73     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()  local
92       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
99       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
118       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
119       vi3x0123 = vi3x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
136       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
155       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
156       vi3x4567 = vi3x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
199       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
206       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
225       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
226       vi3x0123 = vi3x4567;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
243       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
262       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
263       vi3x4567 = vi3x89AB;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
294       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
303       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
310       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
326       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
343       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()