Detailed description |
|
The existing L-BFGS minimizer works very well, but is quite slow for CUDA and OpenCL due to the need to push positions and pull forces over the host bus every step. It would be great to have a minimizer that ran exclusively on the GPU and was perhaps "good enough" for many tasks, but faster.
I've implemented a very rudimentary CustomIntegrator based version as an example, but more sophisticated possibilities exist, and a custom kernel based approach could be superior.
This is also submitted to the issue tracker in GitHub, where a code example is also provided:
https://github.com/SimTk/openmm/issues/334 |
|