Jaron T. Krogel
Jaron T. Krogel
Thanks @djstaros. Please also post the scalar.dat and dmc.dat files for the runs described above.
Running with `crowd_serialize_walkers=yes` appears to fix the problem. It also introduces no slowdown in the cpu run. @ye-luo what is the impact of this flag internally? Is there any reason...
Yes, agreed. The multi-walker implementation needs looking into. Beyond that, I propose we need a PR to set the default to "yes" for cpu runs. If there is no performance...
How about we make it default at least until the other code path is fixed? I see no disadvantage to this at all, and there is currently real research time...
This PR should go in. If there are any dangling issues, I will resolve them afterwards.
Thanks for thinking this through. I agree the current implementation is a mess. How much effort do you think this would be? Would this still allow parameter derivatives to be...
I don't think this would change the data dependence. This estimator currently (and still would) depend primarily on the trial wavefunction. The only main difference is requiring parameter derivatives to...
Can we get around the linear search by using hash tables instead? Or is some type of index mapping array better?
Yes, agreed. We should prioritize getting to first functioning over speed.
I will note that two parameter derivatives are needed: dPsi/dp and dE_local/dp.