145       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()  local
158       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x3456, vget_high_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
162       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
166       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
170       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
363       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()  local
376       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x3456, vget_high_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
380       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
384       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
388       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
575       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()  local
588       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x3456, vget_high_f32(vw4567), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
592       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
596       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
600       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()