Lines Matching full:state

41 static void GEMMBenchmark(benchmark::State& state,  in GEMMBenchmark()  argument
48 state.SkipWithError("cpuinfo initialization failed"); in GEMMBenchmark()
51 if (isa_check && !isa_check(state)) { in GEMMBenchmark()
55 const size_t mc = state.range(0); in GEMMBenchmark()
56 const size_t nc = state.range(1); in GEMMBenchmark()
57 const size_t kc = state.range(2); in GEMMBenchmark()
91 for (auto _ : state) { in GEMMBenchmark()
92 // Use circular buffers (exceeding cache size) and prefetch to control cache state: in GEMMBenchmark()
96 state.PauseTiming(); in GEMMBenchmark()
99 state.ResumeTiming(); in GEMMBenchmark()
117 state.counters["cpufreq"] = cpu_frequency; in GEMMBenchmark()
120 state.counters["OPS"] = benchmark::Counter( in GEMMBenchmark()
121 uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); in GEMMBenchmark()
158 static void GemmlowpBenchmark(benchmark::State& state, uint32_t threads) in GemmlowpBenchmark() argument
160 const size_t mc = state.range(0); in GemmlowpBenchmark()
161 const size_t nc = state.range(1); in GemmlowpBenchmark()
162 const size_t kc = state.range(2); in GemmlowpBenchmark()
190 for (auto _ : state) { in GemmlowpBenchmark()
191 state.PauseTiming(); in GemmlowpBenchmark()
194 state.ResumeTiming(); in GemmlowpBenchmark()
206 state.counters["cpufreq"] = cpu_frequency; in GemmlowpBenchmark()
209 state.counters["OPS"] = benchmark::Counter( in GemmlowpBenchmark()
210 uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); in GemmlowpBenchmark()
213 static void gemmlowp_st(benchmark::State& state, const char* net) in gemmlowp_st() argument
215 GemmlowpBenchmark(state, 1); in gemmlowp_st()
221 static void RuyBenchmark(benchmark::State& state, size_t threads) in RuyBenchmark() argument
223 const size_t mc = state.range(0); in RuyBenchmark()
224 const size_t nc = state.range(1); in RuyBenchmark()
225 const size_t kc = state.range(2); in RuyBenchmark()
280 for (auto _ : state) { in RuyBenchmark()
281 // Use circular buffers (exceeding cache size) and prefetch to control cache state: in RuyBenchmark()
286 state.PauseTiming(); in RuyBenchmark()
289 state.ResumeTiming(); in RuyBenchmark()
300 state.counters["cpufreq"] = cpu_frequency; in RuyBenchmark()
303 state.counters["OPS"] = benchmark::Counter( in RuyBenchmark()
304 uint64_t(state.iterations()) * 2 * mc * nc * kc, benchmark::Counter::kIsRate); in RuyBenchmark()
307 static void ruy_st(benchmark::State& state, const char* net) in ruy_st() argument
309 RuyBenchmark(state, 1); in ruy_st()
315 …static void qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53(benchmark::State& state, const char* n… in qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53() argument
316 GEMMBenchmark(state, in qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a53()
321 …static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, const ch… in qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53() argument
322 …GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, in qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53()
326 …static void qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7(benchmark::State& state, const char* ne… in qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7() argument
327 GEMMBenchmark(state, in qu8_gemm_4x8__aarch32_neon_mlal_lane_cortex_a7()
332 …static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7(benchmark::State& state, const cha… in qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7() argument
333 … GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, in qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7()
337 static void qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64() argument
338 GEMMBenchmark(state, in qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64()
343 …static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* ne… in qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64() argument
344 GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, in qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64()
