Lines Matching full:benchmark
13 #include <benchmark/benchmark.h>
26 static void IGEMMBenchmark(benchmark::State& state, in IGEMMBenchmark()
29 benchmark::utils::IsaCheckFunction isa_check = nullptr) in IGEMMBenchmark()
61 const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr); in IGEMMBenchmark()
62 const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr); in IGEMMBenchmark()
63 const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr); in IGEMMBenchmark()
78 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), in IGEMMBenchmark()
125 benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); in IGEMMBenchmark()
143 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); in IGEMMBenchmark()
148 state.counters["FLOPS"] = benchmark::Counter( in IGEMMBenchmark()
153 benchmark::Counter::kIsRate); in IGEMMBenchmark()
157 static void f32_igemm_1x8__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neon_lane_ld64()
158 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64, 1, 8, 1, 1, benchmark::uti… in f32_igemm_1x8__neon_lane_ld64()
160 static void f32_igemm_4x2__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x2__neon_lane_ld64()
161 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x2__neon_lane_ld64, 4, 2, 1, 1, benchmark::uti… in f32_igemm_4x2__neon_lane_ld64()
163 static void f32_igemm_4x4__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x4__neon_lane_ld64()
164 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64, 4, 4, 1, 1, benchmark::uti… in f32_igemm_4x4__neon_lane_ld64()
166 static void f32_igemm_4x8__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_lane_ld64()
167 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld64, 4, 8, 1, 1, benchmark::uti… in f32_igemm_4x8__neon_lane_ld64()
169 static void f32_igemm_4x8__neon_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_lane_ld128()
170 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld128, 4, 8, 1, 1, benchmark::ut… in f32_igemm_4x8__neon_lane_ld128()
172 static void f32_igemm_6x8__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_lane_ld64()
173 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64, 6, 8, 1, 1, benchmark::uti… in f32_igemm_6x8__neon_lane_ld64()
175 static void f32_igemm_6x8__neon_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_lane_ld128()
176 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld128, 6, 8, 1, 1, benchmark::ut… in f32_igemm_6x8__neon_lane_ld128()
178 static void f32_igemm_1x8__neon_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neon_dup_ld64()
179 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neon_dup_ld64, 1, 8, 1, 1, benchmark::util… in f32_igemm_1x8__neon_dup_ld64()
181 static void f32_igemm_4x8__neon_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_dup_ld128()
182 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld128, 4, 8, 1, 1, benchmark::uti… in f32_igemm_4x8__neon_dup_ld128()
184 static void f32_igemm_4x8__neon_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_dup_ld64()
185 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64, 4, 8, 1, 1, benchmark::util… in f32_igemm_4x8__neon_dup_ld64()
187 static void f32_igemm_6x8__neon_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_dup_ld64()
188 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64, 6, 8, 1, 1, benchmark::util… in f32_igemm_6x8__neon_dup_ld64()
190 static void f32_igemm_6x8__neon_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_dup_ld128()
191 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128, 6, 8, 1, 1, benchmark::uti… in f32_igemm_6x8__neon_dup_ld128()
193 static void f32_igemm_1x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neonfma_dup_ld64()
194 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neonfma_dup_ld64, 1, 8, 1, 1, benchmark::u… in f32_igemm_1x8__neonfma_dup_ld64()
196 static void f32_igemm_4x8__neonfma_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_dup_ld128()
197 …state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld128, 4, 8, 1, 1, benchmark::utils::CheckNEO… in f32_igemm_4x8__neonfma_dup_ld128()
199 static void f32_igemm_4x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_dup_ld64()
200 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64, 4, 8, 1, 1, benchmark::u… in f32_igemm_4x8__neonfma_dup_ld64()
202 static void f32_igemm_6x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_dup_ld64()
203 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64, 6, 8, 1, 1, benchmark::u… in f32_igemm_6x8__neonfma_dup_ld64()
205 static void f32_igemm_6x8__neonfma_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_dup_ld128()
206 …state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128, 6, 8, 1, 1, benchmark::utils::CheckNEO… in f32_igemm_6x8__neonfma_dup_ld128()
208 static void f32_igemm_1x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_1x8s4__neon()
209 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8s4__neon, 1, 8, 1, 4, benchmark::utils::Chec… in f32_igemm_1x8s4__neon()
211 static void f32_igemm_1x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_1x8s4__neonfma()
212 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8s4__neonfma, 1, 8, 1, 4, benchmark::utils::C… in f32_igemm_1x8s4__neonfma()
214 static void f32_igemm_4x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__neon()
215 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__neon, 4, 8, 1, 4, benchmark::utils::Chec… in f32_igemm_4x8s4__neon()
217 static void f32_igemm_4x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__neonfma()
218 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__neonfma, 4, 8, 1, 4, benchmark::utils::C… in f32_igemm_4x8s4__neonfma()
220 static void f32_igemm_6x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__neon()
221 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__neon, 6, 8, 1, 4, benchmark::utils::Chec… in f32_igemm_6x8s4__neon()
223 static void f32_igemm_6x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__neonfma()
