FChebSymM2LHandler.hpp 22 KB
Newer Older
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1
// ===================================================================================
2
// Copyright ScalFmm 2011 INRIA
3 4 5 6 7 8 9 10 11 12 13 14
// 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".
BRAMAS Berenger's avatar
BRAMAS Berenger committed
15
// ===================================================================================
16 17 18 19
#ifndef FCHEBSYMM2LHANDLER_HPP
#define FCHEBSYMM2LHANDLER_HPP

#include <climits>
COULAUD Olivier's avatar
COULAUD Olivier committed
20
#include <sstream>
21

COULAUD Olivier's avatar
COULAUD Olivier committed
22
#include "Utils/FBlas.hpp"
23

24 25
#include "Utils/FAca.hpp"

COULAUD Olivier's avatar
COULAUD Olivier committed
26
#include "FChebTensor.hpp"
27
#include "../Interpolation/FInterpSymmetries.hpp"
COULAUD Olivier's avatar
COULAUD Olivier committed
28
#include "FChebM2LHandler.hpp"
29 30 31 32 33 34 35

/**
 * @author Matthias Messner (matthias.matthias@inria.fr)
 * Please read the license
 */


Matthias Messner's avatar
Matthias Messner committed
36
/*!  Choose either \a FULLY_PIVOTED_ACASVD or \a PARTIALLY_PIVOTED_ACASVD or
37
    \a ONLY_SVD.
COULAUD Olivier's avatar
COULAUD Olivier committed
38
 */
39
//#define ONLY_SVD
Matthias Messner's avatar
Matthias Messner committed
40
//#define FULLY_PIVOTED_ACASVD
41
#define PARTIALLY_PIVOTED_ACASVD
Matthias Messner's avatar
Matthias Messner committed
42 43


Matthias Messner's avatar
Matthias Messner committed
44

Matthias Messner's avatar
Matthias Messner committed
45 46 47 48 49
/*!  Precomputes the 16 far-field interactions (due to symmetries in their
  arrangement all 316 far-field interactions can be represented by
  permutations of the 16 we compute in this function). Depending on whether
  FACASVD is defined or not, either ACA+SVD or only SVD is used to compress
  them. */
50
template <class FReal, int ORDER, typename MatrixKernelClass>
51
static void precompute(const MatrixKernelClass *const MatrixKernel, const FReal CellWidth,
52
        const FReal Epsilon, FReal* K[343], int LowRank[343])
53
{
54 55
    //  std::cout << "\nComputing 16 far-field interactions (l=" << ORDER << ", eps=" << Epsilon
    //                      << ") for cells of width w = " << CellWidth << std::endl;
56

57
    static const unsigned int nnodes = ORDER*ORDER*ORDER;
58

59
    // interpolation points of source (Y) and target (X) cell
60
    FPoint<FReal> X[nnodes], Y[nnodes];
61
    // set roots of target cell (X)
62
    FChebTensor<FReal, ORDER>::setRoots(FPoint<FReal>(0.,0.,0.), CellWidth, X);
63 64
    // temporary matrix
    FReal* U = new FReal [nnodes*nnodes];
65

66 67 68 69 70 71
    // needed for the SVD
     int INFO;
    const unsigned int LWORK = 2 * (3*nnodes + nnodes);
    FReal *const WORK = new FReal [LWORK];
    FReal *const VT = new FReal [nnodes*nnodes];
    FReal *const S = new FReal [nnodes];
Matthias Messner's avatar
Matthias Messner committed
72 73


74 75 76 77
    // initialize timer
    FTic time;
    double overall_time(0.);
    double elapsed_time(0.);
78

79 80
    // initialize rank counter
    unsigned int overall_rank = 0;
Matthias Messner's avatar
Matthias Messner committed
81

82 83 84 85
    unsigned int counter = 0;
    for (int i=2; i<=3; ++i) {
        for (int j=0; j<=i; ++j) {
            for (int k=0; k<=j; ++k) {
86