348 …static void qu8_gemm_1x8__aarch32_neon_mlal_lane_cortex_a7(benchmark::State& state, const char* ne… in qu8_gemm_1x8__aarch32_neon_mlal_lane_cortex_a7() argument
349 GEMMBenchmark(state, in qu8_gemm_1x8__aarch32_neon_mlal_lane_cortex_a7()
354 …static void qu8_gemm_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7(benchmark::State& state, const cha… in qu8_gemm_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7() argument
355 … GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7, in qu8_gemm_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7()
371 …static void qu8_gemm_4x16c4__aarch64_neondot_cortex_a55(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
372 GEMMBenchmark(state, in BENCHMARK_GEMM()
377 static void qu8_gemm_4x16c4__aarch64_neondot_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x16c4__aarch64_neondot_ld128() argument
378 GEMMBenchmark(state, in qu8_gemm_4x16c4__aarch64_neondot_ld128()
383 static void qu8_gemm_4x8c4__aarch64_neondot_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x8c4__aarch64_neondot_ld128() argument
384 GEMMBenchmark(state, in qu8_gemm_4x8c4__aarch64_neondot_ld128()
390 static void qu8_gemm_4x8c4__aarch64_neondot_cortex_a55(benchmark::State& state, const char* net) { in qu8_gemm_4x8c4__aarch64_neondot_cortex_a55() argument
391 GEMMBenchmark(state, in qu8_gemm_4x8c4__aarch64_neondot_cortex_a55()
396 …static void qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53(benchmark::State& state, const char* … in qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53() argument
397 GEMMBenchmark(state, in qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a53()
403 …static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53(benchmark::State& state, const c… in qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53() argument
404 GEMMBenchmark(state, in qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53()
410 static void qu8_gemm_4x16__aarch64_neon_mlal_lane_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x16__aarch64_neon_mlal_lane_ld64() argument
411 GEMMBenchmark(state, in qu8_gemm_4x16__aarch64_neon_mlal_lane_ld64()
417 …static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* n… in qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64() argument
418 GEMMBenchmark(state, in qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_ld64()
424 …static void qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a75(benchmark::State& state, const char* … in qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a75() argument
425 GEMMBenchmark(state, in qu8_gemm_4x16__aarch64_neon_mlal_lane_cortex_a75()
431 …static void qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75(benchmark::State& state, const c… in qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75() argument
432 GEMMBenchmark(state, in qu8_gemm_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75()
452 static void qu8_gemm_1x8c4__neondot(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
453 GEMMBenchmark(state, in BENCHMARK_GEMM()
458 static void qu8_gemm_2x8c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_2x8c4__neondot() argument
459 GEMMBenchmark(state, in qu8_gemm_2x8c4__neondot()
464 static void qu8_gemm_3x8c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_3x8c4__neondot() argument
465 GEMMBenchmark(state, in qu8_gemm_3x8c4__neondot()
470 static void qu8_gemm_4x8c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_4x8c4__neondot() argument
471 GEMMBenchmark(state, in qu8_gemm_4x8c4__neondot()
476 static void qu8_gemm_5x8c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_5x8c4__neondot() argument
477 GEMMBenchmark(state, in qu8_gemm_5x8c4__neondot()
482 static void qu8_gemm_6x8c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_6x8c4__neondot() argument
483 GEMMBenchmark(state, in qu8_gemm_6x8c4__neondot()
488 static void qu8_gemm_8x8c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_8x8c4__neondot() argument
489 GEMMBenchmark(state, in qu8_gemm_8x8c4__neondot()
494 static void qu8_gemm_1x16c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_1x16c4__neondot() argument
