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> |
Frank Barchard | 31328cb | 2020-10-12 11:55:18 -0700 | [diff] [blame] | 19 | #ifdef BENCHMARK_RUY |
| 20 | #include "ruy/ruy.h" |
| 21 | #endif // BENCHMARK_RUY |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 22 | #include "bench/gemm.h" |
| 23 | #include "bench/utils.h" |
| 24 | #include <xnnpack/AlignedAllocator.h> |
| 25 | #include <xnnpack/common.h> |
| 26 | #include <xnnpack/gemm.h> |
| 27 | #include <xnnpack/pack.h> |
| 28 | #include <xnnpack/params-init.h> |
| 29 | #include <xnnpack/params.h> |
| 30 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 31 | static void GEMMBenchmark(benchmark::State& state, |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 32 | xnn_qs8_gemm_minmax_ukernel_function gemm, |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 33 | size_t mr, size_t nr, size_t kr, size_t sr, |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 34 | xnn_init_qs8_conv_minmax_params_fn init_params, |
Marat Dukhan | d5694df | 2021-05-20 17:10:40 -0700 | [diff] [blame] | 35 | benchmark::utils::IsaCheckFunction isa_check = nullptr, |
| 36 | bool extended_weights = false) |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 37 | { |
| 38 | if (!cpuinfo_initialize()) { |
| 39 | state.SkipWithError("cpuinfo initialization failed"); |
| 40 | return; |
| 41 | } |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 42 | if (isa_check && !isa_check(state)) { |
| 43 | return; |
| 44 | } |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 45 | |
| 46 | const size_t mc = state.range(0); |
| 47 | const size_t nc = state.range(1); |
| 48 | const size_t kc = state.range(2); |
| 49 | |
| 50 | const size_t nc_stride = benchmark::utils::RoundUp(nc, nr); |
Marat Dukhan | fbd67a7 | 2022-01-31 18:03:50 -0800 | [diff] [blame] | 51 | const size_t kc_stride = benchmark::utils::RoundUp(kc, kr * sr); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 52 | |
| 53 | std::random_device random_device; |
| 54 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 55 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), std::ref(rng)); |
| 56 | auto i8rng = std::bind( |
Marat Dukhan | b657605 | 2021-07-26 22:13:32 -0700 | [diff] [blame] | 57 | std::uniform_int_distribution<int32_t>(-std::numeric_limits<int8_t>::max(), std::numeric_limits<int8_t>::max()), std::ref(rng)); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 58 | |
Marat Dukhan | 91351ef | 2021-08-04 16:32:28 -0700 | [diff] [blame] | 59 | std::vector<int8_t> a(mc * kc + XNN_EXTRA_BYTES / sizeof(int8_t)); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 60 | std::generate(a.begin(), a.end(), std::ref(i8rng)); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 61 | std::vector<int8_t> k(nc * kc); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 62 | std::generate(k.begin(), k.end(), std::ref(i8rng)); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 63 | std::vector<int32_t> b(nc); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 64 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 65 | |
Marat Dukhan | d5694df | 2021-05-20 17:10:40 -0700 | [diff] [blame] | 66 | const size_t w_element_size = extended_weights ? sizeof(int16_t) : sizeof(int8_t); |
| 67 | const size_t w_size = nc_stride * sizeof(int32_t) + kc_stride * nc_stride * w_element_size; |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 68 | const size_t c_elements = mc * nc; |
| 69 | const size_t num_buffers = 1 + |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 70 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), w_size + c_elements * sizeof(int8_t)); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 71 | |
Marat Dukhan | e13e639 | 2021-07-26 22:22:35 -0700 | [diff] [blame] | 72 | std::vector<char, AlignedAllocator<char, 64>> w(w_size * num_buffers); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 73 | std::fill(w.begin(), w.end(), 0); |
| 74 | const xnn_qs8_packing_params packing_params = { 127 }; |
Marat Dukhan | d5694df | 2021-05-20 17:10:40 -0700 | [diff] [blame] | 75 | if (extended_weights) { |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 76 | xnn_pack_qs8_gemm_xw_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data(), 0, &packing_params); |
Marat Dukhan | d5694df | 2021-05-20 17:10:40 -0700 | [diff] [blame] | 77 | } else { |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 78 | xnn_pack_qs8_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data(), 0, &packing_params); |
Marat Dukhan | d5694df | 2021-05-20 17:10:40 -0700 | [diff] [blame] | 79 | } |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 80 | std::vector<int8_t> c(c_elements * num_buffers); |
| 81 | std::fill(c.begin(), c.end(), 0xA5); |
| 82 | |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 83 | union xnn_qs8_conv_minmax_params quantization_params; |
Marat Dukhan | 725f47e | 2021-05-22 10:06:19 -0700 | [diff] [blame] | 84 | init_params(&quantization_params, 0.75f, 127, -127, 126); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 85 | |
| 86 | size_t buffer_index = 0; |
| 87 | for (auto _ : state) { |
| 88 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 89 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 90 | // - W is not in cache (for any cache level) |
| 91 | // - C is not in cache (for any cache level) |
| 92 | state.PauseTiming(); |
| 93 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(int8_t)); |
| 94 | buffer_index = (buffer_index + 1) % num_buffers; |
| 95 | state.ResumeTiming(); |
| 96 | |
| 97 | for (uint32_t m = 0; m < mc; m += mr) { |
| 98 | const uint32_t mb = min(mc - m, mr); |
| 99 | for (uint32_t n = 0; n < nc; n += nr) { |
| 100 | const uint32_t nb = min(nc - n, nr); |
| 101 | gemm( |
| 102 | mb, nb, kc * sizeof(int8_t), |
| 103 | a.data() + m * kc, kc * sizeof(int8_t), |
Marat Dukhan | d5694df | 2021-05-20 17:10:40 -0700 | [diff] [blame] | 104 | w.data() + w_size * buffer_index + n * (kc_stride * w_element_size + sizeof(int32_t)), |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 105 | c.data() + (mc * buffer_index + m) * nc + n, nc * sizeof(int8_t), nr * sizeof(int8_t), |
| 106 | &quantization_params); |
| 107 | } |
| 108 | } |
| 109 | } |
| 110 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 111 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 112 | if (cpu_frequency != 0) { |
| 113 | state.counters["cpufreq"] = cpu_frequency; |
| 114 | } |
| 115 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 116 | state.counters["OPS"] = benchmark::Counter( |
| 117 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 118 | } |
| 119 | |
Frank Barchard | 31328cb | 2020-10-12 11:55:18 -0700 | [diff] [blame] | 120 | #ifdef BENCHMARK_RUY |
| 121 | static void RuyBenchmark(benchmark::State& state, size_t threads) |
| 122 | { |
| 123 | const size_t mc = state.range(0); |
| 124 | const size_t nc = state.range(1); |
| 125 | const size_t kc = state.range(2); |
| 126 | |
| 127 | std::random_device random_device; |
| 128 | auto rng = std::mt19937(random_device()); |
| 129 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), std::ref(rng)); |
| 130 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
| 131 | |
| 132 | const size_t num_buffers = 1 + |
| 133 | benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), |
| 134 | nc * (sizeof(int8_t) * (mc + kc) + sizeof(int32_t))); |
| 135 | |
| 136 | std::vector<int8_t> a(mc * kc); |
| 137 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 138 | std::vector<int8_t> k(num_buffers * nc * kc); |
| 139 | std::generate(k.begin(), k.end(), std::ref(u8rng)); |
| 140 | std::vector<int32_t> b(num_buffers * nc); |
| 141 | std::generate(b.begin(), b.end(), std::ref(i32rng)); |
| 142 | std::vector<int8_t> c(num_buffers * nc * mc); |
| 143 | std::fill(c.begin(), c.end(), std::nanf("")); |
| 144 | |
| 145 | // Note: context must be static to avoid the cost of re-creating it for each benchmark. |
| 146 | static ruy::Context context; |
| 147 | context.set_max_num_threads(threads); |
| 148 | |
| 149 | ruy::Matrix<int8_t> ruy_a; |
| 150 | ruy::MakeSimpleLayout(nc, kc, ruy::Order::kRowMajor, ruy_a.mutable_layout()); |
| 151 | ruy_a.set_zero_point(127); |
| 152 | ruy::Matrix<int8_t> ruy_b; |
| 153 | ruy::MakeSimpleLayout(kc, mc, ruy::Order::kColMajor, ruy_b.mutable_layout()); |
| 154 | ruy_b.set_data(a.data()); |
| 155 | ruy_b.set_zero_point(127); |
| 156 | ruy::Matrix<int8_t> ruy_c; |
| 157 | ruy::MakeSimpleLayout(nc, mc, ruy::Order::kColMajor, ruy_c.mutable_layout()); |
| 158 | ruy_c.set_zero_point(127); |
| 159 | |
| 160 | ruy::MulParams<int32_t, int8_t> mul_params; |
| 161 | mul_params.set_multiplier_fixedpoint(0x40000000); |
| 162 | |
| 163 | // ruy::Context uses deferred initialization, which affects percieved GEMM performance. Initialization happens during |
| 164 | // the first GEMM calls, and per Benoit Jacob it takes up to ~250 milliseconds for performance to stabilize. |
| 165 | // Thus, on the first benchmark, we compute GEMM for 500 milliseconds (to be safe) without recording performance, and |
| 166 | // keep the ruy::Context object initialized (by being static) between subsequent benchmarks. |
| 167 | static std::once_flag warmup; |
| 168 | std::call_once(warmup, [&](){ |
| 169 | auto start = std::chrono::steady_clock::now(); |
| 170 | do { |
| 171 | ruy_a.set_data(k.data()); |
| 172 | ruy_c.set_data(c.data()); |
| 173 | mul_params.set_bias(b.data()); |
| 174 | |
| 175 | ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); |
| 176 | } while (std::chrono::duration<double>(std::chrono::steady_clock::now() - start).count() < 0.5); |
| 177 | }); |
| 178 | |
| 179 | size_t buffer_index = 0; |
| 180 | for (auto _ : state) { |
| 181 | // Use circular buffers (exceeding cache size) and prefetch to control cache state: |
| 182 | // - A is always in L1 cache (if fits, otherwise L2, L3, etc) |
| 183 | // - K is not in cache (for any cache level) |
| 184 | // - B is not in cache (for any cache level) |
| 185 | // - C is not in cache (for any cache level) |
| 186 | state.PauseTiming(); |
| 187 | benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(int8_t)); |
| 188 | buffer_index = (buffer_index + 1) % num_buffers; |
| 189 | state.ResumeTiming(); |
| 190 | |
| 191 | ruy_a.set_data(k.data() + buffer_index * nc * kc); |
| 192 | ruy_c.set_data(c.data() + buffer_index * mc * nc); |
| 193 | mul_params.set_bias(b.data() + buffer_index * nc); |
| 194 | |
| 195 | ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); |
| 196 | } |
| 197 | |
Marat Dukhan | d713e8a | 2020-12-04 14:23:12 -0800 | [diff] [blame] | 198 | const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| 199 | if (cpu_frequency != 0) { |
| 200 | state.counters["cpufreq"] = cpu_frequency; |
| 201 | } |
| 202 | |
Frank Barchard | 31328cb | 2020-10-12 11:55:18 -0700 | [diff] [blame] | 203 | state.counters["OPS"] = benchmark::Counter( |
| 204 | uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); |
| 205 | } |
| 206 | |
| 207 | static void ruy_st(benchmark::State& state, const char* net) |
| 208 | { |
| 209 | RuyBenchmark(state, 1); |
| 210 | } |
| 211 | #endif // BENCHMARK_RUY |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 212 | |
Zhi An Ng | 1bef0f2 | 2022-01-07 16:13:31 -0800 | [diff] [blame] | 213 | #if XNN_ARCH_ARM && XNN_PLATFORM_JIT && XNN_ENABLE_JIT |
| 214 | static void GEMMBenchmark(benchmark::State& state, |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 215 | xnn_jit_gemm_code_generator_function generator, |
Zhi An Ng | 1bef0f2 | 2022-01-07 16:13:31 -0800 | [diff] [blame] | 216 | size_t mr, size_t nr, size_t kr, size_t sr, |
| 217 | xnn_init_qs8_conv_minmax_params_fn init_params, |
| 218 | benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| 219 | { |
| 220 | xnn_code_buffer code_buffer; |
| 221 | xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE); |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 222 | const size_t nc = state.range(1); |
| 223 | const size_t kc = state.range(2); |
| 224 | generator(&code_buffer, nc, kc, nullptr); |
Zhi An Ng | 1bef0f2 | 2022-01-07 16:13:31 -0800 | [diff] [blame] | 225 | GEMMBenchmark( |
| 226 | state, |
| 227 | reinterpret_cast<xnn_qs8_gemm_minmax_ukernel_function>(code_buffer.code), |
Frank Barchard | ca51090 | 2022-02-02 23:21:53 -0800 | [diff] [blame] | 228 | mr, nr, kr, sr, init_params, isa_check); |
Zhi An Ng | 1bef0f2 | 2022-01-07 16:13:31 -0800 | [diff] [blame] | 229 | xnn_release_code_memory(&code_buffer); |
| 230 | } |
| 231 | |
| 232 | static void jit_qs8_gemm_4x8c4__aarch32_neondot_ld64(benchmark::State& state, const char* net) { |
| 233 | GEMMBenchmark(state, xnn_generate_qs8_gemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, 4, 8, 4, 1, |
| 234 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
| 235 | } |
| 236 | static void jit_qs8_gemm_4x8__aarch32_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { |
| 237 | GEMMBenchmark(state, xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, 4, 8, 1, 1, |
| 238 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 239 | } |
| 240 | static void jit_qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* net) { |
| 241 | GEMMBenchmark(state, xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, 4, 8, 1, 1, |
| 242 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 243 | } |
| 244 | BENCHMARK_GEMM(jit_qs8_gemm_4x8c4__aarch32_neondot_ld64) |
| 245 | BENCHMARK_GEMM(jit_qs8_gemm_4x8__aarch32_neon_mlal_lane_ld64) |
| 246 | BENCHMARK_GEMM(jit_qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
| 247 | #endif // XNN_ARCH_ARM && XNN_PLATFORM_JIT && XNN_ENABLE_JIT |
| 248 | |
Frank Barchard | da7b2e2 | 2021-12-13 23:50:53 -0800 | [diff] [blame] | 249 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
Frank Barchard | 9f3f420 | 2021-12-16 18:13:51 -0800 | [diff] [blame] | 250 | static void qs8_gemm_4x8c4__aarch32_neondot_ld64(benchmark::State& state, const char* net) { |
| 251 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, 4, 8, 4, 1, |
| 252 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
| 253 | } |
Frank Barchard | 0f294ad | 2022-01-24 10:48:38 -0800 | [diff] [blame] | 254 | static void qs8_gemm_4x8c4__aarch32_neondot_cortex_a55(benchmark::State& state, const char* net) { |
| 255 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_cortex_a55, 4, 8, 4, 1, |
| 256 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
| 257 | } |
Frank Barchard | 6cb0fd0 | 2022-02-02 23:36:02 -0800 | [diff] [blame] | 258 | static void qs8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53(benchmark::State& state, const char* net) { |
| 259 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, 4, 8, 1, 1, |
| 260 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 261 | } |
| 262 | static void qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, const char* net) { |
| 263 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, 4, 8, 1, 1, |
| 264 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 265 | } |
| 266 | static void qs8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7(benchmark::State& state, const char* net) { |
| 267 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, 4, 8, 1, 1, |
| 268 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 269 | } |
| 270 | static void qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7(benchmark::State& state, const char* net) { |
| 271 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, 4, 8, 1, 1, |
| 272 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 273 | } |
Frank Barchard | da7b2e2 | 2021-12-13 23:50:53 -0800 | [diff] [blame] | 274 | static void qs8_gemm_4x8__aarch32_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { |
| 275 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, 4, 8, 1, 1, |
| 276 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 277 | } |
| 278 | static void qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* net) { |
| 279 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, 4, 8, 1, 1, |
| 280 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 281 | } |
Frank Barchard | 9f3f420 | 2021-12-16 18:13:51 -0800 | [diff] [blame] | 282 | BENCHMARK_GEMM(qs8_gemm_4x8c4__aarch32_neondot_ld64) |
Frank Barchard | 0f294ad | 2022-01-24 10:48:38 -0800 | [diff] [blame] | 283 | BENCHMARK_GEMM(qs8_gemm_4x8c4__aarch32_neondot_cortex_a55) |
Frank Barchard | 6cb0fd0 | 2022-02-02 23:36:02 -0800 | [diff] [blame] | 284 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53) |
| 285 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53) |
| 286 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7) |
| 287 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7) |
Frank Barchard | da7b2e2 | 2021-12-13 23:50:53 -0800 | [diff] [blame] | 288 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch32_neon_mlal_lane_ld64) |
| 289 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
| 290 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 291 | |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 292 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 293 | static void qs8_gemm_4x16c4__aarch64_neondot_cortex_a55(benchmark::State& state, const char* net) { |
