94       float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE9BDF.val[0], vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()  local
100       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x8ACE9BDF.val[0], vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
104       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x8ACE9BDF.val[1], vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
108       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x8ACE9BDF.val[1], vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
112       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x8ACE9BDF.val[1], vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
127       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x68AC, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
131       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
152       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
156       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
160       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
175       vo0p1 = vfmaq_lane_f32(vo0p1, vi1xACEG, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
179       vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
183       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
209       float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE, vget_high_f32(vw0123), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()  local
215       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x8ACE, vget_high_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
219       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x9BDF, vget_low_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
223       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x9BDF, vget_high_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
227       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x9BDF, vwOP, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
237       vo0p1 = vfmaq_lane_f32(vo0p1, vi1x68AC, vget_high_f32(vw4567), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
241       vo0p1 = vfmaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
251       vo0p1 = vfmaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
255       vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
259       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
270       vo0p1 = vfmaq_lane_f32(vo0p1, vi1xACEG, vget_high_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
274       vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
278       vo0p0 = vaddq_f32(vo0p0, vo0p1);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()