495 GEMMBenchmark(state, in qu8_gemm_1x16c4__neondot()
500 static void qu8_gemm_2x16c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_2x16c4__neondot() argument
501 GEMMBenchmark(state, in qu8_gemm_2x16c4__neondot()
506 static void qu8_gemm_3x16c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_3x16c4__neondot() argument
507 GEMMBenchmark(state, in qu8_gemm_3x16c4__neondot()
512 static void qu8_gemm_4x16c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_4x16c4__neondot() argument
513 GEMMBenchmark(state, in qu8_gemm_4x16c4__neondot()
518 static void qu8_gemm_5x16c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_5x16c4__neondot() argument
519 GEMMBenchmark(state, in qu8_gemm_5x16c4__neondot()
524 static void qu8_gemm_6x16c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_6x16c4__neondot() argument
525 GEMMBenchmark(state, in qu8_gemm_6x16c4__neondot()
530 static void qu8_gemm_8x16c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_8x16c4__neondot() argument
531 GEMMBenchmark(state, in qu8_gemm_8x16c4__neondot()
536 static void qu8_gemm_1x32c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_1x32c4__neondot() argument
537 GEMMBenchmark(state, in qu8_gemm_1x32c4__neondot()
542 static void qu8_gemm_2x32c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_2x32c4__neondot() argument
543 GEMMBenchmark(state, in qu8_gemm_2x32c4__neondot()
548 static void qu8_gemm_3x32c4__neondot(benchmark::State& state, const char* net) { in qu8_gemm_3x32c4__neondot() argument
549 GEMMBenchmark(state, in qu8_gemm_3x32c4__neondot()
576 static void qu8_gemm_1x8__neon_mlal_lane(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
577 GEMMBenchmark(state, in BENCHMARK_GEMM()
582 static void qu8_gemm_2x8__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_2x8__neon_mlal_lane() argument
583 GEMMBenchmark(state, in qu8_gemm_2x8__neon_mlal_lane()
588 static void qu8_gemm_3x8__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_3x8__neon_mlal_lane() argument
589 GEMMBenchmark(state, in qu8_gemm_3x8__neon_mlal_lane()
594 static void qu8_gemm_4x8__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_4x8__neon_mlal_lane() argument
595 GEMMBenchmark(state, in qu8_gemm_4x8__neon_mlal_lane()
600 static void qu8_gemm_6x8__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_6x8__neon_mlal_lane() argument
601 GEMMBenchmark(state, in qu8_gemm_6x8__neon_mlal_lane()
606 static void qu8_gemm_1x16__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_1x16__neon_mlal_lane() argument
607 GEMMBenchmark(state, in qu8_gemm_1x16__neon_mlal_lane()
612 static void qu8_gemm_2x16__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_2x16__neon_mlal_lane() argument
613 GEMMBenchmark(state, in qu8_gemm_2x16__neon_mlal_lane()
618 static void qu8_gemm_3x16__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_3x16__neon_mlal_lane() argument
619 GEMMBenchmark(state, in qu8_gemm_3x16__neon_mlal_lane()
624 static void qu8_gemm_4x16__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_4x16__neon_mlal_lane() argument
625 GEMMBenchmark(state, in qu8_gemm_4x16__neon_mlal_lane()
630 static void qu8_gemm_6x16__neon_mlal_lane(benchmark::State& state, const char* net) { in qu8_gemm_6x16__neon_mlal_lane() argument
631 GEMMBenchmark(state, in qu8_gemm_6x16__neon_mlal_lane()
651 static void qu8_gemm_1x1c4__armsimd32(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
652 GEMMBenchmark(state, in BENCHMARK_GEMM()
657 static void qu8_gemm_2x1c4__armsimd32(benchmark::State& state, const char* net) { in qu8_gemm_2x1c4__armsimd32() argument
658 GEMMBenchmark(state, in qu8_gemm_2x1c4__armsimd32()
663 static void qu8_gemm_1x2c4__armsimd32(benchmark::State& state, const char* net) { in qu8_gemm_1x2c4__armsimd32() argument
664 GEMMBenchmark(state, in qu8_gemm_1x2c4__armsimd32()
669 static void qu8_gemm_2x2c4__armsimd32(benchmark::State& state, const char* net) { in qu8_gemm_2x2c4__armsimd32() argument
670 GEMMBenchmark(state, in qu8_gemm_2x2c4__armsimd32()
684 static void qu8_gemm_1x16c8__avx512skx(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
685 GEMMBenchmark(state, in BENCHMARK_GEMM()
