NekRS: GPU Computation in Single or Double Precision?

[ztdepztdep]

I am running NekRS on a GPU and would like to know whether the computations are performed in single precision (FP32) or double precision (FP64) by default.

Does NekRS always use double precision on GPUs, or is there an option to switch to single precision?
If FP32 is supported, how can I enable it?
Are there any performance considerations when choosing between FP32 and FP64, especially for different GPU architectures (e.g., NVIDIA vs. AMD GPUs)?
Any insights or recommendations would be greatly appreciated!

[yslan]

By default, nekrs runs with double precision for the solvers (dfloat = double) but it uses single precision in preconditioners (pfloat = float).

The current repo/next branch does support single precision for the whole solver.
It's built into an independent executable:

$NEKRS_HOME/bin/nekrs
$NEKRS_HOME/bin/nekrs-fp32

You can use it by setting FP32=1 with the script like this and it should print something into the logfile.

FP32=1 $NEKRS_HOME/bin/nrsmpi turbPipe 1

[stgeke]

@ztdepztdep Can we close this discussion?