| 294 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, 4, 16, 4, 1, |
| 295 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 296 | } |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 297 | static void qs8_gemm_1x16c4__aarch64_neondot_ld32(benchmark::State& state, const char* net) { |
| 298 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__aarch64_neondot_ld32, 1, 16, 4, 1, |
| 299 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 300 | } |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 301 | static void qs8_gemm_1x16c4__aarch64_neondot_ld64(benchmark::State& state, const char* net) { |
| 302 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__aarch64_neondot_ld64, 1, 16, 4, 1, |
| 303 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 304 | } |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 305 | static void qs8_gemm_4x16c4__aarch64_neondot_ld32(benchmark::State& state, const char* net) { |
| 306 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld32, 4, 16, 4, 1, |
| 307 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 308 | } |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 309 | static void qs8_gemm_4x16c4__aarch64_neondot_ld64(benchmark::State& state, const char* net) { |
| 310 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, 4, 16, 4, 1, |
| 311 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 312 | } |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 313 | static void qs8_gemm_4x16c4__aarch64_neondot_ld128(benchmark::State& state, const char* net) { |
| 314 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, 4, 16, 4, 1, |
| 315 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 316 | } |
Frank Barchard | 914f57b | 2021-12-13 12:31:42 -0800 | [diff] [blame] | 317 | static void qs8_gemm_4x8__aarch64_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { |
| 318 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, 4, 8, 1, 1, |
| 319 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 320 | } |
| 321 | static void qs8_gemm_4x8__aarch64_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* net) { |
| 322 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_prfm_ld64, 4, 8, 1, 1, |
| 323 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 324 | } |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 325 | static void qs8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53(benchmark::State& state, const char* net) { |
| 326 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, 4, 16, 1, 1, |
| 327 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 328 | } |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 329 | static void qs8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, const char* net) { |
| 330 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, 4, 16, 1, 1, |
| 331 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 332 | } |
Frank Barchard | 5cffb64 | 2021-11-22 13:59:43 -0800 | [diff] [blame] | 333 | static void qs8_gemm_4x16__aarch64_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { |
| 334 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64, 4, 16, 1, 1, |
| 335 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 336 | } |
| 337 | static void qs8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* net) { |
| 338 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, 4, 16, 1, 1, |
| 339 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 340 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 341 | static void qs8_gemm_1x8c8__aarch64_neon_mlal_prfm(benchmark::State& state, const char* net) { |
| 342 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm, 1, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 343 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 344 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 345 | static void qs8_gemm_1x8c8__aarch64_neon_mlal(benchmark::State& state, const char* net) { |
| 346 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, 1, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 347 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 348 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 349 | static void qs8_gemm_1x8c8__aarch64_neon_mlal_cortex_a53(benchmark::State& state, const char* net) { |
| 350 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, 1, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 351 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 352 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 353 | static void qs8_gemm_1x8c8__aarch64_neon_mlal_prfm_cortex_a53(benchmark::State& state, const char* net) { |
| 354 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53, 1, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 355 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 356 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 357 | static void qs8_gemm_2x8c8__aarch64_neon_mull(benchmark::State& state, const char* net) { |
| 358 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, 2, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 359 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 360 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 361 | static void qs8_gemm_2x8c8__aarch64_neon_mlal(benchmark::State& state, const char* net) { |
| 362 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, 2, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 363 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 364 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 365 | static void qs8_gemm_2x8c8__aarch64_neon_mlal_prfm(benchmark::State& state, const char* net) { |
| 366 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, 2, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 367 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 368 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 369 | static void qs8_gemm_2x8c8__aarch64_neon_mlal_cortex_a53(benchmark::State& state, const char* net) { |
| 370 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, 2, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 371 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 372 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 373 | static void qs8_gemm_2x8c8__aarch64_neon_mlal_prfm_cortex_a53(benchmark::State& state, const char* net) { |
| 374 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, 2, 8, 8, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 375 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 376 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 377 | static void qs8_gemm_2x8c16__aarch64_neon_mlal(benchmark::State& state, const char* net) { |
| 378 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__aarch64_neon_mlal, 2, 8, 16, 1, |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 379 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 380 | } |
| 381 | |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 382 | BENCHMARK_GEMM(qs8_gemm_1x16c4__aarch64_neondot_ld32) |
| 383 | BENCHMARK_GEMM(qs8_gemm_1x16c4__aarch64_neondot_ld64) |
| 384 | BENCHMARK_GEMM(qs8_gemm_4x16c4__aarch64_neondot_ld32) |
| 385 | BENCHMARK_GEMM(qs8_gemm_4x16c4__aarch64_neondot_ld64) |
| 386 | BENCHMARK_GEMM(qs8_gemm_4x16c4__aarch64_neondot_ld128) |
| 387 | BENCHMARK_GEMM(qs8_gemm_4x16c4__aarch64_neondot_cortex_a55) |
Frank Barchard | 914f57b | 2021-12-13 12:31:42 -0800 | [diff] [blame] | 388 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch64_neon_mlal_lane_ld64) |
| 389 | BENCHMARK_GEMM(qs8_gemm_4x8__aarch64_neon_mlal_lane_prfm_ld64) |
Frank Barchard | 889ed10 | 2021-08-20 15:01:29 -0700 | [diff] [blame] | 390 | BENCHMARK_GEMM(qs8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53) |
| 391 | BENCHMARK_GEMM(qs8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53) |
Frank Barchard | 5cffb64 | 2021-11-22 13:59:43 -0800 | [diff] [blame] | 392 | BENCHMARK_GEMM(qs8_gemm_4x16__aarch64_neon_mlal_lane_ld64) |
| 393 | BENCHMARK_GEMM(qs8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64) |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 394 | BENCHMARK_GEMM(qs8_gemm_1x8c8__aarch64_neon_mlal_prfm) |
| 395 | BENCHMARK_GEMM(qs8_gemm_1x8c8__aarch64_neon_mlal) |
| 396 | BENCHMARK_GEMM(qs8_gemm_1x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 397 | BENCHMARK_GEMM(qs8_gemm_1x8c8__aarch64_neon_mlal_cortex_a53) |
| 398 | BENCHMARK_GEMM(qs8_gemm_2x8c8__aarch64_neon_mull) |
| 399 | BENCHMARK_GEMM(qs8_gemm_2x8c8__aarch64_neon_mlal) |
| 400 | BENCHMARK_GEMM(qs8_gemm_2x8c8__aarch64_neon_mlal_prfm) |
| 401 | BENCHMARK_GEMM(qs8_gemm_2x8c8__aarch64_neon_mlal_cortex_a53) |
| 402 | BENCHMARK_GEMM(qs8_gemm_2x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 403 | BENCHMARK_GEMM(qs8_gemm_2x8c16__aarch64_neon_mlal) |
Frank Barchard | 4c3e5a9 | 2021-08-16 19:17:39 -0700 | [diff] [blame] | 404 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 405 | |
| 406 | |
Marat Dukhan | ab67142 | 2020-08-05 16:39:04 -0700 | [diff] [blame] | 407 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 408 | static void qs8_gemm_1x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 409 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, 1, 8, 1, 1, |
| 410 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | a93765f | 2021-01-27 16:25:34 -0800 | [diff] [blame] | 411 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 412 | static void qs8_gemm_2x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 413 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, 2, 8, 1, 1, |
| 414 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Marat Dukhan | ab67142 | 2020-08-05 16:39:04 -0700 | [diff] [blame] | 415 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 416 | static void qs8_gemm_3x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 417 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, 3, 8, 1, 1, |
| 418 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | ec0bf14 | 2021-01-15 10:14:05 -0800 | [diff] [blame] | 419 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 420 | static void qs8_gemm_4x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 421 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane, 4, 8, 1, 1, |
| 422 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | cfbc849 | 2021-01-12 15:43:19 -0800 | [diff] [blame] | 423 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 424 | static void qs8_gemm_6x8__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 425 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, 6, 8, 1, 1, |
| 426 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5549735 | 2021-04-30 11:47:56 -0700 | [diff] [blame] | 427 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 428 | static void qs8_gemm_1x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 429 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, 1, 16, 1, 1, |
| 430 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | a93765f | 2021-01-27 16:25:34 -0800 | [diff] [blame] | 431 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 432 | static void qs8_gemm_2x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 433 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, 2, 16, 1, 1, |
| 434 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Marat Dukhan | ab67142 | 2020-08-05 16:39:04 -0700 | [diff] [blame] | 435 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 436 | static void qs8_gemm_3x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 437 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, 3, 16, 1, 1, |
| 438 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | ec0bf14 | 2021-01-15 10:14:05 -0800 | [diff] [blame] | 439 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 440 | static void qs8_gemm_4x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 441 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, 4, 16, 1, 1, |
| 442 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | cfbc849 | 2021-01-12 15:43:19 -0800 | [diff] [blame] | 443 | } |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 444 | static void qs8_gemm_6x16__neon_mlal_lane(benchmark::State& state, const char* net) { |
| 445 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, 6, 16, 1, 1, |
| 446 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5549735 | 2021-04-30 11:47:56 -0700 | [diff] [blame] | 447 | } |
Frank Barchard | f82ea82 | 2021-12-01 15:43:37 -0800 | [diff] [blame] | 448 | static void qs8_gemm_1x8__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 449 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm, 1, 8, 1, 1, |
| 450 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 451 | } |
| 452 | static void qs8_gemm_2x8__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 453 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane_prfm, 2, 8, 1, 1, |
| 454 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 455 | } |
| 456 | static void qs8_gemm_3x8__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 457 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane_prfm, 3, 8, 1, 1, |
| 458 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 459 | } |
| 460 | static void qs8_gemm_4x8__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 461 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane_prfm, 4, 8, 1, 1, |
| 462 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 463 | } |
| 464 | static void qs8_gemm_6x8__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 465 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, 6, 8, 1, 1, |
| 466 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 467 | } |
| 468 | static void qs8_gemm_1x16__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 469 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane_prfm, 1, 16, 1, 1, |
| 470 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 471 | } |
| 472 | static void qs8_gemm_2x16__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 473 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, 2, 16, 1, 1, |
| 474 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 475 | } |
| 476 | static void qs8_gemm_3x16__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 477 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane_prfm, 3, 16, 1, 1, |
| 478 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 479 | } |
| 480 | static void qs8_gemm_4x16__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 481 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm, 4, 16, 1, 1, |
| 482 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 483 | } |
| 484 | static void qs8_gemm_6x16__neon_mlal_lane_prfm(benchmark::State& state, const char* net) { |
| 485 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, 6, 16, 1, 1, |
| 486 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 487 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 488 | static void qs8_gemm_1x8c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 489 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_dup, 1, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 490 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | a93765f | 2021-01-27 16:25:34 -0800 | [diff] [blame] | 491 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 492 | static void qs8_gemm_2x8c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 493 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup, 2, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 494 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 495 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 496 | static void qs8_gemm_3x8c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 497 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup, 3, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 498 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 499 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 500 | static void qs8_gemm_4x8c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 501 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, 4, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 502 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 503 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 504 | static void qs8_gemm_1x16c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 505 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, 1, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 506 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | a93765f | 2021-01-27 16:25:34 -0800 | [diff] [blame] | 507 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 508 | static void qs8_gemm_2x16c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 509 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, 2, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 510 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 511 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 512 | static void qs8_gemm_3x16c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 513 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, 3, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 514 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 515 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 516 | static void qs8_gemm_4x16c2__neon_mull_dup(benchmark::State& state, const char* net) { |
