239       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
276       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x2345, vk10));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
280       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));  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()
288       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
492       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
529       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x2345, vk10));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
533       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));  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()
541       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
734       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
759       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x2345, vk10));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
763       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));  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()
771       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()