AMD GPU (ROCm) programming in Julia
Documentation | Build Status |
---|---|
AMDGPU.jl can be installed with the Julia package manager.
From the Julia REPL, type ]
to enter the Pkg REPL mode and run:
pkg> add AMDGPU
Or, equivalently, via the Pkg
API:
julia> import Pkg; Pkg.add("AMDGPU")
The package is tested against and being developed for Julia 1.9, and above. Julia 1.8 is not supported and should not be used with AMDGPU.jl, as Julia 1.8 cannot properly handle the code generated by AMDGPU.jl.
Only 64-bit Linux is supported and working at this time, until ROCm is ported to other platforms.
This package is under active maintenance and is reasonably complete, however not all features (and especially performance) are up to par with CUDA.jl.
Feature | Supported | Notes |
---|---|---|
Host-side kernel launches | ✔️ | See #58 |
Device-side kernel launches | ❌ | |
Local (shared) memory (static) | ✔️ | |
Local (shared) memory (dynamic) | ✔️ | |
Fine-grained (UVM) memory | ✔️ | |
Page-locked (pinned) memory | ✔️ |
Usage questions can be posted on the Julia Discourse forum under the GPU domain and/or in the #gpu channel of the Julia Slack.
Contributions are very welcome, as are feature requests and suggestions. Please open an issue if you encounter any problems.
AMDGPU would not have been possible without the work by Tim Besard and contributors to CUDA.jl and LLVM.jl.
AMDGPU.jl is licensed under the MIT License.