87
                // assemble matrix and apply weighting matrices
88 89
                const FPoint<FReal> cy(CellWidth*FReal(i), CellWidth*FReal(j), CellWidth*FReal(k));
                FChebTensor<FReal, ORDER>::setRoots(cy, CellWidth, Y);
90
                FReal weights[nnodes];
91
                FChebTensor<FReal, ORDER>::setRootOfWeights(weights);
92

93
                // now the entry-computer is responsible for weighting the matrix entries
94
                EntryComputer<FReal, MatrixKernelClass> Computer(MatrixKernel, nnodes, X, nnodes, Y, weights);
Matthias Messner's avatar
Matthias Messner committed
95

96 97
                // start timer
                time.tic();
Matthias Messner's avatar
Matthias Messner committed
98 99

#if (defined ONLY_SVD || defined FULLY_PIVOTED_ACASVD)
100
                Computer(0, nnodes, 0, nnodes, U);
Matthias Messner's avatar
Matthias Messner committed
101
#endif
102 103 104
                /*
                // applying weights ////////////////////////////////////////
                FReal weights[nnodes];
105
                FChebTensor<FReal,ORDER>::setRootOfWeights(weights);
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
                for (unsigned int n=0; n<nnodes; ++n) {
                    FBlas::scal(nnodes, weights[n], U + n,  nnodes); // scale rows
                    FBlas::scal(nnodes, weights[n], U + n * nnodes); // scale cols
                }
                 */

                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                // ALL PREPROC FLAGS ARE SET ON TOP OF THIS FILE !!! /////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////



                //////////////////////////////////////////////////////////////
Matthias Messner's avatar
Matthias Messner committed
123
#if (defined FULLY_PIVOTED_ACASVD || defined PARTIALLY_PIVOTED_ACASVD) ////////////
124 125
                FReal *UU, *VV;
                unsigned int rank;
Matthias Messner's avatar
Matthias Messner committed
126 127

#ifdef FULLY_PIVOTED_ACASVD
128
                FAca::pACA(U,        nnodes, nnodes, Epsilon, UU, VV, rank);
Matthias Messner's avatar
Matthias Messner committed
129
#else
130
                FAca::fACA(Computer, nnodes, nnodes, Epsilon, UU, VV, rank);
Matthias Messner's avatar
Matthias Messner committed
131
#endif 
Matthias Messner's avatar
Matthias Messner committed
132

133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
                // QR decomposition
                FReal* phi = new FReal [rank*rank];
                {
                    // QR of U and V
                    FReal* tauU = new FReal [rank];
                    INFO = FBlas::geqrf(nnodes, rank, UU, tauU, LWORK, WORK);
                    assert(INFO==0);
                    FReal* tauV = new FReal [rank];
                    INFO = FBlas::geqrf(nnodes, rank, VV, tauV, LWORK, WORK);
                    assert(INFO==0);
                    // phi = Ru Rv'
                    FReal* rU = new FReal [2 * rank*rank];
                    FReal* rV = rU + rank*rank;
                    FBlas::setzero(2 * rank*rank, rU);
                    for (unsigned int l=0; l<rank; ++l) {
                        FBlas::copy(l+1, UU + l*nnodes, rU + l*rank);
                        FBlas::copy(l+1, VV + l*nnodes, rV + l*rank);
                    }
                    FBlas::gemmt(rank, rank, rank, FReal(1.), rU, rank, rV, rank, phi, rank);
                    delete [] rU;
                    // get Qu and Qv
                    INFO = FBlas::orgqr(nnodes, rank, UU, tauU, LWORK, WORK);
                    assert(INFO==0);
                    INFO = FBlas::orgqr(nnodes, rank, VV, tauV, LWORK, WORK);
                    assert(INFO==0);
                    delete [] tauU;
                    delete [] tauV;
                }

                const unsigned int aca_rank = rank;

                // SVD
                {
                    INFO = FBlas::gesvd(aca_rank, aca_rank, phi, S, VT, aca_rank, LWORK, WORK);
                    if (INFO!=0){
                        std::stringstream stream;
                        stream << INFO;
                        throw std::runtime_error("SVD did not converge with " + stream.str());
                    }
                    rank = getRank(S, aca_rank, Epsilon);
                }                   

                const unsigned int idx = (i+3)*7*7 + (j+3)*7 + (k+3);

