CUDA Driver dependency

[bentsherman]

KINC v3.3.0 is the first version to have CUDA support via ACE. One issue that has emerged as a result is that KINC (and ACE) require the NVIDIA drivers to be installed on the host machine, even if the user doesn't want to use GPUs.

Furthermore, we can't get around this requirement with the Docker image because the NVIDIA Docker images (upon which the KINC image is based) also require the NVIDIA drivers to be installed on the host machine. So even on containerized environments like k8s, it seems that KINC can only be used on machines with the NVIDIA drivers installed.

I'm honestly not sure if we can really get around this at all. Our paradigm of an application that can decide at runtime to use CPUs or GPUs seems to be outside the norm. But here are a couple of ideas to look into later:

• Use the CUDA runtime API instead of the driver API, because I think the runtime API doesn't require the NVIDIA drivers in the same way. This option isn't really viable though because I think I had to use the driver API in ACE in order to use NVRTC.
• Link the CUDA driver API statically to KINC, which might allow KINC to run without linking dynamically with NVIDIA drivers. It would just make the kinc binary a bit larger.
• See if the NVIDIA drivers can be installed easily as a package (especially Ubuntu 18.04). This might at least allow the Docker image to work on CPU nodes.

[bentsherman]

The Dockerfile has been updated using the third option from above. It works, although one shortcoming is that KINC will still fail if either the cuda or opencl settings are set to a device but none is present. So now if you run the docker image without the NVIDIA runtime, you must explicitly set these settings to none.

[bentsherman]

There are now two Docker images for KINC 3.3.0, 3.3.0-cpu and 3.3.0-gpu, so you can use either one depending on what you need.