39   const float32x2_t vw89 = vld1_f32(weights + 8);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()  local
123       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
124       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
125       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
126       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
127       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
183       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
184       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
185       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
186       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
187       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
246       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
247       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
248       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
249       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
250       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
299       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
300       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
301       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
302       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
303       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()