73     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()  local
92       float32x4_t vo0p4 = vmulq_lane_f32(vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
99       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
118       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
119       vi3x0123 = vi3x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
136       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
155       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
156       vi3x4567 = vi3x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
203       float32x4_t vo0p4 = vmulq_lane_f32(vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
210       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
229       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
230       vi3x0123 = vi3x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
247       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
266       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
267       vi3x4567 = vi3x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
302       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
311       float32x4_t vo0p4 = vmulq_lane_f32(vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
318       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
334       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
351       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()