GPU support

[mikegerber]

We current run the workflow in a GPU-less container. From what I have tested, this could get GPU support by:

• basing on a CUDA image, e.g. FROM nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04, and
• running docker run --gpus all

It should still work on CPU-only systems.

[mikegerber]

8eba477 introduced rudimentary support. This could be a bit cleaner though, i.e. no duplication in Dockerfile-core*

[mikegerber]

Now the question is instrumentation, i.e running the containers with docker run --gpus=all.

• Run with --gpus=all if possible
  • (On CPU-only systems this fails, so we need a way to detect the availability of nvidia-container-runtime)
• Use CUDA_VISIBLE_DEVICES to support disabling running on GPU
  • We need that sometimes to make use of the CPU ressources available

TL&DR: Use GPU by default if possible, use CPU if told by the user

[mikegerber]

Related: https://stackoverflow.com/questions/59008295/add-nvidia-runtime-to-docker-runtimes

[mikegerber]

https://github.com/NVIDIA/nvidia-container-runtime#nvidia_visible_devices