                // store
                {
                    // allocate
                    assert(K[idx]==nullptr);
                    K[idx] = new FReal [2*rank*nnodes];

                    // set low rank
                    LowRank[idx] = static_cast<int>(rank);

                    // (U Sigma)
                    for (unsigned int r=0; r<rank; ++r)
                        FBlas::scal(aca_rank, S[r], phi + r*aca_rank);

                    // Qu (U Sigma) 
                    FBlas::gemm(nnodes, aca_rank, rank, FReal(1.), UU, nnodes, phi, aca_rank, K[idx], nnodes);
                    delete [] phi;

                    // Vt -> V and then Qu V
                    FReal *const V = new FReal [aca_rank * rank];
                    for (unsigned int r=0; r<rank; ++r)
                        FBlas::copy(aca_rank, VT + r, aca_rank, V + r*aca_rank, 1);
                    FBlas::gemm(nnodes, aca_rank, rank, FReal(1.), VV, nnodes, V, aca_rank, K[idx] + rank*nnodes, nnodes);
                    delete [] V;
                }

                //// store recompressed UV
                //const unsigned int idx = (i+3)*7*7 + (j+3)*7 + (k+3);
                //assert(K[idx]==NULL);
                //K[idx] = new FReal [2*rank*nnodes];
                //LowRank[idx] = rank;
                //FBlas::copy(rank*nnodes, UU,  K[idx]);
                //FBlas::copy(rank*nnodes, VV,  K[idx] + rank*nnodes);

                delete [] UU;
                delete [] VV;

                elapsed_time = time.tacAndElapsed(); 
                overall_time += elapsed_time;
                overall_rank += rank;
                // std::cout << "(" << i << "," << j << "," << k << ") " << idx <<
                //  ", low rank = " << rank << " (" << aca_rank << ") in " << elapsed_time << "s" << std::endl;

                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                // ALL PREPROC FLAGS ARE SET ON TOP OF THIS FILE !!! /////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
Matthias Messner's avatar
Matthias Messner committed
226 227

#elif defined ONLY_SVD
228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252
                // truncated singular value decomposition of matrix
                INFO = FBlas::gesvd(nnodes, nnodes, U, S, VT, nnodes, LWORK, WORK);
                if (INFO!=0){
                    std::stringstream stream;
                    stream << INFO;
                    throw std::runtime_error("SVD did not converge with " + stream.str());
                }
                const unsigned int rank = getRank<ORDER>(S, Epsilon);

                // store 
                const unsigned int idx = (i+3)*7*7 + (j+3)*7 + (k+3);
                assert(K[idx]==nullptr);
                K[idx] = new FReal [2*rank*nnodes];
                LowRank[idx] = rank;
                for (unsigned int r=0; r<rank; ++r)
                    FBlas::scal(nnodes, S[r], U + r*nnodes);
                FBlas::copy(rank*nnodes, U,  K[idx]);
                for (unsigned int r=0; r<rank; ++r)
                    FBlas::copy(nnodes, VT + r, nnodes, K[idx] + rank*nnodes + r*nnodes, 1);

                elapsed_time = time.tacAndElapsed(); 
                overall_time += elapsed_time;
                overall_rank += rank;
                //              std::cout << "(" << i << "," << j << "," << k << ") " << idx <<
                //  ", low rank = " << rank << " in " << elapsed_time << "s" << std::endl;
Matthias Messner's avatar
Matthias Messner committed
253 254
#else
#error Either fully-, partially pivoted ACA or only SVD must be defined!
Matthias Messner's avatar
Matthias Messner committed
255
#endif ///////////////////////////////////////////////////////////////
256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291
                //////////////////////////////////////////////////////////////


                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                // ALL PREPROC FLAGS ARE SET ON TOP OF THIS FILE !!! /////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////


                // un-weighting ////////////////////////////////////////////
                for (unsigned int n=0; n<nnodes; ++n) {
                    FBlas::scal(rank, FReal(1.) / weights[n], K[idx] + n,               nnodes); // scale rows
                    FBlas::scal(rank, FReal(1.) / weights[n], K[idx] + rank*nnodes + n, nnodes); // scale rows
                }
                //////////////////////////////////////////////////////////      

