Commit 0273208
[AutoWS] Subtile Operator: Support partitioning subtiles with an implicit buffer + Identity Ops (4/N) (#1280)
Summary:
Adds support for partitioning subtiles into two clear regions when there is no explicit SMEM buffer (e.g. regular TMA store without early lowering).
This also supports generating tiles when regions only differ by an "identity op" used to compute an address. An example is when we compute the address and we add an offset, the base tile won't match, but it could if we `addi 0`.
Pull Request resolved: #1280
Reviewed By: Sibylau
Differential Revision: D101666304
Pulled By: njriasan
fbshipit-source-id: f333d6c1cead2187c4391fa88a8c5cf1891e9fab1 parent 7bee576 commit 0273208
2 files changed
Lines changed: 548 additions & 141 deletions
File tree
- lib/Dialect/TritonNvidiaGPU/Transforms
- test/TritonNvidiaGPU
0 commit comments