FChebSymM2LHandler.hpp 28.7 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

COULAUD Olivier's avatar
COULAUD Olivier committed
24
#include "FChebTensor.hpp"
25
#include "../Interpolation/FInterpSymmetries.hpp"
COULAUD Olivier's avatar
COULAUD Olivier committed
26
#include "FChebM2LHandler.hpp"
27 28 29 30 31 32 33

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


Matthias Messner's avatar
Matthias Messner committed
34
/*!  Choose either \a FULLY_PIVOTED_ACASVD or \a PARTIALLY_PIVOTED_ACASVD or
35
    \a ONLY_SVD.
COULAUD Olivier's avatar
COULAUD Olivier committed
36
 */
37
//#define ONLY_SVD
Matthias Messner's avatar
Matthias Messner committed
38
//#define FULLY_PIVOTED_ACASVD
39
#define PARTIALLY_PIVOTED_ACASVD
Matthias Messner's avatar
Matthias Messner committed
40 41


Matthias Messner's avatar
Matthias Messner committed
42 43


Matthias Messner's avatar
Matthias Messner committed
44
/*!  The fully pivoted adaptive cross approximation (fACA) compresses a
45 46 47 48 49 50 51 52 53 54 55 56
    far-field interaction as \f$K\sim UV^\top\f$. The fACA requires all entries
    to be computed beforehand, then the compression follows in
    \f$\mathcal{O}(2\ell^3k)\f$ operations based on the required accuracy
    \f$\varepsilon\f$. The matrix K will be destroyed as a result.

    @param[in] K far-field to be approximated
    @param[in] nx number of rows
    @param[in] ny number of cols
    @param[in] eps prescribed accuracy
    @param[out] U matrix containing \a k column vectors
    @param[out] V matrix containing \a k row vectors
    @param[out] k final low-rank depends on prescribed accuracy \a eps
COULAUD Olivier's avatar
COULAUD Olivier committed
57
 */
58
template <class FReal>
59
void fACA(FReal *const K,
60 61
          const unsigned int nx, const unsigned int ny,
          const double eps, FReal* &U, FReal* &V, unsigned int &k)
Matthias Messner's avatar
Matthias Messner committed
62
{
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136
    // control vectors (true if not used, false if used)
    bool *const r = new bool[nx];
    bool *const c = new bool[ny];
    for (unsigned int i=0; i<nx; ++i) r[i] = true;
    for (unsigned int j=0; j<ny; ++j) c[j] = true;

    // compute Frobenius norm of original Matrix K
    FReal norm2K = 0;
    for (unsigned int j=0; j<ny; ++j) {
        const FReal *const colK = K + j*nx;
        norm2K += FBlas::scpr(nx, colK, colK);
    }

    // initialize rank k and UV'
    k = 0;
    const unsigned int maxk = (nx + ny) / 2;
    U = new FReal[nx * maxk];
    V = new FReal[ny * maxk];
    FBlas::setzero(nx*maxk, U);
    FBlas::setzero(ny*maxk, V);
    FReal norm2R;

    ////////////////////////////////////////////////
    // start fully pivoted ACA
    do {

        // find max(K) and argmax(K)
        FReal maxK = 0.;
        unsigned int pi=0, pj=0;
        for (unsigned int j=0; j<ny; ++j)
            if (c[j]) {
                const FReal *const colK = K + j*nx;
                for (unsigned int i=0; i<nx; ++i)
                    if (r[i] && maxK < FMath::Abs(colK[i])) {
                        maxK = FMath::Abs(colK[i]);
                        pi = i; 
                        pj = j;
                    }
            }

        // copy pivot cross into U and V
        FReal *const colU = U + k*nx;
        FReal *const colV = V + k*ny;
        const FReal pivot = K[pj*nx + pi];
        for (unsigned int i=0; i<nx; ++i) if (r[i]) colU[i] = K[pj*nx + i];
        for (unsigned int j=0; j<ny; ++j) if (c[j]) colV[j] = K[j *nx + pi] / pivot;

        // don't use these cols and rows anymore
        c[pj] = false;
        r[pi] = false;

        // subtract k-th outer product from K
        for (unsigned int j=0; j<ny; ++j)
            if (c[j]) {
                FReal *const colK = K + j*nx;
                FBlas::axpy(nx, FReal(-1. * colV[j]), colU, colK);
            }