| 517 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, 4, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 518 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 519 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 520 | static void qs8_gemm_1x8c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 521 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, 1, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 522 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 523 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 524 | static void qs8_gemm_2x8c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 525 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, 2, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 526 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 527 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 528 | static void qs8_gemm_3x8c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 529 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, 3, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 530 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 531 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 532 | static void qs8_gemm_4x8c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 533 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, 4, 8, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 534 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 535 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 536 | static void qs8_gemm_1x16c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 537 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, 1, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 538 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 539 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 540 | static void qs8_gemm_2x16c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 541 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup, 2, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 542 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 543 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 544 | static void qs8_gemm_3x16c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 545 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup, 3, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 546 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 547 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 548 | static void qs8_gemm_4x16c2__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 549 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup, 4, 16, 2, 1, |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 550 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 8247e21 | 2021-02-03 18:12:33 -0800 | [diff] [blame] | 551 | } |
Frank Barchard | 15eec02 | 2021-11-17 13:26:20 -0800 | [diff] [blame] | 552 | static void qs8_gemm_1x8c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 553 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, 1, 8, 2, 1, |
| 554 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 555 | } |
| 556 | static void qs8_gemm_2x8c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 557 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r, 2, 8, 2, 1, |
| 558 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 559 | } |
| 560 | static void qs8_gemm_3x8c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 561 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r, 3, 8, 2, 1, |
| 562 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 563 | } |
| 564 | static void qs8_gemm_4x8c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 565 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, 4, 8, 2, 1, |
| 566 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 567 | } |
| 568 | static void qs8_gemm_1x16c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 569 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, 1, 16, 2, 1, |
| 570 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 571 | } |
| 572 | static void qs8_gemm_2x16c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 573 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r, 2, 16, 2, 1, |
| 574 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 575 | } |
| 576 | static void qs8_gemm_3x16c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 577 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r, 3, 16, 2, 1, |
| 578 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 579 | } |
| 580 | static void qs8_gemm_4x16c2__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 581 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, 4, 16, 2, 1, |
| 582 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 583 | } |
| 584 | static void qs8_gemm_1x8c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 585 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, 1, 8, 2, 1, |
| 586 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 587 | } |
| 588 | static void qs8_gemm_2x8c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 589 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld1r, 2, 8, 2, 1, |
| 590 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 591 | } |
| 592 | static void qs8_gemm_3x8c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 593 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r, 3, 8, 2, 1, |
| 594 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 595 | } |
| 596 | static void qs8_gemm_4x8c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 597 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, 4, 8, 2, 1, |
| 598 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 599 | } |
| 600 | static void qs8_gemm_1x16c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 601 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, 1, 16, 2, 1, |
| 602 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 603 | } |
| 604 | static void qs8_gemm_2x16c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 605 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r, 2, 16, 2, 1, |
| 606 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 607 | } |
| 608 | static void qs8_gemm_3x16c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 609 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r, 3, 16, 2, 1, |
| 610 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 611 | } |
| 612 | static void qs8_gemm_4x16c2__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 613 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, 4, 16, 2, 1, |
| 614 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 615 | } |
| 616 | static void qs8_gemm_1x8c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 617 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld2r, 1, 8, 2, 1, |
| 618 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 619 | } |
| 620 | static void qs8_gemm_2x8c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 621 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, 2, 8, 2, 1, |
| 622 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 623 | } |
| 624 | static void qs8_gemm_3x8c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 625 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r, 3, 8, 2, 1, |
| 626 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 627 | } |
| 628 | static void qs8_gemm_4x8c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 629 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, 4, 8, 2, 1, |
| 630 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 631 | } |
| 632 | static void qs8_gemm_1x16c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 633 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r, 1, 16, 2, 1, |
| 634 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 635 | } |
| 636 | static void qs8_gemm_2x16c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 637 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r, 2, 16, 2, 1, |
| 638 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 639 | } |
| 640 | static void qs8_gemm_3x16c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 641 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, 3, 16, 2, 1, |
| 642 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 643 | } |
| 644 | static void qs8_gemm_4x16c2__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 645 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r, 4, 16, 2, 1, |
| 646 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 647 | } |
| 648 | static void qs8_gemm_1x8c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 649 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, 1, 8, 2, 1, |
| 650 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 651 | } |
| 652 | static void qs8_gemm_2x8c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 653 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, 2, 8, 2, 1, |
| 654 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 655 | } |
| 656 | static void qs8_gemm_3x8c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 657 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r, 3, 8, 2, 1, |
| 658 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 659 | } |
| 660 | static void qs8_gemm_4x8c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 661 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r, 4, 8, 2, 1, |
| 662 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 663 | } |
| 664 | static void qs8_gemm_1x16c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 665 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r, 1, 16, 2, 1, |
| 666 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 667 | } |
| 668 | static void qs8_gemm_2x16c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 669 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, 2, 16, 2, 1, |
| 670 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 671 | } |
| 672 | static void qs8_gemm_3x16c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 673 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, 3, 16, 2, 1, |
| 674 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 675 | } |
| 676 | static void qs8_gemm_4x16c2__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 677 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r, 4, 16, 2, 1, |
| 678 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 679 | } |
Frank Barchard | 42f5c50 | 2021-11-16 10:04:21 -0800 | [diff] [blame] | 680 | static void qs8_gemm_1x8c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 681 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, 1, 8, 2, 1, |
| 682 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 683 | } |
| 684 | static void qs8_gemm_2x8c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 685 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, 2, 8, 2, 1, |
| 686 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 687 | } |
| 688 | static void qs8_gemm_3x8c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 689 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, 3, 8, 2, 1, |
| 690 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 691 | } |
| 692 | static void qs8_gemm_4x8c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 693 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, 4, 8, 2, 1, |
| 694 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 695 | } |
| 696 | static void qs8_gemm_1x16c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 697 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, 1, 16, 2, 1, |
| 698 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 699 | } |
| 700 | static void qs8_gemm_2x16c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 701 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, 2, 16, 2, 1, |
| 702 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 703 | } |
| 704 | static void qs8_gemm_3x16c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 705 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, 3, 16, 2, 1, |
| 706 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 707 | } |
| 708 | static void qs8_gemm_4x16c2__neon_mull_ld4r(benchmark::State& state, const char* net) { |
| 709 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, 4, 16, 2, 1, |
| 710 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 711 | } |
| 712 | static void qs8_gemm_1x8c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 713 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld4r, 1, 8, 2, 1, |
| 714 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 715 | } |
| 716 | static void qs8_gemm_2x8c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 717 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r, 2, 8, 2, 1, |
| 718 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 719 | } |
| 720 | static void qs8_gemm_3x8c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 721 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, 3, 8, 2, 1, |
| 722 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 723 | } |
| 724 | static void qs8_gemm_4x8c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 725 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, 4, 8, 2, 1, |
| 726 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 727 | } |
| 728 | static void qs8_gemm_1x16c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 729 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r, 1, 16, 2, 1, |
| 730 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 731 | } |
| 732 | static void qs8_gemm_2x16c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 733 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r, 2, 16, 2, 1, |
| 734 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 735 | } |
| 736 | static void qs8_gemm_3x16c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 737 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r, 3, 16, 2, 1, |
| 738 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 739 | } |
| 740 | static void qs8_gemm_4x16c2__neon_mlal_ld4r(benchmark::State& state, const char* net) { |
| 741 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, 4, 16, 2, 1, |
| 742 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 743 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 744 | static void qs8_gemm_1x8c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 745 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, 1, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 746 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 747 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 748 | static void qs8_gemm_2x8c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 749 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, 2, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 750 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 751 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 752 | static void qs8_gemm_3x8c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 753 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, 3, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 754 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 755 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 756 | static void qs8_gemm_4x8c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 757 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, 4, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 758 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 759 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 760 | static void qs8_gemm_1x16c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 761 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, 1, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 762 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 763 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 764 | static void qs8_gemm_2x16c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 765 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, 2, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 766 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 767 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 768 | static void qs8_gemm_3x16c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 769 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, 3, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 770 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 771 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 772 | static void qs8_gemm_4x16c2s4__neon_mull(benchmark::State& state, const char* net) { |