224 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__neonfma, 6, 8, 1, 4, benchmark::utils::C… in f32_igemm_6x8s4__neonfma()
226 static void f32_igemm_8x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_8x8s4__neon()
227 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x8s4__neon, 8, 8, 1, 4, benchmark::utils::Chec… in f32_igemm_8x8s4__neon()
229 static void f32_igemm_8x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_8x8s4__neonfma()
230 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x8s4__neonfma, 8, 8, 1, 4, benchmark::utils::C… in f32_igemm_8x8s4__neonfma()
261 static void f32_igemm_4x8__aarch32_neon_ld64(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
264 static void f32_igemm_4x8__aarch32_neon_cortex_a7(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a7()
267 static void f32_igemm_4x8__aarch32_neon_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a53()
270 static void f32_igemm_4x8__aarch32_neon_cortex_a55(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a55()
273 static void f32_igemm_4x8__aarch32_neon_pld_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_pld_cortex_a75()
276 static void f32_igemm_4x8__aarch32_neon_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a75()
289 static void f32_igemm_1x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
293 static void f32_igemm_1x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_1x8__aarch64_neonfma_cortex_a53()
297 static void f32_igemm_1x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { in f32_igemm_1x8__aarch64_neonfma_cortex_a57()
301 static void f32_igemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_1x8__aarch64_neonfma_cortex_a75()
305 static void f32_igemm_4x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_cortex_a53()
309 static void f32_igemm_4x8__aarch64_neonfma_cortex_a55(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_cortex_a55()
313 static void f32_igemm_4x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_cortex_a57()
317 static void f32_igemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_cortex_a75()
321 static void f32_igemm_5x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { in f32_igemm_5x8__aarch64_neonfma_cortex_a57()
325 static void f32_igemm_5x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_5x8__aarch64_neonfma_cortex_a75()
329 static void f32_igemm_4x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_4x12__aarch64_neonfma_cortex_a53()
333 static void f32_igemm_6x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a53()
337 static void f32_igemm_6x8__aarch64_neonfma_cortex_a55(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a55()
341 static void f32_igemm_6x8__aarch64_neonfma_cortex_a73(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a73()
345 static void f32_igemm_6x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a57()
349 static void f32_igemm_6x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a75()
353 static void f32_igemm_1x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neonfma_lane_ld64()
357 static void f32_igemm_4x2__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x2__neonfma_lane_ld64()
361 static void f32_igemm_4x4__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x4__neonfma_lane_ld64()
365 static void f32_igemm_4x8__neonfma_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_lane_ld128()
369 static void f32_igemm_4x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_lane_ld64()
373 static void f32_igemm_6x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_lane_ld64()
377 static void f32_igemm_6x8__neonfma_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_lane_ld128()
407 static void f32_igemm_1x8__sse_load1(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
410 static void f32_igemm_3x8__sse_load1(benchmark::State& state, const char* net) { in f32_igemm_3x8__sse_load1()
413 static void f32_igemm_4x8__sse_load1(benchmark::State& state, const char* net) { in f32_igemm_4x8__sse_load1()
416 static void f32_igemm_5x8__sse_load1(benchmark::State& state, const char* net) { in f32_igemm_5x8__sse_load1()
420 static void f32_igemm_1x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_1x8__sse_dup()
423 static void f32_igemm_3x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_3x8__sse_dup()
426 static void f32_igemm_4x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_4x8__sse_dup()
429 static void f32_igemm_5x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_5x8__sse_dup()
433 static void f32_igemm_1x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_1x8s4__sse()
436 static void f32_igemm_3x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_3x8s4__sse()
439 static void f32_igemm_4x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__sse()
442 static void f32_igemm_5x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_5x8s4__sse()
446 static void f32_igemm_1x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_1x8__sse2_dup()
449 static void f32_igemm_3x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_3x8__sse2_dup()
452 static void f32_igemm_4x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_4x8__sse2_dup()
455 static void f32_igemm_5x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_5x8__sse2_dup()
459 static void f32_igemm_1x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_1x8__avx_broadcast()
460 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__avx_broadcast, 1, 8, 1, 1, benchmark::util… in f32_igemm_1x8__avx_broadcast()
462 static void f32_igemm_4x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_4x8__avx_broadcast()
463 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__avx_broadcast, 4, 8, 1, 1, benchmark::util… in f32_igemm_4x8__avx_broadcast()
465 static void f32_igemm_5x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_5x8__avx_broadcast()
466 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__avx_broadcast, 5, 8, 1, 1, benchmark::util… in f32_igemm_5x8__avx_broadcast()
468 static void f32_igemm_6x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_6x8__avx_broadcast()
