[Prune] Remove fly-gpu-to-llvm pass in favor of upstream gpu-to-llvm, use async_dependency to track stream argument

[sjfeng1999]

Now that async_dependencies is used to track the stream parameter, the upstream gpu-to-llvm pass correctly handles stream propagation. The custom fly-gpu-to-llvm hack is no longer needed.

Made-with: Cursor

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

Removes the project-specific fly-gpu-to-llvm conversion pass and switches the Python JIT compilation pipeline to rely on upstream MLIR gpu-to-llvm, with stream propagation now modeled via async_dependencies.

Changes:

• Drop fly-gpu-to-llvm pass registration/build targets/C API exports and delete the pass implementation.
• Update Python kernel launch emission to pass the stream through gpu.launch_func via async_dependencies.
• Switch the JIT pipeline fragment from fly-gpu-to-llvm{...} to gpu-to-llvm{...} and remove the corresponding MLIR conversion test + docs reference.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tools/fly-opt/fly-opt.cpp	Stops registering the removed fly-gpu-to-llvm pass.
tools/fly-opt/CMakeLists.txt	Removes linkage to MLIRFlyGpuToLLVM.
tests/mlir/Conversion/fly_gpu_to_llvm.mlir	Deletes FileCheck coverage for the removed custom pass.
python/mlir_flydsl/FlyRegisterEverything.cpp	Stops registering the removed pass in Python bindings.
python/flydsl/compiler/kernel_function.py	Emits gpu.launch_func with async_dependencies derived from the stream.
python/flydsl/compiler/jit_function.py	Uses upstream gpu-to-llvm in the pipeline and includes formatting-only edits.
lib/Conversion/FlyGpuToLLVM/FlyGpuToLLVM.cpp	Removes the custom pass implementation.
lib/Conversion/FlyGpuToLLVM/CMakeLists.txt	Removes the conversion library target.
lib/Conversion/CMakeLists.txt	Stops building the removed conversion subdirectory.
lib/CAPI/Dialect/FlyROCDL/FlyROCDLDialect.cpp	Removes C API registration wrapper for the removed pass.
lib/CAPI/Dialect/FlyROCDL/CMakeLists.txt	Removes linkage to MLIRFlyGpuToLLVM.
include/flydsl/Conversion/Passes.td	Removes ODS pass definition for fly-gpu-to-llvm.
include/flydsl/Conversion/Passes.h	Drops include of the removed pass header.
include/flydsl/Conversion/FlyGpuToLLVM/FlyGpuToLLVM.h	Removes the public header for the deleted pass.
include/flydsl-c/FlyROCDLDialect.h	Removes the exported C API declaration for registering the deleted pass.
docs/testing_benchmarking_guide.md	Removes the deleted test from the documentation table.

Comments suppressed due to low confidence (1)

python/flydsl/compiler/jit_function.py:312

• The PR removes the custom fly-gpu-to-llvm pass and deletes its dedicated FileCheck coverage, but there is no replacement test to ensure the upstream gpu-to-llvm continues to (a) propagate the stream argument via async_dependencies and (b) avoids creating/destroying streams when a stream is provided. Consider adding a new tests/mlir/Conversion/* test that runs --gpu-to-llvm (or the full pipeline fragment used here) and checks the resulting LLVM lowering to guard against regressions when MLIR is updated.

            f"rocdl-attach-target{{O=2 abi=600 chip={chip} correct-sqrt=true daz=false fast=false features= "
            f"finite-only=false module= triple=amdgcn-amd-amdhsa unsafe-math=false wave64=true}}",
            "convert-scf-to-cf",
            "convert-cf-to-llvm",
            "gpu-to-llvm{use-bare-pointers-for-host=true use-bare-pointers-for-kernels=true}",
            "convert-arith-to-llvm",
            "convert-func-to-llvm",
            "reconcile-unrealized-casts",
            "gpu-module-to-binary{format=fatbin}",

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[coderfeli]

Looks good. + @jli-melchior

[jli-melchior]

LGTM