| 773 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull, 4, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 774 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 775 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 776 | static void qs8_gemm_1x8c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 777 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, 1, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 778 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 779 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 780 | static void qs8_gemm_2x8c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 781 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal, 2, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 782 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 783 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 784 | static void qs8_gemm_3x8c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 785 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal, 3, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 786 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 787 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 788 | static void qs8_gemm_4x8c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 789 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal, 4, 8, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 790 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 791 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 792 | static void qs8_gemm_1x16c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 793 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, 1, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 794 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 795 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 796 | static void qs8_gemm_2x16c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 797 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal, 2, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 798 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 799 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 800 | static void qs8_gemm_3x16c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 801 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal, 3, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 802 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 803 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 804 | static void qs8_gemm_4x16c2s4__neon_mlal(benchmark::State& state, const char* net) { |
| 805 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, 4, 16, 2, 4, |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 806 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 807 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 808 | static void qs8_gemm_1x8c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 809 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mull_dup, 1, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 810 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 811 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 812 | static void qs8_gemm_2x8c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 813 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, 2, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 814 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 815 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 816 | static void qs8_gemm_3x8c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 817 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, 3, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 818 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 819 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 820 | static void qs8_gemm_4x8c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 821 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, 4, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 822 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 823 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 824 | static void qs8_gemm_1x16c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 825 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, 1, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 826 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 827 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 828 | static void qs8_gemm_2x16c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 829 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, 2, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 830 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 831 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 832 | static void qs8_gemm_3x16c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 833 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, 3, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 834 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 835 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 836 | static void qs8_gemm_4x16c4__neon_mull_dup(benchmark::State& state, const char* net) { |
| 837 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup, 4, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 838 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 839 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 840 | static void qs8_gemm_1x8c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 841 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, 1, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 842 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 843 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 844 | static void qs8_gemm_2x8c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 845 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, 2, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 846 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 847 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 848 | static void qs8_gemm_3x8c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 849 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, 3, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 850 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 851 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 852 | static void qs8_gemm_4x8c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 853 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, 4, 8, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 854 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 855 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 856 | static void qs8_gemm_1x16c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 857 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, 1, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 858 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 859 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 860 | static void qs8_gemm_2x16c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 861 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup, 2, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 862 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 863 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 864 | static void qs8_gemm_3x16c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 865 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup, 3, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 866 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 867 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 868 | static void qs8_gemm_4x16c4__neon_mlal_dup(benchmark::State& state, const char* net) { |
| 869 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup, 4, 16, 4, 1, |
Frank Barchard | 5132010 | 2021-11-05 16:01:30 -0700 | [diff] [blame] | 870 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 871 | } |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 872 | static void qs8_gemm_1x8c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 873 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, 1, 8, 4, 1, |
| 874 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 875 | } |
| 876 | static void qs8_gemm_2x8c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 877 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, 2, 8, 4, 1, |
| 878 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 879 | } |
| 880 | static void qs8_gemm_3x8c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 881 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, 3, 8, 4, 1, |
| 882 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 883 | } |
| 884 | static void qs8_gemm_4x8c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 885 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, 4, 8, 4, 1, |
| 886 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 887 | } |
| 888 | static void qs8_gemm_1x16c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 889 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r, 1, 16, 4, 1, |
| 890 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 891 | } |
| 892 | static void qs8_gemm_2x16c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 893 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r, 2, 16, 4, 1, |
| 894 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 895 | } |
| 896 | static void qs8_gemm_3x16c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 897 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, 3, 16, 4, 1, |
| 898 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 899 | } |
| 900 | static void qs8_gemm_4x16c4__neon_mull_ld1r(benchmark::State& state, const char* net) { |
| 901 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, 4, 16, 4, 1, |
| 902 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 903 | } |
| 904 | static void qs8_gemm_1x8c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 905 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, 1, 8, 4, 1, |
| 906 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 907 | } |
| 908 | static void qs8_gemm_2x8c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 909 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r, 2, 8, 4, 1, |
| 910 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 911 | } |
| 912 | static void qs8_gemm_3x8c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 913 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, 3, 8, 4, 1, |
| 914 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 915 | } |
| 916 | static void qs8_gemm_4x8c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 917 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r, 4, 8, 4, 1, |
| 918 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 919 | } |
| 920 | static void qs8_gemm_1x16c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 921 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r, 1, 16, 4, 1, |
| 922 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 923 | } |
| 924 | static void qs8_gemm_2x16c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 925 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r, 2, 16, 4, 1, |
| 926 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 927 | } |
| 928 | static void qs8_gemm_3x16c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 929 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r, 3, 16, 4, 1, |
| 930 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 931 | } |
| 932 | static void qs8_gemm_4x16c4__neon_mlal_ld1r(benchmark::State& state, const char* net) { |
| 933 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, 4, 16, 4, 1, |
| 934 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 935 | } |
| 936 | static void qs8_gemm_1x8c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 937 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld2r, 1, 8, 4, 1, |
| 938 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 939 | } |
| 940 | static void qs8_gemm_2x8c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 941 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, 2, 8, 4, 1, |
| 942 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 943 | } |
| 944 | static void qs8_gemm_3x8c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 945 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, 3, 8, 4, 1, |
| 946 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 947 | } |
| 948 | static void qs8_gemm_4x8c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 949 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r, 4, 8, 4, 1, |
| 950 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 951 | } |
| 952 | static void qs8_gemm_1x16c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 953 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r, 1, 16, 4, 1, |
| 954 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 955 | } |
| 956 | static void qs8_gemm_2x16c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 957 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, 2, 16, 4, 1, |
| 958 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 959 | } |
| 960 | static void qs8_gemm_3x16c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 961 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, 3, 16, 4, 1, |
| 962 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 963 | } |
| 964 | static void qs8_gemm_4x16c4__neon_mull_ld2r(benchmark::State& state, const char* net) { |
| 965 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r, 4, 16, 4, 1, |
| 966 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 967 | } |
| 968 | static void qs8_gemm_1x8c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 969 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld2r, 1, 8, 4, 1, |
| 970 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 971 | } |
| 972 | static void qs8_gemm_2x8c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 973 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, 2, 8, 4, 1, |
| 974 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 975 | } |
| 976 | static void qs8_gemm_3x8c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 977 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r, 3, 8, 4, 1, |
| 978 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 979 | } |
| 980 | static void qs8_gemm_4x8c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 981 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r, 4, 8, 4, 1, |
| 982 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 983 | } |
| 984 | static void qs8_gemm_1x16c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 985 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r, 1, 16, 4, 1, |
| 986 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 987 | } |
| 988 | static void qs8_gemm_2x16c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 989 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, 2, 16, 4, 1, |
| 990 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 991 | } |
| 992 | static void qs8_gemm_3x16c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 993 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, 3, 16, 4, 1, |
| 994 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 995 | } |
| 996 | static void qs8_gemm_4x16c4__neon_mlal_ld2r(benchmark::State& state, const char* net) { |
| 997 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, 4, 16, 4, 1, |
| 998 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
| 999 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1000 | static void qs8_gemm_1x8c8__neon_mull(benchmark::State& state, const char* net) { |
| 1001 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mull, 1, 8, 8, 1, |
| 1002 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1003 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1004 | static void qs8_gemm_2x8c8__neon_mull(benchmark::State& state, const char* net) { |
| 1005 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, 2, 8, 8, 1, |
| 1006 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1007 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1008 | static void qs8_gemm_3x8c8__neon_mull(benchmark::State& state, const char* net) { |
| 1009 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, 3, 8, 8, 1, |
| 1010 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1011 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1012 | static void qs8_gemm_4x8c8__neon_mull(benchmark::State& state, const char* net) { |
| 1013 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, 4, 8, 8, 1, |
| 1014 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1015 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1016 | static void qs8_gemm_1x16c8__neon_mull(benchmark::State& state, const char* net) { |
| 1017 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull, 1, 16, 8, 1, |
| 1018 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1019 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1020 | static void qs8_gemm_2x16c8__neon_mull(benchmark::State& state, const char* net) { |
| 1021 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull, 2, 16, 8, 1, |
| 1022 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1023 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1024 | static void qs8_gemm_3x16c8__neon_mull(benchmark::State& state, const char* net) { |
| 1025 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull, 3, 16, 8, 1, |
| 1026 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1027 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1028 | static void qs8_gemm_4x16c8__neon_mull(benchmark::State& state, const char* net) { |
| 1029 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull, 4, 16, 8, 1, |
