META: CUDA external language implementation

[maleadt]

Summary: make inference and codegen modular and configurable for packages to reuse, eg., when generating code for a different platform. This way, we avoid bloating the compiler, and make it possible to develop and import new hardware support without requiring modifications to the compiler/base.

Concretely, this PR will track the necessary additions to Julia master and host the remaining diff to support the CUDA compiler over at CUDAnative.jl.

Inference

We need to influence inference in order to select alternative functions for some stdlib functionality (think sin, which now calls out to libm but needs to call another library), sometimes even depending on the GPU version.

• Allow modifying the inference-limiting parameters to perform more exhaustive search: Put inference parameters in a dedicated object #18496
• Properly handle caching (either uncached, or using multiple method caches)
• Hooks for eg. overriding method selection: tb/inference_hooks

Codegen

Much of this is similar to inference:

• Allow specifying certain parameters, eg. disabling optimization: RFC: CodegenParams for external language implementations #19046
• Allow disabling language features: RFC: CodegenParams for external language implementations #19046
• Handle caches: partially implemented in RFC: CodegenParams for external language implementations #19046
• Hooks for changing code emission: tb/codegen_hooks

Longer-term:

• Make codegen sane: pure entry-points, less global state

Support

• Wrap LLVM C API: LLVM.jl

[maleadt]

Small update: LLVM.jl is now powerful enough to fully implement the PTX JIT (which is pretty simple).

[datnamer]

Would be cool to keep in mind a Webassembly usecase for this:

https://github.com/WebAssembly/wasm-jit-prototype
https://github.com/WebAssembly/binaryen

[maleadt]

Yes, definitely. Implementation wise, we can look at Rust for inspiration as they have a functional LLVM-based wasm target nowadays. Do you have any use-cases in mind?

[datnamer]

Well my (dream/pony) high level usecases include distributing fully client side interactive reports and machine learning mobile apps.

[maleadt]

Rebased on top of #18496. I've been working on similar 'codegen params' in the JuliaGPU/julia repo, but that still needs to be fleshed out.

[maleadt]

I've removed commits from this PR, making it a tracking issue instead. Code has moved to tb/cuda.

[maleadt]

This is outdated.