yade-dev team mailing list archive

[Anton Gladky <gladky.anton@xxxxxxxxx>]

Hi all,

sorry for the noise. This change is reverted.

Anton


2014-05-16 11:48 GMT+02:00 Anton Gladky <gladky.anton@xxxxxxxxx>:
> Dear users/devs, who is working with clumps,
>
> I have done a tiny modification in NewtonIntegrator [1], which
> should fix numerical instability with clumps in a rare cases.
>
> ============================================
> ============================================
> Subject: [PATCH] Do not use getTorqueUnsynced and forces.getForceUnsynced in
>  clumps
>
> It seems both of these commands are unsafe. It causes NaN velocities
> in a very specific cases. From my point of view it is OK, because
> addForceTorqueFromMembers does not modify data itself.
> ---
>  pkg/dem/NewtonIntegrator.cpp | 6 ------
>  1 file changed, 6 deletions(-)
>
> diff --git a/pkg/dem/NewtonIntegrator.cpp b/pkg/dem/NewtonIntegrator.cpp
> index e961bf2..082add8 100644
> --- a/pkg/dem/NewtonIntegrator.cpp
> +++ b/pkg/dem/NewtonIntegrator.cpp
> @@ -146,14 +146,8 @@ void NewtonIntegrator::action()
>   // clumps forces
>   if(b->isClump()) {
>   b->shape->cast<Clump>().addForceTorqueFromMembers(state,scene,f,m);
> - #ifdef YADE_OPENMP
> - //it is safe here, since only one thread will read/write
> - scene->forces.getTorqueUnsynced(id)+=m;
> - scene->forces.getForceUnsynced(id)+=f;
> - #else
>   scene->forces.addTorque(id,m);
>   scene->forces.addForce(id,f);
> - #endif
>   }
>   //in most cases, the initial force on clumps will be zero and next
> line is not changing f and m, but make sure we don't miss something
> (e.g. user defined forces on clumps)
>   f=scene->forces.getForce(id); m=scene->forces.getTorque(id);
> --
> 1.9.3
> ============================================
> ============================================
>
> Please, test you scripts with the newest version. Yadedaily should
> get this fix tonight.
>
> Comments are very welcome.
>
> [1] https://github.com/yade/trunk/commit/5707bb55fb9a3605e10a82cdecbec8efcc7d906f
>
> Best regards
>
> Anton