265       vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
266       vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
267       vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
268       vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
269       vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
270       vi5x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
271       vi6x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
449       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
450       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
451       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
452       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
453       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
454       vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
455       vi6x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
811       vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
812       vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
813       vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
814       vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
815       vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
816       vi5x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
958       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
959       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
960       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
961       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
962       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
963       vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1206       vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1207       vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1208       vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1209       vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1210       vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1293       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1294       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1295       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1296       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1297       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()