cuda.core v0.1.0

[leofang]

We are pleased to introduce a new subpackage cuda.core to Python users, the first very step toward "pythonic CUDA"! 🚀

This is the initial beta release, with more beta releases planned to follow.

Mission

The new module cuda.core is aimed to

• provide idiomatic ("pythonic") access to CUDA Driver, Runtime, and JIT compiler toolchain
• focus on developer productivity by ensuring end-to-end CUDA development can be performed quickly and entirely in Python
• avoid homegrown Python abstractions for CUDA for new Python GPU libraries starting from scratch
• ease developer burden of maintaining and catching up with latest CUDA features
• flatten the learning curve for current and future generations of CUDA developers

Development model

cuda.core adopts an iterative development process to gradually cover major CUDA features in a pythonic fashion, while collecting/incorporating early adopters' feedback for the design. The cuda.core design can and may change across beta releases, and will be declared stable when we reach GA (General Access, v1.0.0).

Key Features and Enhancements

• Idiomatic/pythonic access to CUDA Driver, objects covered
  • Device
  • Stream
    • Also introduce the __cuda_stream__ protocol
  • Event
  • Buffer
  • launch() & LaunchConfig
    • support Python scalars and NumPy/ctypes fixed-width scalars as kernel arguments
  • Program
  • ObjectCode
  • Kernel

Code samples are available on GitHub: https://github.com/NVIDIA/cuda-python/tree/main/cuda_core/examples

All functionalities are currently hosted under the cuda.core.experimental namespace. Once the features become stable they will be moved out of experimental.

Release notes

See the release notes (and full documentation) at https://nvidia.github.io/cuda-python/cuda-core/0.1.0/.

---

This discussion was created from the release cuda.core v0.1.0.

[jianruyin]

Congratulations on your release. Could you please introduce more about the necessary whole stack of CUDA and where is cuda-python, and where will it go? Thanks.

[leofang]

Yes, thanks @jianruyin for asking. I've added a bit more detailed post to the OP above.

Regarding the question

and where will it go?

as noted in the updated OP, we plan to gradually cover major CUDA features in a pythonic fashion, but we would love to hear about what users need too. Our plan/roadmap is tracked in various milestones.