FChebSymM2LHandler.hpp 23.6 KB
Newer Older
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1 2 3 4 5 6 7 8 9 10
// ===================================================================================
// Logiciel initial: ScalFmm Version 0.5
// Co-auteurs : Olivier Coulaud, Matthias Messner.
// Propriétaires : INRIA.
// Copyright © 2011-2012, diffusé sous les termes et conditions d’une licence propriétaire.
// Initial software: ScalFmm Version 0.5
// Co-authors: Olivier Coulaud, Matthias Messner.
// Owners: INRIA.
// Copyright © 2011-2012, spread under the terms and conditions of a proprietary license.
// ===================================================================================
11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
#ifndef FCHEBSYMM2LHANDLER_HPP
#define FCHEBSYMM2LHANDLER_HPP

#include <climits>

#include "../../Utils/FBlas.hpp"

#include "./FChebTensor.hpp"
#include "./FChebSymmetries.hpp"
#include "./FChebM2LHandler.hpp"

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


Matthias Messner's avatar
Matthias Messner committed
28 29
/*!  Choose either \a FULLY_PIVOTED_ACASVD or \a PARTIALLY_PIVOTED_ACASVD or
	\a ONLY_SVD.
Matthias Messner's avatar
Matthias Messner committed
30
*/
31
//#define ONLY_SVD
Matthias Messner's avatar
Matthias Messner committed
32
//#define FULLY_PIVOTED_ACASVD
33
#define PARTIALLY_PIVOTED_ACASVD
Matthias Messner's avatar
Matthias Messner committed
34 35


Matthias Messner's avatar
Matthias Messner committed
36 37


Matthias Messner's avatar
Matthias Messner committed
38 39 40 41
/*!  The fully pivoted adaptive cross approximation (fACA) compresses a
	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
Matthias Messner's avatar
Matthias Messner committed
42
	\f$\varepsilon\f$. The matrix K will be destroyed as a result.
Matthias Messner's avatar
Matthias Messner committed
43 44

	@param[in] K far-field to be approximated
45 46
	@param[in] nx number of rows
	@param[in] ny number of cols
Matthias Messner's avatar
Matthias Messner committed
47 48 49 50 51
	@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
*/
52 53 54
void fACA(FReal *const K,
					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
55 56
{
	// control vectors (true if not used, false if used)
57 58 59 60
	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;
Matthias Messner's avatar
Matthias Messner committed
61 62 63

	// compute Frobenius norm of original Matrix K
	FReal norm2K = 0;
64 65 66
	for (unsigned int j=0; j<ny; ++j) {
		const FReal *const colK = K + j*nx;
		norm2K += FBlas::scpr(nx, colK, colK);
Matthias Messner's avatar
Matthias Messner committed
67 68 69 70
	}

	// initialize rank k and UV'
	k = 0;
71 72 73 74 75
	const 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);
Matthias Messner's avatar
Matthias Messner committed
76 77 78 79 80 81 82 83 84
	FReal norm2R;

	////////////////////////////////////////////////
	// start fully pivoted ACA
	do {
		
		// find max(K) and argmax(K)
		FReal maxK = 0.;
		int pi=0, pj=0;
85
		for (unsigned int j=0; j<ny; ++j)
Matthias Messner's avatar
Matthias Messner committed
86
			if (c[j]) {
87 88
				const FReal *const colK = K + j*nx;
				for (unsigned int i=0; i<nx; ++i)
Matthias Messner's avatar
Matthias Messner committed
89 90 91 92 93 94 95 96
					if (r[i] && maxK < FMath::Abs(colK[i])) {
						maxK = FMath::Abs(colK[i]);
						pi = i; 
						pj = j;
					}
			}

		// copy pivot cross into U and V
97 98 99 100 101
		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;
Matthias Messner's avatar
Matthias Messner committed
102 103 104 105 106 107
		
		// dont use these cols and rows anymore
		c[pj] = false;
		r[pi] = false;
		
		// subtract k-th outer product from K
108
		for (unsigned int j=0; j<ny; ++j)
Matthias Messner's avatar
Matthias Messner committed
109
			if (c[j]) {
110 111
				FReal *const colK = K + j*nx;
				FBlas::axpy(nx, FReal(-1. * colV[j]), colU, colK);
Matthias Messner's avatar
Matthias Messner committed
112 113 114 115
			}

		// compute Frobenius norm of updated K
		norm2R = 0.;
116
		for (unsigned int j=0; j<ny; ++j)
Matthias Messner's avatar
Matthias Messner committed
117
			if (c[j]) {
118 119
				const FReal *const colK = K + j*nx;
				norm2R += FBlas::scpr(nx, colK, colK);
Matthias Messner's avatar
Matthias Messner committed
120 121 122 123 124 125 126
			}

