[Feature Request] Efficient neighborlist construction/usage plugin for nonbonded force calculation

[TablewareBox]

Summary

Construct neighborlist using non-JITable code (for speed), and use it in downstream nonbonded force calculation.

Motivation

The construction of nblist is the most time-consuming part of DMFF. This is because the code is written in python and JITted, but automatic differentiation is useless in the usage. To optimize the code, we can provide a python/c++/CUDA plugin for nblist construction.

Suggested Solutions

Possible solutions are:

• Code migrated from Gromacs N×N×M clustered neighborlist
• Code migrated from Openmm
• A CUDA kernel developed by @TablewareBox at 2020

Further Information, Files, and Links

No response

[dingye18]

FYI: https://github.com/openmm/NNPOps/tree/master/src/pytorch/neighbors