88       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()  local
94       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
108       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x3456, vget_low_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
127       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
141       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x5678, vget_low_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
160       vo0p2 = vmlaq_lane_f32(vo0p2, vi0x6789, vget_low_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
168       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
171       vo0p2 = vaddq_f32(vo0p2, vo0p3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
172       vo0p0 = vaddq_f32(vo0p0, vo0p2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
198       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()  local
204       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
218       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x3456, vget_low_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
237       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
251       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x5678, vget_low_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
270       vo0p2 = vmlaq_lane_f32(vo0p2, vi0x6789, vget_low_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
278       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
281       vo0p2 = vaddq_f32(vo0p2, vo0p3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
282       vo0p0 = vaddq_f32(vo0p0, vo0p2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
305       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()  local
311       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x4567, vget_high_f32(vwKLMN), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
325       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x3456, vget_low_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
339       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
354       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x5678, vget_low_f32(vw89AB), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
368       vo0p2 = vmlaq_lane_f32(vo0p2, vi0x6789, vget_low_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
376       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
379       vo0p2 = vaddq_f32(vo0p2, vo0p3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
380       vo0p0 = vaddq_f32(vo0p0, vo0p2);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()