334       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
348       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
352       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
356       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
360       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
587       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
601       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
605       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
609       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
613       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
817       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
831       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
835       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
839       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
843       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()