39   const float32x2_t vw89 = vld1_f32(weights + 8);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()  local
135       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
136       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
137       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
138       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
139       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
140       vo5p0 = vmlaq_lane_f32(vo5p0, vi7x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
204       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
205       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
206       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
207       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
208       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
209       vo5p0 = vmlaq_lane_f32(vo5p0, vi7x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
276       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
277       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
278       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
279       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
280       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
281       vo5p0 = vmlaq_lane_f32(vo5p0, vi7x4567, vw89, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
337       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
338       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
339       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
340       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
341       vo4p0 = vmlaq_lane_f32(vo4p0, vi6x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
342       vo5p0 = vmlaq_lane_f32(vo5p0, vi7x5678, vw89, 1);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()