MLIR-based GEMM, convolution, attention, GEMM+GEMM, and CONV+GEMM kernel generator for AMD GPUs.

rocMLIR is an MLIR-based GPU kernel generator targeting AMD GPUs. The high-level lowering is migraphx -> tosa / linalg -> rock, which then continues through MLIR's amdgpu and rocdl dialects to HSACO via the LLVM AMDGPU backend (vendored under external/llvm-project/).

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.

• An AMD GPU and a working ROCm installation (with rocminfo on PATH).
• A reasonably recent clang / clang++ (the ROCm-shipped compiler at /opt/rocm/llvm/bin/clang++ is the standard development toolchain).
• lld, ninja, and CMake >= 3.20.
• Python 3 (>= 3.8 if you build with LLVM_INCLUDE_TESTS=ON, the default; >= 3.0 otherwise). Required at configure time for the vendored LLVM build, plus in-tree development scripts and the LIT test runner. Not needed by downstream consumers (e.g. MIGraphX) that only link against the prebuilt librockCompiler.
• Git.

To not actually run the tests, use check-rocmlir-build-only.

To build the static librockCompiler library used by MIGraphX:

To install librockCompiler so MIGraphX can find it:

Additional developer documentation lives under mlir/docs/.

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 LLVM build directory under external/llvm-project/, and links the right runtime libraries (libmlir_rocm_runtime, libconv-validation-wrappers, runner utils, etc.):

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.
• -pv_with_gpu -- validate against a GPU reference instead.
• -pr -- print kernel results.
• -mfma=on|off -- explicitly enable/disable MFMA (or -wmma=on|off on WMMA targets).

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

rocmlir-driver is a wrapper around the kernel generation pipeline. Use -c (or --kernel-pipeline=full --host-pipeline=runner) to run the default pipeline. Adding --debug-only=serialize-to-isa will dump the GCN assembly for the executed kernels to standard error.

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/fusion/e2e/. To build and run the full in-tree test suite (from the build directory):

By default, we infer the use of GPU-specific acceleration instructions (MFMA or WMMA) based on the features of the currently available GPU. To disable this, add -DROCMLIR_GEN_FLAGS="-mfma=off -wmma=off" to the cmake invocation above. Note that this will not affect behavior in production/static library builds, which do not use rocmlir-gen.

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

For bugs and feature requests, open a GitHub Issue.

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

For questions, issues, or contributions, please reach out to the maintainers:

• Chris Austen — @causten · chausten@amd.com

See CODEOWNERS for the full ownership list.

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