zhouronghua wrote:

> In offloading languages like CUDA / HIP my understanding is that the 
> device-side compilations are not independent build targets, so we do not want 
> to create dependencies for them separate from the host. This is why all of 
> them are currently attached to the host compilation, which is why I'm leaning 
> towards saying that the proper solution is to list these files as 
> dependencies on the one we already have. There is no separate 'device' build 
> to re-run as far as make is concerned, it's just the host object.

I understand your point. Since no one compiles CUDA/HIP kernel files separately 
to generate cubin/fatbin—instead, we get a host object file with embedded 
cubin/fatbin—we need to trigger incremental builds for the .cu file whenever 
kernel dependencies change. This ensures both the host code and the injected 
cubin/fatbin are regenerated.

Just like the example above, if we compile a host object file containing the 
SM70 cubin, and then sm_70_features.h is modified, we need a way to inform the 
build system that an incremental rebuild should be triggered, correct?

https://github.com/llvm/llvm-project/pull/176072
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to