                ++counter;
            }
        }
    }
    //std::cout << "The approximation of the " << counter
    //      << " far-field interactions (overall rank " << overall_rank
    //      << " / " << 16*nnodes
    //      << " , sizeM2L= " << 2*overall_rank*nnodes*sizeof(FReal) << ""
    //      << " / " << 16*nnodes*nnodes*sizeof(FReal) << " B"
    //      << ") took " << overall_time << "s\n" << std::endl;

    std::cout << "Compressed and set M2L operators (" << 2*overall_rank*nnodes*sizeof(FReal) << " B) in " << overall_time << "sec." << std::endl;

    delete [] U;
    delete [] WORK;
    delete [] VT;
    delete [] S;
292 293 294 295 296 297 298 299 300 301
}









Matthias Messner's avatar
Matthias Messner committed
302
/*!  \class SymmetryHandler 
303

304
    \brief Deals with all the symmetries in the arrangement of the far-field interactions
Matthias Messner's avatar
Matthias Messner committed
305

306
    Stores permutation indices and permutation vectors to reduce 316 (7^3-3^3)
Matthias Messner's avatar
Matthias Messner committed
307 308 309 310
  different far-field interactions to 16 only. We use the number 343 (7^3)
  because it allows us to use to associate the far-field interactions based on
  the index \f$t = 7^2(i+3) + 7(j+3) + (k+3)\f$ where \f$(i,j,k)\f$ denotes
  the relative position of the source cell to the target cell. */
311
template <class FReal, int ORDER, KERNEL_FUNCTION_TYPE TYPE> class SymmetryHandler;
312 313

/*! Specialization for homogeneous kernel functions */
314 315
template <class FReal, int ORDER>
class SymmetryHandler<FReal, ORDER, HOMOGENEOUS>
316
{
317
    static const unsigned int nnodes = ORDER*ORDER*ORDER;
318

319 320 321
    // M2L operators
    FReal*    K[343];
    int LowRank[343];
322 323

public:
COULAUD Olivier's avatar
COULAUD Olivier committed
324

325 326 327
    // permutation vectors and permutated indices
    unsigned int pvectors[343][nnodes];
    unsigned int pindices[343];
328 329


330 331 332 333 334 335 336 337 338 339
    /** Constructor: with 16 small SVDs */
    template <typename MatrixKernelClass>
    SymmetryHandler(const MatrixKernelClass *const MatrixKernel, const FReal Epsilon,
                    const FReal, const unsigned int)
    {
        // init all 343 item to zero, because effectively only 16 exist
        for (unsigned int t=0; t<343; ++t) {
            K[t]            = nullptr;
            LowRank[t] = 0;
        }
COULAUD Olivier's avatar
COULAUD Olivier committed
340

341 342 343 344 345 346 347 348 349 350
        // set permutation vector and indices
        const FInterpSymmetries<ORDER> Symmetries;
        for (int i=-3; i<=3; ++i)
            for (int j=-3; j<=3; ++j)
                for (int k=-3; k<=3; ++k) {
                    const unsigned int idx = ((i+3) * 7 + (j+3)) * 7 + (k+3);
                    pindices[idx] = 0;
                    if (abs(i)>1 || abs(j)>1 || abs(k)>1)
                        pindices[idx] = Symmetries.getPermutationArrayAndIndex(i,j,k, pvectors[idx]);
                }
351

352 353
        // precompute 16 M2L operators
        const FReal ReferenceCellWidth = FReal(2.0);
354
        precompute<FReal, ORDER>(MatrixKernel, ReferenceCellWidth, Epsilon, K, LowRank);
355
    }
356 357 358



359 360 361 362 363
    /** Destructor */
    ~SymmetryHandler()
    {
        for (unsigned int t=0; t<343; ++t) if (K[t]!=nullptr) delete [] K[t];
    }
364

365

366
    /*! return the t-th approximated far-field interactions*/
367
    const FReal * getK(const  int, const unsigned int t) const
368
    {   return K[t]; }
369

370
    /*! return the t-th approximated far-field interactions*/
371
    int getLowRank(const int, const unsigned int t) const
372
    {   return LowRank[t]; }
373 374 375 376 377 378 379 380 381

