[backend-cuda]

Parameterises the CUDA backend with

  1. device-id — method for selecting which device(s) to run on:

    int | round-robin | local-rank | uuid

  2. mpi-type — type of MPI library that is being used:

    standard | cuda-aware

  3. cflags — additional NVIDIA realtime compiler (nvrtc) flags:

    string

  4. cublas-nkerns — number of kernel algorithms to try when benchmarking, defaults to 512:

    int

  5. gimmik-nkerns — number of kernel algorithms to try when benchmarking, defaults to 8:

    int

  6. gimmik-nbench — number of benchmarking runs for each kernel, defaults to 5:

    int

Example:

[backend-cuda]
device-id = round-robin
mpi-type = standard