Re: Gadget 4 Single/double precision performance

From: Volker Springel <vspringel_at_MPA-Garching.MPG.DE>
Date: Fri, 4 Dec 2020 15:31:41 +0100

Hi Tiago,

> On 3. Dec 2020, at 18:58, Tiago Castro <tiagobscastro_at_gmail.com> wrote:
>
> Hi Volker,
>
> Many thanks. In fact, this was my intention, thanks for noticing my mistake. Re-running the code I noticed that I made another mistake compiling the code with EXTRA_HIGH_EWALD_ACCURACY, PMGRID, and FMM. I understood that for the PM-FMM gravity solver Ewald corrections were not needed and was expecting the sims with and without EXTRA_HIGH_EWALD_ACCURACY to be the same at machine precision. This is not the case, have I misunderstood something? Many thanks for the help.

Note that Ewald corrections may still be used even if you have PMGRID, provided you use TREEPM_NOTIMESPLIT and you have set ActivePartFracForPMinsteadOfEwald to a number larger than 0.

Regardless of this, binary invariance of the code's results between runs with the same settings, and the same MPI rank number and number of compute nodes, can in general only be expected if you activate PRESERVE_SHMEM_BINARY_INVARIANCE

Regards,
Volker
Received on 2020-12-04 15:31:42