CUDA sin is broken

[wsmoses]

There was some change such (? to CUDA.jl) that nv_sin/etc is not always ebing linked in now, which means we cannot find it to generate the derivative of nv_cos.

[wsmoses]

@vchuravy if you have any insights here

[wsmoses]

So this was broken by https://github.com/EnzymeAD/Enzyme.jl/pull/915/files aka GPUCompiler 0.20

cc @maleadt if you might have any insights here.

Essentially what used to happen is that for cuda code using device functions Enzyme/GPUCompiler would link in all of libdevice (aka sin, cos, etc). This is necessary since the derivative of nv_sin is nv_cos.

Since ?GPUCompiler 0.20, it appears that by the time Enzyme gets it, the (from the source) unused nv_cos is dropped, leaving Enzyme.jl no implementation to use on CUDA, leading to issues.

[wsmoses]

Yes, this PR: JuliaGPU/GPUCompiler.jl#458 specifically broke Enzyme CUDA compilation here, for the reasons above.

@maleadt can we go back to having linking before codegen (otherwise we are unable to do the relevant force preservation and llvm.sin -> __nv_sin) ?

[maleadt]

@maleadt can we go back to having linking before codegen (otherwise we are unable to do the relevant force preservation and llvm.sin -> __nv_sin) ?

Not easily, as that was a correctness fix for deferred-mode compilation. Changing it would re-introduce a bug.

You should be able to sufficiently customize the compilation already though. Where were you currently detecting __nv_sin?