cuda.compute: Expose a way to pass additional compilation options to NVRTC

[shwina]

https://docs.nvidia.com/cuda/nvrtc/index.html#supported-compile-options

It would be useful to be able to pass additional compiler flags to NVRTC to control e.g., floating point math.

The CCCL.c layer already provides this facility, we just need to plumb it through to the user.

The API could look something like:

set_nvrtc_compile_options({"fmad": "true"})

Or on a per algorithm basis:

reducer = make_reduce_into(d_in, d_out, op, h_init, num_items, nvrtc_compile_options={"fmad": "true"})

[jrhemstad]

We could also consider doing something simpler and just allow someone to pass us a string that we forward as the compile flags to the NVRTC invocation.

That would insulate us from if NVRTC adds new flags we don't have to worry about exposing them.

[NaderAlAwar]

We already support custom build options in c.parallel through the *_build_ex functions

cccl/c/parallel/include/cccl/c/types.h

Lines 108 to 114 in 286842c

	typedef struct cccl_build_config
	{
	const char** extra_compile_flags; // e.g., {"-DENABLE_FAST_MATH", "-O3"}
	size_t num_extra_compile_flags;
	const char** extra_include_dirs; // e.g., {"/path/to/my/headers"}
	size_t num_extra_include_dirs;
	} cccl_build_config;

We don't expose them to python yet though. There are examples of their usage in the c.parallel tests.

[shwina]

That's almost what we're doing with nvrtc_compile_options={"fmad": "true"}. We would just translate the user-provide dictionary to --fmad=true.

Although now that I think about it, a string would probably be less of a hassle (we don't have to worry about things like -switch v/s --option=value etc.,)

[leofang]

Would it be possible for cuda.compute to just take a cuda.core.ProgramOptions and feed it to cccl-c? It seems like reinventing wheel to me.

[shwina]

Would it be possible for cuda.compute to just take a cuda.core.ProgramOptions and feed it to cccl-c? It seems like reinventing wheel to me.

Hmm - are the ProgramOptions supposed to be 1:1 with NVRTC compiler options or something else? There are certain options I can't find (e.g., fmad, maxrregcount).

[shwina]

OK, I see that the options are named differently (e.g., fma= and max_register_count=).

• Can cuda.core also support the verbatim options? This way, users don't have to worry about translating from the NVRTC option name to the corresponding cuda.core option name.
• Can the _as_bytes() method be made public?

I do like passing Python values much better than strings.

[leofang]

@shwina could you create an issue in cuda-python?

ProgramOptions is a superset of all options accepted by NVRTC, NVVM, and nvJitLink. Unfortunately, there will be no names that can make everyone happy, because even within these libraries they don't agree with each other 🥲 Example: -fmad vs -fma.

[shwina]

Example: -fmad vs -fma

So support both? ;)