| 1030 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 1031 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1032 | static void qs8_gemm_1x8c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1033 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, 1, 8, 8, 1, |
| 1034 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1035 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1036 | static void qs8_gemm_2x8c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1037 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal, 2, 8, 8, 1, |
| 1038 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1039 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1040 | static void qs8_gemm_3x8c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1041 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, 3, 8, 8, 1, |
| 1042 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1043 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1044 | static void qs8_gemm_4x8c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1045 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal, 4, 8, 8, 1, |
| 1046 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1047 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1048 | static void qs8_gemm_1x16c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1049 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal, 1, 16, 8, 1, |
| 1050 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1051 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1052 | static void qs8_gemm_2x16c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1053 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal, 2, 16, 8, 1, |
| 1054 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1055 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1056 | static void qs8_gemm_3x16c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1057 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, 3, 16, 8, 1, |
| 1058 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1059 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1060 | static void qs8_gemm_4x16c8__neon_mlal(benchmark::State& state, const char* net) { |
| 1061 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, 4, 16, 8, 1, |
| 1062 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1063 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1064 | static void qs8_gemm_1x8c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1065 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal, 1, 8, 16, 1, |
| 1066 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1067 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1068 | static void qs8_gemm_2x8c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1069 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal, 2, 8, 16, 1, |
| 1070 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1071 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1072 | static void qs8_gemm_3x8c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1073 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, 3, 8, 16, 1, |
| 1074 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1075 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1076 | static void qs8_gemm_4x8c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1077 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal, 4, 8, 16, 1, |
| 1078 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1079 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1080 | static void qs8_gemm_1x16c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1081 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal, 1, 16, 16, 1, |
| 1082 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1083 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1084 | static void qs8_gemm_2x16c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1085 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal, 2, 16, 16, 1, |
| 1086 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1087 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1088 | static void qs8_gemm_3x16c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1089 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, 3, 16, 16, 1, |
| 1090 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1091 | } |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1092 | static void qs8_gemm_4x16c16__neon_mlal(benchmark::State& state, const char* net) { |
| 1093 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, 4, 16, 16, 1, |
| 1094 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEON); |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1095 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1096 | static void qs8_gemm_1x8c4__neondot(benchmark::State& state, const char* net) { |
| 1097 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neondot, 1, 8, 4, 1, |
| 1098 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | f1fd89e | 2020-10-06 23:27:11 -0700 | [diff] [blame] | 1099 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1100 | static void qs8_gemm_4x8c4__neondot(benchmark::State& state, const char* net) { |
| 1101 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, 4, 8, 4, 1, |
| 1102 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1103 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1104 | static void qs8_gemm_6x8c4__neondot(benchmark::State& state, const char* net) { |
| 1105 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot, 6, 8, 4, 1, |
| 1106 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1107 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1108 | static void qs8_gemm_8x8c4__neondot(benchmark::State& state, const char* net) { |
| 1109 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot, 8, 8, 4, 1, |
| 1110 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1111 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1112 | static void qs8_gemm_1x16c4__neondot(benchmark::State& state, const char* net) { |
| 1113 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, 1, 16, 4, 1, |
| 1114 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | f1fd89e | 2020-10-06 23:27:11 -0700 | [diff] [blame] | 1115 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1116 | static void qs8_gemm_4x16c4__neondot(benchmark::State& state, const char* net) { |
| 1117 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, 4, 16, 4, 1, |
| 1118 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1119 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1120 | static void qs8_gemm_6x16c4__neondot(benchmark::State& state, const char* net) { |
| 1121 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, 6, 16, 4, 1, |
| 1122 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | 2fa1745 | 2020-09-22 09:51:46 -0700 | [diff] [blame] | 1123 | } |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1124 | static void qs8_gemm_8x16c4__neondot(benchmark::State& state, const char* net) { |
| 1125 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, 8, 16, 4, 1, |
| 1126 | xnn_init_qs8_conv_minmax_rndnu_neon_params, benchmark::utils::CheckNEONDOT); |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1127 | } |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1128 | BENCHMARK_GEMM(qs8_gemm_1x8c4__neon_mull_dup) |
| 1129 | BENCHMARK_GEMM(qs8_gemm_2x8c4__neon_mull_dup) |
| 1130 | BENCHMARK_GEMM(qs8_gemm_3x8c4__neon_mull_dup) |
| 1131 | BENCHMARK_GEMM(qs8_gemm_4x8c4__neon_mull_dup) |
| 1132 | BENCHMARK_GEMM(qs8_gemm_1x16c4__neon_mull_dup) |
| 1133 | BENCHMARK_GEMM(qs8_gemm_2x16c4__neon_mull_dup) |
| 1134 | BENCHMARK_GEMM(qs8_gemm_3x16c4__neon_mull_dup) |
| 1135 | BENCHMARK_GEMM(qs8_gemm_4x16c4__neon_mull_dup) |
| 1136 | BENCHMARK_GEMM(qs8_gemm_1x8c4__neon_mlal_dup) |
| 1137 | BENCHMARK_GEMM(qs8_gemm_2x8c4__neon_mlal_dup) |
| 1138 | BENCHMARK_GEMM(qs8_gemm_3x8c4__neon_mlal_dup) |
| 1139 | BENCHMARK_GEMM(qs8_gemm_4x8c4__neon_mlal_dup) |
| 1140 | BENCHMARK_GEMM(qs8_gemm_1x16c4__neon_mlal_dup) |
| 1141 | BENCHMARK_GEMM(qs8_gemm_2x16c4__neon_mlal_dup) |
| 1142 | BENCHMARK_GEMM(qs8_gemm_3x16c4__neon_mlal_dup) |
| 1143 | BENCHMARK_GEMM(qs8_gemm_4x16c4__neon_mlal_dup) |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 1144 | BENCHMARK_GEMM(qs8_gemm_1x8c4__neon_mull_ld1r) |
| 1145 | BENCHMARK_GEMM(qs8_gemm_2x8c4__neon_mull_ld1r) |
| 1146 | BENCHMARK_GEMM(qs8_gemm_3x8c4__neon_mull_ld1r) |
| 1147 | BENCHMARK_GEMM(qs8_gemm_4x8c4__neon_mull_ld1r) |
| 1148 | BENCHMARK_GEMM(qs8_gemm_1x16c4__neon_mull_ld1r) |
| 1149 | BENCHMARK_GEMM(qs8_gemm_2x16c4__neon_mull_ld1r) |
| 1150 | BENCHMARK_GEMM(qs8_gemm_3x16c4__neon_mull_ld1r) |
| 1151 | BENCHMARK_GEMM(qs8_gemm_4x16c4__neon_mull_ld1r) |
| 1152 | BENCHMARK_GEMM(qs8_gemm_1x8c4__neon_mlal_ld1r) |
| 1153 | BENCHMARK_GEMM(qs8_gemm_2x8c4__neon_mlal_ld1r) |
| 1154 | BENCHMARK_GEMM(qs8_gemm_3x8c4__neon_mlal_ld1r) |
| 1155 | BENCHMARK_GEMM(qs8_gemm_4x8c4__neon_mlal_ld1r) |
| 1156 | BENCHMARK_GEMM(qs8_gemm_1x16c4__neon_mlal_ld1r) |
| 1157 | BENCHMARK_GEMM(qs8_gemm_2x16c4__neon_mlal_ld1r) |
| 1158 | BENCHMARK_GEMM(qs8_gemm_3x16c4__neon_mlal_ld1r) |
| 1159 | BENCHMARK_GEMM(qs8_gemm_4x16c4__neon_mlal_ld1r) |
| 1160 | BENCHMARK_GEMM(qs8_gemm_1x8c4__neon_mull_ld2r) |
| 1161 | BENCHMARK_GEMM(qs8_gemm_2x8c4__neon_mull_ld2r) |
| 1162 | BENCHMARK_GEMM(qs8_gemm_3x8c4__neon_mull_ld2r) |
| 1163 | BENCHMARK_GEMM(qs8_gemm_4x8c4__neon_mull_ld2r) |
| 1164 | BENCHMARK_GEMM(qs8_gemm_1x16c4__neon_mull_ld2r) |
| 1165 | BENCHMARK_GEMM(qs8_gemm_2x16c4__neon_mull_ld2r) |
| 1166 | BENCHMARK_GEMM(qs8_gemm_3x16c4__neon_mull_ld2r) |
| 1167 | BENCHMARK_GEMM(qs8_gemm_4x16c4__neon_mull_ld2r) |
| 1168 | BENCHMARK_GEMM(qs8_gemm_1x8c4__neon_mlal_ld2r) |
| 1169 | BENCHMARK_GEMM(qs8_gemm_2x8c4__neon_mlal_ld2r) |
| 1170 | BENCHMARK_GEMM(qs8_gemm_3x8c4__neon_mlal_ld2r) |
| 1171 | BENCHMARK_GEMM(qs8_gemm_4x8c4__neon_mlal_ld2r) |
| 1172 | BENCHMARK_GEMM(qs8_gemm_1x16c4__neon_mlal_ld2r) |
| 1173 | BENCHMARK_GEMM(qs8_gemm_2x16c4__neon_mlal_ld2r) |
| 1174 | BENCHMARK_GEMM(qs8_gemm_3x16c4__neon_mlal_ld2r) |
| 1175 | BENCHMARK_GEMM(qs8_gemm_4x16c4__neon_mlal_ld2r) |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1176 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mull_dup) |
| 1177 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mull_dup) |
| 1178 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mull_dup) |
| 1179 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mull_dup) |
| 1180 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mull_dup) |
| 1181 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mull_dup) |
| 1182 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mull_dup) |
| 1183 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mull_dup) |
| 1184 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mlal_dup) |
| 1185 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mlal_dup) |
| 1186 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mlal_dup) |
| 1187 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mlal_dup) |
| 1188 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mlal_dup) |
| 1189 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mlal_dup) |
| 1190 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mlal_dup) |
| 1191 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mlal_dup) |
Frank Barchard | 15eec02 | 2021-11-17 13:26:20 -0800 | [diff] [blame] | 1192 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mull_ld1r) |
| 1193 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mull_ld1r) |
| 1194 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mull_ld1r) |
| 1195 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mull_ld1r) |
| 1196 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mull_ld1r) |
| 1197 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mull_ld1r) |
| 1198 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mull_ld1r) |
| 1199 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mull_ld1r) |
| 1200 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mlal_ld1r) |
| 1201 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mlal_ld1r) |
| 1202 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mlal_ld1r) |
| 1203 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mlal_ld1r) |
| 1204 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mlal_ld1r) |
| 1205 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mlal_ld1r) |
| 1206 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mlal_ld1r) |
| 1207 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mlal_ld1r) |
| 1208 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mull_ld2r) |
| 1209 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mull_ld2r) |
| 1210 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mull_ld2r) |
| 1211 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mull_ld2r) |
| 1212 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mull_ld2r) |
| 1213 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mull_ld2r) |
| 1214 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mull_ld2r) |
| 1215 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mull_ld2r) |
| 1216 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mlal_ld2r) |
| 1217 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mlal_ld2r) |
| 1218 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mlal_ld2r) |
| 1219 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mlal_ld2r) |
| 1220 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mlal_ld2r) |
| 1221 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mlal_ld2r) |
| 1222 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mlal_ld2r) |
| 1223 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mlal_ld2r) |
Frank Barchard | 42f5c50 | 2021-11-16 10:04:21 -0800 | [diff] [blame] | 1224 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mull_ld4r) |
| 1225 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mull_ld4r) |
| 1226 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mull_ld4r) |
| 1227 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mull_ld4r) |
| 1228 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mull_ld4r) |
| 1229 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mull_ld4r) |
| 1230 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mull_ld4r) |
| 1231 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mull_ld4r) |
| 1232 | BENCHMARK_GEMM(qs8_gemm_1x8c2__neon_mlal_ld4r) |
| 1233 | BENCHMARK_GEMM(qs8_gemm_2x8c2__neon_mlal_ld4r) |
| 1234 | BENCHMARK_GEMM(qs8_gemm_3x8c2__neon_mlal_ld4r) |
| 1235 | BENCHMARK_GEMM(qs8_gemm_4x8c2__neon_mlal_ld4r) |
| 1236 | BENCHMARK_GEMM(qs8_gemm_1x16c2__neon_mlal_ld4r) |
| 1237 | BENCHMARK_GEMM(qs8_gemm_2x16c2__neon_mlal_ld4r) |
| 1238 | BENCHMARK_GEMM(qs8_gemm_3x16c2__neon_mlal_ld4r) |
| 1239 | BENCHMARK_GEMM(qs8_gemm_4x16c2__neon_mlal_ld4r) |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1240 | BENCHMARK_GEMM(qs8_gemm_1x8c2s4__neon_mull) |
| 1241 | BENCHMARK_GEMM(qs8_gemm_2x8c2s4__neon_mull) |
| 1242 | BENCHMARK_GEMM(qs8_gemm_3x8c2s4__neon_mull) |
| 1243 | BENCHMARK_GEMM(qs8_gemm_4x8c2s4__neon_mull) |
| 1244 | BENCHMARK_GEMM(qs8_gemm_1x16c2s4__neon_mull) |
| 1245 | BENCHMARK_GEMM(qs8_gemm_2x16c2s4__neon_mull) |
| 1246 | BENCHMARK_GEMM(qs8_gemm_3x16c2s4__neon_mull) |
| 1247 | BENCHMARK_GEMM(qs8_gemm_4x16c2s4__neon_mull) |
| 1248 | BENCHMARK_GEMM(qs8_gemm_1x8c2s4__neon_mlal) |
| 1249 | BENCHMARK_GEMM(qs8_gemm_2x8c2s4__neon_mlal) |
| 1250 | BENCHMARK_GEMM(qs8_gemm_3x8c2s4__neon_mlal) |
| 1251 | BENCHMARK_GEMM(qs8_gemm_4x8c2s4__neon_mlal) |
| 1252 | BENCHMARK_GEMM(qs8_gemm_1x16c2s4__neon_mlal) |
| 1253 | BENCHMARK_GEMM(qs8_gemm_2x16c2s4__neon_mlal) |
| 1254 | BENCHMARK_GEMM(qs8_gemm_3x16c2s4__neon_mlal) |
| 1255 | BENCHMARK_GEMM(qs8_gemm_4x16c2s4__neon_mlal) |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 1256 | BENCHMARK_GEMM(qs8_gemm_1x8__neon_mlal_lane) |
| 1257 | BENCHMARK_GEMM(qs8_gemm_2x8__neon_mlal_lane) |
| 1258 | BENCHMARK_GEMM(qs8_gemm_3x8__neon_mlal_lane) |
| 1259 | BENCHMARK_GEMM(qs8_gemm_4x8__neon_mlal_lane) |
| 1260 | BENCHMARK_GEMM(qs8_gemm_6x8__neon_mlal_lane) |
| 1261 | BENCHMARK_GEMM(qs8_gemm_1x16__neon_mlal_lane) |
| 1262 | BENCHMARK_GEMM(qs8_gemm_2x16__neon_mlal_lane) |
| 1263 | BENCHMARK_GEMM(qs8_gemm_3x16__neon_mlal_lane) |
| 1264 | BENCHMARK_GEMM(qs8_gemm_4x16__neon_mlal_lane) |
| 1265 | BENCHMARK_GEMM(qs8_gemm_6x16__neon_mlal_lane) |
Frank Barchard | f82ea82 | 2021-12-01 15:43:37 -0800 | [diff] [blame] | 1266 | BENCHMARK_GEMM(qs8_gemm_1x8__neon_mlal_lane_prfm) |
| 1267 | BENCHMARK_GEMM(qs8_gemm_2x8__neon_mlal_lane_prfm) |
| 1268 | BENCHMARK_GEMM(qs8_gemm_3x8__neon_mlal_lane_prfm) |
| 1269 | BENCHMARK_GEMM(qs8_gemm_4x8__neon_mlal_lane_prfm) |
| 1270 | BENCHMARK_GEMM(qs8_gemm_6x8__neon_mlal_lane_prfm) |
| 1271 | BENCHMARK_GEMM(qs8_gemm_1x16__neon_mlal_lane_prfm) |
| 1272 | BENCHMARK_GEMM(qs8_gemm_2x16__neon_mlal_lane_prfm) |
| 1273 | BENCHMARK_GEMM(qs8_gemm_3x16__neon_mlal_lane_prfm) |
| 1274 | BENCHMARK_GEMM(qs8_gemm_4x16__neon_mlal_lane_prfm) |
| 1275 | BENCHMARK_GEMM(qs8_gemm_6x16__neon_mlal_lane_prfm) |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1276 | BENCHMARK_GEMM(qs8_gemm_1x8c8__neon_mull) |
| 1277 | BENCHMARK_GEMM(qs8_gemm_2x8c8__neon_mull) |
| 1278 | BENCHMARK_GEMM(qs8_gemm_3x8c8__neon_mull) |
| 1279 | BENCHMARK_GEMM(qs8_gemm_4x8c8__neon_mull) |
