feat: Add global fast_math flag to CudaBuilder

[koreaygj]

Hi
I try to implemente a fast_math() method to solve this issue. It provides a convenient way to enable fast math approximations globally, equivalent to NVCC's --use_fast_math option.

Description

Implement a global fast_math flag for CudaBuilder as a convenience method equivalent to NVCC's --use_fast_math option.

According to the NVCC official documentation, the --use_fast_math flag enables fast approximations for floating-point operations by internally setting ftz, prec-sqrt, and prec-div.

Implementation

I implemented a fast_math() method that sets these three options internally:

pubfnfast_math(mutself,fast_math:bool) -> Self{if fast_math {self.ftz = true;self.fast_sqrt = true;self.fast_div = true;self.fma_contraction = true;}self}

Usage

Users can now enable fast math globally with a single method call:

CudaBuilder::new("kernel").fast_math(true).build()?

Instead of manually setting each flag:

CudaBuilder::new("kernel").ftz(true).fast_sqrt(true).fast_div(true).build()?

Implementation Details

• The fast_math() method is a convenience wrapper that internally enables ftz, fast_sqrt, and fast_div
• Leverages existing LLVM argument generation in invoke_rustc()
• No changes to the rustc_codegen_nvvm backend required
• Maintains backward compatibility (default is false)

Relates to #262

[nnethercote]

The implementation is confused. How do the individual flags interact with the new flag? What happens if you specify fast_math and then also fast_sqrt(false)? Seems like either:

• CudaBuilder should handle all the fast_math stuff, and pass individual flags to NVCC, but not --use_fast_math
• CudaBuilder should just record fast_math and then pass --use_fast_math through to NVCC

But currently it's doing a mixture of both.

Also, was this PR generated by AI? The description is very long, with subheadings and bullet points that are typical for AI.

[koreaygj]

Thank you for your review.
I used AI to generate the PR description, next time I write myself...

Do you think delegating to NVCC would be the better approach? I need your advices.
Instead of CudaBuilder managing the individual flags, we could just pass --use_fast_math directly to NVCC and let it handle the optimizations.

[LegNeato]

How does this jive with https://simonbyrne.github.io/notes/fastmath/ ?

[koreaygj]

Sorry to late! @nnethercote@LegNeato
I've reworked the implementation based on your review:

• Refactor CudaBuilder: Remove redundant fast_math field and LLVM args.
• Direct Flag Passing: Replace --use_fast_math with individual LLVM flags for better precision.
• Add doc comment to know IEEE 754 tradeoffs.

How does this jive with https://simonbyrne.github.io/notes/fastmath/ ?

Regarding IEEE 754 concerns — fast_math() doesn't add any new capability. It's a convenience wrapper over the same four individual flags (ftz, fast_sqrt, fast_div, fma_contraction) that already exist.