};






/*! Specialization for non-homogeneous kernel functions */
382 383
template <class FReal, int ORDER>
class SymmetryHandler<FReal, ORDER, NON_HOMOGENEOUS>
384
{
385
    static const unsigned int nnodes = ORDER*ORDER*ORDER;
386

387 388
    // Height of octree; needed only in the case of non-homogeneous kernel functions
    const unsigned int TreeHeight;
389

390 391 392
    // M2L operators for all levels in the octree
    FReal***    K;
    int** LowRank;
393 394

public:
COULAUD Olivier's avatar
COULAUD Olivier committed
395

396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458
    // permutation vectors and permutated indices
    unsigned int pvectors[343][nnodes];
    unsigned int pindices[343];


    /** Constructor: with 16 small SVDs */
    template <typename MatrixKernelClass>
    SymmetryHandler(const MatrixKernelClass *const MatrixKernel, const double Epsilon,
                    const FReal RootCellWidth, const unsigned int inTreeHeight)
    : TreeHeight(inTreeHeight)
    {
        // init all 343 item to zero, because effectively only 16 exist
        K       = new FReal** [TreeHeight];
        LowRank = new int*    [TreeHeight];
        K[0]       = nullptr; K[1]       = nullptr;
        LowRank[0] = nullptr; LowRank[1] = nullptr;
        for (unsigned int l=2; l<TreeHeight; ++l) {
            K[l]       = new FReal* [343];
            LowRank[l] = new int    [343];
            for (unsigned int t=0; t<343; ++t) {
                K[l][t]       = nullptr;
                LowRank[l][t] = 0;
            }
        }


        // set permutation vector and indices
        const FInterpSymmetries<ORDER> Symmetries;
        for (int i=-3; i<=3; ++i)
            for (int j=-3; j<=3; ++j)
                for (int k=-3; k<=3; ++k) {
                    const unsigned int idx = ((i+3) * 7 + (j+3)) * 7 + (k+3);
                    pindices[idx] = 0;
                    if (abs(i)>1 || abs(j)>1 || abs(k)>1)
                        pindices[idx] = Symmetries.getPermutationArrayAndIndex(i,j,k, pvectors[idx]);
                }

        // precompute 16 M2L operators at all levels having far-field interactions
        FReal CellWidth = RootCellWidth / FReal(2.); // at level 1
        CellWidth /= FReal(2.);                      // at level 2
        for (unsigned int l=2; l<TreeHeight; ++l) {
            precompute<ORDER>(MatrixKernel, CellWidth, Epsilon, K[l], LowRank[l]);
            CellWidth /= FReal(2.);                    // at level l+1 
        }
    }



    /** Destructor */
    ~SymmetryHandler()
    {
        for (unsigned int l=0; l<TreeHeight; ++l) {
            if (K[l]!=nullptr) {
                for (unsigned int t=0; t<343; ++t) if (K[l][t]!=nullptr) delete [] K[l][t];
                delete [] K[l];
            }
            if (LowRank[l]!=nullptr)    delete [] LowRank[l];
        }
        delete [] K;
        delete [] LowRank;
    }

    /*! return the t-th approximated far-field interactions*/
459
    const FReal * getK(const  int l, const unsigned int t) const
460 461 462
    {   return K[l][t]; }

    /*! return the t-th approximated far-field interactions*/
463
    int getLowRank(const  int l, const unsigned int t) const
464
    {   return LowRank[l][t]; }
465

466 467 468 469 470 471 472 473 474 475 476 477 478
};








#include <fstream>
#include <sstream>


Matthias Messner's avatar
Matthias Messner committed
479 480 481
/**
 * Computes, compresses and stores the 16 M2L kernels in a binary file.
 */
482
template <class FReal, int ORDER, typename MatrixKernelClass>
483 484
static void ComputeAndCompressAndStoreInBinaryFile(const MatrixKernelClass *const MatrixKernel, const FReal Epsilon)
{
485 486 487 488 489 490
    static const unsigned int nnodes = ORDER*ORDER*ORDER;

