File size: 21,620 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 |
// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <algorithm>
#include <cfloat>
#include <cmath>
#include <functional>
#include <random>
#include <vector>
#include <benchmark/benchmark.h>
#include <fp16/fp16.h>
#include "bench/gemm.h"
#include "bench/utils.h"
#include <xnnpack.h>
#include <xnnpack/aligned-allocator.h>
#include <xnnpack/common.h>
#include <xnnpack/gemm.h>
#include <xnnpack/math.h>
#include <xnnpack/pack.h>
#include <xnnpack/microfnptr.h>
#include <xnnpack/microparams-init.h>
static void f16_gemm(benchmark::State& state,
xnn_f16_gemm_minmax_ukernel_fn gemm,
xnn_init_f16_minmax_params_fn init_params,
size_t mr, size_t nr, size_t kr, size_t sr,
benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
if (isa_check != nullptr && !isa_check(state)) {
return;
}
const size_t mc = state.range(0);
const size_t nc = state.range(1);
const size_t kc = state.range(2);
const size_t nc_stride = benchmark::utils::RoundUp(nc, nr);
const size_t kc_stride = benchmark::utils::RoundUp(kc, kr * sr);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng));
auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
std::vector<uint16_t> a(mc * kc + XNN_EXTRA_BYTES / sizeof(uint16_t));
std::generate(a.begin(), a.end(), std::ref(f16rng));
std::vector<uint16_t> k(nc * kc);
std::generate(k.begin(), k.end(), std::ref(f16rng));
std::vector<uint16_t> b(nc);
std::generate(b.begin(), b.end(), std::ref(f16rng));
const size_t w_elements = nc_stride * kc_stride + nc_stride;
const size_t c_elements = mc * nc;
const size_t num_buffers = 1 +
benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
sizeof(uint16_t) * (w_elements + c_elements));
std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> w(w_elements * num_buffers);
std::fill(w.begin(), w.end(), 0);
xnn_pack_f16_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data(), 0, nullptr);
std::vector<uint16_t> c(c_elements * num_buffers);
std::fill(c.begin(), c.end(), UINT16_C(0x7E00) /* NaN */);
// Prepare minmax parameters.
xnn_f16_minmax_params params;
init_params(¶ms,
UINT16_C(0xFC00) /* -inf */, UINT16_C(0x7C00) /* inf */);
size_t buffer_index = 0;
for (auto _ : state) {
// Use circular buffers (exceeding cache size) and prefetch to control cache state:
// - A is always in L1 cache (if fits, otherwise L2, L3, etc)
// - W is not in cache (for any cache level)
// - C is not in cache (for any cache level)
state.PauseTiming();
benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint16_t));
buffer_index = (buffer_index + 1) % num_buffers;
state.ResumeTiming();
for (uint32_t m = 0; m < mc; m += mr) {
const uint32_t mb = min(mc - m, mr);
for (uint32_t n = 0; n < nc; n += nr) {
const uint32_t nb = min(nc - n, nr);
gemm(
mb, nb, kc * sizeof(uint16_t),
a.data() + m * kc, kc * sizeof(uint16_t),
w.data() + (nc_stride * buffer_index + n) * (kc_stride + 1),
c.data() + (mc * buffer_index + m) * nc + n, nc * sizeof(uint16_t), nr * sizeof(uint16_t),
¶ms);
}
}
}
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
state.counters["FLOPS"] = benchmark::Counter(
uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate);
}
#if XNN_PLATFORM_JIT
static void f16_gemm(benchmark::State& state,
xnn_jit_gemm_code_generator_fn generator,
xnn_init_f16_minmax_params_fn init_params,
size_t mr, size_t nr, size_t kr, size_t sr,
benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
if (isa_check != nullptr && !isa_check(state)) {
return;
}
const size_t mc = state.range(0);
const size_t nc = state.range(1);
const size_t kc = state.range(2);
const size_t nc_stride = benchmark::utils::RoundUp(nc, nr);
const size_t kc_stride = benchmark::utils::RoundUp(kc, kr * sr);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng));
auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
std::vector<uint16_t> a(mc * kc + XNN_EXTRA_BYTES / sizeof(uint16_t));
std::generate(a.begin(), a.end(), std::ref(f16rng));
std::vector<uint16_t> k(nc * kc);
std::generate(k.begin(), k.end(), std::ref(f16rng));
std::vector<uint16_t> b(nc);
std::generate(b.begin(), b.end(), std::ref(f16rng));
const size_t w_elements = nc_stride * kc_stride + nc_stride;
const size_t c_elements = mc * nc;
const size_t num_buffers = 1 +
benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
sizeof(uint16_t) * (w_elements + c_elements));
std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> w(w_elements * num_buffers);
std::fill(w.begin(), w.end(), 0);
xnn_pack_f16_gemm_goi_w(1 /* groups */, nc, kc, nr, kr, sr, k.data(), b.data(), w.data(), 0, nullptr);
std::vector<uint16_t> c(c_elements * num_buffers);
std::fill(c.begin(), c.end(), UINT16_C(0x7E00) /* NaN */);
// Prepare minmax parameters.
