Skip to content

CUDA for Free Energy Dynamics

Viewing 2 posts - 1 through 2 (of 2 total)
  • Author
    Posts
  • #7431

    Hi OpenLB developers,

    Would you please advise whether OpenLB 1.6 supports GPU usage via CUDA for Free Energy dynamics?

    Thanks,
    Mehrdad

    #7432
    Adrian
    Keymaster

    The local dynamics used for the Free Energy model are adapted to the new style introduced in 1.5 so they will work on GPUs [1]. The only remaining obstacle is that the post processors used for the free energy boundary conditions are not yet adapted to the operator style. We’ll get around to it as we continue closing the remaining holes in our GPU support but this is mostly done alongside our active research projects as we require. So I can not give you a clear date on when this will become available.

    If you want I can guide you in adapting the boundary conditions by yourself (this is mostly refactoring, i.e. extracting the per-cell logic to an generic apply functions and explicitly declaring all parameters). e.g. changing

    
    void FreeEnergyWallProcessor3D<T,DESCRIPTOR>::
    processSubDomain(BlockLattice<T,DESCRIPTOR>& blockLattice, int x0_, int x1_, int y0_, int y1_, int z0_, int z1_)
    {
      int newX0, newX1, newY0, newY1, newZ0, newZ1;
      if ( util::intersect (
             x0, x1, y0, y1, z0, z1,
             x0_, x1_, y0_, y1_, z0_, z1_,
             newX0, newX1, newY0, newY1, newZ0, newZ1 ) ) {
    
        for (int iX=newX0; iX<=newX1; ++iX) {
          for (int iY=newY0; iY<=newY1; ++iY) {
            for (int iZ=newZ0; iZ<=newZ1; ++iZ) {
              T rhoBulk = blockLattice.get(iX-discreteNormalX, iY-discreteNormalY, iZ-discreteNormalZ).computeRho();
              T rhoTmp = 0;
              for (int iPop = 1; iPop < DESCRIPTOR::q; ++iPop) {
                rhoTmp += blockLattice.get(iX,iY,iZ)[iPop];
              }
              T rho = rhoBulk + addend;
              blockLattice.get(iX,iY,iZ)[0] = rho - rhoTmp - 1;
            }
          }
        }
      }
    }
    

    to roughly

    
    template <typename T, typename DESCRIPTOR, int normalX, int normalY, int normalZ>
    struct FreeEnergyWallProcessor3D {
    
    static constexpr OperatorScope scope = OperatorScope::PerCellWithParameters;
    
    using parameters = meta::list<ADDEND>();
    
    template <typename CELL, typename PARAMETERS>
    void  apply(CELL& cell, PARAMETERS& params) any_platform {
       T rhoBulk = cell.neighbor(normalX, normalY, normalZ).computeRho();
       T rhoTmp = 0;
       for (int iPop = 1; iPop < DESCRIPTOR::q; ++iPop) {
         rhoTmp += cell[iPop];
        }
        T rho = rhoBulk + params.template get<ADDEND>();
        cell[0] = rho - rhoTmp - 1;
    }
    
    };
    

    [1]: I just checked, we’ll need to add the any_platform macro to the methods in src/dynamics/freeEnergyDynamics.h

Viewing 2 posts - 1 through 2 (of 2 total)
  • You must be logged in to reply to this topic.