691 static void qu8_gemm_2x16c8__avx512skx(benchmark::State& state, const char* net) { in qu8_gemm_2x16c8__avx512skx() argument
692 GEMMBenchmark(state, in qu8_gemm_2x16c8__avx512skx()
698 static void qu8_gemm_3x16c8__avx512skx(benchmark::State& state, const char* net) { in qu8_gemm_3x16c8__avx512skx() argument
699 GEMMBenchmark(state, in qu8_gemm_3x16c8__avx512skx()
705 static void qu8_gemm_4x16c8__avx512skx(benchmark::State& state, const char* net) { in qu8_gemm_4x16c8__avx512skx() argument
706 GEMMBenchmark(state, in qu8_gemm_4x16c8__avx512skx()
712 static void qu8_gemm_1x8c8__avx2(benchmark::State& state, const char* net) { in qu8_gemm_1x8c8__avx2() argument
713 GEMMBenchmark(state, in qu8_gemm_1x8c8__avx2()
719 static void qu8_gemm_2x8c8__avx2(benchmark::State& state, const char* net) { in qu8_gemm_2x8c8__avx2() argument
720 GEMMBenchmark(state, in qu8_gemm_2x8c8__avx2()
726 static void qu8_gemm_3x8c8__avx2(benchmark::State& state, const char* net) { in qu8_gemm_3x8c8__avx2() argument
727 GEMMBenchmark(state, in qu8_gemm_3x8c8__avx2()
733 static void qu8_gemm_1x4c2__xop_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__xop_ld64() argument
734 GEMMBenchmark(state, in qu8_gemm_1x4c2__xop_ld64()
740 static void qu8_gemm_2x4c2__xop_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__xop_ld64() argument
741 GEMMBenchmark(state, in qu8_gemm_2x4c2__xop_ld64()
747 static void qu8_gemm_3x4c2__xop_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__xop_ld64() argument
748 GEMMBenchmark(state, in qu8_gemm_3x4c2__xop_ld64()
754 static void qu8_gemm_4x4c2__xop_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__xop_ld64() argument
755 GEMMBenchmark(state, in qu8_gemm_4x4c2__xop_ld64()
761 static void qu8_gemm_1x4c2__xop_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__xop_ld128() argument
762 GEMMBenchmark(state, in qu8_gemm_1x4c2__xop_ld128()
768 static void qu8_gemm_2x4c2__xop_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__xop_ld128() argument
769 GEMMBenchmark(state, in qu8_gemm_2x4c2__xop_ld128()
775 static void qu8_gemm_3x4c2__xop_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__xop_ld128() argument
776 GEMMBenchmark(state, in qu8_gemm_3x4c2__xop_ld128()
782 static void qu8_gemm_4x4c2__xop_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__xop_ld128() argument
783 GEMMBenchmark(state, in qu8_gemm_4x4c2__xop_ld128()
789 static void qu8_gemm_1x4c8__xop_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__xop_ld64() argument
790 GEMMBenchmark(state, in qu8_gemm_1x4c8__xop_ld64()
796 static void qu8_gemm_2x4c8__xop_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__xop_ld64() argument
797 GEMMBenchmark(state, in qu8_gemm_2x4c8__xop_ld64()
803 static void qu8_gemm_3x4c8__xop_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__xop_ld64() argument
804 GEMMBenchmark(state, in qu8_gemm_3x4c8__xop_ld64()
810 static void qu8_gemm_1x4c8__xop_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__xop_ld128() argument
811 GEMMBenchmark(state, in qu8_gemm_1x4c8__xop_ld128()
817 static void qu8_gemm_2x4c8__xop_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__xop_ld128() argument
818 GEMMBenchmark(state, in qu8_gemm_2x4c8__xop_ld128()
824 static void qu8_gemm_3x4c8__xop_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__xop_ld128() argument
825 GEMMBenchmark(state, in qu8_gemm_3x4c8__xop_ld128()
831 static void qu8_gemm_1x4c2__avx_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__avx_ld64() argument
832 GEMMBenchmark(state, in qu8_gemm_1x4c2__avx_ld64()
838 static void qu8_gemm_2x4c2__avx_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__avx_ld64() argument
839 GEMMBenchmark(state, in qu8_gemm_2x4c2__avx_ld64()
845 static void qu8_gemm_3x4c2__avx_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__avx_ld64() argument
846 GEMMBenchmark(state, in qu8_gemm_3x4c2__avx_ld64()
852 static void qu8_gemm_4x4c2__avx_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__avx_ld64() argument
853 GEMMBenchmark(state, in qu8_gemm_4x4c2__avx_ld64()
