include_run.yml

parameterset:
  # The runtime options
  - name: run_set
    parameter:
      # Number of MPI ranks to launch (or iterate over NUMA domains)
      - name: nproc
        type: int
        mode: python
        _: |  # Powers of two until total number of available cores
          ','.join(
            [
              str(2**i)
              for i in range(int($numa_domains*$cores_per_numa_domain).bit_length()-1)
            ] + [str($numa_domains*$cores_per_numa_domain)]
            if $mpi == 1 else [
              str(i+1) for i in range($numa_domains)
            ]
          )

      # Number of OpenMP threads to use (per rank)
      - name: numomp
        type: int
        mode: python
        _: "$numa_domains*$cores_per_numa_domain//$nproc"

      # Number of horizontal columns
      - {name: ngptotg, type: int, _: 163840}

      # Default NPROMA value per variant
      - name: default_nproma
        mode: python
        type: int
        _: >
          {
            'dwarf-cloudsc-fortran': {'sp': 32, 'dp': 16}['$precision'],
            'dwarf-cloudsc-c': {'sp': 32, 'dp': 16}['$precision'],
            'dwarf-cloudsc-gpu-claw': 8192,
            'dwarf-cloudsc-gpu-scc': 256,
            'dwarf-cloudsc-gpu-scc-hoist': 128,
            'dwarf-cloudsc-gpu-omp-scc-hoist': 128,
          }['$target']

      # Optionally sweep neighbouring NPROMA values
      - name: nproma
        tag: "sweep_nproma"
        type: int
        mode: python
        _: "','.join(str(int($default_nproma * 2**i)) for i in range(-2, 3))"
      - {name: nproma, tag: "!sweep_nproma", _: $default_nproma}

      # Executable(s) to benchmark
      - name: target
        _: "dwarf-cloudsc-fortran"