xnn_f16_minmax_params params;
init_params(¶ms,
UINT16_C(0xFC00) /* -inf */, UINT16_C(0x7C00) /* inf */);
jit_gemm_params jit_params = {};
jit_params.f16_minmax.min = UINT16_C(0xFC00); /* -inf */
jit_params.f16_minmax.max = UINT16_C(0x7C00); /* inf */
xnn_code_buffer code_buffer;
xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE);
generator(&code_buffer, mr, nc % nr, kc * sizeof(float), &jit_params);
xnn_finalize_code_memory(&code_buffer);
xnn_f16_gemm_minmax_ukernel_fn gemm = reinterpret_cast<xnn_f16_gemm_minmax_ukernel_fn>(code_buffer.start);
size_t buffer_index = 0;
for (auto _ : state) {
// Use circular buffers (exceeding cache size) and prefetch to control cache state:
// - A is always in L1 cache (if fits, otherwise L2, L3, etc)
// - W is not in cache (for any cache level)
// - C is not in cache (for any cache level)
state.PauseTiming();
benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint16_t));
buffer_index = (buffer_index + 1) % num_buffers;
state.ResumeTiming();
for (uint32_t m = 0; m < mc; m += mr) {
const uint32_t mb = min(mc - m, mr);
for (uint32_t n = 0; n < nc; n += nr) {
const uint32_t nb = min(nc - n, nr);
gemm(
mb, nb, kc * sizeof(uint16_t),
a.data() + m * kc, kc * sizeof(uint16_t),
w.data() + (nc_stride * buffer_index + n) * (kc_stride + 1),
c.data() + (mc * buffer_index + m) * nc + n, nc * sizeof(uint16_t), nr * sizeof(uint16_t),
¶ms);
}
}
}
xnn_release_code_memory(&code_buffer);
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
state.counters["FLOPS"] = benchmark::Counter(
uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate);
}
#endif // XNN_PLATFORM_JIT
#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
static void f16_gemm_1x16__asm_aarch64_neonfp16arith_ld32(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld32,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_1x16__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_4x16__asm_aarch64_neonfp16arith_ld32(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld32,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_4x16__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16__asm_aarch64_neonfp16arith_cortex_a55(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16__asm_aarch64_neonfp16arith_cortex_a55r0(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55r0,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16__asm_aarch64_neonfp16arith_cortex_a75(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a75,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16__asm_aarch64_neonfp16arith_ld32(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_ld32,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_1x8__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_1x8__asm_aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_4x8__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_4x8__asm_aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x8__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x8__asm_aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_8x8__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_8x8__asm_aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/8, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
BENCHMARK_GEMM(f16_gemm_1x16__asm_aarch64_neonfp16arith_ld32)
BENCHMARK_GEMM(f16_gemm_1x16__asm_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_4x16__asm_aarch64_neonfp16arith_ld32)
BENCHMARK_GEMM(f16_gemm_4x16__asm_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_6x16__asm_aarch64_neonfp16arith_cortex_a55)
BENCHMARK_GEMM(f16_gemm_6x16__asm_aarch64_neonfp16arith_cortex_a55r0)
BENCHMARK_GEMM(f16_gemm_6x16__asm_aarch64_neonfp16arith_cortex_a75)
BENCHMARK_GEMM(f16_gemm_6x16__asm_aarch64_neonfp16arith_ld32)
BENCHMARK_GEMM(f16_gemm_6x16__asm_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_1x8__asm_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_4x8__asm_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_6x8__asm_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_8x8__asm_aarch64_neonfp16arith_ld64)
#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
#if XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
static void f16_gemm_1x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_1x8__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_4x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_8x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/8, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_1x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_4x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_8x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/8, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
BENCHMARK_GEMM(f16_gemm_1x8__neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_4x8__neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_6x8__neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_8x8__neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_1x16__neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_4x16__neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_6x16__neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_8x16__neonfp16arith_ld64)
#endif // XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
static void f16_gemm_1x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_1x8__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_4x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_4x8__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_5x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_5x8__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/5, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_6x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_6x8__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_7x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_7x8__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/7, /*nr=*/8, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_1x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_1x16__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_3x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_3x16__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/3, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_4x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_4x16__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
static void f16_gemm_5x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_f16_gemm_minmax_ukernel_5x16__avx2_broadcast,
xnn_init_f16_minmax_avx_params,
/*mr=*/5, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckAVX2);
}
BENCHMARK_GEMM(f16_gemm_1x8__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_4x8__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_5x8__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_6x8__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_7x8__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_1x16__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_3x16__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_4x16__avx2_broadcast)
BENCHMARK_GEMM(f16_gemm_5x16__avx2_broadcast)
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
#if XNN_ARCH_ARM64 && XNN_PLATFORM_JIT
static void f16_gemm_1x16_1x16__jit_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_generate_f16_gemm_ukernel_1x16__aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_4x16_4x16__jit_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_generate_f16_gemm_ukernel_4x16__aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_generate_f16_gemm_ukernel_6x16__aarch64_neonfp16arith_ld64,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_cortex_a55(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_generate_f16_gemm_ukernel_6x16__aarch64_neonfp16arith_cortex_a55,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_cortex_a55r0(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_generate_f16_gemm_ukernel_6x16__aarch64_neonfp16arith_cortex_a55r0,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
static void f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_cortex_a75(benchmark::State& state, const char* net) {
f16_gemm(state,
xnn_generate_f16_gemm_ukernel_6x16__aarch64_neonfp16arith_cortex_a75,
xnn_init_f16_minmax_fp16arith_params,
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
benchmark::utils::CheckNEONFP16ARITH);
}
BENCHMARK_GEMM(f16_gemm_1x16_1x16__jit_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_4x16_4x16__jit_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_ld64)
BENCHMARK_GEMM(f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_cortex_a55)
BENCHMARK_GEMM(f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_cortex_a55r0)
BENCHMARK_GEMM(f16_gemm_6x16_6x16__jit_aarch64_neonfp16arith_cortex_a75)
#endif // XNN_ARCH_ARM && XNN_PLATFORM_JIT
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif
|