        // compute Frobenius norm of updated K
        norm2R = 0.0;
        for (unsigned int j=0; j<ny; ++j)
            if (c[j]) {
                const FReal *const colK = K + j*nx;
                norm2R += FBlas::scpr(nx, colK, colK);
            }

        // increment rank k
        ++k ;

    } while (norm2R > eps*eps * norm2K);
    ////////////////////////////////////////////////

    delete [] r;
    delete [] c;
Matthias Messner's avatar
Matthias Messner committed
137 138 139
}


Matthias Messner's avatar
Matthias Messner committed
140 141 142 143




144 145 146 147 148





Matthias Messner's avatar
Matthias Messner committed
149
/*!  The partially pivoted adaptive cross approximation (pACA) compresses a
150 151 152 153 154 155 156 157 158 159 160 161 162 163
    far-field interaction as \f$K\sim UV^\top\f$. The pACA computes the matrix
    entries on the fly, as they are needed. The compression follows in
    \f$\mathcal{O}(2\ell^3k)\f$ operations based on the required accuracy
    \f$\varepsilon\f$. The matrix K will be destroyed as a result.

    @tparam ComputerType the functor type which allows to compute matrix entries

    @param[in] Computer the entry-computer functor
    @param[in] eps prescribed accuracy
    @param[in] nx number of rows
    @param[in] ny number of cols
    @param[out] U matrix containing \a k column vectors
    @param[out] V matrix containing \a k row vectors
    @param[out] k final low-rank depends on prescribed accuracy \a eps
COULAUD Olivier's avatar
COULAUD Olivier committed
164
 */
165
template <class FReal, typename ComputerType>
166
void pACA(const ComputerType& Computer,
167 168
        const unsigned int nx, const unsigned int ny,
        const FReal eps, FReal* &U, FReal* &V, unsigned int &k)
Matthias Messner's avatar
Matthias Messner committed
169
{
170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
    // control vectors (true if not used, false if used)
    bool *const r = new bool[nx];
    bool *const c = new bool[ny];
    for (unsigned int i=0; i<nx; ++i) r[i] = true;
    for (unsigned int j=0; j<ny; ++j) c[j] = true;

    // initialize rank k and UV'
    k = 0;
    const FReal eps2 = eps * eps;
    const unsigned int maxk = (nx + ny) / 2;
    U = new FReal[nx * maxk];
    V = new FReal[ny * maxk];

    // initialize norm
    FReal norm2S(0.);
    FReal norm2uv(0.);

    ////////////////////////////////////////////////
    // start partially pivoted ACA
    unsigned int J = 0, I = 0;

    do {
        FReal *const colU = U + nx*k;
        FReal *const colV = V + ny*k;

        ////////////////////////////////////////////
        // compute row I and its residual
        Computer(I, I+1, 0, ny, colV);
        r[I] = false;
        for (unsigned int l=0; l<k; ++l) {
            FReal *const u = U + nx*l;
            FReal *const v = V + ny*l;
            FBlas::axpy(ny, FReal(-1. * u[I]), v, colV);
        }

        // find max of residual and argmax
        FReal maxval = 0.;
        for (unsigned int j=0; j<ny; ++j) {
            const FReal abs_val = FMath::Abs(colV[j]);
            if (c[j] && maxval < abs_val) {
                maxval = abs_val;
                J = j;
            }
        }
        // find pivot and scale column of V
        const FReal pivot = FReal(1.) / colV[J];
        FBlas::scal(ny, pivot, colV);

        ////////////////////////////////////////////
        // compute col J and its residual
        Computer(0, nx, J, J+1, colU);
        c[J] = false;
        for (unsigned int l=0; l<k; ++l) {
            FReal *const u = U + nx*l;
            FReal *const v = V + ny*l;
            FBlas::axpy(nx, FReal(-1. * v[J]), u, colU);
        }

        // find max of residual and argmax
        maxval = 0.0;
        for (unsigned int i=0; i<nx; ++i) {
            const FReal abs_val = FMath::Abs(colU[i]);
            if (r[i] && maxval < abs_val) {
                maxval = abs_val;
                I = i;
            }
        }