    // compute and compress ////////////
    FReal* K[343];
    int LowRank[343];
    for (unsigned int idx=0; idx<343; ++idx) { K[idx] = nullptr; LowRank[idx] = 0;  }
491
    precompute<FReal,ORDER>(MatrixKernel, FReal(2.), Epsilon, K, LowRank);
492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524

    // write to binary file ////////////
    FTic time; time.tic();
    // start computing process
    const char precision = (typeid(FReal)==typeid(double) ? 'd' : 'f');
    std::stringstream sstream;
    sstream << "sym2l_" << precision << "_o" << ORDER << "_e" << Epsilon << ".bin";
    const std::string filename(sstream.str());
    std::ofstream stream(filename.c_str(),
            std::ios::out | std::ios::binary | std::ios::trunc);
    if (stream.good()) {
        stream.seekp(0);
        for (unsigned int idx=0; idx<343; ++idx)
            if (K[idx]!=nullptr) {
                // 1) write index
                stream.write(reinterpret_cast<char*>(&idx), sizeof(int));
                // 2) write low rank (int)
                int rank = LowRank[idx];
                stream.write(reinterpret_cast<char*>(&rank), sizeof(int));
                // 3) write U and V (both: rank*nnodes * FReal)
                FReal *const U = K[idx];
                FReal *const V = K[idx] + rank*nnodes;
                stream.write(reinterpret_cast<char*>(U), sizeof(FReal)*rank*nnodes);
                stream.write(reinterpret_cast<char*>(V), sizeof(FReal)*rank*nnodes);
            }
    } else throw std::runtime_error("File could not be opened to write");
    stream.close();
    // write info
    //  std::cout << "Compressed M2L operators stored in binary file " << filename
    //                  << " in " << time.tacAndElapsed() << "sec." << std::endl;

    // free memory /////////////////////
    for (unsigned int t=0; t<343; ++t) if (K[t]!=nullptr) delete [] K[t];
525 526 527
}


Matthias Messner's avatar
Matthias Messner committed
528 529 530 531
/**
 * Reads the 16 compressed M2L kernels from the binary files and writes them
 * in K and the respective low-rank in LowRank.
 */
532
template <class FReal, int ORDER>
533 534
void ReadFromBinaryFile(const FReal Epsilon, FReal* K[343], int LowRank[343])
{
535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577
    // compile time constants
    const unsigned int nnodes = ORDER*ORDER*ORDER;

    // find filename
    const char precision = (typeid(FReal)==typeid(double) ? 'd' : 'f');
    std::stringstream sstream;
    sstream << "sym2l_" << precision << "_o" << ORDER << "_e" << Epsilon << ".bin";
    const std::string filename(sstream.str());

    // read binary file
    std::ifstream istream(filename.c_str(),
            std::ios::in | std::ios::binary | std::ios::ate);
    const std::ifstream::pos_type size = istream.tellg();
    if (size<=0) throw std::runtime_error("The requested binary file does not yet exist. Exit.");

    if (istream.good()) {
        istream.seekg(0);
        // 1) read index (int)
        int _idx;
        istream.read(reinterpret_cast<char*>(&_idx), sizeof(int));
        // loop to find 16 compressed m2l operators
        for (int idx=0; idx<343; ++idx) {
            K[idx] = nullptr;
            LowRank[idx] = 0;
            // if it exists
            if (idx == _idx) {
                // 2) read low rank (int)
                int rank;
                istream.read(reinterpret_cast<char*>(&rank), sizeof(int));
                LowRank[idx] = rank;
                // 3) read U and V (both: rank*nnodes * FReal)
                K[idx] = new FReal [2*rank*nnodes];
                FReal *const U = K[idx];
                FReal *const V = K[idx] + rank*nnodes;
                istream.read(reinterpret_cast<char*>(U), sizeof(FReal)*rank*nnodes);
                istream.read(reinterpret_cast<char*>(V), sizeof(FReal)*rank*nnodes);

                // 1) read next index
                istream.read(reinterpret_cast<char*>(&_idx), sizeof(int));
            }
        }
    }   else throw std::runtime_error("File could not be opened to read");
    istream.close();
578 579 580 581 582 583 584
}





#endif