spmm80_4x1__neonfma_x2) |
| 328 | BENCHMARK_SPMM(spmm80_4x2__neonfma) |
| 329 | BENCHMARK_SPMM(spmm80_4x4__neonfma) |
| 330 | BENCHMARK_SPMM(spmm80_8x1__neonfma) |
| 331 | BENCHMARK_SPMM(spmm80_8x1__neonfma_pipelined) |
| 332 | BENCHMARK_SPMM(spmm80_8x1__neonfma_x2) |
| 333 | BENCHMARK_SPMM(spmm80_8x2__neonfma) |
| 334 | BENCHMARK_SPMM(spmm80_8x4__neonfma) |
| 335 | BENCHMARK_SPMM(spmm80_12x1__neonfma) |
| 336 | BENCHMARK_SPMM(spmm80_12x2__neonfma) |
| 337 | BENCHMARK_SPMM(spmm80_12x4__neonfma) |
| 338 | BENCHMARK_SPMM(spmm80_16x1__neonfma) |
| 339 | BENCHMARK_SPMM(spmm80_16x1__neonfma_pipelined) |
| 340 | BENCHMARK_SPMM(spmm80_16x1__neonfma_x2) |
| 341 | BENCHMARK_SPMM(spmm80_16x2__neonfma) |
| 342 | BENCHMARK_SPMM(spmm80_16x4__neonfma) |
| 343 | BENCHMARK_SPMM(spmm80_32x1__neonfma) |
| 344 | BENCHMARK_SPMM(spmm80_32x1__neonfma_pipelined) |
| 345 | BENCHMARK_SPMM(spmm80_32x1__neonfma_x2) |
| 346 | BENCHMARK_SPMM(spmm80_32x2__neonfma) |
| 347 | BENCHMARK_SPMM(spmm80_32x4__neonfma) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 348 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 349 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 350 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 351 | static void spmm80_4x1__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 352 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__sse, 4, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 353 | } |
| 354 | |
| 355 | static void spmm80_8x1__sse(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 356 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__sse, 8, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 357 | } |
| 358 | |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 359 | static void spmm80_16x1__sse(benchmark::State& state, const char* net) { |
| 360 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__sse, 16, 1, 0.8f); |
| 361 | } |
| 362 | |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 363 | static void spmm80_32x1__sse(benchmark::State& state, const char* net) { |
| 364 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__sse, 32, 1, 0.8f); |
| 365 | } |
| 366 | |
Marat Dukhan | 1631e3e | 2020-12-06 19:29:31 -0800 | [diff] [blame] | 367 | BENCHMARK_SPMM(spmm80_4x1__sse) |
| 368 | BENCHMARK_SPMM(spmm80_8x1__sse) |
| 369 | BENCHMARK_SPMM(spmm80_16x1__sse) |
| 370 | BENCHMARK_SPMM(spmm80_32x1__sse) |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 371 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 372 | |
| 373 | static void spmm80_1x1__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 374 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_1x1__scalar, 1, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 375 | } |
| 376 | |
| 377 | static void spmm80_2x1__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 378 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_2x1__scalar, 2, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 379 | } |
| 380 | |
| 381 | static void spmm80_4x1__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 382 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__scalar, 4, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 383 | } |
| 384 | |
| 385 | static void spmm80_8x1__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 386 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__scalar, 8, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 387 | } |
| 388 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 389 | static void spmm80_8x2__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 390 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x2__scalar, 8, 2, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 391 | } |
| 392 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 393 | static void spmm80_8x4__scalar(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 394 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x4__scalar, 8, 4, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 395 | } |
| 396 | |
| 397 | static void spmm80_1x1__scalar_pipelined(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 398 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_1x1__scalar_pipelined, 1, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 399 | } |
| 400 | |
| 401 | static void spmm80_2x1__scalar_pipelined(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 402 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_2x1__scalar_pipelined, 2, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 403 | } |
| 404 | |
| 405 | static void spmm80_4x1__scalar_pipelined(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 406 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__scalar_pipelined, 4, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 407 | } |
| 408 | |
