234       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()  local
248       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
250       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
258       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
412       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()  local
426       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
428       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
436       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
573       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()  local
587       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
589       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
597       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()