299       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
317       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
321       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));  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()
335       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
552       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
570       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
574       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));  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()
588       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
782       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
800       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
804       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));  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()
818       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()