FUnifTensorialKernel.hpp 9.82 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
// ===================================================================================
// Copyright ScalFmm 2011 INRIA, Olivier Coulaud, Bérenger Bramas, Matthias Messner
// olivier.coulaud@inria.fr, berenger.bramas@inria.fr
// This software is a computer program whose purpose is to compute the FMM.
//
// This software is governed by the CeCILL-C and LGPL licenses and
// abiding by the rules of distribution of free software.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
// GNU General Public and CeCILL-C Licenses for more details.
// "http://www.cecill.info".
// "http://www.gnu.org/licenses".
// ===================================================================================
#ifndef FUNIFTENSORIALKERNEL_HPP
#define FUNIFTENSORIALKERNEL_HPP

#include "../../Utils/FGlobal.hpp"
#include "../../Utils/FTrace.hpp"
#include "../../Utils/FSmartPointer.hpp"

#include "./FAbstractUnifKernel.hpp"
24 25
#include "./FUnifM2LHandler.hpp"
#include "./FUnifTensorialM2LHandler.hpp" //PB: temporary version
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

class FTreeCoordinate;

/**
 * @author Pierre Blanchard (pierre.blanchard@inria.fr)
 * @class FUnifTensorialKernel
 * @brief
 * Please read the license
 *
 * This kernels implement the Lagrange interpolation based FMM operators. It
 * implements all interfaces (P2P,P2M,M2M,M2L,L2L,L2P) which are required by
 * the FFmmAlgorithm and FFmmAlgorithmThread.
 *
 * PB: 3 IMPORTANT remarks !!!
 *
41 42
 * 1) Handling tensorial kernels (DIM,NRHS,NLHS) and having multiple rhs 
 * (NVALS) are considered 2 distinct features and are currently combined.
43
 *
44 45 46 47 48 49
 * 2) When it comes to applying M2L it is NOT much faster to loop over 
 * NRHSxNLHS inside applyM2L (at least for the Lagrange case).
 * 2-bis) During precomputation the tensorial matrix kernels are evaluated 
 * blockwise, but this is not always possible. 
 * In fact, in the ChebyshevSym variant the matrix kernel needs to be 
 * evaluated compo-by-compo since we currently use a scalar ACA.
50
 *
51 52 53
 * 3) We currently use multiple 1D FFT instead of multidim FFT since embedding
 * is circulant. Multidim FFT could be used if embedding were block circulant.
 * TODO investigate possibility of block circulant embedding
54 55 56 57 58 59 60 61 62 63
 *
 * @tparam CellClass Type of cell
 * @tparam ContainerClass Type of container to store particles
 * @tparam MatrixKernelClass Type of matrix kernel function
 * @tparam ORDER Lagrange interpolation order
 */
template < class CellClass,	class ContainerClass,	class MatrixKernelClass, int ORDER, int NVALS = 1>
class FUnifTensorialKernel
  : public FAbstractUnifKernel< CellClass, ContainerClass, MatrixKernelClass, ORDER, NVALS>
{
64
  enum {nRhs = MatrixKernelClass::NRHS,
65
        nLhs = MatrixKernelClass::NLHS};
66 67 68

protected://PB: for OptiDis

69
  // private types
70
  typedef FUnifTensorialM2LHandler<ORDER,MatrixKernelClass,MatrixKernelClass::Type> M2LHandlerClass;
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85

  // using from
  typedef FAbstractUnifKernel< CellClass, ContainerClass, MatrixKernelClass, ORDER, NVALS>
  AbstractBaseClass;

