Lines Matching full:benchmark

14 #include <benchmark/benchmark.h>
29 static void f32_igemm(benchmark::State& state, in f32_igemm()
33 benchmark::utils::IsaCheckFunction isa_check = nullptr) in f32_igemm()
65 const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr); in f32_igemm()
66 const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr); in f32_igemm()
67 const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr * sr); in f32_igemm()
82 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), in f32_igemm()
130 benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); in f32_igemm()
145 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); in f32_igemm()
150 state.counters["FLOPS"] = benchmark::Counter( in f32_igemm()
155 benchmark::Counter::kIsRate); in f32_igemm()
159 static void f32_igemm(benchmark::State& state, in f32_igemm()
163 benchmark::utils::IsaCheckFunction isa_check = nullptr) in f32_igemm()
195 const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr); in f32_igemm()
196 const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr); in f32_igemm()
197 const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr * sr); in f32_igemm()
212 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(), in f32_igemm()
279 benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float)); in f32_igemm()
295 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); in f32_igemm()
300 state.counters["FLOPS"] = benchmark::Counter( in f32_igemm()
305 benchmark::Counter::kIsRate); in f32_igemm()
311 …static void jit_f32_igemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net… in jit_f32_igemm_1x8__aarch64_neonfma_cortex_a75()
315 …static void jit_f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char… in jit_f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75()
319 …static void jit_f32_igemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net… in jit_f32_igemm_4x8__aarch64_neonfma_cortex_a75()
323 …static void jit_f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char… in jit_f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a75()
327 static void jit_f32_igemm_6x8__aarch64_neonfma_ld128(benchmark::State& state, const char* net) { in jit_f32_igemm_6x8__aarch64_neonfma_ld128()
339 static void name(benchmark::State &state, const char *net) { \
357 static void jit_f32_igemm_4x8__aarch32_neon_ld64(benchmark::State& state, const char* net) { in jit_f32_igemm_4x8__aarch32_neon_ld64()
361 static void jit_f32_igemm_4x8__aarch32_neon_cortex_a7(benchmark::State& state, const char* net) { in jit_f32_igemm_4x8__aarch32_neon_cortex_a7()
365 static void jit_f32_igemm_4x8__aarch32_neon_cortex_a53(benchmark::State& state, const char* net) { in jit_f32_igemm_4x8__aarch32_neon_cortex_a53()
369 static void jit_f32_igemm_4x8__aarch32_neon_cortex_a55(benchmark::State& state, const char* net) { in jit_f32_igemm_4x8__aarch32_neon_cortex_a55()
373 …static void jit_f32_igemm_4x8__aarch32_neon_prfm_cortex_a75(benchmark::State& state, const char* n… in jit_f32_igemm_4x8__aarch32_neon_prfm_cortex_a75()
377 static void jit_f32_igemm_4x8__aarch32_neon_cortex_a75(benchmark::State& state, const char* net) { in jit_f32_igemm_4x8__aarch32_neon_cortex_a75()
391 static void f32_igemm_4x8__aarch32_neon_ld64(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
395 static void f32_igemm_4x8__aarch32_neon_cortex_a7(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a7()
399 static void f32_igemm_4x8__aarch32_neon_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a53()
403 …static void f32_igemm_4x8__aarch32_neon_prfm_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_prfm_cortex_a53()
407 static void f32_igemm_4x8__aarch32_neon_cortex_a55(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a55()
411 …static void f32_igemm_4x8__aarch32_neon_prfm_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_prfm_cortex_a75()
415 static void f32_igemm_4x8__aarch32_neon_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch32_neon_cortex_a75()
431 static void f32_igemm_1x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
435 static void f32_igemm_1x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_1x8__aarch64_neonfma_cortex_a53()
439 …static void f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a53(benchmark::State& state, const char* ne… in f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a53()
443 static void f32_igemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_1x8__aarch64_neonfma_cortex_a75()
447 …static void f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* ne… in f32_igemm_1x8__aarch64_neonfma_prfm_cortex_a75()
451 static void f32_igemm_4x2__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_4x2__aarch64_neonfma_cortex_a75()
455 …static void f32_igemm_4x2__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* ne… in f32_igemm_4x2__aarch64_neonfma_prfm_cortex_a75()
459 static void f32_igemm_4x2__aarch64_neonfma_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x2__aarch64_neonfma_ld64()
463 static void f32_igemm_4x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_cortex_a53()
467 …static void f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a53(benchmark::State& state, const char* ne… in f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a53()
