Lines Matching refs:vo0p0
115 v128_t vo0p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
125 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
128 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
131 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
134 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
137 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
147 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
153 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
156 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
159 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
175 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
178 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
181 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
184 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
187 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
197 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
200 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
203 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
206 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
209 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
225 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
228 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
231 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
234 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
237 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
241 v128_t vo0 = wasm_f32x4_max(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
251 v128_t vo0p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
268 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
271 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
274 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
277 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
280 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
290 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
293 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
296 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
299 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
302 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
318 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
321 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
324 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
327 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
330 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
340 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
343 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
346 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
349 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
352 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
368 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
371 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
374 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
377 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
380 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
384 v128_t vo0 = wasm_f32x4_max(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
397 v128_t vo0p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
407 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
410 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
413 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
416 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
419 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
429 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
432 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
435 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
438 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
441 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
451 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
454 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
457 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
460 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
463 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
473 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
476 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
479 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
482 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
485 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
495 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
498 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
501 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
504 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
507 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
511 v128_t vo0 = wasm_f32x4_max(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()