rocmlirTriton

MLIR-based GEMM, convolution, attention, GEMM+GEMM, and CONV+GEMM kernel generator for AMD GPUs, built on a Triton compilation backend.

rocmlirTriton is a Triton-backed GPU kernel generator derived from rocMLIR. Both share, for the most part, the same high-level lowering: migraphx -> tosa (or linalg) -> rock, and diverge only at the codegen step: rocMLIR lowers rock to MLIR's AMDGPU and ROCDL dialects to HSACO via the LLVM AMDGPU backend, while rocmlirTriton hands off to Triton's TTIR -> TTGIR -> LLIR pipeline (see Pipelines.cpp / TritonToHsaco.cpp), with vendored Triton (under external/triton) and Triton-pinned LLVM/MLIR (under external/llvm-project) producing the final HSACO.

It targets AMD CDNA and RDNA GPUs (gfx9xx / gfx10xx / gfx11xx / gfx12xx), and is primarily consumed as the static librockCompiler library by MIGraphX, though it can also be driven standalone for kernel generation, validation, and performance tuning.

Prerequisites

• An AMD GPU and a working ROCm installation (with rocminfo on PATH).
• clang / clang++ 20 (defaults to clang-20 / clang++-20; override via the C_COMPILER / CXX_COMPILER environment variables).
• lld, ninja, and CMake >= 3.20.
• Python 3 (only needed for in-tree development scripts and the LIT test runner; not required for production builds or MIGraphX integration).
• Git

Installation

Triton and LLVM/MLIR are vendored in the repo (under external/triton and external/llvm-project), imported via git subtree. The build is driven by cmake.sh, which configures and builds LLVM/MLIR, Triton, and rocmlirTriton from those vendored trees. Patch files under llvm-patches/ and triton-patches/ are kept for provenance and to simplify the next upstream bump; they are not applied during CMake configure.

git clone https://github.com/ROCm/rocmlirTriton.git
cd rocmlirTriton
bash cmake.sh

cmake.sh wipes build/ before configuring. Pass --no-clean to reconfigure an existing build directory instead, which avoids a full LLVM rebuild when you only want to change a few -D flags. Any extra arguments are forwarded to cmake.

To install librockCompiler so MIGraphX can find it:

cmake --install build --prefix /path/to/MIGraphX/deps

Additional developer documentation lives under docs/.

Usage

A typical standalone pipeline generates a kernel with rocmlir-gen, lowers it with rocmlir-driver -c, and runs it via rocm-run -- a wrapper around mlir-runner that auto-locates the rocMLIR build and the Triton-pinned LLVM build directory, and links the right runtime libraries (libmlir_rocm_runtime, libconv-validation-wrappers, runner utils, etc.):

ARCH=$(rocminfo | grep -o 'gfx[0-9a-z]*' | head -1)

build/bin/rocmlir-gen -pv -operation gemm -t f16 -out_datatype f32 \
    --arch "$ARCH" -g 1 -m 64 -k 256 -n 128 \
  | build/bin/rocmlir-driver -c \
  | build/bin/rocm-run

Useful rocmlir-gen flags:

• --arch -- target AMDGPU architecture (e.g. gfx942, gfx950, gfx1100); MFMA/WMMA support is inferred from the chosen architecture.
• -t / --dtype -- data type selector (e.g. f16, f32, bf16, i8, fp8_fp8).
• -out_datatype / --out_dtype / -tc -- override the output data type independently of -t (e.g. f16 input with f32 output).
• --perf_config -- supply a serialized tuning configuration.
• -ph -- emit host code alongside the kernel.
• -pv -- validate kernel results against a CPU reference.
• -pr -- print kernel results.

Run build/bin/rocmlir-gen --help for the full, current option list.

More examples live under mlir/test/rocmlir-driver/ (notably sanity.mlir), with end-to-end PR tests under mlir/test/fusion/pr-e2e/ (including the MIGraphX-dialect mixr-* tests) and mlir/test/e2e/. To build and run the full in-tree test suite (from the build directory):

cd build && ninja check-rocmlir

Contributing

We welcome contributions! Please read CONTRIBUTING.md for the issue-reporting and pull-request workflow.

For bugs and feature requests, open a GitHub Issue.

---

Security

To report a security vulnerability, do not open a public GitHub issue. See SECURITY.md for our responsible disclosure policy.

---

Contact

For questions, bug reports, or feature requests, please open a GitHub issue. Contributions are welcome through pull requests.

See CODEOWNERS for the full ownership list.

---

License

This project is licensed under the Apache License 2.0 with LLVM Exceptions.