| 409 | static void spmm80_8x1__scalar_pipelined(benchmark::State& state, const char* net) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 410 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__scalar_pipelined, 8, 1, 0.8f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 411 | } |
| 412 | |
Marat Dukhan | 1631e3e | 2020-12-06 19:29:31 -0800 | [diff] [blame] | 413 | BENCHMARK_SPMM(spmm80_1x1__scalar) |
| 414 | BENCHMARK_SPMM(spmm80_2x1__scalar) |
| 415 | BENCHMARK_SPMM(spmm80_4x1__scalar) |
| 416 | BENCHMARK_SPMM(spmm80_8x1__scalar) |
| 417 | BENCHMARK_SPMM(spmm80_8x2__scalar) |
| 418 | BENCHMARK_SPMM(spmm80_8x4__scalar) |
| 419 | BENCHMARK_SPMM(spmm80_1x1__scalar_pipelined) |
| 420 | BENCHMARK_SPMM(spmm80_2x1__scalar_pipelined) |
| 421 | BENCHMARK_SPMM(spmm80_4x1__scalar_pipelined) |
| 422 | BENCHMARK_SPMM(spmm80_8x1__scalar_pipelined) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 423 | |
Frank Barchard | 9e05340 | 2020-10-19 15:29:08 -0700 | [diff] [blame] | 424 | #if XNN_ARCH_WASMSIMD |
| 425 | static void spmm80_4x1__wasmsimd_arm(benchmark::State& state, const char* net) { |
| 426 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_arm, 4, 1, 0.8f); |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 427 | } |
Marat Dukhan | a199d49 | 2020-07-24 15:01:25 -0700 | [diff] [blame] | 428 | |
Frank Barchard | 9e05340 | 2020-10-19 15:29:08 -0700 | [diff] [blame] | 429 | static void spmm80_8x1__wasmsimd_arm(benchmark::State& state, const char* net) { |
| 430 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm, 8, 1, 0.8f); |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 431 | } |
Marat Dukhan | a199d49 | 2020-07-24 15:01:25 -0700 | [diff] [blame] | 432 | |
Frank Barchard | 9e05340 | 2020-10-19 15:29:08 -0700 | [diff] [blame] | 433 | static void spmm80_16x1__wasmsimd_arm(benchmark::State& state, const char* net) { |
| 434 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm, 16, 1, 0.8f); |
| 435 | } |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 436 | |
| 437 | static void spmm80_32x1__wasmsimd_arm(benchmark::State& state, const char* net) { |
| 438 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm, 32, 1, 0.8f); |
| 439 | } |
| 440 | |
Frank Barchard | 9e05340 | 2020-10-19 15:29:08 -0700 | [diff] [blame] | 441 | static void spmm80_4x1__wasmsimd_x86(benchmark::State& state, const char* net) { |
| 442 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_x86, 4, 1, 0.8f); |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 443 | } |
Marat Dukhan | a199d49 | 2020-07-24 15:01:25 -0700 | [diff] [blame] | 444 | |
Frank Barchard | 9e05340 | 2020-10-19 15:29:08 -0700 | [diff] [blame] | 445 | static void spmm80_8x1__wasmsimd_x86(benchmark::State& state, const char* net) { |
| 446 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86, 8, 1, 0.8f); |
| 447 | } |
| 448 | |
| 449 | static void spmm80_16x1__wasmsimd_x86(benchmark::State& state, const char* net) { |
| 450 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86, 16, 1, 0.8f); |
| 451 | } |
| 452 | |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 453 | static void spmm80_32x1__wasmsimd_x86(benchmark::State& state, const char* net) { |
| 454 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86, 32, 1, 0.8f); |
| 455 | } |
| 456 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 457 | static void spmm80_32x1__wasmsimd_arm_x2(benchmark::State& state, const char* net) { |
| 458 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x2, 32, 1, 0.8f); |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 459 | } |
| 460 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 461 | static void spmm80_4x1__wasmsimd_arm_x2(benchmark::State& state, const char* net) { |
| 462 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_arm_x2, 4, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 463 | } |
| 464 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 465 | static void spmm80_8x1__wasmsimd_arm_x2(benchmark::State& state, const char* net) { |
| 466 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm_x2, 8, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 467 | } |
| 468 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 469 | static void spmm80_16x1__wasmsimd_arm_x2(benchmark::State& state, const char* net) { |
| 470 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_x2, 16, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 471 | } |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 472 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 473 | static void spmm80_32x1__wasmsimd_x86_x2(benchmark::State& state, const char* net) { |
| 474 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2, 32, 1, 0.8f); |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 475 | } |
| 476 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 477 | static void spmm80_4x1__wasmsimd_x86_x2(benchmark::State& state, const char* net) { |
| 478 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_x86_x2, 4, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 479 | } |