859 static void qu8_gemm_1x4c2__avx_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__avx_ld128() argument
860 GEMMBenchmark(state, in qu8_gemm_1x4c2__avx_ld128()
866 static void qu8_gemm_2x4c2__avx_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__avx_ld128() argument
867 GEMMBenchmark(state, in qu8_gemm_2x4c2__avx_ld128()
873 static void qu8_gemm_3x4c2__avx_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__avx_ld128() argument
874 GEMMBenchmark(state, in qu8_gemm_3x4c2__avx_ld128()
880 static void qu8_gemm_4x4c2__avx_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__avx_ld128() argument
881 GEMMBenchmark(state, in qu8_gemm_4x4c2__avx_ld128()
887 static void qu8_gemm_1x4c8__avx_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__avx_ld64() argument
888 GEMMBenchmark(state, in qu8_gemm_1x4c8__avx_ld64()
894 static void qu8_gemm_2x4c8__avx_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__avx_ld64() argument
895 GEMMBenchmark(state, in qu8_gemm_2x4c8__avx_ld64()
901 static void qu8_gemm_3x4c8__avx_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__avx_ld64() argument
902 GEMMBenchmark(state, in qu8_gemm_3x4c8__avx_ld64()
908 static void qu8_gemm_1x4c8__avx_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__avx_ld128() argument
909 GEMMBenchmark(state, in qu8_gemm_1x4c8__avx_ld128()
915 static void qu8_gemm_2x4c8__avx_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__avx_ld128() argument
916 GEMMBenchmark(state, in qu8_gemm_2x4c8__avx_ld128()
922 static void qu8_gemm_3x4c8__avx_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__avx_ld128() argument
923 GEMMBenchmark(state, in qu8_gemm_3x4c8__avx_ld128()
929 static void qu8_gemm_1x4c2__sse41_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__sse41_ld64() argument
930 GEMMBenchmark(state, in qu8_gemm_1x4c2__sse41_ld64()
936 static void qu8_gemm_2x4c2__sse41_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__sse41_ld64() argument
937 GEMMBenchmark(state, in qu8_gemm_2x4c2__sse41_ld64()
943 static void qu8_gemm_3x4c2__sse41_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__sse41_ld64() argument
944 GEMMBenchmark(state, in qu8_gemm_3x4c2__sse41_ld64()
950 static void qu8_gemm_4x4c2__sse41_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__sse41_ld64() argument
951 GEMMBenchmark(state, in qu8_gemm_4x4c2__sse41_ld64()
957 static void qu8_gemm_1x4c2__sse41_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__sse41_ld128() argument
958 GEMMBenchmark(state, in qu8_gemm_1x4c2__sse41_ld128()
964 static void qu8_gemm_2x4c2__sse41_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__sse41_ld128() argument
965 GEMMBenchmark(state, in qu8_gemm_2x4c2__sse41_ld128()
971 static void qu8_gemm_3x4c2__sse41_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__sse41_ld128() argument
972 GEMMBenchmark(state, in qu8_gemm_3x4c2__sse41_ld128()
978 static void qu8_gemm_4x4c2__sse41_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__sse41_ld128() argument
979 GEMMBenchmark(state, in qu8_gemm_4x4c2__sse41_ld128()
985 static void qu8_gemm_1x4c8__sse41_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__sse41_ld64() argument
986 GEMMBenchmark(state, in qu8_gemm_1x4c8__sse41_ld64()
992 static void qu8_gemm_2x4c8__sse41_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__sse41_ld64() argument
993 GEMMBenchmark(state, in qu8_gemm_2x4c8__sse41_ld64()
999 static void qu8_gemm_3x4c8__sse41_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__sse41_ld64() argument
1000 GEMMBenchmark(state, in qu8_gemm_3x4c8__sse41_ld64()
1006 static void qu8_gemm_1x4c8__sse41_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__sse41_ld128() argument
1007 GEMMBenchmark(state, in qu8_gemm_1x4c8__sse41_ld128()
1013 static void qu8_gemm_2x4c8__sse41_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__sse41_ld128() argument
1014 GEMMBenchmark(state, in qu8_gemm_2x4c8__sse41_ld128()
1020 static void qu8_gemm_3x4c8__sse41_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__sse41_ld128() argument
1021 GEMMBenchmark(state, in qu8_gemm_3x4c8__sse41_ld128()