| 1280 | BENCHMARK_GEMM(qs8_gemm_1x16c8__neon_mull) |
| 1281 | BENCHMARK_GEMM(qs8_gemm_2x16c8__neon_mull) |
| 1282 | BENCHMARK_GEMM(qs8_gemm_3x16c8__neon_mull) |
| 1283 | BENCHMARK_GEMM(qs8_gemm_4x16c8__neon_mull) |
| 1284 | BENCHMARK_GEMM(qs8_gemm_1x8c8__neon_mlal) |
| 1285 | BENCHMARK_GEMM(qs8_gemm_2x8c8__neon_mlal) |
| 1286 | BENCHMARK_GEMM(qs8_gemm_3x8c8__neon_mlal) |
| 1287 | BENCHMARK_GEMM(qs8_gemm_4x8c8__neon_mlal) |
| 1288 | BENCHMARK_GEMM(qs8_gemm_1x16c8__neon_mlal) |
| 1289 | BENCHMARK_GEMM(qs8_gemm_2x16c8__neon_mlal) |
| 1290 | BENCHMARK_GEMM(qs8_gemm_3x16c8__neon_mlal) |
| 1291 | BENCHMARK_GEMM(qs8_gemm_4x16c8__neon_mlal) |
| 1292 | BENCHMARK_GEMM(qs8_gemm_1x8c16__neon_mlal) |
| 1293 | BENCHMARK_GEMM(qs8_gemm_2x8c16__neon_mlal) |
| 1294 | BENCHMARK_GEMM(qs8_gemm_3x8c16__neon_mlal) |
| 1295 | BENCHMARK_GEMM(qs8_gemm_4x8c16__neon_mlal) |
| 1296 | BENCHMARK_GEMM(qs8_gemm_1x16c16__neon_mlal) |
| 1297 | BENCHMARK_GEMM(qs8_gemm_2x16c16__neon_mlal) |
| 1298 | BENCHMARK_GEMM(qs8_gemm_3x16c16__neon_mlal) |
| 1299 | BENCHMARK_GEMM(qs8_gemm_4x16c16__neon_mlal) |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1300 | |
| 1301 | BENCHMARK_GEMM(qs8_gemm_1x8c4__neondot) |
| 1302 | BENCHMARK_GEMM(qs8_gemm_4x8c4__neondot) |
| 1303 | BENCHMARK_GEMM(qs8_gemm_6x8c4__neondot) |
| 1304 | BENCHMARK_GEMM(qs8_gemm_8x8c4__neondot) |
| 1305 | BENCHMARK_GEMM(qs8_gemm_1x16c4__neondot) |
| 1306 | BENCHMARK_GEMM(qs8_gemm_4x16c4__neondot) |
| 1307 | BENCHMARK_GEMM(qs8_gemm_6x16c4__neondot) |
| 1308 | BENCHMARK_GEMM(qs8_gemm_8x16c4__neondot) |
Marat Dukhan | ab67142 | 2020-08-05 16:39:04 -0700 | [diff] [blame] | 1309 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1310 | |
Marat Dukhan | 725f47e | 2021-05-22 10:06:19 -0700 | [diff] [blame] | 1311 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 1312 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1313 | static void qs8_gemm_2x16c8__avx512skx(benchmark::State& state, const char* net) { |
Marat Dukhan | 8102593 | 2021-05-26 09:01:05 -0700 | [diff] [blame] | 1314 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx, 2, 16, 8, 1, |
| 1315 | xnn_init_qs8_conv_minmax_fp32_avx512_params, benchmark::utils::CheckAVX512SKX); |
| 1316 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1317 | static void qs8_gemm_3x16c8__avx512skx(benchmark::State& state, const char* net) { |
Marat Dukhan | 8102593 | 2021-05-26 09:01:05 -0700 | [diff] [blame] | 1318 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx, 3, 16, 8, 1, |
| 1319 | xnn_init_qs8_conv_minmax_fp32_avx512_params, benchmark::utils::CheckAVX512SKX); |
| 1320 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1321 | static void qs8_gemm_4x16c8__avx512skx(benchmark::State& state, const char* net) { |
Marat Dukhan | 8102593 | 2021-05-26 09:01:05 -0700 | [diff] [blame] | 1322 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx, 4, 16, 8, 1, |
| 1323 | xnn_init_qs8_conv_minmax_fp32_avx512_params, benchmark::utils::CheckAVX512SKX); |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 1324 | } |
| 1325 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1326 | static void qs8_gemm_2x8c8__avx2(benchmark::State& state, const char* net) { |
Marat Dukhan | 8102593 | 2021-05-26 09:01:05 -0700 | [diff] [blame] | 1327 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__avx2, 2, 8, 8, 1, |
| 1328 | xnn_init_qs8_conv_minmax_fp32_avx2_params, benchmark::utils::CheckAVX2); |
| 1329 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1330 | static void qs8_gemm_3x8c8__avx2(benchmark::State& state, const char* net) { |
Marat Dukhan | 8102593 | 2021-05-26 09:01:05 -0700 | [diff] [blame] | 1331 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x8c8__avx2, 3, 8, 8, 1, |
| 1332 | xnn_init_qs8_conv_minmax_fp32_avx2_params, benchmark::utils::CheckAVX2); |
| 1333 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1334 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1335 | static void qs8_gemm_xw_2x8c8__avx2(benchmark::State& state, const char* net) { |
| 1336 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x8c8__avx2, 2, 8, 8, 1, |
| 1337 | xnn_init_qs8_conv_minmax_fp32_avx2_params, benchmark::utils::CheckAVX2, true); |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1338 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1339 | static void qs8_gemm_xw_3x8c8__avx2(benchmark::State& state, const char* net) { |
| 1340 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x8c8__avx2, 3, 8, 8, 1, |
| 1341 | xnn_init_qs8_conv_minmax_fp32_avx2_params, benchmark::utils::CheckAVX2, true); |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1342 | } |
Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 1343 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1344 | static void qs8_gemm_2x4c2__xop_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1345 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64, 2, 4, 2, 1, |
| 1346 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1347 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1348 | static void qs8_gemm_3x4c2__xop_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1349 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64, 3, 4, 2, 1, |
| 1350 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1351 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1352 | static void qs8_gemm_4x4c2__xop_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1353 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64, 4, 4, 2, 1, |
| 1354 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1355 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1356 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1357 | static void qs8_gemm_2x4c2__xop_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1358 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128, 2, 4, 2, 1, |
| 1359 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1360 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1361 | static void qs8_gemm_3x4c2__xop_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1362 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128, 3, 4, 2, 1, |
| 1363 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1364 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1365 | static void qs8_gemm_4x4c2__xop_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1366 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128, 4, 4, 2, 1, |
| 1367 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1368 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1369 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1370 | static void qs8_gemm_xw_2x4c2__xop(benchmark::State& state, const char* net) { |
| 1371 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__xop, 2, 4, 2, 1, |
| 1372 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1373 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1374 | static void qs8_gemm_xw_3x4c2__xop(benchmark::State& state, const char* net) { |
| 1375 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__xop, 3, 4, 2, 1, |
| 1376 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1377 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1378 | static void qs8_gemm_xw_4x4c2__xop(benchmark::State& state, const char* net) { |
| 1379 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__xop, 4, 4, 2, 1, |
| 1380 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1381 | } |
| 1382 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1383 | static void qs8_gemm_2x4c8__xop_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1384 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64, 2, 4, 8, 1, |
| 1385 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1386 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1387 | static void qs8_gemm_3x4c8__xop_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1388 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64, 3, 4, 8, 1, |
| 1389 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1390 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1391 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1392 | static void qs8_gemm_2x4c8__xop_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1393 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128, 2, 4, 8, 1, |
| 1394 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1395 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1396 | static void qs8_gemm_3x4c8__xop_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1397 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128, 3, 4, 8, 1, |
| 1398 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP); |
| 1399 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1400 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1401 | static void qs8_gemm_xw_2x4c8__xop(benchmark::State& state, const char* net) { |
| 1402 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__xop, 2, 4, 8, 1, |
| 1403 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1404 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1405 | static void qs8_gemm_xw_3x4c8__xop(benchmark::State& state, const char* net) { |
| 1406 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__xop, 3, 4, 8, 1, |
| 1407 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckXOP, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1408 | } |
| 1409 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1410 | static void qs8_gemm_2x4c2__avx_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1411 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64, 2, 4, 2, 1, |
| 1412 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1413 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1414 | static void qs8_gemm_3x4c2__avx_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1415 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64, 3, 4, 2, 1, |
| 1416 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1417 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1418 | static void qs8_gemm_4x4c2__avx_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1419 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64, 4, 4, 2, 1, |
| 1420 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1421 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1422 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1423 | static void qs8_gemm_2x4c2__avx_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1424 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128, 2, 4, 2, 1, |
| 1425 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1426 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1427 | static void qs8_gemm_3x4c2__avx_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1428 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128, 3, 4, 2, 1, |
| 1429 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1430 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1431 | static void qs8_gemm_4x4c2__avx_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1432 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128, 4, 4, 2, 1, |
| 1433 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1434 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1435 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1436 | static void qs8_gemm_xw_2x4c2__avx(benchmark::State& state, const char* net) { |
| 1437 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__avx, 2, 4, 2, 1, |
| 1438 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1439 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1440 | static void qs8_gemm_xw_3x4c2__avx(benchmark::State& state, const char* net) { |
| 1441 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__avx, 3, 4, 2, 1, |
| 1442 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1443 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1444 | static void qs8_gemm_xw_4x4c2__avx(benchmark::State& state, const char* net) { |
| 1445 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__avx, 4, 4, 2, 1, |
| 1446 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1447 | } |
| 1448 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1449 | static void qs8_gemm_2x4c8__avx_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1450 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64, 2, 4, 8, 1, |
| 1451 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1452 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1453 | static void qs8_gemm_3x4c8__avx_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1454 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64, 3, 4, 8, 1, |
| 1455 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1456 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1457 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1458 | static void qs8_gemm_2x4c8__avx_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1459 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128, 2, 4, 8, 1, |
| 1460 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1461 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1462 | static void qs8_gemm_3x4c8__avx_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1463 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128, 3, 4, 8, 1, |
| 1464 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX); |
| 1465 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1466 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1467 | static void qs8_gemm_xw_2x4c8__avx(benchmark::State& state, const char* net) { |
| 1468 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__avx, 2, 4, 8, 1, |
| 1469 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1470 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1471 | static void qs8_gemm_xw_3x4c8__avx(benchmark::State& state, const char* net) { |
| 1472 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__avx, 3, 4, 8, 1, |
| 1473 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckAVX, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1474 | } |
| 1475 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1476 | static void qs8_gemm_2x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1477 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64, 2, 4, 2, 1, |
| 1478 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1479 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1480 | static void qs8_gemm_3x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1481 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64, 3, 4, 2, 1, |
| 1482 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1483 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1484 | static void qs8_gemm_4x4c2__sse41_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1485 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64, 4, 4, 2, 1, |
| 1486 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1487 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1488 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1489 | static void qs8_gemm_2x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1490 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128, 2, 4, 2, 1, |
| 1491 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1492 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1493 | static void qs8_gemm_3x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1494 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128, 3, 4, 2, 1, |
| 1495 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1496 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1497 | static void qs8_gemm_4x4c2__sse41_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1498 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128, 4, 4, 2, 1, |
| 1499 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1500 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1501 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1502 | static void qs8_gemm_xw_2x4c2__sse41(benchmark::State& state, const char* net) { |
| 1503 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__sse41, 2, 4, 2, 1, |
| 1504 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1505 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1506 | static void qs8_gemm_xw_3x4c2__sse41(benchmark::State& state, const char* net) { |
| 1507 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse41, 3, 4, 2, 1, |
| 1508 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1509 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1510 | static void qs8_gemm_xw_4x4c2__sse41(benchmark::State& state, const char* net) { |
| 1511 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse41, 4, 4, 2, 1, |
| 1512 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1513 | } |
| 1514 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1515 | static void qs8_gemm_2x4c8__sse41_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1516 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64, 2, 4, 8, 1, |
| 1517 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1518 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1519 | static void qs8_gemm_3x4c8__sse41_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1520 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64, 3, 4, 8, 1, |