| 480 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 481 | static void spmm80_8x1__wasmsimd_x86_x2(benchmark::State& state, const char* net) { |
| 482 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86_x2, 8, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 483 | } |
| 484 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 485 | static void spmm80_16x1__wasmsimd_x86_x2(benchmark::State& state, const char* net) { |
| 486 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2, 16, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 487 | } |
| 488 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 489 | static void spmm80_32x1__wasmsimd_arm_x4(benchmark::State& state, const char* net) { |
| 490 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x4, 32, 1, 0.8f); |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 491 | } |
| 492 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 493 | static void spmm80_4x1__wasmsimd_arm_x4(benchmark::State& state, const char* net) { |
| 494 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_arm_x4, 4, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 495 | } |
| 496 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 497 | static void spmm80_8x1__wasmsimd_arm_x4(benchmark::State& state, const char* net) { |
| 498 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm_x4, 8, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 499 | } |
| 500 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 501 | static void spmm80_16x1__wasmsimd_arm_x4(benchmark::State& state, const char* net) { |
| 502 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_x4, 16, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 503 | } |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 504 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 505 | static void spmm80_32x1__wasmsimd_x86_x4(benchmark::State& state, const char* net) { |
| 506 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x4, 32, 1, 0.8f); |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 507 | } |
| 508 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 509 | static void spmm80_4x1__wasmsimd_x86_x4(benchmark::State& state, const char* net) { |
| 510 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_x86_x4, 4, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 511 | } |
| 512 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 513 | static void spmm80_8x1__wasmsimd_x86_x4(benchmark::State& state, const char* net) { |
| 514 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86_x4, 8, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 515 | } |
| 516 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 517 | static void spmm80_16x1__wasmsimd_x86_x4(benchmark::State& state, const char* net) { |
| 518 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x4, 16, 1, 0.8f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 519 | } |
Frank Barchard | c2c0d16 | 2020-11-03 15:46:30 -0800 | [diff] [blame] | 520 | static void spmm80_4x1__wasmsimd_arm_pipelined(benchmark::State& state, const char* net) { |
| 521 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_arm_pipelined, 4, 1, 0.8f); |
| 522 | } |
| 523 | |
| 524 | static void spmm80_8x1__wasmsimd_arm_pipelined(benchmark::State& state, const char* net) { |
| 525 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm_pipelined, 8, 1, 0.8f); |
| 526 | } |
| 527 | |
| 528 | static void spmm80_16x1__wasmsimd_arm_pipelined(benchmark::State& state, const char* net) { |
| 529 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined, 16, 1, 0.8f); |
| 530 | } |
| 531 | |
| 532 | static void spmm80_32x1__wasmsimd_arm_pipelined(benchmark::State& state, const char* net) { |
| 533 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined, 32, 1, 0.8f); |
| 534 | } |
| 535 | |
| 536 | static void spmm80_4x1__wasmsimd_x86_pipelined(benchmark::State& state, const char* net) { |
| 537 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_x86_pipelined, 4, 1, 0.8f); |
| 538 | } |
| 539 | |
| 540 | static void spmm80_8x1__wasmsimd_x86_pipelined(benchmark::State& state, const char* net) { |
| 541 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86_pipelined, 8, 1, 0.8f); |
| 542 | } |
| 543 | |
| 544 | static void spmm80_16x1__wasmsimd_x86_pipelined(benchmark::State& state, const char* net) { |
| 545 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined, 16, 1, 0.8f); |
| 546 | } |
| 547 | |
| 548 | static void spmm80_32x1__wasmsimd_x86_pipelined(benchmark::State& state, const char* net) { |
| 549 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined, 32, 1, 0.8f); |
| 550 | } |
| 551 | |
| 552 | static void spmm80_32x1__wasmsimd_arm_pipelined_x2(benchmark::State& state, const char* net) { |
| 553 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2, 32, 1, 0.8f); |
| 554 | } |