1027 static void qu8_gemm_1x4c2__sse2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__sse2_ld64() argument
1028 GEMMBenchmark(state, in qu8_gemm_1x4c2__sse2_ld64()
1033 static void qu8_gemm_2x4c2__sse2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__sse2_ld64() argument
1034 GEMMBenchmark(state, in qu8_gemm_2x4c2__sse2_ld64()
1039 static void qu8_gemm_3x4c2__sse2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__sse2_ld64() argument
1040 GEMMBenchmark(state, in qu8_gemm_3x4c2__sse2_ld64()
1045 static void qu8_gemm_4x4c2__sse2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__sse2_ld64() argument
1046 GEMMBenchmark(state, in qu8_gemm_4x4c2__sse2_ld64()
1051 static void qu8_gemm_1x4c2__sse2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__sse2_ld128() argument
1052 GEMMBenchmark(state, in qu8_gemm_1x4c2__sse2_ld128()
1057 static void qu8_gemm_2x4c2__sse2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__sse2_ld128() argument
1058 GEMMBenchmark(state, in qu8_gemm_2x4c2__sse2_ld128()
1063 static void qu8_gemm_3x4c2__sse2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__sse2_ld128() argument
1064 GEMMBenchmark(state, in qu8_gemm_3x4c2__sse2_ld128()
1069 static void qu8_gemm_4x4c2__sse2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__sse2_ld128() argument
1070 GEMMBenchmark(state, in qu8_gemm_4x4c2__sse2_ld128()
1075 static void qu8_gemm_1x4c8__sse2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__sse2_ld64() argument
1076 GEMMBenchmark(state, in qu8_gemm_1x4c8__sse2_ld64()
1081 static void qu8_gemm_2x4c8__sse2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__sse2_ld64() argument
1082 GEMMBenchmark(state, in qu8_gemm_2x4c8__sse2_ld64()
1087 static void qu8_gemm_3x4c8__sse2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__sse2_ld64() argument
1088 GEMMBenchmark(state, in qu8_gemm_3x4c8__sse2_ld64()
1093 static void qu8_gemm_1x4c8__sse2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__sse2_ld128() argument
1094 GEMMBenchmark(state, in qu8_gemm_1x4c8__sse2_ld128()
1099 static void qu8_gemm_2x4c8__sse2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__sse2_ld128() argument
1100 GEMMBenchmark(state, in qu8_gemm_2x4c8__sse2_ld128()
1105 static void qu8_gemm_3x4c8__sse2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__sse2_ld128() argument
1106 GEMMBenchmark(state, in qu8_gemm_3x4c8__sse2_ld128()
1184 static void qu8_gemm_1x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
1185 GEMMBenchmark(state, in BENCHMARK_GEMM()
1190 static void qu8_gemm_2x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__wasmsimd_dot16x2_ld64() argument
1191 GEMMBenchmark(state, in qu8_gemm_2x4c2__wasmsimd_dot16x2_ld64()
1196 static void qu8_gemm_3x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__wasmsimd_dot16x2_ld64() argument
1197 GEMMBenchmark(state, in qu8_gemm_3x4c2__wasmsimd_dot16x2_ld64()
1202 static void qu8_gemm_4x4c2__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__wasmsimd_dot16x2_ld64() argument
1203 GEMMBenchmark(state, in qu8_gemm_4x4c2__wasmsimd_dot16x2_ld64()
1209 static void qu8_gemm_1x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2__wasmsimd_dot16x2_ld128() argument
1210 GEMMBenchmark(state, in qu8_gemm_1x4c2__wasmsimd_dot16x2_ld128()
1215 static void qu8_gemm_2x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2__wasmsimd_dot16x2_ld128() argument
1216 GEMMBenchmark(state, in qu8_gemm_2x4c2__wasmsimd_dot16x2_ld128()
1221 static void qu8_gemm_3x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2__wasmsimd_dot16x2_ld128() argument
1222 GEMMBenchmark(state, in qu8_gemm_3x4c2__wasmsimd_dot16x2_ld128()
1227 static void qu8_gemm_4x4c2__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2__wasmsimd_dot16x2_ld128() argument
1228 GEMMBenchmark(state, in qu8_gemm_4x4c2__wasmsimd_dot16x2_ld128()
1234 static void qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld64() argument
1235 GEMMBenchmark(state, in qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld64()
1240 static void qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64() argument
1241 GEMMBenchmark(state, in qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld64()
