61   const __m128 vk43 = _mm_load1_ps(weights + 24);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
323       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
324       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
325       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
326       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
327       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
576       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
577       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
578       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
579       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
580       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
806       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
807       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
808       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
809       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
810       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()