libcublas-dev should depend on cuda-crt-dev_*

[adibbley]

cublas headers depend on cudart headers, which depend on nvcc (crt) headers. We should add these dependencies once those packages are available.

[bdice]

When compilers are available, we can also add tests that compile a sample executable (but can't run it without a GPU). Suggesting this, which was proposed by @robertmaynard.

// test.cpp
#include <cublas_v2.h>

int main() {
    cublasStatus_t stat;
    cublasHandle_t handle;
    stat = cublasCreate(&handle);
    return stat != CUBLAS_STATUS_SUCCESS;
}

Then we'd depend on compilers in the test and execute something like g++ -I... -lcublas test.cpp.

[jakirkham]

Since {{ compiler("cuda") }} is going to be needed as a requirement/build by any recipe build CUDA packages, this should be picked up via cuda-nvcc. If not, we may need to tweak the dependencies of cuda-nvcc and/or cuda-cudart to make sure that is happening

[bdice]

That’s fine for downstream recipes, but we still need these requirements on the libcublas recipe if users are building applications from source with libcublas.

[jakirkham]

Maybe I'm missing something.

Thought this was for libcublas-dev. Understand mitigating the libcublas dlopen issue ( #9 ) requires a runtime dependency.

How is a runtime dependency used here?

[robertmaynard]

How is a runtime dependency used here?

The headers of libcublas require headers from cuda-cudart-dev ( cuda_runtime.h) which requires headers from cuda-nvcc ( crt/host_defines.h ).

So to compile/use libcublas with a C or C++ compiler you still need the above packages installed to include cublas.h.

[jakirkham]

cc @billysuh7 (to look into this further)

[jakirkham]

Now that the CRT headers are split out of NVCC as of CUDA 12.2 ( conda-forge/cuda-nvcc-feedstock#32 ), this should be doable

It sounds like cudart is not needed for cuBLAS. So that doesn't need to be added (confirmed by running ldd on libcublas.so to check for libcudart linkages)

In terms of cuBLAS packages, these make most sense to add to the -dev packages

[bdice]

@jakirkham Running ldd doesn’t check for dlopen, right? Is there a possibility of libcudart being dlopened, like nvrtc in #9?

Ultimately I think only library maintainers can provide accurate dependency lists, ldd is “necessary but not sufficient” for accuracy.

[leofang]

All math libs must link to cudart statically. It's a requirement.

[bdice]

Great, then I agree we can make progress on this when someone has time.

[billysuh7]

I will take this. Thanks.

[jakirkham]

Should add the above comment ( #8 (comment) ) was a summary of our discussion in the meeting yesterday ( with feedback from @adibbley and @robertmaynard )

[jakirkham]

Retitled the issue to hopefully make this clearer

Thanks Billy! 🙏