dlopen libcuda.so

[madsbk]

Like we do with cuFile, dlopen the CUDA driver library

@shwina

[robertmaynard]

Overall looks great!

We also need to drop the following line from the CMakeLists.txt since we don't need to link to libcuda.so anymore.

target_link_libraries(kvikio INTERFACE CUDA::cuda_driver)

[madsbk]

We also need to drop the following line from the CMakeLists.txt since we don't need to link to libcuda.so anymore.

@robertmaynard since we still need cuda.h, do you know if we can do something similar to:

target_link_libraries(kvikio INTERFACE cufile::cuFile_interface)

Or do we just assume that cuda.h in the include path?

[robertmaynard]

@robertmaynard since we still need cuda.h, do you know if we can do something similar to:

target_link_libraries(kvikio INTERFACE cufile::cuFile_interface)

Yes, if you link to the CUDA::toolkit target you will get the required include directories with no actual linking to a cuda library

[madsbk]

Thanks for the review @robertmaynard, we now return references and linking like:

target_link_libraries(kvikio INTERFACE CUDA::toolkit)

[jakirkham]

Should we add some kind of test to ensure we are not linking to libcuda.so?

Is there anything we need to do on the Python side? Or for nvCOMP bindings?

[madsbk]

Should we add some kind of test to ensure we are not linking to libcuda.so?

I checked this previously f837214 but now that we don't link to libcuda.so, we get a compile error if using the CUDA API directly:

Is there anything we need to do on the Python side? Or for nvCOMP bindings?

I don't think so, nvCOMP is using the Runtime CUDA API.

[robertmaynard]

Why did we drop the following check, as it also confirmed that we didn't bring in a hard-dependency transitively?

# Check that `libcuda.so` is NOT being linked
LDD_BASIC_IO=`ldd ${WORKSPACE}/libkvikio-debug-build/examples/basic_io`
if [[ "$LDD_BASIC_IO" == *"libcuda.so"* ]]; then
  echo "[ERROR] examples/basic_io shouln't link to libcuda.so: ${LDD_BASIC_IO}"
  return 1
fi

[madsbk]

Why did we drop the following check, as it also confirmed that we didn't bring in a hard-dependency transitively?

Good point, adding the check again :)

[jakirkham]

@robertmaynard what do you think about the changes above? 🙂

[robertmaynard]

@robertmaynard what do you think about the changes above? slightly_smiling_face

Approved

[jakirkham]

@gpucibot merge

[madsbk]

Thanks @robertmaynard and @jakirkham !