469 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__avx_broadcast, 6, 8, 1, 1, benchmark::util… in f32_igemm_6x8__avx_broadcast()
471 static void f32_igemm_7x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_7x8__avx_broadcast()
472 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_7x8__avx_broadcast, 7, 8, 1, 1, benchmark::util… in f32_igemm_7x8__avx_broadcast()
475 static void f32_igemm_1x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_1x8__fma3_broadcast()
476 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__fma3_broadcast, 1, 8, 1, 1, benchmark::uti… in f32_igemm_1x8__fma3_broadcast()
478 static void f32_igemm_4x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_4x8__fma3_broadcast()
479 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__fma3_broadcast, 4, 8, 1, 1, benchmark::uti… in f32_igemm_4x8__fma3_broadcast()
481 static void f32_igemm_5x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_5x8__fma3_broadcast()
482 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__fma3_broadcast, 5, 8, 1, 1, benchmark::uti… in f32_igemm_5x8__fma3_broadcast()
484 static void f32_igemm_6x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_6x8__fma3_broadcast()
485 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__fma3_broadcast, 6, 8, 1, 1, benchmark::uti… in f32_igemm_6x8__fma3_broadcast()
487 static void f32_igemm_7x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_7x8__fma3_broadcast()
488 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_7x8__fma3_broadcast, 7, 8, 1, 1, benchmark::uti… in f32_igemm_7x8__fma3_broadcast()
490 static void f32_igemm_8x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_8x8__fma3_broadcast()
491 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x8__fma3_broadcast, 8, 8, 1, 1, benchmark::uti… in f32_igemm_8x8__fma3_broadcast()
494 static void f32_igemm_1x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_1x16__avx512f_broadcast()
495 …ate, xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast, 1, 16, 1, 1, benchmark::utils::CheckFMA… in f32_igemm_1x16__avx512f_broadcast()
497 static void f32_igemm_4x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_4x16__avx512f_broadcast()
498 …ate, xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast, 4, 16, 1, 1, benchmark::utils::CheckFMA… in f32_igemm_4x16__avx512f_broadcast()
500 static void f32_igemm_5x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_5x16__avx512f_broadcast()
501 …ate, xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast, 5, 16, 1, 1, benchmark::utils::CheckFMA… in f32_igemm_5x16__avx512f_broadcast()
503 static void f32_igemm_6x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_6x16__avx512f_broadcast()
504 …ate, xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast, 6, 16, 1, 1, benchmark::utils::CheckFMA… in f32_igemm_6x16__avx512f_broadcast()
506 static void f32_igemm_7x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_7x16__avx512f_broadcast()
507 …ate, xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast, 7, 16, 1, 1, benchmark::utils::CheckFMA… in f32_igemm_7x16__avx512f_broadcast()
509 static void f32_igemm_8x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_8x16__avx512f_broadcast()
510 …ate, xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast, 8, 16, 1, 1, benchmark::utils::CheckFMA… in f32_igemm_8x16__avx512f_broadcast()
555 static void f32_igemm_3x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
559 static void f32_igemm_4x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_arm_loadsplat()
563 static void f32_igemm_5x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_arm_loadsplat()
567 static void f32_igemm_6x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_arm_loadsplat()
571 static void f32_igemm_3x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_3x8__wasmsimd_x86_loadsplat()
575 static void f32_igemm_4x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_x86_loadsplat()
579 static void f32_igemm_5x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_x86_loadsplat()
583 static void f32_igemm_6x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_x86_loadsplat()
587 static void f32_igemm_3x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_3x8__wasmsimd_arm_splat()
591 static void f32_igemm_4x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_arm_splat()
595 static void f32_igemm_5x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_arm_splat()
599 static void f32_igemm_6x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_arm_splat()
603 static void f32_igemm_3x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_3x8__wasmsimd_x86_splat()
607 static void f32_igemm_4x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_x86_splat()
611 static void f32_igemm_5x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_x86_splat()
615 static void f32_igemm_6x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_x86_splat()
619 static void f32_igemm_3x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_3x8s4__wasmsimd_arm()
623 static void f32_igemm_4x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__wasmsimd_arm()
627 static void f32_igemm_5x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_5x8s4__wasmsimd_arm()
631 static void f32_igemm_6x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__wasmsimd_arm()
635 static void f32_igemm_3x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_3x8s4__wasmsimd_x86()
639 static void f32_igemm_4x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__wasmsimd_x86()
643 static void f32_igemm_5x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_5x8s4__wasmsimd_x86()
647 static void f32_igemm_6x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__wasmsimd_x86()
677 static void f32_igemm_1x4__scalar(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
681 static void f32_igemm_2x4__scalar(benchmark::State& state, const char* net) { in f32_igemm_2x4__scalar()
685 static void f32_igemm_4x4__scalar(benchmark::State& state, const char* net) { in f32_igemm_4x4__scalar()