Closed vyasr closed 2 years ago
I'd be interested in contributing to this.
Given the difficulties with LLVM, I'd like to investigate whether a non-LLVM implementation on the CPU could perform well.
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.
This issue has been automatically closed because it has not had recent activity.
Description
The ability to specify new pair potentials via a JIT kernel rather than having to implement a full new Evaluator.
Proposed solution
The API for this could be similar to the current
jit.patch
. We could also take inspiration from OpenMM, which uses Sympy to accept arbitrary mathematical expressions and JIT-compile them. The hard part will be ensuring that CUDA-JIT (and potentially HIP-JIT) plays nicely with the rest of our infrastructure.Additional context
Josh and I have discussed this possibility before, I'm just creating the issue now so that we can keep track of it.