471 static void f32_igemm_4x8__aarch64_neonfma_cortex_a55(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_cortex_a55()
475 static void f32_igemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_cortex_a75()
479 …static void f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* ne… in f32_igemm_4x8__aarch64_neonfma_prfm_cortex_a75()
483 static void f32_igemm_4x8__aarch64_neonfma_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_ld64()
487 static void f32_igemm_4x8__aarch64_neonfma_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__aarch64_neonfma_ld128()
491 static void f32_igemm_5x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_5x8__aarch64_neonfma_cortex_a75()
495 …static void f32_igemm_5x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* ne… in f32_igemm_5x8__aarch64_neonfma_prfm_cortex_a75()
499 static void f32_igemm_4x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_4x12__aarch64_neonfma_cortex_a53()
503 static void f32_igemm_6x8__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a53()
507 …static void f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a53(benchmark::State& state, const char* ne… in f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a53()
511 static void f32_igemm_6x8__aarch64_neonfma_cortex_a55(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a55()
515 static void f32_igemm_6x8__aarch64_neonfma_cortex_a73(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a73()
519 static void f32_igemm_6x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_cortex_a75()
523 …static void f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a75(benchmark::State& state, const char* ne… in f32_igemm_6x8__aarch64_neonfma_prfm_cortex_a75()
527 static void f32_igemm_6x8__aarch64_neonfma_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_ld64()
531 static void f32_igemm_6x8__aarch64_neonfma_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__aarch64_neonfma_ld128()
535 static void f32_igemm_1x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neonfma_lane_ld64()
539 static void f32_igemm_4x2__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x2__neonfma_lane_ld64()
543 static void f32_igemm_6x2__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x2__neonfma_lane_ld64()
547 static void f32_igemm_4x4__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x4__neonfma_lane_ld64()
551 static void f32_igemm_4x8__neonfma_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_lane_ld128()
555 static void f32_igemm_4x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_lane_ld64()
559 static void f32_igemm_6x8__neonfma_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_lane_ld64()
563 static void f32_igemm_6x8__neonfma_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_lane_ld128()
606 static void f32_igemm_1x8__neon_lane_ld64(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
608 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in BENCHMARK_CONV()
610 static void f32_igemm_4x2__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x2__neon_lane_ld64()
612 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_4x2__neon_lane_ld64()
614 static void f32_igemm_6x2__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x2__neon_lane_ld64()
616 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_6x2__neon_lane_ld64()
618 static void f32_igemm_4x4__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x4__neon_lane_ld64()
620 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_4x4__neon_lane_ld64()
622 static void f32_igemm_4x8__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_lane_ld64()
624 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_4x8__neon_lane_ld64()
626 static void f32_igemm_4x8__neon_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_lane_ld128()
628 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_4x8__neon_lane_ld128()
630 static void f32_igemm_6x8__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_lane_ld64()
632 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_6x8__neon_lane_ld64()
634 static void f32_igemm_6x8__neon_lane_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_lane_ld128()
636 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_6x8__neon_lane_ld128()
638 static void f32_igemm_1x8__neon_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neon_dup_ld64()
640 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_1x8__neon_dup_ld64()
642 static void f32_igemm_4x8__neon_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_dup_ld128()
644 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_4x8__neon_dup_ld128()
646 static void f32_igemm_4x8__neon_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neon_dup_ld64()
648 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_4x8__neon_dup_ld64()
650 static void f32_igemm_6x8__neon_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_dup_ld64()
652 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_6x8__neon_dup_ld64()
654 static void f32_igemm_6x8__neon_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neon_dup_ld128()