| 1521 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1522 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1523 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1524 | static void qs8_gemm_2x4c8__sse41_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1525 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128, 2, 4, 8, 1, |
| 1526 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1527 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1528 | static void qs8_gemm_3x4c8__sse41_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1529 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128, 3, 4, 8, 1, |
| 1530 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41); |
| 1531 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1532 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1533 | static void qs8_gemm_xw_2x4c8__sse41(benchmark::State& state, const char* net) { |
| 1534 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse41, 2, 4, 8, 1, |
| 1535 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1536 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1537 | static void qs8_gemm_xw_3x4c8__sse41(benchmark::State& state, const char* net) { |
| 1538 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse41, 3, 4, 8, 1, |
| 1539 | xnn_init_qs8_conv_minmax_fp32_sse4_params, benchmark::utils::CheckSSE41, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1540 | } |
| 1541 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1542 | static void qs8_gemm_2x4c8__ssse3_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1543 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld64, 2, 4, 8, 1, |
| 1544 | xnn_init_qs8_conv_minmax_fp32_sse2_params, benchmark::utils::CheckSSSE3); |
| 1545 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1546 | static void qs8_gemm_3x4c8__ssse3_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1547 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld64, 3, 4, 8, 1, |
| 1548 | xnn_init_qs8_conv_minmax_fp32_sse2_params, benchmark::utils::CheckSSSE3); |
| 1549 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1550 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1551 | static void qs8_gemm_2x4c8__ssse3_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1552 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld128, 2, 4, 8, 1, |
| 1553 | xnn_init_qs8_conv_minmax_fp32_sse2_params, benchmark::utils::CheckSSSE3); |
| 1554 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1555 | static void qs8_gemm_3x4c8__ssse3_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1556 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld128, 3, 4, 8, 1, |
| 1557 | xnn_init_qs8_conv_minmax_fp32_sse2_params, benchmark::utils::CheckSSSE3); |
| 1558 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1559 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1560 | static void qs8_gemm_xw_2x4c8__ssse3(benchmark::State& state, const char* net) { |
| 1561 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__ssse3, 2, 4, 8, 1, |
| 1562 | xnn_init_qs8_conv_minmax_fp32_sse2_params, benchmark::utils::CheckSSSE3, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1563 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1564 | static void qs8_gemm_xw_3x4c8__ssse3(benchmark::State& state, const char* net) { |
| 1565 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__ssse3, 3, 4, 8, 1, |
| 1566 | xnn_init_qs8_conv_minmax_fp32_sse2_params, benchmark::utils::CheckSSSE3, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1567 | } |
| 1568 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1569 | static void qs8_gemm_2x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1570 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64, 2, 4, 2, 1, |
| 1571 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1572 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1573 | static void qs8_gemm_3x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1574 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64, 3, 4, 2, 1, |
| 1575 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1576 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1577 | static void qs8_gemm_4x4c2__sse2_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1578 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64, 4, 4, 2, 1, |
| 1579 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1580 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1581 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1582 | static void qs8_gemm_2x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1583 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128, 2, 4, 2, 1, |
| 1584 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1585 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1586 | static void qs8_gemm_3x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1587 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128, 3, 4, 2, 1, |
| 1588 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1589 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1590 | static void qs8_gemm_4x4c2__sse2_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1591 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128, 4, 4, 2, 1, |
| 1592 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1593 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1594 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1595 | static void qs8_gemm_xw_2x4c2__sse2(benchmark::State& state, const char* net) { |
| 1596 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__sse2, 2, 4, 2, 1, |
| 1597 | xnn_init_qs8_conv_minmax_fp32_sse2_params, nullptr, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1598 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1599 | static void qs8_gemm_xw_3x4c2__sse2(benchmark::State& state, const char* net) { |
| 1600 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse2, 3, 4, 2, 1, |
| 1601 | xnn_init_qs8_conv_minmax_fp32_sse2_params, nullptr, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1602 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1603 | static void qs8_gemm_xw_4x4c2__sse2(benchmark::State& state, const char* net) { |
| 1604 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse2, 4, 4, 2, 1, |
| 1605 | xnn_init_qs8_conv_minmax_fp32_sse2_params, nullptr, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1606 | } |
| 1607 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1608 | static void qs8_gemm_2x4c8__sse2_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1609 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64, 2, 4, 8, 1, |
| 1610 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1611 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1612 | static void qs8_gemm_3x4c8__sse2_ld64(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1613 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64, 3, 4, 8, 1, |
| 1614 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1615 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1616 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1617 | static void qs8_gemm_2x4c8__sse2_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1618 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128, 2, 4, 8, 1, |
| 1619 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1620 | } |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1621 | static void qs8_gemm_3x4c8__sse2_ld128(benchmark::State& state, const char* net) { |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1622 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128, 3, 4, 8, 1, |
| 1623 | xnn_init_qs8_conv_minmax_fp32_sse2_params); |
| 1624 | } |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1625 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1626 | static void qs8_gemm_xw_2x4c8__sse2(benchmark::State& state, const char* net) { |
| 1627 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse2, 2, 4, 8, 1, |
| 1628 | xnn_init_qs8_conv_minmax_fp32_sse2_params, nullptr, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1629 | } |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1630 | static void qs8_gemm_xw_3x4c8__sse2(benchmark::State& state, const char* net) { |
| 1631 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse2, 3, 4, 8, 1, |
| 1632 | xnn_init_qs8_conv_minmax_fp32_sse2_params, nullptr, true); |
Marat Dukhan | 2e42787 | 2021-04-09 23:40:07 -0700 | [diff] [blame] | 1633 | } |
| 1634 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1635 | BENCHMARK_GEMM(qs8_gemm_2x16c8__avx512skx) |
| 1636 | BENCHMARK_GEMM(qs8_gemm_3x16c8__avx512skx) |
| 1637 | BENCHMARK_GEMM(qs8_gemm_4x16c8__avx512skx) |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 1638 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1639 | BENCHMARK_GEMM(qs8_gemm_2x8c8__avx2) |
| 1640 | BENCHMARK_GEMM(qs8_gemm_3x8c8__avx2) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1641 | BENCHMARK_GEMM(qs8_gemm_xw_2x8c8__avx2) |
| 1642 | BENCHMARK_GEMM(qs8_gemm_xw_3x8c8__avx2) |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1643 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1644 | BENCHMARK_GEMM(qs8_gemm_2x4c2__xop_ld64) |
| 1645 | BENCHMARK_GEMM(qs8_gemm_3x4c2__xop_ld64) |
| 1646 | BENCHMARK_GEMM(qs8_gemm_4x4c2__xop_ld64) |
| 1647 | BENCHMARK_GEMM(qs8_gemm_2x4c2__xop_ld128) |
| 1648 | BENCHMARK_GEMM(qs8_gemm_3x4c2__xop_ld128) |
| 1649 | BENCHMARK_GEMM(qs8_gemm_4x4c2__xop_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1650 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c2__xop) |
| 1651 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c2__xop) |
| 1652 | BENCHMARK_GEMM(qs8_gemm_xw_4x4c2__xop) |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1653 | BENCHMARK_GEMM(qs8_gemm_2x4c8__xop_ld64) |
| 1654 | BENCHMARK_GEMM(qs8_gemm_3x4c8__xop_ld64) |
| 1655 | BENCHMARK_GEMM(qs8_gemm_2x4c8__xop_ld128) |
| 1656 | BENCHMARK_GEMM(qs8_gemm_3x4c8__xop_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1657 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c8__xop) |
| 1658 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c8__xop) |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1659 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1660 | BENCHMARK_GEMM(qs8_gemm_2x4c2__avx_ld64) |
| 1661 | BENCHMARK_GEMM(qs8_gemm_3x4c2__avx_ld64) |
| 1662 | BENCHMARK_GEMM(qs8_gemm_4x4c2__avx_ld64) |
| 1663 | BENCHMARK_GEMM(qs8_gemm_2x4c2__avx_ld128) |
| 1664 | BENCHMARK_GEMM(qs8_gemm_3x4c2__avx_ld128) |
| 1665 | BENCHMARK_GEMM(qs8_gemm_4x4c2__avx_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1666 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c2__avx) |
| 1667 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c2__avx) |
| 1668 | BENCHMARK_GEMM(qs8_gemm_xw_4x4c2__avx) |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1669 | BENCHMARK_GEMM(qs8_gemm_2x4c8__avx_ld64) |
| 1670 | BENCHMARK_GEMM(qs8_gemm_3x4c8__avx_ld64) |
| 1671 | BENCHMARK_GEMM(qs8_gemm_2x4c8__avx_ld128) |
| 1672 | BENCHMARK_GEMM(qs8_gemm_3x4c8__avx_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1673 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c8__avx) |
| 1674 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c8__avx) |
Marat Dukhan | a3c1633 | 2021-04-02 15:03:27 -0700 | [diff] [blame] | 1675 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1676 | BENCHMARK_GEMM(qs8_gemm_2x4c2__sse41_ld64) |
| 1677 | BENCHMARK_GEMM(qs8_gemm_3x4c2__sse41_ld64) |
| 1678 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse41_ld64) |
| 1679 | BENCHMARK_GEMM(qs8_gemm_2x4c2__sse41_ld128) |
| 1680 | BENCHMARK_GEMM(qs8_gemm_3x4c2__sse41_ld128) |
| 1681 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse41_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1682 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c2__sse41) |
| 1683 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c2__sse41) |
| 1684 | BENCHMARK_GEMM(qs8_gemm_xw_4x4c2__sse41) |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1685 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse41_ld64) |
| 1686 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse41_ld64) |
| 1687 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse41_ld128) |
| 1688 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse41_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1689 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c8__sse41) |
| 1690 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c8__sse41) |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1691 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1692 | BENCHMARK_GEMM(qs8_gemm_2x4c8__ssse3_ld64) |
| 1693 | BENCHMARK_GEMM(qs8_gemm_3x4c8__ssse3_ld64) |
| 1694 | BENCHMARK_GEMM(qs8_gemm_2x4c8__ssse3_ld128) |
| 1695 | BENCHMARK_GEMM(qs8_gemm_3x4c8__ssse3_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1696 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c8__ssse3) |
| 1697 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c8__ssse3) |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1698 | |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1699 | BENCHMARK_GEMM(qs8_gemm_2x4c2__sse2_ld64) |
| 1700 | BENCHMARK_GEMM(qs8_gemm_3x4c2__sse2_ld64) |
| 1701 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse2_ld64) |
| 1702 | BENCHMARK_GEMM(qs8_gemm_2x4c2__sse2_ld128) |
| 1703 | BENCHMARK_GEMM(qs8_gemm_3x4c2__sse2_ld128) |
| 1704 | BENCHMARK_GEMM(qs8_gemm_4x4c2__sse2_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1705 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c2__sse2) |
| 1706 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c2__sse2) |
| 1707 | BENCHMARK_GEMM(qs8_gemm_xw_4x4c2__sse2) |
Marat Dukhan | 529d2c1 | 2021-08-06 15:37:03 -0700 | [diff] [blame] | 1708 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse2_ld64) |
| 1709 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse2_ld64) |
| 1710 | BENCHMARK_GEMM(qs8_gemm_2x4c8__sse2_ld128) |
| 1711 | BENCHMARK_GEMM(qs8_gemm_3x4c8__sse2_ld128) |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1712 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c8__sse2) |
| 1713 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c8__sse2) |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1714 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1715 | |
Marat Dukhan | 725f47e | 2021-05-22 10:06:19 -0700 | [diff] [blame] | 1716 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1717 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1718 | static void qs8_gemm_2x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1719 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64, 2, 4, 2, 1, |
| 1720 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1721 | } |
| 1722 | static void qs8_gemm_3x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1723 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64, 3, 4, 2, 1, |
| 1724 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1725 | } |
| 1726 | static void qs8_gemm_4x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1727 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64, 4, 4, 2, 1, |
| 1728 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1729 | } |
| 1730 | |
| 1731 | static void qs8_gemm_2x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1732 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128, 2, 4, 2, 1, |
| 1733 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1734 | } |
| 1735 | static void qs8_gemm_3x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1736 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128, 3, 4, 2, 1, |
| 1737 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1738 | } |
| 1739 | static void qs8_gemm_4x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1740 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128, 4, 4, 2, 1, |
| 1741 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1742 | } |
| 1743 | |
| 1744 | static void qs8_gemm_xw_2x4c2__wasmsimd_dot16x2(benchmark::State& state, const char* net) { |
| 1745 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2, 2, 4, 2, 1, |
| 1746 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
| 1747 | } |
| 1748 | static void qs8_gemm_xw_3x4c2__wasmsimd_dot16x2(benchmark::State& state, const char* net) { |
| 1749 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2, 3, 4, 2, 1, |
| 1750 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
| 1751 | } |
| 1752 | static void qs8_gemm_xw_4x4c2__wasmsimd_dot16x2(benchmark::State& state, const char* net) { |
| 1753 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2, 4, 4, 2, 1, |
| 1754 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
| 1755 | } |
| 1756 | |
Marat Dukhan | 348c377 | 2022-02-01 00:36:50 -0800 | [diff] [blame] | 1757 | static void qs8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1758 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64, 2, 4, 2, 4, |
| 1759 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1760 | } |
| 1761 | static void qs8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1762 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64, 3, 4, 2, 4, |
| 1763 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1764 | } |
