Use target attached on gpu module when lowering #1666
Conversation
| @@ -42,7 +42,6 @@ void configureGpuToROCDLConversionLegality(ConversionTarget &target); | |||
| /// is configurable. | |||
| std::unique_ptr<OperationPass<gpu::GPUModuleOp>> | |||
| createLowerGpuOpsToROCDLOpsPass( | |||
| const std::string &chipset = "gfx900", | |||
There was a problem hiding this comment.
For backwards compatibity/other people using the pass/..., keep this, but use a default value that's something like target or infer
|
Since this is NFC, I d rather land it upstream first -- so to avoid further changes (if any) from upstream comments that'd be upon the next upstream merge. |
| } | ||
| const ROCDL::ROCDLTargetAttr targetAttr = | ||
| mlir::dyn_cast<ROCDL::ROCDLTargetAttr>(targets.getValue().front()); | ||
| maybeChipset = amdgpu::Chipset::parse(targetAttr.getChip()); |
There was a problem hiding this comment.
this and failed(maybeChipset) is almost the same as the "else" code. Could we set chipset=targetAttr.getChip() and move it out of if/else?
umangyadav
force-pushed
the
target_attr
branch
from
efd63a4
to
c966d3b
Compare
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## develop #1666 +/- ##
===========================================
- Coverage 78.00% 77.96% -0.05%
===========================================
Files 98 98
Lines 26498 26499 +1
Branches 3809 3809
===========================================
- Hits 20671 20660 -11
+ Misses 4311 4301 -10
- Partials 1516 1538 +22
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. |
Moves
rocdl-attach-targetbeforeconvert-gpu-to-rocdlpass.rocdl-attach-targetwould add target attributes on gpu module which are then read duringconvert-gpu-to-rocdlSolves Task 1 of https://github.com/ROCm/rocMLIR-internal/issues/1534