  /// Needed for M2L operator
  FSmartPointer<  M2LHandlerClass,FSmartPointerMemory> M2LHandler;

public:
  /**
   * The constructor initializes all constant attributes and it reads the
   * precomputed and compressed M2L operators from a binary file (an
   * runtime_error is thrown if the required file is not valid).
   */
  FUnifTensorialKernel(const int inTreeHeight,
86
                       const FReal inBoxWidth,
87 88 89 90 91 92 93
                       const FPoint& inBoxCenter,
                       const double inMatParam = 0.0)
    : FAbstractUnifKernel< CellClass, ContainerClass, MatrixKernelClass, ORDER, NVALS>(inTreeHeight,inBoxWidth,inBoxCenter,inMatParam),
      M2LHandler(new M2LHandlerClass(AbstractBaseClass::MatrixKernel.getPtr(),
                                     inTreeHeight,
                                     inBoxWidth))// PB: for non homogeneous case
  { }
94 95 96 97 98 99


  void P2M(CellClass* const LeafCell,
           const ContainerClass* const SourceParticles)
  {
    const FPoint LeafCellCenter(AbstractBaseClass::getLeafCellCenter(LeafCell->getCoordinate())); 
100

101
    for(int idxV = 0 ; idxV < NVALS ; ++idxV){
102 103 104 105 106

      // 1) apply Sy
      AbstractBaseClass::Interpolator->applyP2M(LeafCellCenter, AbstractBaseClass::BoxWidthLeaf,
                                                LeafCell->getMultipole(idxV*nRhs), SourceParticles);

107 108 109 110 111 112 113 114 115 116
      for(int idxRhs = 0 ; idxRhs < nRhs ; ++idxRhs){
        // update multipole index
        int idxMul = idxV*nRhs + idxRhs;

        // 2) apply Discrete Fourier Transform
        M2LHandler->applyZeroPaddingAndDFT(LeafCell->getMultipole(idxMul), 
                                           LeafCell->getTransformedMultipole(idxMul));

      }
    }// NVALS
117 118 119 120 121 122 123
  }


  void M2M(CellClass* const FRestrict ParentCell,
           const CellClass*const FRestrict *const FRestrict ChildCells,
           const int /*TreeLevel*/)
  {
124 125 126 127 128 129 130 131 132 133 134 135
    for(int idxV = 0 ; idxV < NVALS ; ++idxV){
      for(int idxRhs = 0 ; idxRhs < nRhs ; ++idxRhs){
        // update multipole index
        int idxMul = idxV*nRhs + idxRhs;

        // 1) apply Sy
        FBlas::scal(AbstractBaseClass::nnodes, FReal(0.), ParentCell->getMultipole(idxMul));
        for (unsigned int ChildIndex=0; ChildIndex < 8; ++ChildIndex){
          if (ChildCells[ChildIndex]){
            AbstractBaseClass::Interpolator->applyM2M(ChildIndex, ChildCells[ChildIndex]->getMultipole(idxMul),
                                                      ParentCell->getMultipole(idxMul));
          }
136
        }
137 138 139
        // 2) Apply Discete Fourier Transform
        M2LHandler->applyZeroPaddingAndDFT(ParentCell->getMultipole(idxMul), 
                                           ParentCell->getTransformedMultipole(idxMul));
140
      }
141
    }// NVALS
142 143 144 145 146 147 148 149
  }


  void M2L(CellClass* const FRestrict TargetCell,
           const CellClass* SourceCells[343],
           const int /*NumSourceCells*/,
           const int TreeLevel)
  {
150
    const FReal CellWidth(AbstractBaseClass::BoxWidth / FReal(FMath::pow(2, TreeLevel)));
151
    const FReal scale(AbstractBaseClass::MatrixKernel.getPtr()->getScaleFactor(CellWidth));
152 153 154 155 156 157 158 159 160 161 162 163 164

    for(int idxV = 0 ; idxV < NVALS ; ++idxV){
      for (int idxLhs=0; idxLhs < nLhs; ++idxLhs){
        // update local index
        int idxLoc = idxV*nLhs + idxLhs;
        // load transformed local expansion
        FComplexe *const TransformedLocalExpansion = TargetCell->getTransformedLocal(idxLoc);

        for (int idxRhs=0; idxRhs < nRhs; ++idxRhs){
          // update multipole index
          int idxMul = idxV*nRhs + idxRhs;
          // update kernel index such that: x_i = K_{ij}y_j 
          int idxK = idxLhs*nRhs + idxRhs;
165 166 167
          // get index in matrix kernel
          unsigned int d 
            = AbstractBaseClass::MatrixKernel.getPtr()->getPosition(idxK);
168 169 170

          for (int idx=0; idx<343; ++idx){
            if (SourceCells[idx]){
171
              M2LHandler->applyFC(idx, TreeLevel, scale, d,
172
                                  SourceCells[idx]->getTransformedMultipole(idxMul),
173
                                  TransformedLocalExpansion);
174 175