1246 static void qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64() argument
1247 GEMMBenchmark(state, in qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld64()
1252 static void qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64() argument
1253 GEMMBenchmark(state, in qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld64()
1259 static void qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld128() argument
1260 GEMMBenchmark(state, in qu8_gemm_1x4c2s4__wasmsimd_dot16x2_ld128()
1265 static void qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128() argument
1266 GEMMBenchmark(state, in qu8_gemm_2x4c2s4__wasmsimd_dot16x2_ld128()
1271 static void qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128() argument
1272 GEMMBenchmark(state, in qu8_gemm_3x4c2s4__wasmsimd_dot16x2_ld128()
1277 static void qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128() argument
1278 GEMMBenchmark(state, in qu8_gemm_4x4c2s4__wasmsimd_dot16x2_ld128()
1284 static void qu8_gemm_1x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__wasmsimd_dot16x2_ld64() argument
1285 GEMMBenchmark(state, in qu8_gemm_1x4c8__wasmsimd_dot16x2_ld64()
1290 static void qu8_gemm_2x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__wasmsimd_dot16x2_ld64() argument
1291 GEMMBenchmark(state, in qu8_gemm_2x4c8__wasmsimd_dot16x2_ld64()
1296 static void qu8_gemm_3x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__wasmsimd_dot16x2_ld64() argument
1297 GEMMBenchmark(state, in qu8_gemm_3x4c8__wasmsimd_dot16x2_ld64()
1302 static void qu8_gemm_4x4c8__wasmsimd_dot16x2_ld64(benchmark::State& state, const char* net) { in qu8_gemm_4x4c8__wasmsimd_dot16x2_ld64() argument
1303 GEMMBenchmark(state, in qu8_gemm_4x4c8__wasmsimd_dot16x2_ld64()
1309 static void qu8_gemm_1x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_1x4c8__wasmsimd_dot16x2_ld128() argument
1310 GEMMBenchmark(state, in qu8_gemm_1x4c8__wasmsimd_dot16x2_ld128()
1315 static void qu8_gemm_2x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_2x4c8__wasmsimd_dot16x2_ld128() argument
1316 GEMMBenchmark(state, in qu8_gemm_2x4c8__wasmsimd_dot16x2_ld128()
1321 static void qu8_gemm_3x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_3x4c8__wasmsimd_dot16x2_ld128() argument
1322 GEMMBenchmark(state, in qu8_gemm_3x4c8__wasmsimd_dot16x2_ld128()
1327 static void qu8_gemm_4x4c8__wasmsimd_dot16x2_ld128(benchmark::State& state, const char* net) { in qu8_gemm_4x4c8__wasmsimd_dot16x2_ld128() argument
1328 GEMMBenchmark(state, in qu8_gemm_4x4c8__wasmsimd_dot16x2_ld128()
1364 static void qu8_gemm_1x2__wasm_fmagic(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
1365 GEMMBenchmark(state, in BENCHMARK_GEMM()
1370 static void qu8_gemm_2x2__wasm_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_2x2__wasm_fmagic() argument
1371 GEMMBenchmark(state, in qu8_gemm_2x2__wasm_fmagic()
1376 static void qu8_gemm_3x2__wasm_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_3x2__wasm_fmagic() argument
1377 GEMMBenchmark(state, in qu8_gemm_3x2__wasm_fmagic()
1382 static void qu8_gemm_4x2__wasm_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_4x2__wasm_fmagic() argument
1383 GEMMBenchmark(state, in qu8_gemm_4x2__wasm_fmagic()
1388 static void qu8_gemm_1x4__wasm_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_1x4__wasm_fmagic() argument
1389 GEMMBenchmark(state, in qu8_gemm_1x4__wasm_fmagic()
1394 static void qu8_gemm_2x4__wasm_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_2x4__wasm_fmagic() argument
1395 GEMMBenchmark(state, in qu8_gemm_2x4__wasm_fmagic()
1400 static void qu8_gemm_3x4__wasm_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_3x4__wasm_fmagic() argument
1401 GEMMBenchmark(state, in qu8_gemm_3x4__wasm_fmagic()
1406 static void qu8_gemm_4x4__wasm_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_4x4__wasm_fmagic() argument
1407 GEMMBenchmark(state, in qu8_gemm_4x4__wasm_fmagic()
1424 static void qu8_gemm_1x2__scalar_fmagic(benchmark::State& state, const char* net) { in BENCHMARK_GEMM()
1425 GEMMBenchmark(state, in BENCHMARK_GEMM()
