You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 20c0651
Browse filesBrowse the repository at this point in the historyBrowse files
Arm backend: Generalize MatchArgDtypePass (#24227)
The pass is responsible for matching dtypes
of operators inserted after executorch's type-promotion.
It currently just handles sub and where. Additionally, it does not
handle all args of sub correctly.
- Expand targets of MatchArgDtypePass
- Use torch.result_type to cast both operands of supported binary
operations. This hanldes cases like uint8+int8 which previously
incorrectly cast to int8 (uint8 value of 200 can't be represented by
int8)
- Move the pass later to make sure constants are folded.
- Allow promotable mixed inputs in the combined TOSA profile.
- Add mixed-dtype tests and update VGF matmul expectations. The removed
strict xfails covered vector matmul cases that decompose to mul and sum;
casting mul's integer operand now lets those cases lower successfully.
Keep xfails for the backends and A16W8 cases that still need them.
cc @digantdesai@freddan80@per@zingo@oscarandersson8218@mansnils@Sebastian-Larsson@robell@rascani
Signed-off-by: Erik Lundell <erik.lundell@arm.com>
0 commit comments