Lines Matching full:state
26 static void IGEMMBenchmark(benchmark::State& state, in IGEMMBenchmark() argument
31 if (isa_check && !isa_check(state)) { in IGEMMBenchmark()
35 const size_t input_height = state.range(0); in IGEMMBenchmark()
36 const size_t input_width = state.range(1); in IGEMMBenchmark()
37 const size_t kernel_height = state.range(2); in IGEMMBenchmark()
38 const size_t kernel_width = state.range(3); in IGEMMBenchmark()
40 const size_t padding_height = state.range(4); in IGEMMBenchmark()
41 const size_t padding_width = state.range(5); in IGEMMBenchmark()
42 const size_t subsampling = state.range(6); in IGEMMBenchmark()
43 const size_t dilation = state.range(7); in IGEMMBenchmark()
44 const size_t group_input_channels = state.range(8); in IGEMMBenchmark()
45 const size_t group_output_channels = state.range(9); in IGEMMBenchmark()
123 for (auto _ : state) { in IGEMMBenchmark()
124 state.PauseTiming(); in IGEMMBenchmark()
127 state.ResumeTiming(); in IGEMMBenchmark()
145 state.counters["cpufreq"] = cpu_frequency; in IGEMMBenchmark()
148 state.counters["FLOPS"] = benchmark::Counter( in IGEMMBenchmark()
149 uint64_t(state.iterations()) * 2 * in IGEMMBenchmark()
157 static void f32_igemm_1x8__neon_lane_ld64(benchmark::State& state, const char* net) { in f32_igemm_1x8__neon_lane_ld64() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
197 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld128, 4, 8, 1, 1, benchmark::… 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() argument
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() argument
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() argument
206 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128, 6, 8, 1, 1, benchmark::… 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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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()
262 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_ld64, 4, 8, 1, 1); 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() argument
265 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7, 4, 8, 1, 1); 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() argument
268 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53, 4, 8, 1, 1); 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() argument
271 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55, 4, 8, 1, 1); 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() argument
274 … IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_pld_cortex_a75, 4, 8, 1, 1); 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() argument
277 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75, 4, 8, 1, 1); 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()
290 … IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53, 1, 12, 1, 1); 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() argument
294 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53, 1, 8, 1, 1); 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() argument
298 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57, 1, 8, 1, 1); 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() argument
302 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75, 1, 8, 1, 1); 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() argument
306 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53, 4, 8, 1, 1); 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() argument
310 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55, 4, 8, 1, 1); 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() argument
314 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a57, 4, 8, 1, 1); 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() argument
318 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75, 4, 8, 1, 1); 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() argument
322 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a57, 5, 8, 1, 1); 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() argument
326 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75, 5, 8, 1, 1); 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() argument
330 … IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53, 4, 12, 1, 1); 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() argument
334 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53, 6, 8, 1, 1); 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() argument
338 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55, 6, 8, 1, 1); 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() argument
342 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73, 6, 8, 1, 1); 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() argument
346 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a57, 6, 8, 1, 1); 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() argument
350 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75, 6, 8, 1, 1); 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() argument
354 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__neonfma_lane_ld64, 1, 8, 1, 1); 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() argument
358 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64, 4, 2, 1, 1); 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() argument
362 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64, 4, 4, 1, 1); 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() argument
366 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld128, 4, 8, 1, 1); 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() argument
370 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld64, 4, 8, 1, 1); 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() argument
374 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64, 6, 8, 1, 1); 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() argument
378 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld128, 6, 8, 1, 1); in f32_igemm_6x8__neonfma_lane_ld128()
407 static void f32_igemm_1x8__sse_load1(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
408 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__sse_load1, 1, 8, 1, 1); in BENCHMARK_CONV()
410 static void f32_igemm_3x8__sse_load1(benchmark::State& state, const char* net) { in f32_igemm_3x8__sse_load1() argument
411 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__sse_load1, 3, 8, 1, 1); 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() argument
414 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__sse_load1, 4, 8, 1, 1); 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() argument
417 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__sse_load1, 5, 8, 1, 1); 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() argument
421 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__sse_dup, 1, 8, 1, 1); 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() argument
424 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__sse_dup, 3, 8, 1, 1); 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() argument
