Skip to content

Remove duplicate wait with GPU PME and GPU Update

Andrey Alekseenko requested to merge aa-3988-remove-duplicate-wait-with-gpu-pme-and-gpu-update into master

With GPU PME on a separate rank and GPU update, we had duplicate waitCoordinatesReadyOnHost in do_force.

First, we were waiting for X on host before sending X to PME nodes in https://gitlab.com/gromacs/gromacs/-/blob/a125c30d34300506974b2a0202aafaf249681dc4/src/gromacs/mdlib/sim_util.cpp#L1469

Then we did a second wait before a bunch of host tasks: https://gitlab.com/gromacs/gromacs/-/blob/a125c30d34300506974b2a0202aafaf249681dc4/src/gromacs/mdlib/sim_util.cpp#L1812

I also added a new StepWorload flag to indicate that we have active work on a separate PME rank this step, to get rid of repeated simulationWork.haveSeparatePmeRank && stepWork.computeSlowForces.

Refs #3988.

Merge request reports