add_tile_offset cross-term is not scaled by the element size in most transform/threadblock access iterators

Which component has the problem?

CUTLASS C++ - include/cutlass/transform/threadblock/predicated_tile_access_iterator.h and siblings

Describe the bug

add_tile_offset(TensorCoord const &tile_offset) moves the iterator by whole tiles along both ranks. When a caller passes a non-zero contiguous() component while advancing along the strided rank (or vice versa), that cross-term must be scaled by the element size because the iterators keep byte-based pointers. The residue-handling branch of the plain predicated tile access iterator does scale it, but every other path applies the raw tile count:

include/cutlass/transform/threadblock/predicated_tile_access_iterator.h, residue path (correct):

pointer_ += Shape::kContiguous * tile_offset.contiguous() * sizeof_bits<Element>::value / 8;
...
pointer_ += Shape::kStrided * tile_offset.strided() * sizeof_bits<Element>::value / 8;

same function, steady-state path a few lines below (unscaled):

pointer_ += Shape::kContiguous * tile_offset.contiguous();
...
pointer_ += Shape::kStrided * tile_offset.strided();

So the same call arguments take differently-scaled code paths depending on is_residual_tile_. The same unscaled pattern appears in:

• predicated_tile_access_iterator_2dthreadtile.h (both paths)
• ell_predicated_tile_access_iterator.h (both paths)
• predicated_tile_access_iterator_triangular_matrix.h (both paths)
• predicated_tile_access_iterator.h, AffineRankN<2> specialization, which additionally ignores stride_[0] for the contiguous rank

I believe no in-tree caller currently trips this: every add_tile_offset call site on these global-memory iterators passes exactly one non-zero component ({1,0} / {0,1} / {0,-1} style), so the cross term is always zero and the missing scaling is invisible. It is nevertheless a broken API contract for any external code moving tiles diagonally with Element wider than one byte, and the intra-file inconsistency shows at least one of the two paths must be wrong.

Steps to reproduce

Static: diff the two branches of PredicatedTileAccessIterator::add_tile_offset in the file above. A host-side demonstration needs only instantiating the iterator with a two-component tile offset, which no current example does.

Expected behavior

Both paths should apply the same, correctly scaled cross-term (element size, and for the AffineRankN<2> case the layout stride).

Environment details

• Current main, header-only observation
• Found while auditing the transform/threadblock iterator family; no functional failure observed in any in-tree kernel

Start by comparing add_tile_offset in include/cutlass/transform/threadblock/predicated_tile_access_iterator.h and its residue and steady-state branches. Then inspect predicated_tile_access_iterator_2dthreadtile.h, ell_predicated_tile_access_iterator.h, predicated_tile_access_iterator_triangular_matrix.h, and the AffineRankN<2> specialization. Done means all identified paths handle diagonal offsets with element-size scaling and the specialization respects layout stride; a host-side two-component instantiation can validate the behavior.