427 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__sse_dup, 4, 8, 1, 1); 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() argument
430 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__sse_dup, 5, 8, 1, 1); in f32_igemm_5x8__sse_dup()
433 static void f32_igemm_1x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_1x8s4__sse() argument
434 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8s4__sse, 1, 8, 1, 4); in f32_igemm_1x8s4__sse()
436 static void f32_igemm_3x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_3x8s4__sse() argument
437 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8s4__sse, 3, 8, 1, 4); in f32_igemm_3x8s4__sse()
439 static void f32_igemm_4x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_4x8s4__sse() argument
440 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__sse, 4, 8, 1, 4); in f32_igemm_4x8s4__sse()
442 static void f32_igemm_5x8s4__sse(benchmark::State& state, const char* net) { in f32_igemm_5x8s4__sse() argument
443 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8s4__sse, 5, 8, 1, 4); 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() argument
447 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x8__sse2_dup, 1, 8, 1, 1); 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() argument
450 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup, 3, 8, 1, 1); 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() argument
453 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup, 4, 8, 1, 1); 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() argument
456 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup, 5, 8, 1, 1); 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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
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() argument
495 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast, 1, 16, 1, 1, benchmark… 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() argument
498 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast, 4, 16, 1, 1, benchmark… 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() argument
501 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast, 5, 16, 1, 1, benchmark… 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() argument
504 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast, 6, 16, 1, 1, benchmark… 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() argument
507 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast, 7, 16, 1, 1, benchmark… 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() argument
510 …IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast, 8, 16, 1, 1, benchmark… in f32_igemm_8x16__avx512f_broadcast()
555 static void f32_igemm_3x8__wasmsimd_arm_loadsplat(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
556 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_arm_loadsplat, 3, 8, 1, 1); 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() argument
560 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_arm_loadsplat, 4, 8, 1, 1); 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() argument
564 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_loadsplat, 5, 8, 1, 1); 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() argument
568 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_loadsplat, 6, 8, 1, 1); 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() argument
572 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_x86_loadsplat, 3, 8, 1, 1); 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() argument
576 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_x86_loadsplat, 4, 8, 1, 1); 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() argument
580 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat, 5, 8, 1, 1); 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() argument
584 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat, 6, 8, 1, 1); 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() argument
588 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_arm_splat, 3, 8, 1, 1); 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() argument
592 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_arm_splat, 4, 8, 1, 1); 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() argument
596 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_splat, 5, 8, 1, 1); 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() argument
600 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat, 6, 8, 1, 1); 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() argument
604 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8__wasmsimd_x86_splat, 3, 8, 1, 1); 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() argument
608 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_x86_splat, 4, 8, 1, 1); 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() argument
612 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_splat, 5, 8, 1, 1); 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() argument
616 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat, 6, 8, 1, 1); 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() argument
620 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm, 3, 8, 1, 4); 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() argument
624 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__wasmsimd_arm, 4, 8, 1, 4); 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() argument
628 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm, 5, 8, 1, 4); 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() argument
632 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__wasmsimd_arm, 6, 8, 1, 4); 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() argument
636 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_x86, 3, 8, 1, 4); 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() argument
640 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x8s4__wasmsimd_x86, 4, 8, 1, 4); 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() argument
644 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_x86, 5, 8, 1, 4); 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() argument
648 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_6x8s4__wasmsimd_x86, 6, 8, 1, 4); in f32_igemm_6x8s4__wasmsimd_x86()
677 static void f32_igemm_1x4__scalar(benchmark::State& state, const char* net) { in BENCHMARK_CONV()
678 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_1x4__scalar, 1, 4, 1, 1); in BENCHMARK_CONV()
681 static void f32_igemm_2x4__scalar(benchmark::State& state, const char* net) { in f32_igemm_2x4__scalar() argument
682 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_2x4__scalar, 2, 4, 1, 1); in f32_igemm_2x4__scalar()
685 static void f32_igemm_4x4__scalar(benchmark::State& state, const char* net) { in f32_igemm_4x4__scalar() argument
686 IGEMMBenchmark(state, xnn_f32_igemm_minmax_ukernel_4x4__scalar, 4, 4, 1, 1); in f32_igemm_4x4__scalar()