Pip install booth

[Zaneham]

Eventual goal is to make this pip installable so it can be interactable within python. This may be particularly useful to people especially as Booth comes with a CPU backend supporting x86 and RISC-V architectures alongside existing GPU support.

Currently Booth is a CLI today which is fine. However, a sizeable portion of people who write kernels usually live in the world of Pytorch and within Python. The end goal is to make it interactable for those users.

Not a goal:

• Not a Triton backend. It's MLIR based and contains LLVM.
• Not a CUDA runtime reimplementation.
• There will be no initial performance claims. I need to actually get this running correctly first.

It will be shipped into a few distinct phases/groups

1: Wheels

• setup.py shim over the existing Makefile
• cibuildwheel, which shouldn't be an issue because this has no dependencies.
• console_scripts entrypoint. Wheels on releases with checksums

2: In process compilation

libbooth.so plus a ctypes wrapper. There'll be no CPython headers. One wheel per platform across all Python versions

-booth_ctx_create() / booth_ctx_destroy() - no global state; callers compile in parallel. -Caller-frees results -Bringing over Booths structured diagnostics across the ABI -booth.compile (src, target=.....) -booth.targets()` which shows what we currently support.

3 Execution

• Launch a compiled kernel with a grid and arguments
• Buffer transfer in and out
• Uniformity across targets so the same script runs anywhere

If anyones reading and has any ideas:

Uniform launch is hard because different stacks have different support.

Package name, booth may be taken. May have to think of an easy-to-understand name.

Thanks for reading!