diff --git a/include/cute/atom/copy_traits_xe_2d.hpp b/include/cute/atom/copy_traits_xe_2d.hpp index 2df8ae0a38..7c526a405e 100644 --- a/include/cute/atom/copy_traits_xe_2d.hpp +++ b/include/cute/atom/copy_traits_xe_2d.hpp @@ -500,7 +500,8 @@ make_block_2d_copy(const CopyOp& op, using LayoutCopy_TV = typename SGCopy::TiledLayout_TV; // Expand the shape. - auto x_shape = elem_scale(ShapeTiler_MN{}, atom_shape); + auto x_atom_shape = append>(atom_shape, _1{}); + auto x_shape = elem_scale(ShapeTiler_MN{}, x_atom_shape); // Expand the single-SG TV layout to the full shape, then tile. auto x_tv_layout1 = composition(make_layout(ShapeTiler_MN{}, make_layout(x_shape).stride()), LayoutCopy_TV{});