pyscamp.autotune

pyscamp.autotune(devices=None, cache_path='')

Run the SCAMP GPU kernel autotuner for the selected device(s) and persist the chosen kernel configurations to disk. Future pyscamp calls on the same machine will read these configurations from the cache and use them when launching GPU kernels.

A full sweep takes a few minutes on a recent GPU. The output is verbose so you can follow progress; pass cache_path to redirect the write elsewhere (e.g. for a sandboxed run).

Parameters:

  • devices (list[int], optional) – List of CUDA device IDs to tune. If empty (default), only device 0 is tuned – a full sweep takes O(minutes) and most multi-GPU boxes hold identical devices, so sweeping them all wastes wall time on identical configs. Pass devices=[0, 1, ...] explicitly to tune more than one (e.g. if you have two different GPU models).

  • cache_path (str, optional) – Filesystem path to read/write the cache from. Empty (default) resolves in this order: $SCAMP_AUTOTUNE_CACHE (if set, used verbatim), then $XDG_CACHE_HOME/scamp/autotune.txt (if set), then a platform-specific user dir ($HOME/.cache/scamp/autotune.txt on Linux/macOS; %LOCALAPPDATA%\scamp\autotune.txt on Windows). Parent directories are created automatically.

Returns:

Number of devices that were tuned.

Return type:

int

Raises:

  • RuntimeError – If pyscamp was built without CUDA support.

  • ValueError – If no CUDA devices are available.