Add dependency information to manifest json

[jumpywizards]

In order to enable easy packaging in different contexts it would be extremely helpfull if dependency information for all cuda-toolkit components/binaries would be included in the manifest .json file.

Can you please add those to the manifest?

[kmittman]

Hi @jumpywizards can you provide any more detail on what format you are looking for?

For example, an array of $component names? Inclusion of >= $version ?
Or something more like Depends: $SONAME ?

Keep in mind, there are both "soft" dependencies and "fixed" ones.

It might be helpful if you could mention what sorts of package formats that you intend to maintain.

For reference here is the current JSON schema: https://developer.download.nvidia.com/compute/redist/redistrib-v2.schema.json

[jumpywizards]

Hi @kmittman, thank you for your quick response!

I am trying to build a set of fine grained Conan packages for the CUDA-Toolkit, mainly to be used in CMake projects. In my case I would need (toolkit-internal) dependency information on two levels: i) For the Conan packages, I need inter-component dependency information (i.e. libcusolver -> [libcublas, libcusparse]), and ii) for the build system, I need inter-binary dependency information (i.e. libnvblas.so -> libcublas.so). The latter would be especially helpfull for static libraries where the dependency information is not part of the binary and cannot be determined via tools like ojdump.

As for the format, I would expect for every component and platform something like a list of binaries paired with their respective binary dependencies + component source (maybe relative path?). I would leave the details of the format up to you.

[jumpywizards]

Btw., I noticed that libOpenCL is part of the nvidia_driver component as well as the cuda_cudart component and that both differ from each other.

Is there a reason for that?

[kmittman]

Some overlap between the toolkit and driver is common/expected. They install in different paths, so generally it should not be an issue except for ldconfig. However in this case it should only exist in the CUDA runtime (cuda_cudart).

==> cuda_cudart-linux-x86_64-11.5.50-archive/
lib/libOpenCL.so -> libOpenCL.so.1
lib/libOpenCL.so.1 -> libOpenCL.so.1.0
lib/libOpenCL.so.1.0 -> libOpenCL.so.1.0.0
lib/libOpenCL.so.1.0.0
==> nvidia_driver-linux-x86_64-495.29.05-archive/
lib/libnvidia-opencl.so.495.29.05
lib/libOpenCL.so.1.0.0

versus (in this case the file is pruned from the .deb/.rpm NVIDIA driver packaging)

==> cuda-cudart-dev-11-5_11.5.50-1_amd64.deb
/usr/local/cuda-11.5/targets/x86_64-linux/lib/libOpenCL.so -> libOpenCL.so.1
==> cuda-cudart-11-5_11.5.50-1_amd64.deb
/usr/local/cuda-11.5/targets/x86_64-linux/lib/libOpenCL.so.1 -> libOpenCL.so.1.0
/usr/local/cuda-11.5/targets/x86_64-linux/lib/libOpenCL.so.1.0 -> libOpenCL.so.1.0.0
/usr/local/cuda-11.5/targets/x86_64-linux/lib/libOpenCL.so.1.0.0
==> libnvidia-compute-495_495.29.05-0ubuntu1_amd64.deb
/usr/lib/x86_64-linux-gnu/libnvidia-opencl.so.1 -> libnvidia-opencl.so.495.29.05
/usr/lib/x86_64-linux-gnu/libnvidia-opencl.so.495.29.05

These tarballs are not installers and the file shuffling exercises are left up to the package maintainer.