I'm excited to announce that basic ROCm support was added to [community-testing] yesterday!
What is ROCm? ROCm is AMD's open source framework for GPU programming. It comprises an OpenCL implementation, HIP (AMD's CUDA) and OpenMP offloading.
Which GPUs are supported? Official support is limited, see .
Which parts were added to [community-testing]? The low level components (rocminfo, rocm-smi-lib), the OpenCL runtime (rocm-opencl-runtime) and the HIP runtime / compiler (hip-runtime-amd).
How can I use it? Install (hip-runtime-amd rocm-smi-lib rocm-opencl-runtime) to use all currently available packages. First check if the amdgpu kernel module is loaded and add your login user to the video and render groups. Run /opt/rocm/bin/rocminfo to validate that the HSA runtime correctly detects your hardware. You can also check the output of clinfo for OpenCL support.
Use case: blender First, install hip-runtime-amd and add /opt/rocm/bin to PATH. Open blender, go to Edit -> Preferences and select "HIP" in "System / Cycles Render Devices". Open your favorite scene and select "GPU compute" in "Render Properties". You can monitor the GPU load with
watch -n.1 rocm-smi
if you've installed rocm-smi-lib.
What's next? Add the ROCm math libraries (rocblas, rocsparse, rocfft, ..) that are needed for Machine Learning and many other applications.