History

[MLIR][XeGPU] Enhance XeGPU lane layout to support "wrap-around" distribution (#186958 )

This PR extends XeGPU lane layout to support wrap-around distribution,
enabling replication of lane-level tensor tiles across all lanes when
the tile size matches lane_data along a given dimension. Previously,
distribution required the tile size to exceed the number of lanes ×
lane_data for even partitioning.

This PR also refactors layout attribute interface functions:

computeDistributedShape() computes the distributed vector shape and is
shared by work-to-subgroup and subgroup-to-lane distribution, which
follow the same distribution rule (even or wrap-around).

computeStaticDistributedCoords() computes compile-time distributed
coordinates of sub-tiles per subgroup/lane. It is the compile-time
counterpart of computeDistributedCoords() and is used by
isCompatibleWith().

2026-03-20 17:42:25 -07:00

benchmark/python

…

cmake/modules

[mlir-python] Fix duplicate EnumAttr builder registration across dialects. (#187191 )

2026-03-19 21:02:23 -07:00

docs

[mlir][gpu] Fix typo in documentation (#156619 )

2026-03-18 13:05:58 +00:00

examples

[MLIR] Add missing dialects to C API (#82190 )

2026-01-07 12:51:33 -08:00