GitLab

I think they need to live in their own file, since the main application
has to be compiled with mpicc, not nvcc.

It's 40. This is surprising given that each node has 160 CPUs.

Now the calculation runs for about 2 seconds. Previously it was a
fraction of a second, so GFLOPS between runs varied wildly.

I get 1600 GFLOPS on a single GPU (V100) as expected.

This seems to fix things.

This gives you a unique index within a single node. Will be useful for
assigning GPUs to threads.