device-selection.rst

CUDA device selection

DEME runs a dynamic worker (dT) and a kinematic/contact-detection worker (kT). Their logical CUDA devices are selected when DEMSolver is constructed.

Count-based selection

import deme

solver = deme.DEMSolver(1)  # dT and kT both use logical device 0

DEMSolver() is equivalent to DEMSolver(2). When at least two devices are visible, it assigns dT to device 0 and kT to device 1. If only one is visible, DEME warns and assigns both workers to device 0. Only counts 1 and 2 are accepted.

Explicit logical device IDs

Use a list to choose devices rather than relying on the first visible devices:

.. literalinclude:: examples/device_selection.py
   :language: python
   :linenos:

A one-element list assigns both workers to that device. In a two-element list, the first ID is assigned to dT and the second to kT. Repeated IDs are valid, so DEMSolver([2, 2]) places both workers on logical device 2.

GetGPUDeviceIDs() returns [dT_device, kT_device] and is useful for logging or choosing a destination device for direct data retrieval.

Logical versus physical IDs

IDs are CUDA logical device IDs as seen by this process. Environment-level masking and reordering such as CUDA_VISIBLE_DEVICES therefore affect the numbering. For example:

CUDA_VISIBLE_DEVICES=3,1 python my_simulation.py

Inside that process, logical device 0 denotes physical device 3 and logical device 1 denotes physical device 1. An explicit ID must be non-negative and smaller than the number of visible devices; invalid IDs fail during construction.

The constructor allocates CUDA resources. Select visibility before starting Python, and construct the solver only after the intended CUDA environment is in place.