267       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()  local
285       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
288       vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
297       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
485       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()  local
503       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
506       vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
515       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
683       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()  local
701       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
704       vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi6x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
713       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()