125       v128_t vo2p0 = vbias;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()  local
137       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, vk02));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
141       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
145       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
149       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, vk32));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
153       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
165       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk01));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
169       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
173       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk21));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
177       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
181       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, vk41));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
200       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
204       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, vk10));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
208       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
212       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
216       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
228       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
232       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
236       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
240       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
244       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
263       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
267       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
271       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
275       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
279       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
284       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
297       v128_t vo2p0 = vbias;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()  local
317       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, vk02));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
321       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
325       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
329       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, vk32));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
333       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
345       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk01));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
349       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
353       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk21));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
357       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
361       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, vk41));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
380       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
384       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, vk10));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
388       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
392       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
396       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
408       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
412       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
416       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
420       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
424       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
443       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
447       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
451       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
455       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
459       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
464       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
480       v128_t vo2p0 = vbias;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()  local
492       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, vk02));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
496       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
500       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
504       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, vk32));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
508       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
520       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk01));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
524       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
528       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk21));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
532       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
536       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, vk41));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
548       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
552       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, vk10));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
556       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
560       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
564       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
576       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
580       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
584       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
588       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
592       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
604       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
608       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
612       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
616       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
620       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
625       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()