656 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_6x8__neon_dup_ld128()
658 static void f32_igemm_1x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neonfma_dup_ld64()
660 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_1x8__neonfma_dup_ld64()
662 static void f32_igemm_4x8__neonfma_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_dup_ld128()
664 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_4x8__neonfma_dup_ld128()
666 static void f32_igemm_4x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_4x8__neonfma_dup_ld64()
668 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_4x8__neonfma_dup_ld64()
670 static void f32_igemm_6x8__neonfma_dup_ld64(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_dup_ld64()
672 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_6x8__neonfma_dup_ld64()
674 static void f32_igemm_6x8__neonfma_dup_ld128(benchmark::State& state, const char* net) { in f32_igemm_6x8__neonfma_dup_ld128()
676 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_6x8__neonfma_dup_ld128()
678 static void f32_igemm_1x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_1x8s4__neon()
680 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_1x8s4__neon()
682 static void f32_igemm_4x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__neon()
684 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_4x8s4__neon()
686 static void f32_igemm_6x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__neon()
688 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_6x8s4__neon()
690 static void f32_igemm_8x8s4__neon(benchmark::State& state, const char* net) { in f32_igemm_8x8s4__neon()
692 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEON); in f32_igemm_8x8s4__neon()
694 static void f32_igemm_1x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_1x8s4__neonfma()
696 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_1x8s4__neonfma()
698 static void f32_igemm_4x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__neonfma()
700 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_4x8s4__neonfma()
702 static void f32_igemm_6x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__neonfma()
704 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_6x8s4__neonfma()
706 static void f32_igemm_8x8s4__neonfma(benchmark::State& state, const char* net) { in f32_igemm_8x8s4__neonfma()
708 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckNEONFMA); in f32_igemm_8x8s4__neonfma()
742 static void f32_igemm_1x8__sse_load1(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
746 static void f32_igemm_3x8__sse_load1(benchmark::State& state, const char* net) { in f32_igemm_3x8__sse_load1()
750 static void f32_igemm_4x8__sse_load1(benchmark::State& state, const char* net) { in f32_igemm_4x8__sse_load1()
754 static void f32_igemm_5x8__sse_load1(benchmark::State& state, const char* net) { in f32_igemm_5x8__sse_load1()
759 static void f32_igemm_1x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_1x8__sse_dup()
763 static void f32_igemm_3x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_3x8__sse_dup()
767 static void f32_igemm_4x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_4x8__sse_dup()
771 static void f32_igemm_5x8__sse_dup(benchmark::State& state, const char* net) { in f32_igemm_5x8__sse_dup()
776 static void f32_igemm_1x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_1x8s4__sse()
780 static void f32_igemm_3x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_3x8s4__sse()
784 static void f32_igemm_4x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__sse()
788 static void f32_igemm_5x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_5x8s4__sse()
793 static void f32_igemm_1x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_1x8__sse2_dup()
797 static void f32_igemm_3x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_3x8__sse2_dup()
801 static void f32_igemm_4x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_4x8__sse2_dup()
805 static void f32_igemm_5x8__sse2_dup(benchmark::State& state, const char* net) { in f32_igemm_5x8__sse2_dup()
810 static void f32_igemm_1x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_1x8__avx_broadcast()
812 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); in f32_igemm_1x8__avx_broadcast()
814 static void f32_igemm_4x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_4x8__avx_broadcast()
816 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); in f32_igemm_4x8__avx_broadcast()
818 static void f32_igemm_5x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_5x8__avx_broadcast()
820 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); in f32_igemm_5x8__avx_broadcast()
822 static void f32_igemm_6x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_6x8__avx_broadcast()
824 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); in f32_igemm_6x8__avx_broadcast()
826 static void f32_igemm_7x8__avx_broadcast(benchmark::State& state, const char* net) { in f32_igemm_7x8__avx_broadcast()
828 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckAVX); in f32_igemm_7x8__avx_broadcast()
831 static void f32_igemm_1x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_1x8__fma3_broadcast()
