201       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()  local
211       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
212       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
219       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
339       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()  local
349       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
350       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
357       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
463       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()  local
473       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
474       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
481       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()