rocWMMA

October 27, 2025 ยท View on GitHub

Caution

The rocWMMA repository is retired, please use the ROCm/rocm-libraries repository

Welcome! rocWMMA is a C++ library for accelerating mixed-precision matrix multiply-accumulate (MMA) operations leveraging AMD GPU hardware. rocWMMA makes it easier to break down MMA problems into fragments and distribute block-wise MMA operations in parallel across GPU wavefronts. The API consists of a header library, that can be used to compile MMA acceleration directly into GPU kernel device code. This can benefit from compiler optimization in the generation of kernel assembly, and doesn't incur additional overhead costs of linking to external runtime libraries or having to launch separate kernels.

rocWMMA includes sample projects to validate and demonstrate API usage. These include simple GEMMs, performant GEMMs, DLRM, GEMV and hipRTC integration.

The test suite includes validation and benchmarking projects that focus on unit testing, GEMMs and DLRM.

Note

The published rocWMMA documentation is available at rocWMMA in an organized, easy-to-read format, with search and a table of contents. The documentation source files reside in the rocWMMA/docs folder of this repository. As with all ROCm projects, the documentation is open source. For more information, see Contribute to ROCm documentation.

Requirements

rocWMMA currently supports the following AMD GPU architectures:

  • CDNA class GPU featuring matrix core support: gfx908, gfx90a, gfx942, gfx950 as 'gfx9'
  • RDNA class GPU featuring AI acceleration support: gfx1100, gfx1101, gfx1102, gfx1151 as 'gfx11'; gfx1200, gfx1201 as 'gfx12'

Dependencies:

  • Minimum ROCm version support is 6.4.
  • Minimum cmake version support is 3.14.
  • Minimum ROCm-cmake version support is 0.8.0.
  • Minimum rocBLAS version support is rocBLAS 4.0.0 for ROCm 6.0* (or ROCm packages rocblas and rocblas-dev).
  • Minimum ROCm SMI version support is 7.6.0** (or ROCm packages rocm-smi-lib and librocm-smi-dev).
  • Minimum HIP runtime version support is 4.3.0 (or ROCm package hip-runtime-amd).
  • Minimum LLVM OpenMP runtime dev package version support is 10.0 (available as ROCm package rocm-llvm-dev).
    * = if using rocBLAS for validation.
    ** = if building benchmark tests (configuring with ROCWMMA_BUILD_BENCHMARK_TESTS=ON).

    It is best to use available ROCm packages from the same release where applicable.

Build with CMake

For more detailed information, please refer to the rocWMMA installation guide.

Project options

OptionDescriptionDefault value
GPU_TARGETSBuild code for specific GPU target(s)gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201
AMDGPU_TARGETS(Deprecated) Build code for specific GPU target(s)gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201
ROCWMMA_BUILD_TESTSBuild TestsON
ROCWMMA_BUILD_SAMPLESBuild SamplesON
ROCWMMA_BUILD_DOCSBuild doxygen documentation from codeOFF
ROCWMMA_BUILD_ASSEMBLYGenerate assembly filesOFF
ROCWMMA_BUILD_VALIDATION_TESTSBuild validation testsON (requires ROCWMMA_BUILD_TESTS=ON)
ROCWMMA_BUILD_BENCHMARK_TESTSBuild benchmark testsOFF (requires ROCWMMA_BUILD_TESTS=ON)
ROCWMMA_BUILD_EXTENDED_TESTSBuild extended testing coverageOFF (requires ROCWMMA_BUILD_TESTS=ON)
ROCWMMA_VALIDATE_WITH_ROCBLASUse rocBLAS for validation testsON (requires ROCWMMA_BUILD_VALIDATION_TESTS=ON)
ROCWMMA_BENCHMARK_WITH_ROCBLASInclude rocBLAS benchmarking dataOFF (requires ROCWMMA_BUILD_BENCHMARK_TESTS=ON)
ROCWMMA_USE_SYSTEM_GOOGLETESTUse system Google Test library instead of downloading and building itOFF (requires ROCWMMA_BUILD_TESTS=ON)

Example configurations

By default, the project is configured in release mode and is linked against rocBLAS for validating results. Here are some configuration examples:

ConfigurationCommand
BasicCC=/opt/rocm/bin/amdclang CXX=/opt/rocm/bin/amdclang++ cmake -B<build_dir> .
Targeting gfx908CC=/opt/rocm/bin/amdclang CXX=/opt/rocm/bin/amdclang++ cmake -B<build_dir> . -DGPU_TARGETS=gfx908:xnack-
Debug buildCC=/opt/rocm/bin/amdclang CXX=/opt/rocm/bin/amdclang++ cmake -B<build_dir> . -DCMAKE_BUILD_TYPE=Debug
Build without rocBLAS (default on)CC=/opt/rocm/bin/amdclang CXX=/opt/rocm/bin/amdclang++ cmake -B<build_dir> . -DROCWMMA_VALIDATE_WITH_ROCBLAS=OFF -DROCWMMA_BENCHMARK_WITH_ROCBLAS=OFF

After configuration, build with cmake --build <build_dir> -- -j<nproc>

Documentation

For more comprehensive documentation on installation, samples and test contents, API reference and programmer's guide you can build the documentation locally in different ways.

Html

cd docs

pip3 install -r sphinx/requirements.txt

python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html

The HTML documentation can be viewed in your browser by opening docs/_build/html/index.html result.

Pdf

cd docs

sudo apt-get update
sudo apt-get install doxygen
sudo apt-get install texlive-latex-base texlive-latex-extra

pip3 install -r sphinx/requirements.txt

python3 -m sphinx -T -E -b latex -d _build/doctrees -D language=en . _build/latex

cd _build/latex

pdflatex rocwmma.tex

Running the above commands generates rocwmma.pdf.

The latest official documentation for rocWMMA is available at: https://rocm.docs.amd.com/projects/rocWMMA/en/latest/index.html.

Contributing to the rocWMMA Library

Community collaboration is encouraged! If you are considering contributing, please follow the rocWMMA Contribution Guide to get started.