1430 static void qu8_gemm_2x2__scalar_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_2x2__scalar_fmagic() argument
1431 GEMMBenchmark(state, in qu8_gemm_2x2__scalar_fmagic()
1436 static void qu8_gemm_3x2__scalar_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_3x2__scalar_fmagic() argument
1437 GEMMBenchmark(state, in qu8_gemm_3x2__scalar_fmagic()
1442 static void qu8_gemm_4x2__scalar_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_4x2__scalar_fmagic() argument
1443 GEMMBenchmark(state, in qu8_gemm_4x2__scalar_fmagic()
1448 static void qu8_gemm_1x4__scalar_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_1x4__scalar_fmagic() argument
1449 GEMMBenchmark(state, in qu8_gemm_1x4__scalar_fmagic()
1454 static void qu8_gemm_2x4__scalar_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_2x4__scalar_fmagic() argument
1455 GEMMBenchmark(state, in qu8_gemm_2x4__scalar_fmagic()
1460 static void qu8_gemm_3x4__scalar_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_3x4__scalar_fmagic() argument
1461 GEMMBenchmark(state, in qu8_gemm_3x4__scalar_fmagic()
1466 static void qu8_gemm_4x4__scalar_fmagic(benchmark::State& state, const char* net) { in qu8_gemm_4x4__scalar_fmagic() argument
1467 GEMMBenchmark(state, in qu8_gemm_4x4__scalar_fmagic()
1473 static void qu8_gemm_1x2__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_1x2__scalar_imagic() argument
1474 GEMMBenchmark(state, in qu8_gemm_1x2__scalar_imagic()
1479 static void qu8_gemm_2x2__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_2x2__scalar_imagic() argument
1480 GEMMBenchmark(state, in qu8_gemm_2x2__scalar_imagic()
1485 static void qu8_gemm_3x2__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_3x2__scalar_imagic() argument
1486 GEMMBenchmark(state, in qu8_gemm_3x2__scalar_imagic()
1491 static void qu8_gemm_4x2__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_4x2__scalar_imagic() argument
1492 GEMMBenchmark(state, in qu8_gemm_4x2__scalar_imagic()
1497 static void qu8_gemm_1x4__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_1x4__scalar_imagic() argument
1498 GEMMBenchmark(state, in qu8_gemm_1x4__scalar_imagic()
1503 static void qu8_gemm_2x4__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_2x4__scalar_imagic() argument
1504 GEMMBenchmark(state, in qu8_gemm_2x4__scalar_imagic()
1509 static void qu8_gemm_3x4__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_3x4__scalar_imagic() argument
1510 GEMMBenchmark(state, in qu8_gemm_3x4__scalar_imagic()
1515 static void qu8_gemm_4x4__scalar_imagic(benchmark::State& state, const char* net) { in qu8_gemm_4x4__scalar_imagic() argument
1516 GEMMBenchmark(state, in qu8_gemm_4x4__scalar_imagic()
1522 static void qu8_gemm_1x2__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_1x2__scalar_lrintf() argument
1523 GEMMBenchmark(state, in qu8_gemm_1x2__scalar_lrintf()
1528 static void qu8_gemm_2x2__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_2x2__scalar_lrintf() argument
1529 GEMMBenchmark(state, in qu8_gemm_2x2__scalar_lrintf()
1534 static void qu8_gemm_3x2__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_3x2__scalar_lrintf() argument
1535 GEMMBenchmark(state, in qu8_gemm_3x2__scalar_lrintf()
1540 static void qu8_gemm_4x2__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_4x2__scalar_lrintf() argument
1541 GEMMBenchmark(state, in qu8_gemm_4x2__scalar_lrintf()
1546 static void qu8_gemm_1x4__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_1x4__scalar_lrintf() argument
1547 GEMMBenchmark(state, in qu8_gemm_1x4__scalar_lrintf()
1552 static void qu8_gemm_2x4__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_2x4__scalar_lrintf() argument
1553 GEMMBenchmark(state, in qu8_gemm_2x4__scalar_lrintf()
1558 static void qu8_gemm_3x4__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_3x4__scalar_lrintf() argument
1559 GEMMBenchmark(state, in qu8_gemm_3x4__scalar_lrintf()
1564 static void qu8_gemm_4x4__scalar_lrintf(benchmark::State& state, const char* net) { in qu8_gemm_4x4__scalar_lrintf() argument
1565 GEMMBenchmark(state, in qu8_gemm_4x4__scalar_lrintf()