74     float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()  local
94       vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
100       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
120       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
121       vi4x0123 = vi4x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
137       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
157       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
158       vi4x4567 = vi4x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
201       vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
207       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
227       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
228       vi4x0123 = vi4x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
244       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
264       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
265       vi4x4567 = vi4x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
295       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
305       vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
311       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
327       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
344       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()