[ANN] cudajit: Bindings to the `cuda` and `nvrtc` libraries

Community
1

Hi! I’m happy to share cudajit 0.4.0: manually-selected bindings for Nvidia GPU programming. cudajit should soon propagate to the opam repository.
Bindings to the cuda and nvrtc libraries with a unified interface
API documentation
Currently supported:

  • Compiling a kernel with conversion to PTX, launching a kernel.
  • Synchronous and asynchronous memory copying.
  • Contexts and streams.
  • (GPU) device attributes.

Currently not supported:

  • Events.
  • CUDA graph features.
  • Cooperative kernel launch.

Let me know your needs so I can prioritize. PRs are also welcome!

11 Likes
2

cudajit 0.5.0 is now available in the opam repository. It’s organized into modules, and it adds support for CUDA events.

5 Likes
3

cudajit 0.7.0 is now available in the opam repository. It is now split into separate libraries covering NVRTC bindings and CUDA bindings, so that Nvrtc doesn’t need CUDA drivers to run.

Version 0.7.0 brings full native Windows compatibility.

Version 0.6.0 improves memory safety and debuggability.

cudajit.0.7.0 can be used with OCANNL neural_nets_lib.0.5.2 also in the opam repository. Enjoy!

1 Like