        ////////////////////////////////////////////
        // increment Frobenius norm: |Sk|^2 += |uk|^2 |vk|^2 + 2 sumj ukuj vjvk
        FReal normuuvv(0.);
        for (unsigned int l=0; l<k; ++l)
            normuuvv += FBlas::scpr(nx, colU, U + nx*l) * FBlas::scpr(ny, V + ny*l, colV);
        norm2uv = FBlas::scpr(nx, colU, colU) * FBlas::scpr(ny, colV, colV);
        norm2S += norm2uv + 2*normuuvv;

        ////////////////////////////////////////////
        // increment low-rank
        ++k;

    } while (norm2uv > eps2 * norm2S);

    delete [] r;
    delete [] c;
Matthias Messner's avatar
Matthias Messner committed
254 255 256 257
}



Matthias Messner's avatar
Matthias Messner committed
258 259 260 261 262
/*!  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. */
263
template <class FReal, int ORDER, typename MatrixKernelClass>
264
static void precompute(const MatrixKernelClass *const MatrixKernel, const FReal CellWidth,
265
        const FReal Epsilon, FReal* K[343], int LowRank[343])
266
{
267 268
    //  std::cout << "\nComputing 16 far-field interactions (l=" << ORDER << ", eps=" << Epsilon
    //                      << ") for cells of width w = " << CellWidth << std::endl;
269

270
    static const unsigned int nnodes = ORDER*ORDER*ORDER;
271

272
    // interpolation points of source (Y) and target (X) cell
273
    FPoint<FReal> X[nnodes], Y[nnodes];
274
    // set roots of target cell (X)
275
    FChebTensor<FReal, ORDER>::setRoots(FPoint<FReal>(0.,0.,0.), CellWidth, X);
276 277
    // temporary matrix
    FReal* U = new FReal [nnodes*nnodes];
278

279 280 281 282 283 284
    // 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
285 286


287 288 289 290
    // initialize timer
    FTic time;
    double overall_time(0.);
    double elapsed_time(0.);
291

292 293
    // initialize rank counter
    unsigned int overall_rank = 0;
Matthias Messner's avatar
Matthias Messner committed
294

295 296 297 298
    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) {
299

300
                // assemble matrix and apply weighting matrices
301 302
                const FPoint<FReal> cy(CellWidth*FReal(i), CellWidth*FReal(j), CellWidth*FReal(k));
                FChebTensor<FReal, ORDER>::setRoots(cy, CellWidth, Y);
303
                FReal weights[nnodes];
304
                FChebTensor<FReal, ORDER>::setRootOfWeights(weights);
305

306
                // now the entry-computer is responsible for weighting the matrix entries
307
                EntryComputer<FReal, MatrixKernelClass> Computer(MatrixKernel, nnodes, X, nnodes, Y, weights);
Matthias Messner's avatar
Matthias Messner committed
308

309 310
                // start timer
                time.tic();
Matthias Messner's avatar
Matthias Messner committed
311 312

#if (defined ONLY_SVD || defined FULLY_PIVOTED_ACASVD)
313
                Computer(0, nnodes, 0, nnodes, U);
Matthias Messner's avatar
Matthias Messner committed
314
#endif
315 316 317
                /*
                // applying weights ////////////////////////////////////////
                FReal weights[nnodes];
318
                FChebTensor<FReal,ORDER>::setRootOfWeights(weights);
319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
                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
336
#if (defined FULLY_PIVOTED_ACASVD || defined PARTIALLY_PIVOTED_ACASVD) ////////////
337 338
                FReal *UU, *VV;
                unsigned int rank;
Matthias Messner's avatar
Matthias Messner committed
339 340

#ifdef FULLY_PIVOTED_ACASVD
341
                fACA(U,        nnodes, nnodes, Epsilon, UU, VV, rank);
Matthias Messner's avatar
Matthias Messner committed
342
#else
343
                pACA(Computer, nnodes, nnodes, Epsilon, UU, VV, rank);
Matthias Messner's avatar
Matthias Messner committed
344
#endif 
Matthias Messner's avatar
Matthias Messner committed
345

346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438
                // 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
439 440

#elif defined ONLY_SVD
441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465
                // 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
466 467
#else
#error Either fully-, partially pivoted ACA or only SVD must be defined!
Matthias Messner's avatar
Matthias Messner committed
468
#endif ///////////////////////////////////////////////////////////////
469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504
                //////////////////////////////////////////////////////////////


