FUnifTensorialKernel.hpp 13.5 KB
Newer Older
1
// ===================================================================================
2 3
// Copyright ScalFmm 2016 INRIA
//
4 5
// This software is a computer program whose purpose is to compute the FMM.
//
6
// This software is governed by Mozilla Public License Version 2.0 (MPL 2.0) and
7 8 9 10 11
// 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
12 13
// Mozilla Public License Version 2.0 (MPL 2.0) for more details.
// https://www.mozilla.org/en-US/MPL/2.0/
14
// ===================================================================================
15 16
// Keep in private GIT

17 18 19 20
#ifndef FUNIFTENSORIALKERNEL_HPP
#define FUNIFTENSORIALKERNEL_HPP

#include "../../Utils/FGlobal.hpp"
21

22 23 24
#include "../../Utils/FSmartPointer.hpp"

#include "./FAbstractUnifKernel.hpp"
25 26
#include "./FUnifM2LHandler.hpp"
#include "./FUnifTensorialM2LHandler.hpp" //PB: temporary version
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41

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 !!!
 *
42 43
 * 1) Handling tensorial kernels (DIM,NRHS,NLHS) and having multiple rhs 
 * (NVALS) are considered 2 distinct features and are currently combined.
44
 *
45 46 47 48 49 50
 * 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.
51
 *
52 53 54
 * 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
55 56 57 58 59 60
 *
 * @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
 */
61
template < class FReal, class CellClass, class ContainerClass,   class MatrixKernelClass, int ORDER, int NVALS = 1>
62
class FUnifTensorialKernel
63
    : public FAbstractUnifKernel<FReal, CellClass, ContainerClass, MatrixKernelClass, ORDER, NVALS>
64
{
65 66 67 68
    enum {nRhs = MatrixKernelClass::NRHS,
          nLhs = MatrixKernelClass::NLHS,
          nPot = MatrixKernelClass::NPOT,
          nPV = MatrixKernelClass::NPV};
69 70 71

protected://PB: for OptiDis

72
    // private types
73
    typedef FUnifTensorialM2LHandler<FReal, ORDER,MatrixKernelClass,MatrixKernelClass::Type> M2LHandlerClass;
74

75
    // using from
76
    typedef FAbstractUnifKernel< FReal, CellClass, ContainerClass, MatrixKernelClass, ORDER, NVALS>
77
    AbstractBaseClass;
78

79 80
    /// Needed for P2P and M2L operators
    const MatrixKernelClass *const MatrixKernel;
81

82 83
    /// Needed for M2L operator
    const M2LHandlerClass M2LHandler;
84

85 86 87
    /// Leaf level separation criterion
    const int LeafLevelSeparationCriterion;

88
public:
89 90 91 92 93 94 95
    /**
     * 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,
                         const FReal inBoxWidth,
96
                         const FPoint<FReal>& inBoxCenter,
97
                         const MatrixKernelClass *const inMatrixKernel,
98 99
                         const FReal inBoxWidthExtension,
                         const int inLeafLevelSeparationCriterion = 1)
100
    : FAbstractUnifKernel< FReal, CellClass, ContainerClass, MatrixKernelClass, ORDER, NVALS>(inTreeHeight,inBoxWidth,inBoxCenter,inBoxWidthExtension),
101 102
      MatrixKernel(inMatrixKernel),
      M2LHandler(MatrixKernel,
103
                 inTreeHeight,
104
                 inBoxWidth,
105 106 107
                 inBoxWidthExtension,
                 inLeafLevelSeparationCriterion), 
      LeafLevelSeparationCriterion(inLeafLevelSeparationCriterion)
108 109 110 111 112 113
    { }


    void P2M(CellClass* const LeafCell,
             const ContainerClass* const SourceParticles)
    {
114
        const FPoint<FReal> LeafCellCenter(AbstractBaseClass::getLeafCellCenter(LeafCell->getCoordinate()));
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
        const FReal ExtendedLeafCellWidth(AbstractBaseClass::BoxWidthLeaf 
                                          + AbstractBaseClass::BoxWidthExtension);

        for(int idxV = 0 ; idxV < NVALS ; ++idxV){

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

            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));
131

132
            }
133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
        }// NVALS
    }


    void M2M(CellClass* const FRestrict ParentCell,
             const CellClass*const FRestrict *const FRestrict ChildCells,
             const int TreeLevel)
    {
        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), 
                                                                  TreeLevel/*Cell width extension specific*/);
                    }
                }
                // 2) Apply Discete Fourier Transform
                M2LHandler.applyZeroPaddingAndDFT(ParentCell->getMultipole(idxMul), 
                                                  ParentCell->getTransformedMultipole(idxMul));
            }
        }// NVALS
    }


164 165
    void M2L(CellClass* const FRestrict TargetCell, const CellClass* SourceCells[],
             const int neighborPositions[], const int inSize, const int TreeLevel)  override {
166 167 168 169 170 171 172 173 174 175 176
        const FReal CellWidth(AbstractBaseClass::BoxWidth / FReal(FMath::pow(2, TreeLevel)));
        const FReal ExtendedCellWidth(CellWidth + AbstractBaseClass::BoxWidthExtension);
        const FReal scale(MatrixKernel->getScaleFactor(ExtendedCellWidth));

        for(int idxV = 0 ; idxV < NVALS ; ++idxV){
            for (int idxLhs=0; idxLhs < nLhs; ++idxLhs){

