233       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()  local
245       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
247       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
249       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
257       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
408       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()  local
420       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
422       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
424       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
432       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
566       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()  local
578       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
580       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
582       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
590       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()