                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                //////////////////////////////////////////////////////////////
                // 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;
505 506 507 508 509 510 511 512 513 514
}









Matthias Messner's avatar
Matthias Messner committed
515
/*!  \class SymmetryHandler 
516

517
    \brief Deals with all the symmetries in the arrangement of the far-field interactions
Matthias Messner's avatar
Matthias Messner committed
518

519
    Stores permutation indices and permutation vectors to reduce 316 (7^3-3^3)
Matthias Messner's avatar
Matthias Messner committed
520 521 522 523
  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. */
524
template <class FReal, int ORDER, KERNEL_FUNCTION_TYPE TYPE> class SymmetryHandler;
525 526

/*! Specialization for homogeneous kernel functions */
527 528
template <class FReal, int ORDER>
class SymmetryHandler<FReal, ORDER, HOMOGENEOUS>
529
{
530
    static const unsigned int nnodes = ORDER*ORDER*ORDER;
531

532 533 534
    // M2L operators
    FReal*    K[343];
    int LowRank[343];
535 536

public:
COULAUD Olivier's avatar
COULAUD Olivier committed
537

538 539 540
    // permutation vectors and permutated indices
    unsigned int pvectors[343][nnodes];
    unsigned int pindices[343];
541 542


543 544 545 546 547 548 549 550 551 552
    /** 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
553

554 555 556 557 558 559 560 561 562 563
        // 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]);
                }
564

565 566
        // precompute 16 M2L operators
        const FReal ReferenceCellWidth = FReal(2.0);
567
        precompute<FReal, ORDER>(MatrixKernel, ReferenceCellWidth, Epsilon, K, LowRank);
568
    }
569 570 571



572 573 574 575 576
    /** Destructor */
    ~SymmetryHandler()
    {
        for (unsigned int t=0; t<343; ++t) if (K[t]!=nullptr) delete [] K[t];
    }
577

578

579
    /*! return the t-th approximated far-field interactions*/
580
    const FReal * getK(const  int, const unsigned int t) const
581
    {   return K[t]; }
582

583
    /*! return the t-th approximated far-field interactions*/
584
    int getLowRank(const int, const unsigned int t) const
585
    {   return LowRank[t]; }
586 587 588 589 590 591 592 593 594

};






/*! Specialization for non-homogeneous kernel functions */
595 596
template <class FReal, int ORDER>
class SymmetryHandler<FReal, ORDER, NON_HOMOGENEOUS>
597
{
598
    static const unsigned int nnodes = ORDER*ORDER*ORDER;
599

600 601
    // Height of octree; needed only in the case of non-homogeneous kernel functions
    const unsigned int TreeHeight;
602

603 604 605
    // M2L operators for all levels in the octree
    FReal***    K;
    int** LowRank;
606 607

public:
COULAUD Olivier's avatar
COULAUD Olivier committed
608

609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671
    // 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*/
672
    const FReal * getK(const  int l, const unsigned int t) const
673 674 675
    {   return K[l][t]; }

    /*! return the t-th approximated far-field interactions*/
676
    int getLowRank(const  int l, const unsigned int t) const
677
    {   return LowRank[l][t]; }
678

679 680 681 682 683 684 685 686 687 688 689 690 691
};








#include <fstream>
#include <sstream>


Matthias Messner's avatar
Matthias Messner committed
692 693 694
/**
 * Computes, compresses and stores the 16 M2L kernels in a binary file.
 */
695
template <class FReal, int ORDER, typename MatrixKernelClass>
696 697
static void ComputeAndCompressAndStoreInBinaryFile(const MatrixKernelClass *const MatrixKernel, const FReal Epsilon)
{
698 699 700 701 702 703
    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;  }
704
    precompute<FReal,ORDER>(MatrixKernel, FReal(2.), Epsilon, K, LowRank);
705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737

    // 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];
738 739 740
}


Matthias Messner's avatar
Matthias Messner committed
741 742 743 744
/**
 * Reads the 16 compressed M2L kernels from the binary files and writes them
 * in K and the respective low-rank in LowRank.
 */
745
template <class FReal, int ORDER>
746 747
void ReadFromBinaryFile(const FReal Epsilon, FReal* K[343], int LowRank[343])
{
748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790
    // 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();
791 792 793 794 795 796 797
}





#endif