165       v128_t vacc0123p0 = wasm_v128_load(w);  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()  local
175       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
184       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
193       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
202       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
211       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
220       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
229       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
238       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
247       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
256       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
265       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
274       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
283       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
292       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
301       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
310       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
319       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
328       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
337       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
346       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
355       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
364       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
373       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
382       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
391       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
397       v128_t vacc0123 = wasm_f32x4_max(vacc0123p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
408       v128_t vacc0123p0 = wasm_v128_load(w);  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()  local
414       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
420       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
426       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
432       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
438       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
444       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
450       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
456       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
462       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
468       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
474       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
480       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
486       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
492       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
498       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
504       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
510       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
516       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
522       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
528       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
534       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
540       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
546       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
552       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
558       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
563       v128_t vacc0123 = wasm_f32x4_max(vacc0123p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
570       v128_t vacc0123p0 = wasm_v128_load(w);  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()  local
574       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
578       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
582       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
586       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
590       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
594       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
598       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
602       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
606       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
610       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
614       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
618       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
622       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
626       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
630       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
634       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
638       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
642       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
646       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
650       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
654       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
658       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
662       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
666       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
670       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
673       v128_t vacc0123 = wasm_f32x4_max(vacc0123p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()