| 555 | |
| 556 | static void spmm80_4x1__wasmsimd_arm_pipelined_x2(benchmark::State& state, const char* net) { |
| 557 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_arm_pipelined_x2, 4, 1, 0.8f); |
| 558 | } |
| 559 | |
| 560 | static void spmm80_8x1__wasmsimd_arm_pipelined_x2(benchmark::State& state, const char* net) { |
| 561 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm_pipelined_x2, 8, 1, 0.8f); |
| 562 | } |
| 563 | |
| 564 | static void spmm80_16x1__wasmsimd_arm_pipelined_x2(benchmark::State& state, const char* net) { |
| 565 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined_x2, 16, 1, 0.8f); |
| 566 | } |
| 567 | |
| 568 | static void spmm80_32x1__wasmsimd_x86_pipelined_x2(benchmark::State& state, const char* net) { |
| 569 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2, 32, 1, 0.8f); |
| 570 | } |
| 571 | |
| 572 | static void spmm80_4x1__wasmsimd_x86_pipelined_x2(benchmark::State& state, const char* net) { |
| 573 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_x86_pipelined_x2, 4, 1, 0.8f); |
| 574 | } |
| 575 | |
| 576 | static void spmm80_8x1__wasmsimd_x86_pipelined_x2(benchmark::State& state, const char* net) { |
| 577 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86_pipelined_x2, 8, 1, 0.8f); |
| 578 | } |
| 579 | |
| 580 | static void spmm80_16x1__wasmsimd_x86_pipelined_x2(benchmark::State& state, const char* net) { |
| 581 | SpMMBenchmark(state, xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined_x2, 16, 1, 0.8f); |
| 582 | } |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 583 | |
Marat Dukhan | 1631e3e | 2020-12-06 19:29:31 -0800 | [diff] [blame] | 584 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_arm) |
| 585 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_arm_x2) |
| 586 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_arm_x4) |
| 587 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_x86) |
| 588 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_x86_x2) |
| 589 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_x86_x4) |
| 590 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_arm) |
| 591 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_arm_x2) |
| 592 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_arm_x4) |
| 593 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_x86) |
| 594 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_x86_x2) |
| 595 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_x86_x4) |
| 596 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_arm) |
| 597 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_arm_x2) |
| 598 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_arm_x4) |
| 599 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_x86) |
| 600 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_x86_x2) |
| 601 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_x86_x4) |
| 602 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_arm) |
| 603 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_arm_x2) |
| 604 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_arm_x4) |
| 605 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_x86) |
| 606 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_x86_x2) |
| 607 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_x86_x4) |
| 608 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_arm_pipelined) |
| 609 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_arm_pipelined_x2) |
| 610 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_x86_pipelined) |
| 611 | BENCHMARK_SPMM(spmm80_4x1__wasmsimd_x86_pipelined_x2) |
| 612 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_arm_pipelined) |
| 613 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_arm_pipelined_x2) |
| 614 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_x86_pipelined) |
| 615 | BENCHMARK_SPMM(spmm80_8x1__wasmsimd_x86_pipelined_x2) |
| 616 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_arm_pipelined) |
| 617 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_arm_pipelined_x2) |
| 618 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_x86_pipelined) |
| 619 | BENCHMARK_SPMM(spmm80_16x1__wasmsimd_x86_pipelined_x2) |
| 620 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_arm_pipelined) |
| 621 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_arm_pipelined_x2) |
| 622 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_x86_pipelined) |
| 623 | BENCHMARK_SPMM(spmm80_32x1__wasmsimd_x86_pipelined_x2) |
Frank Barchard | c2c0d16 | 2020-11-03 15:46:30 -0800 | [diff] [blame] | 624 | |
Frank Barchard | 9e05340 | 2020-10-19 15:29:08 -0700 | [diff] [blame] | 625 | #endif // XNN_ARCH_WASMSIMD |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 626 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 627 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 628 | BENCHMARK_MAIN(); |
| 629 | #endif |