Previously gpu-kernel-outlining pass was also doing index computation sinking into gpu.launch before actual outlining.
Split ops sinking from gpu-kernel-outlining pass into separate pass, so users can use theirs own sinking pass before outlining.
To achieve old behavior users will need to call both passes: -gpu-launch-sink-index-computations -gpu-kernel-outlining.
Details
Details
Diff Detail
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
Comment Actions
Thanks for splitting this out.
Consider adding an explanation to the commit message to let users know what they have to change to get the old behavior.
mlir/include/mlir/Dialect/GPU/Passes.td | ||
---|---|---|
14 | As this is now hardwired to sinking index computations, maybe call it gpu-launch-sink-index-computations? |
Comment Actions
mlir/tools/mlir-vulkan-runner/mlir-vulkan-runner.cpp did use the pass, was it intentional to change behavior, or should it now also have the new pass before outlining?
Comment Actions
mlir/tools/mlir-vulkan-runner/mlir-vulkan-runner.cpp did use the pass, was it intentional to change behavior, or should it now also have the new pass before outlining?
It should be safe to add this new pass before outlining to exactly preserve old behavior but I'm don't know if it really needed there.
As this is now hardwired to sinking index computations, maybe call it gpu-launch-sink-index-computations?