Closed ocmadin closed 1 year ago
@ocmadin, this should be a small edit to the codebase, and I have played around running evaluator with OpenCL in the past. I'll create a PR for this.
Rather than hard-coding the platform, what about using the fastest available platform that supports the desired precision? We have a utility in openmmtools to help with this.
@jchodera the utility in openmmtools
looks great! I'll incorporate this into Evaluator to automatically determine the best possible platform but also give the user the option to override the default.
A small update - this is in the 0.4.x line via #465. It is included in the 0.4.0rc3
builds, which should be online within the hour. It's also expected to be in the full release if all goes well.
In release 0.4.0 via #465 - thanks Jeff!
We've recently (thanks to @mrshirts) got access to some AMD GPUs which would be useful to run evaluator on. After talking with @mattwthompson, it seems like evaluator's interaction with OpenMM supports OpenCL: https://github.com/openforcefield/openff-evaluator/blob/18962755f7663c37430d2baffb6c9bd80bd0556c/openff/evaluator/utils/openmm.py#L63-L68
However, Evaluator currently does not support using OpenCL when scheduling jobs with
dask
: https://github.com/openforcefield/openff-evaluator/blob/21da54363009d83110b54d57e4416ae31df3868b/openff/evaluator/backends/dask.py#L299-L308Is it possible to add support for using Evaluator with OpenCL? This is not a particularly time-sensitive ask, but would be good to have in the future.