            }
176
          }
177 178 179
        }// NRHS
      }// NLHS
    }// NVALS
180 181 182 183 184 185 186
  }


  void L2L(const CellClass* const FRestrict ParentCell,
           CellClass* FRestrict *const FRestrict ChildCells,
           const int /*TreeLevel*/)
  {
187 188 189 190 191 192 193 194 195 196 197
    for(int idxV = 0 ; idxV < NVALS ; ++idxV){
      for(int idxLhs = 0 ; idxLhs < nLhs ; ++idxLhs){
        int idxLoc = idxV*nLhs + idxLhs;
        // 1) Apply Inverse Discete Fourier Transform
        M2LHandler->unapplyZeroPaddingAndDFT(ParentCell->getTransformedLocal(idxLoc),
                                             const_cast<CellClass*>(ParentCell)->getLocal(idxLoc));
        // 2) apply Sx
        for (unsigned int ChildIndex=0; ChildIndex < 8; ++ChildIndex){
          if (ChildCells[ChildIndex]){
            AbstractBaseClass::Interpolator->applyL2L(ChildIndex, ParentCell->getLocal(idxLoc), ChildCells[ChildIndex]->getLocal(idxLoc));
          }
198 199
        }
      }
200
    }// NVALS
201 202 203 204 205 206 207
  }

  void L2P(const CellClass* const LeafCell,
           ContainerClass* const TargetParticles)
  {
    const FPoint LeafCellCenter(AbstractBaseClass::getLeafCellCenter(LeafCell->getCoordinate()));

208 209 210 211 212 213
    for(int idxV = 0 ; idxV < NVALS ; ++idxV){
      for(int idxLhs = 0 ; idxLhs < nLhs ; ++idxLhs){
        int idxLoc = idxV*nLhs + idxLhs;
        // 1)  Apply Inverse Discete Fourier Transform
        M2LHandler->unapplyZeroPaddingAndDFT(LeafCell->getTransformedLocal(idxLoc), 
                                             const_cast<CellClass*>(LeafCell)->getLocal(idxLoc));
214

215 216 217 218 219
      }

      // 2.a) apply Sx
      AbstractBaseClass::Interpolator->applyL2P(LeafCellCenter, AbstractBaseClass::BoxWidthLeaf,
                                                LeafCell->getLocal(idxV*nLhs), TargetParticles);
220

221 222 223
      // 2.b) apply Px (grad Sx)
      AbstractBaseClass::Interpolator->applyL2PGradient(LeafCellCenter, AbstractBaseClass::BoxWidthLeaf,
                                                        LeafCell->getLocal(idxV*nLhs), TargetParticles);
224

225
    }// NVALS
226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
  }

  void P2P(const FTreeCoordinate& /* LeafCellCoordinate */, // needed for periodic boundary conditions
           ContainerClass* const FRestrict TargetParticles,
           const ContainerClass* const FRestrict /*SourceParticles*/,
           ContainerClass* const NeighborSourceParticles[27],
           const int /* size */)
  {
    DirectInteractionComputer<MatrixKernelClass::Identifier, NVALS>::P2P(TargetParticles,NeighborSourceParticles);
  }


  void P2PRemote(const FTreeCoordinate& /*inPosition*/,
                 ContainerClass* const FRestrict inTargets, const ContainerClass* const FRestrict /*inSources*/,
                 ContainerClass* const inNeighbors[27], const int /*inSize*/){
    DirectInteractionComputer<MatrixKernelClass::Identifier, NVALS>::P2PRemote(inTargets,inNeighbors,27);
  }

};


247
#endif //FUNIFTENSORIALKERNEL_HPP
248 249

// [--END--]