AutotunedObject

class hoomd.operation.AutotunedObject

An object with autotuned kernel parameters.

AutotunedObject instances may complete portions of their computation with one or more GPU kernels. Each GPU kernel is executed with a set of parameters (kernel_parameters) that influence the run time of the execution. After initialization, AutotunedObject varies these parameters as the simulation runs and searches for the best performing combination. The optimal parameters depend on your system’s size, density, force field parameters, the specific hardware you execute on, and more.

Check is_tuning_complete to check if the search is complete. After the search is complete, AutotunedObject holds the parameters constant at the optimal values. Typical operations require thousands of time steps to complete tuning. Some may take tens of thousands or more depending on the parameters you set.

Tip

When you significantly change your system during the simulation (e.g. compress to a higher density), then you can tune the parameters again after with tune_kernel_parameters. This step is optional, but may increase performance as the new system parameters may lead to different optimal parameters.

Note

In MPI parallel execution, all methods and attributes of AutotunedObject reference the rank local tuner objects. Different ranks may tune different optimal kernel parameters and may complete tuning at different times. Use only hoomd.Operations.is_tuning_complete in control flow operations, as this method is reduced across all ranks.

See also

• hoomd.Operations.is_tuning_complete

• hoomd.Operations.tune_kernel_parameters

property is_tuning_complete

Check if kernel parameter tuning is complete.

True when tuning is complete and kernel_parameters has locked optimal parameters for all active kernels used by this object.

Example:

while not operation.is_tuning_complete:
    simulation.run(1000)

Type:

bool

property kernel_parameters

Kernel parameters.

The dictionary maps GPU kernel names to tuples of integers that control how the kernel executes on the GPU. These values will change during the tuning process and remain static after tuning completes. Set the kernel parameters for one or more kernels to force specific values and stop tuning.

Warning

The keys and valid values in this dictionary depend on the hardware device, the HOOMD-blue version, and the value of class attributes. Keys and/or valid values may change even with new patch releases of HOOMD-blue.

Provided that you use the same HOOMD-blue binary on the same hardware and execute a script with the same parameters, you may save the tuned values from one run and load them in the next.

Examples:

kernel_parameters = operation.kernel_parameters

operation.kernel_parameters = kernel_parameters

Type:

dict[str, tuple[float]]

tune_kernel_parameters()

Start tuning kernel parameters.

After calling tune_kernel_parameters, AutotunedObject will vary the kernel parameters in subsequent time steps, check the run time of each, and lock to the fastest performing parameters after the scan is complete.

Example:

operation.tune_kernel_parameters()