| 1765 | static void qs8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1766 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64, 4, 4, 2, 4, |
| 1767 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1768 | } |
| 1769 | |
| 1770 | static void qs8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1771 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128, 2, 4, 2, 4, |
| 1772 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1773 | } |
| 1774 | static void qs8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1775 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128, 3, 4, 2, 4, |
| 1776 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1777 | } |
| 1778 | static void qs8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1779 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128, 4, 4, 2, 4, |
| 1780 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1781 | } |
| 1782 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1783 | static void qs8_gemm_2x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1784 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64, 2, 4, 8, 1, |
| 1785 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1786 | } |
| 1787 | static void qs8_gemm_3x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1788 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64, 3, 4, 8, 1, |
| 1789 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1790 | } |
| 1791 | static void qs8_gemm_4x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { |
| 1792 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64, 4, 4, 8, 1, |
| 1793 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1794 | } |
| 1795 | |
| 1796 | static void qs8_gemm_2x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1797 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128, 2, 4, 8, 1, |
| 1798 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1799 | } |
| 1800 | static void qs8_gemm_3x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1801 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128, 3, 4, 8, 1, |
| 1802 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1803 | } |
| 1804 | static void qs8_gemm_4x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { |
| 1805 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128, 4, 4, 8, 1, |
| 1806 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
| 1807 | } |
| 1808 | |
| 1809 | static void qs8_gemm_xw_2x4c8__wasmsimd_dot16x2(benchmark::State& state, const char* net) { |
| 1810 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2, 2, 4, 8, 1, |
| 1811 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
| 1812 | } |
| 1813 | static void qs8_gemm_xw_3x4c8__wasmsimd_dot16x2(benchmark::State& state, const char* net) { |
| 1814 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2, 3, 4, 8, 1, |
| 1815 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
| 1816 | } |
| 1817 | static void qs8_gemm_xw_4x4c8__wasmsimd_dot16x2(benchmark::State& state, const char* net) { |
| 1818 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2, 4, 4, 8, 1, |
| 1819 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
| 1820 | } |
| 1821 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1822 | static void qs8_gemm_2x4c8__wasmsimd_mul16_ld64(benchmark::State& state, const char* net) { |
| 1823 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld64, 2, 4, 8, 1, |
Marat Dukhan | 8674629 | 2021-08-06 17:27:18 -0700 | [diff] [blame] | 1824 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1825 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1826 | static void qs8_gemm_3x4c8__wasmsimd_mul16_ld64(benchmark::State& state, const char* net) { |
| 1827 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld64, 3, 4, 8, 1, |
Marat Dukhan | 8674629 | 2021-08-06 17:27:18 -0700 | [diff] [blame] | 1828 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1829 | } |
| 1830 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1831 | static void qs8_gemm_2x4c8__wasmsimd_mul16_ld128(benchmark::State& state, const char* net) { |
| 1832 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128, 2, 4, 8, 1, |
Marat Dukhan | 8674629 | 2021-08-06 17:27:18 -0700 | [diff] [blame] | 1833 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1834 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1835 | static void qs8_gemm_3x4c8__wasmsimd_mul16_ld128(benchmark::State& state, const char* net) { |
| 1836 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128, 3, 4, 8, 1, |
Marat Dukhan | 8674629 | 2021-08-06 17:27:18 -0700 | [diff] [blame] | 1837 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params); |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1838 | } |
| 1839 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1840 | static void qs8_gemm_xw_2x4c8__wasmsimd_mul16(benchmark::State& state, const char* net) { |
| 1841 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16, 2, 4, 8, 1, |
Marat Dukhan | 8674629 | 2021-08-06 17:27:18 -0700 | [diff] [blame] | 1842 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1843 | } |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1844 | static void qs8_gemm_xw_3x4c8__wasmsimd_mul16(benchmark::State& state, const char* net) { |
| 1845 | GEMMBenchmark(state, xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16, 3, 4, 8, 1, |
Marat Dukhan | 8674629 | 2021-08-06 17:27:18 -0700 | [diff] [blame] | 1846 | xnn_init_qs8_conv_minmax_fp32_wasmsimd_params, nullptr, true); |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1847 | } |
| 1848 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1849 | BENCHMARK_GEMM(qs8_gemm_2x4c2__wasmsimd_dot16x2_ld64) |
| 1850 | BENCHMARK_GEMM(qs8_gemm_3x4c2__wasmsimd_dot16x2_ld64) |
| 1851 | BENCHMARK_GEMM(qs8_gemm_4x4c2__wasmsimd_dot16x2_ld64) |
| 1852 | BENCHMARK_GEMM(qs8_gemm_2x4c2__wasmsimd_dot16x2_ld128) |
| 1853 | BENCHMARK_GEMM(qs8_gemm_3x4c2__wasmsimd_dot16x2_ld128) |
| 1854 | BENCHMARK_GEMM(qs8_gemm_4x4c2__wasmsimd_dot16x2_ld128) |
| 1855 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c2__wasmsimd_dot16x2) |
| 1856 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c2__wasmsimd_dot16x2) |
| 1857 | BENCHMARK_GEMM(qs8_gemm_xw_4x4c2__wasmsimd_dot16x2) |
| 1858 | |
Marat Dukhan | 348c377 | 2022-02-01 00:36:50 -0800 | [diff] [blame] | 1859 | BENCHMARK_GEMM(qs8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64) |
| 1860 | BENCHMARK_GEMM(qs8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64) |
| 1861 | BENCHMARK_GEMM(qs8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64) |
| 1862 | BENCHMARK_GEMM(qs8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128) |
| 1863 | BENCHMARK_GEMM(qs8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128) |
| 1864 | BENCHMARK_GEMM(qs8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128) |
| 1865 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1866 | BENCHMARK_GEMM(qs8_gemm_2x4c8__wasmsimd_dot16x2_ld64) |
| 1867 | BENCHMARK_GEMM(qs8_gemm_3x4c8__wasmsimd_dot16x2_ld64) |
| 1868 | BENCHMARK_GEMM(qs8_gemm_4x4c8__wasmsimd_dot16x2_ld64) |
| 1869 | BENCHMARK_GEMM(qs8_gemm_2x4c8__wasmsimd_dot16x2_ld128) |
| 1870 | BENCHMARK_GEMM(qs8_gemm_3x4c8__wasmsimd_dot16x2_ld128) |
| 1871 | BENCHMARK_GEMM(qs8_gemm_4x4c8__wasmsimd_dot16x2_ld128) |
| 1872 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c8__wasmsimd_dot16x2) |
| 1873 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c8__wasmsimd_dot16x2) |
| 1874 | BENCHMARK_GEMM(qs8_gemm_xw_4x4c8__wasmsimd_dot16x2) |
| 1875 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1876 | BENCHMARK_GEMM(qs8_gemm_2x4c8__wasmsimd_mul16_ld64) |
| 1877 | BENCHMARK_GEMM(qs8_gemm_3x4c8__wasmsimd_mul16_ld64) |
| 1878 | BENCHMARK_GEMM(qs8_gemm_2x4c8__wasmsimd_mul16_ld128) |
| 1879 | BENCHMARK_GEMM(qs8_gemm_3x4c8__wasmsimd_mul16_ld128) |
| 1880 | BENCHMARK_GEMM(qs8_gemm_xw_2x4c8__wasmsimd_mul16) |
| 1881 | BENCHMARK_GEMM(qs8_gemm_xw_3x4c8__wasmsimd_mul16) |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 1882 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 1883 | |
Marat Dukhan | 725f47e | 2021-05-22 10:06:19 -0700 | [diff] [blame] | 1884 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1885 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| 1886 | static void qs8_gemm_2x2__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1887 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x2__wasm_fmagic, 2, 2, 1, 1, |
| 1888 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
| 1889 | } |
| 1890 | static void qs8_gemm_3x2__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1891 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic, 3, 2, 1, 1, |
| 1892 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
| 1893 | } |
| 1894 | static void qs8_gemm_4x2__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1895 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic, 4, 2, 1, 1, |
| 1896 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
| 1897 | } |
| 1898 | static void qs8_gemm_2x4__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1899 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic, 2, 4, 1, 1, |
| 1900 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
| 1901 | } |
| 1902 | static void qs8_gemm_3x4__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1903 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4__wasm_fmagic, 3, 4, 1, 1, |
| 1904 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
| 1905 | } |
| 1906 | static void qs8_gemm_4x4__wasm_fmagic(benchmark::State& state, const char* net) { |
| 1907 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic, 4, 4, 1, 1, |
| 1908 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
| 1909 | } |
| 1910 | |
| 1911 | BENCHMARK_GEMM(qs8_gemm_2x2__wasm_fmagic) |
| 1912 | BENCHMARK_GEMM(qs8_gemm_3x2__wasm_fmagic) |
| 1913 | BENCHMARK_GEMM(qs8_gemm_4x2__wasm_fmagic) |
| 1914 | BENCHMARK_GEMM(qs8_gemm_2x4__wasm_fmagic) |
| 1915 | BENCHMARK_GEMM(qs8_gemm_3x4__wasm_fmagic) |
| 1916 | BENCHMARK_GEMM(qs8_gemm_4x4__wasm_fmagic) |
| 1917 | #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| 1918 | |
| 1919 | |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1920 | static void qs8_gemm_2x2__scalar_fmagic(benchmark::State& state, const char* net) { |
| 1921 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_fmagic, 2, 2, 1, 1, |
| 1922 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
Marat Dukhan | a1a4e78 | 2021-05-07 17:49:47 -0700 | [diff] [blame] | 1923 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1924 | static void qs8_gemm_3x2__scalar_fmagic(benchmark::State& state, const char* net) { |
| 1925 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic, 3, 2, 1, 1, |
| 1926 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
Marat Dukhan | a1a4e78 | 2021-05-07 17:49:47 -0700 | [diff] [blame] | 1927 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1928 | static void qs8_gemm_4x2__scalar_fmagic(benchmark::State& state, const char* net) { |
| 1929 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic, 4, 2, 1, 1, |
| 1930 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
Marat Dukhan | a1a4e78 | 2021-05-07 17:49:47 -0700 | [diff] [blame] | 1931 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1932 | static void qs8_gemm_2x4__scalar_fmagic(benchmark::State& state, const char* net) { |
| 1933 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic, 2, 4, 1, 1, |
| 1934 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
Marat Dukhan | a1a4e78 | 2021-05-07 17:49:47 -0700 | [diff] [blame] | 1935 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1936 | static void qs8_gemm_3x4__scalar_fmagic(benchmark::State& state, const char* net) { |
| 1937 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic, 3, 4, 1, 1, |
| 1938 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
Marat Dukhan | a1a4e78 | 2021-05-07 17:49:47 -0700 | [diff] [blame] | 1939 | } |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1940 | static void qs8_gemm_4x4__scalar_fmagic(benchmark::State& state, const char* net) { |
| 1941 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic, 4, 4, 1, 1, |
| 1942 | xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params); |
Marat Dukhan | a1a4e78 | 2021-05-07 17:49:47 -0700 | [diff] [blame] | 1943 | } |
| 1944 | |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 1945 | static void qs8_gemm_2x2__scalar_imagic(benchmark::State& state, const char* net) { |
| 1946 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, 2, 2, 1, 1, |
| 1947 | xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params); |
| 1948 | } |
| 1949 | static void qs8_gemm_3x2__scalar_imagic(benchmark::State& state, const char* net) { |
| 1950 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, 3, 2, 1, 1, |
| 1951 | xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params); |
| 1952 | } |
| 1953 | static void qs8_gemm_4x2__scalar_imagic(benchmark::State& state, const char* net) { |
| 1954 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, 4, 2, 1, 1, |
| 1955 | xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params); |
| 1956 | } |
| 1957 | static void qs8_gemm_2x4__scalar_imagic(benchmark::State& state, const char* net) { |
| 1958 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, 2, 4, 1, 1, |
| 1959 | xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params); |
| 1960 | } |
| 1961 | static void qs8_gemm_3x4__scalar_imagic(benchmark::State& state, const char* net) { |
| 1962 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, 3, 4, 1, 1, |
| 1963 | xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params); |
| 1964 | } |
| 1965 | static void qs8_gemm_4x4__scalar_imagic(benchmark::State& state, const char* net) { |
| 1966 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, 4, 4, 1, 1, |
| 1967 | xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params); |
| 1968 | } |
| 1969 | |
| 1970 | static void qs8_gemm_2x2__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1971 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, 2, 2, 1, 1, |
| 1972 | xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params); |
| 1973 | } |
| 1974 | static void qs8_gemm_3x2__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1975 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, 3, 2, 1, 1, |
| 1976 | xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params); |
| 1977 | } |
| 1978 | static void qs8_gemm_4x2__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1979 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, 4, 2, 1, 1, |
| 1980 | xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params); |
| 1981 | } |
| 1982 | static void qs8_gemm_2x4__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1983 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, 2, 4, 1, 1, |
| 1984 | xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params); |
| 1985 | } |
| 1986 | static void qs8_gemm_3x4__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1987 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, 3, 4, 1, 1, |
| 1988 | xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params); |
| 1989 | } |
| 1990 | static void qs8_gemm_4x4__scalar_lrintf(benchmark::State& state, const char* net) { |
| 1991 | GEMMBenchmark(state, xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, 4, 4, 1, 1, |
| 1992 | xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params); |
| 1993 | } |
| 1994 | |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1995 | BENCHMARK_GEMM(qs8_gemm_2x2__scalar_fmagic) |
| 1996 | BENCHMARK_GEMM(qs8_gemm_3x2__scalar_fmagic) |
| 1997 | BENCHMARK_GEMM(qs8_gemm_4x2__scalar_fmagic) |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1998 | BENCHMARK_GEMM(qs8_gemm_2x4__scalar_fmagic) |
| 1999 | BENCHMARK_GEMM(qs8_gemm_3x4__scalar_fmagic) |
| 2000 | BENCHMARK_GEMM(qs8_gemm_4x4__scalar_fmagic) |
Marat Dukhan | a1a4e78 | 2021-05-07 17:49:47 -0700 | [diff] [blame] | 2001 | |
Marat Dukhan | 440e8ed | 2022-01-04 15:30:57 -0800 | [diff] [blame] | 2002 | BENCHMARK_GEMM(qs8_gemm_2x2__scalar_imagic) |
| 2003 | BENCHMARK_GEMM(qs8_gemm_3x2__scalar_imagic) |
| 2004 | BENCHMARK_GEMM(qs8_gemm_4x2__scalar_imagic) |
| 2005 | BENCHMARK_GEMM(qs8_gemm_2x4__scalar_imagic) |
| 2006 | BENCHMARK_GEMM(qs8_gemm_3x4__scalar_imagic) |
| 2007 | BENCHMARK_GEMM(qs8_gemm_4x4__scalar_imagic) |
| 2008 | |
| 2009 | BENCHMARK_GEMM(qs8_gemm_2x2__scalar_lrintf) |
| 2010 | BENCHMARK_GEMM(qs8_gemm_3x2__scalar_lrintf) |
| 2011 | BENCHMARK_GEMM(qs8_gemm_4x2__scalar_lrintf) |
| 2012 | BENCHMARK_GEMM(qs8_gemm_2x4__scalar_lrintf) |
| 2013 | BENCHMARK_GEMM(qs8_gemm_3x4__scalar_lrintf) |
| 2014 | BENCHMARK_GEMM(qs8_gemm_4x4__scalar_lrintf) |
| 2015 | |
Marat Dukhan | 725f47e | 2021-05-22 10:06:19 -0700 | [diff] [blame] | 2016 | |
Frank Barchard | 31328cb | 2020-10-12 11:55:18 -0700 | [diff] [blame] | 2017 | #ifdef BENCHMARK_RUY |
| 2018 | BENCHMARK_GEMM(ruy_st) |
| 2019 | #endif // BENCHMARK_RUY |
| 2020 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 2021 | #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| 2022 | BENCHMARK_MAIN(); |
| 2023 | #endif |