GPU¶

class hoomd.device.GPU(communicator=None, message_filename=None, notice_level=2, gpu_id=None)¶

Bases: Device

Select a GPU to execute simulations.

Parameters:

communicator (hoomd.communicator.Communicator) – MPI communicator object. When None, create a default communicator that uses all MPI ranks.
message_filename (str) – Filename to write messages to. When None, use sys.stdout and sys.stderr. Messages from multiple MPI ranks are collected into this file.
notice_level (int) – Minimum level of messages to print.
gpu_id (int) – GPU id to use. Set to None to let the driver auto-select a GPU.

Tip

Call GPU.get_available_devices to get a human readable list of devices. gpu_id = 0 will select the first device in this list, 1 will select the second, and so on.

The ordering of the devices is determined by the GPU driver and runtime.

Device auto-selection

When gpu_id is None, HOOMD will ask the GPU driver to auto-select a GPU. In most cases, this will select device 0. When all devices are set to a compute exclusive mode, the driver will choose a free GPU.

MPI

In MPI execution environments, create a GPU device on every rank. When gpu_id is left None, HOOMD will attempt to detect the MPI local rank environment and choose an appropriate GPU with id = local_rank % num_capable_gpus. Set notice_level to 3 to see status messages from this process. Override this auto-selection by providing appropriate device ids on each rank.

Note

Not all features are optimized to use this code path, and it requires that all GPUs support concurrent managed memory access and have high bandwidth interconnects.

Example:

gpu = hoomd.device.GPU()

Members inherited from Device:

communicator¶: The MPI Communicator. Read more...

property notice_level¶: Minimum level of messages to print. Read more...

property message_filename¶: Filename to write messages to. Read more...

property device¶: Descriptions of the active hardware device. Read more...

notice()¶: Write a notice message. Read more...

Members defined in GPU:

property compute_capability¶

Compute capability of the device.

The tuple includes the major and minor versions of the CUDA compute capability: (major, minor).

Type:: tuple(int, int)

enable_profiling()¶

Enable GPU profiling.

When using GPU profiling tools on HOOMD, select the option to disable profiling on start. Initialize and run a simulation long enough that all autotuners have completed, then open enable_profiling() as a context manager and continue the simulation for a time. Profiling stops when the context manager closes.

Example:

simulation = hoomd.util.make_example_simulation(device=gpu)
with gpu.enable_profiling():
    simulation.run(1000)

static get_available_devices()¶

Get the available GPU devices.

Returns:: Descriptions of the available devices (if any).
Return type:: list[str]

static get_unavailable_device_reasons()¶

Get messages describing the reasons why devices are unavailable.

Returns:: Messages indicating why some devices are unavailable (if any).
Return type:: list[str]

property gpu_error_checking¶

Whether to check for GPU error conditions after every call.

When False (the default), error messages from the GPU may not be noticed immediately. Set to True to increase the accuracy of the GPU error messages at the cost of significantly reduced performance.

Example:

gpu.gpu_error_checking = True

Type:: bool

static is_available()¶

Test if the GPU device is available.

Returns:: True if this build of HOOMD supports GPUs, False if not.
Return type:: bool