53   const __m128 vk30 = _mm_load1_ps(weights + 16);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
282       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
283       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
284       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
285       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
286       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
535       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
536       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
537       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
538       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
539       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
765       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
766       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
767       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
768       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
769       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x2345, vk30));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()