266       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()  local
281       vo3p1 = _mm_add_ps(vo3p1, _mm_mul_ps(vi5x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
284       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
287       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
296       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
484       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()  local
499       vo3p1 = _mm_add_ps(vo3p1, _mm_mul_ps(vi5x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
502       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
505       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
514       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
682       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()  local
697       vo3p1 = _mm_add_ps(vo3p1, _mm_mul_ps(vi5x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
700       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
703       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
712       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()