                // update local index
                const int idxLoc = idxV*nLhs + idxLhs;

                // load transformed local expansion
177
                FComplex<FReal> *const TransformedLocalExpansion = TargetCell->getTransformedLocal(idxLoc);
178 179 180 181 182 183 184 185 186 187

                // update idxRhs
                const int idxRhs = idxLhs % nPV; 

                // update multipole index
                const int idxMul = idxV*nRhs + idxRhs;

                // get index in matrix kernel
                const unsigned int d = MatrixKernel->getPosition(idxLhs);

188 189
                for(int idxExistingNeigh = 0 ; idxExistingNeigh < inSize ; ++idxExistingNeigh){
                    const int idx = neighborPositions[idxExistingNeigh];
190

191 192 193
                    M2LHandler.applyFC(idx, TreeLevel, scale, d,
                                       SourceCells[idxExistingNeigh]->getTransformedMultipole(idxMul),
                                       TransformedLocalExpansion);
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226

                }
            }// NLHS=NPOT*NPV
        }// NVALS
    }


    void L2L(const CellClass* const FRestrict ParentCell,
             CellClass* FRestrict *const FRestrict ChildCells,
             const int TreeLevel)
    {
        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),
                                                                  TreeLevel/*Cell width extension specific*/);
                    }
                }
            }
        }// NVALS
    }

    void L2P(const CellClass* const LeafCell,
             ContainerClass* const TargetParticles)
    {
227
        const FPoint<FReal> LeafCellCenter(AbstractBaseClass::getLeafCellCenter(LeafCell->getCoordinate()));
228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250
        const FReal ExtendedLeafCellWidth(AbstractBaseClass::BoxWidthLeaf 
                                          + AbstractBaseClass::BoxWidthExtension);

        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));

            }

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

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

        }// NVALS
    }

251
    void P2P(const FTreeCoordinate& inPosition,
252
             ContainerClass* const FRestrict inTargets, const ContainerClass* const FRestrict inSources,
253 254
             ContainerClass* const inNeighbors[], const int neighborPositions[],
             const int inSize) override {
255 256 257 258 259 260 261 262 263 264 265
        // Standard FMM separation criterion, i.e. max 27 neighbor clusters per leaf
        if(LeafLevelSeparationCriterion==1) {
            if(inTargets == inSources){
                P2POuter(inPosition, inTargets, inNeighbors, neighborPositions, inSize);
                DirectInteractionComputer<FReal, MatrixKernelClass::NCMP, NVALS>::P2PInner(inTargets,MatrixKernel);
            }
            else{
                const ContainerClass* const srcPtr[1] = {inSources};
                DirectInteractionComputer<FReal, MatrixKernelClass::NCMP, NVALS>::P2PRemote(inTargets,srcPtr,1,MatrixKernel);
                DirectInteractionComputer<FReal, MatrixKernelClass::NCMP, NVALS>::P2PRemote(inTargets,inNeighbors,inSize,MatrixKernel);
            }
266
        }
267 268 269
        // Nearfield interactions are only computed within the target leaf
        else if(LeafLevelSeparationCriterion==0){
            DirectInteractionComputer<FReal,MatrixKernelClass::NCMP, NVALS>::P2PRemote(inTargets,inNeighbors,inSize,MatrixKernel);
270
        }
271
        // If criterion equals -1 then no P2P need to be performed.
272 273 274 275 276 277
    }

    void P2POuter(const FTreeCoordinate& /*inLeafPosition*/,
             ContainerClass* const FRestrict inTargets,
             ContainerClass* const inNeighbors[], const int neighborPositions[],
             const int inSize) override {
278 279 280 281 282 283
        int nbNeighborsToCompute = 0;
        while(nbNeighborsToCompute < inSize
              && neighborPositions[nbNeighborsToCompute] < 14){
            nbNeighborsToCompute += 1;
        }
        DirectInteractionComputer<FReal, MatrixKernelClass::NCMP, NVALS>::P2P(inTargets,inNeighbors,nbNeighborsToCompute,MatrixKernel);
284 285 286 287 288
    }


    void P2PRemote(const FTreeCoordinate& /*inPosition*/,
                   ContainerClass* const FRestrict inTargets, const ContainerClass* const FRestrict /*inSources*/,
289
                   const ContainerClass* const inNeighbors[], const int /*neighborPositions*/[],
290
                   const int inSize) override {
291 292 293 294 295 296 297
        // Standard FMM separation criterion, i.e. max 27 neighbor clusters per leaf
        if(LeafLevelSeparationCriterion==1) 
            DirectInteractionComputer<FReal, MatrixKernelClass::NCMP, NVALS>::P2PRemote(inTargets,inNeighbors,inSize,MatrixKernel);
        // Nearfield interactions are only computed within the target leaf
        if(LeafLevelSeparationCriterion==0) 
            DirectInteractionComputer<FReal, MatrixKernelClass::NCMP, NVALS>::P2PRemote(inTargets,inNeighbors,0,MatrixKernel);
        // If criterion equals -1 then no P2P need to be performed.        
298
    }
299 300 301 302

};


303
#endif //FUNIFTENSORIALKERNEL_HPP
304 305

// [--END--]