833 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); in f32_igemm_1x8__fma3_broadcast()
835 static void f32_igemm_4x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_4x8__fma3_broadcast()
837 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); in f32_igemm_4x8__fma3_broadcast()
839 static void f32_igemm_5x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_5x8__fma3_broadcast()
841 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); in f32_igemm_5x8__fma3_broadcast()
843 static void f32_igemm_6x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_6x8__fma3_broadcast()
845 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); in f32_igemm_6x8__fma3_broadcast()
847 static void f32_igemm_7x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_7x8__fma3_broadcast()
849 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); in f32_igemm_7x8__fma3_broadcast()
851 static void f32_igemm_8x8__fma3_broadcast(benchmark::State& state, const char* net) { in f32_igemm_8x8__fma3_broadcast()
853 xnn_init_f32_minmax_avx_params, benchmark::utils::CheckFMA3); in f32_igemm_8x8__fma3_broadcast()
856 static void f32_igemm_1x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_1x16__avx512f_broadcast()
858 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); in f32_igemm_1x16__avx512f_broadcast()
860 static void f32_igemm_4x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_4x16__avx512f_broadcast()
862 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); in f32_igemm_4x16__avx512f_broadcast()
864 static void f32_igemm_5x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_5x16__avx512f_broadcast()
866 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); in f32_igemm_5x16__avx512f_broadcast()
868 static void f32_igemm_6x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_6x16__avx512f_broadcast()
870 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); in f32_igemm_6x16__avx512f_broadcast()
872 static void f32_igemm_7x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_7x16__avx512f_broadcast()
874 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); in f32_igemm_7x16__avx512f_broadcast()
876 static void f32_igemm_8x16__avx512f_broadcast(benchmark::State& state, const char* net) { in f32_igemm_8x16__avx512f_broadcast()
878 xnn_init_f32_minmax_scalar_params, benchmark::utils::CheckFMA3); in f32_igemm_8x16__avx512f_broadcast()
924 static void f32_igemm_3x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
929 static void f32_igemm_4x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_arm_loadsplat()
934 static void f32_igemm_5x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_arm_loadsplat()
939 static void f32_igemm_6x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_arm_loadsplat()
944 static void f32_igemm_3x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_3x8__wasmsimd_x86_loadsplat()
949 static void f32_igemm_4x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_x86_loadsplat()
954 static void f32_igemm_5x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_x86_loadsplat()
959 static void f32_igemm_6x8__wasmsimd_x86_loadsplat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_x86_loadsplat()
964 static void f32_igemm_3x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_3x8__wasmsimd_arm_splat()
969 static void f32_igemm_4x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_arm_splat()
974 static void f32_igemm_5x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_arm_splat()
979 static void f32_igemm_6x8__wasmsimd_arm_splat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_arm_splat()
984 static void f32_igemm_3x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_3x8__wasmsimd_x86_splat()
989 static void f32_igemm_4x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_4x8__wasmsimd_x86_splat()
994 static void f32_igemm_5x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_5x8__wasmsimd_x86_splat()
999 static void f32_igemm_6x8__wasmsimd_x86_splat(benchmark::State& state, const char* net) { in f32_igemm_6x8__wasmsimd_x86_splat()
1004 static void f32_igemm_3x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_3x8s4__wasmsimd_arm()
1009 static void f32_igemm_4x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__wasmsimd_arm()
1014 static void f32_igemm_5x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_5x8s4__wasmsimd_arm()
1019 static void f32_igemm_6x8s4__wasmsimd_arm(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__wasmsimd_arm()
1024 static void f32_igemm_3x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_3x8s4__wasmsimd_x86()
1029 static void f32_igemm_4x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__wasmsimd_x86()
1034 static void f32_igemm_5x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_5x8s4__wasmsimd_x86()
1039 static void f32_igemm_6x8s4__wasmsimd_x86(benchmark::State& state, const char* net) { in f32_igemm_6x8s4__wasmsimd_x86()
1071 static void f32_igemm_1x4__scalar(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
1076 static void f32_igemm_2x4__scalar(benchmark::State& state, const char* net) { in f32_igemm_2x4__scalar()
1081 static void f32_igemm_4x4__scalar(benchmark::State& state, const char* net) { in f32_igemm_4x4__scalar()