NVIDIA / NVIDIA/cutlass

[QST]iterator store interface[store(Fragment &frag,TensorCoord const &tile_offset)] not offset in units of whole tiles

Open
#1,314 13 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

inactive-30d inactive-90d question
Dominant language
C++
Stars
10.5k
Forks
2.1k
Avg merge
3d 11h
Merged PRs (30d)
7

Description

In file include\cutlass\gemm\warp\mma_tensor_op_tile_iterator.h,
In the store interface of iterator,
As explained in the comments, "stores a tile with a logical offset in units of whole tiles", However, TensorRef is actually used for offset which is offset in units of elements.
store_with_pointer_offset(frag, ref_.offset(tile_offset) + pointer_offset);

If I understand correctly, it should bahave like the add_tile_offset interface, multiplying tensorCoord by warp shape?
ref_.add_coord_offset(tile_offset * make_Coord(Shape::kRow, Shape::kColumn));

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start in include/cutlass/gemm/warp/mma_tensor_op_tile_iterator.h at store, store_with_pointer_offset, and add_tile_offset, then compare their documented units with TensorRef offset behavior. Done means the store interface consistently treats tile_offset as whole-tile units and the relevant CUTLASS checks pass.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.