		// increment rank k
		k++;
		
	} while (norm2R > eps*eps * norm2K);
	////////////////////////////////////////////////
127 128 129

	delete [] r;
	delete [] c;
Matthias Messner's avatar
Matthias Messner committed
130 131 132
}


Matthias Messner's avatar
Matthias Messner committed
133 134 135 136




137 138 139 140 141





Matthias Messner's avatar
Matthias Messner committed
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
/*!  The partially pivoted adaptive cross approximation (pACA) compresses a
	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
*/
template <typename ComputerType>
159
void pACA(const ComputerType& Computer,
Matthias Messner's avatar
Matthias Messner committed
160
					const unsigned int nx, const unsigned int ny,
161
					const FReal eps, FReal* &U, FReal* &V, unsigned int &k)
Matthias Messner's avatar
Matthias Messner committed
162 163 164 165 166 167 168 169 170
{
	// 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;
171
	const FReal eps2 = eps * eps;
Matthias Messner's avatar
Matthias Messner committed
172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194
	const 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;
195
			FBlas::axpy(ny, FReal(-1. * u[I]), v, colV);
Matthias Messner's avatar
Matthias Messner committed
196 197 198 199
		}
		
		// find max of residual and argmax
		FReal maxval = 0.;
200 201 202 203
		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;
Matthias Messner's avatar
Matthias Messner committed
204 205
				J = j;
			}
206 207
		}
		// find pivot and scale column of V
208 209
		const FReal pivot = FReal(1.) / colV[J];
		FBlas::scal(ny, pivot, colV);
Matthias Messner's avatar
Matthias Messner committed
210 211 212 213 214 215 216 217
		
		////////////////////////////////////////////
		// 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;
218
			FBlas::axpy(nx, FReal(-1. * v[J]), u, colU);
Matthias Messner's avatar
Matthias Messner committed
219 220 221 222
		}
		
		// find max of residual and argmax
		maxval = 0.;
223 224 225 226
		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;
Matthias Messner's avatar
Matthias Messner committed
227 228
				I = i;
			}
229
		}
Matthias Messner's avatar
Matthias Messner committed
230 231 232 233 234 235 236 237 238 239 240 241 242
		
		////////////////////////////////////////////
			// 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++;

243
	} while (norm2uv > eps2 * norm2S);
Matthias Messner's avatar
Matthias Messner committed
244 245 246 247 248 249 250
	
	delete [] r;
	delete [] c;
}



Matthias Messner's avatar
Matthias Messner committed
251 252 253 254 255
/*!  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. */
256
template <int ORDER, typename MatrixKernelClass>
257
static void precompute(const MatrixKernelClass *const MatrixKernel, const FReal CellWidth,
258
											 const FReal Epsilon, FReal* K[343], int LowRank[343])
259
{
260 261
	std::cout << "\nComputing 16 far-field interactions (l=" << ORDER << ", eps=" << Epsilon
						<< ") for cells of width w = " << CellWidth << std::endl;
262

263
	static const unsigned int nnodes = ORDER*ORDER*ORDER;
264 265 266 267

	// interpolation points of source (Y) and target (X) cell
	FPoint X[nnodes], Y[nnodes];
	// set roots of target cell (X)
268
	FChebTensor<ORDER>::setRoots(FPoint(0.,0.,0.), CellWidth, X);
269 270 271 272
	// temporary matrix
	FReal* U = new FReal [nnodes*nnodes];

	// needed for the SVD
Matthias Messner's avatar
Matthias Messner committed
273
	unsigned int INFO;
274 275 276 277
	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
278 279 280 281


	// initialize timer
	FTic time;
282 283 284 285 286
	double overall_time(0.);
	double elapsed_time(0.);

	// initialize rank counter
	unsigned int overall_rank = 0;
Matthias Messner's avatar
Matthias Messner committed
287

288 289 290 291 292
	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) {

Matthias Messner's avatar
Matthias Messner committed
293
				// assemble matrix and apply weighting matrices
294 295
				const FPoint cy(CellWidth*FReal(i), CellWidth*FReal(j), CellWidth*FReal(k));
				FChebTensor<ORDER>::setRoots(cy, CellWidth, Y);
Matthias Messner's avatar
Matthias Messner committed
296 297
				FReal weights[nnodes];
				FChebTensor<ORDER>::setRootOfWeights(weights);
298

Matthias Messner's avatar
Matthias Messner committed
299 300 301 302 303 304 305 306 307 308
				// now the entry-computer is responsible for weighting the matrix entries
				EntryComputer<MatrixKernelClass> Computer(nnodes, X, nnodes, Y, weights);

				// start timer
				time.tic();

#if (defined ONLY_SVD || defined FULLY_PIVOTED_ACASVD)
				Computer(0, nnodes, 0, nnodes, U);
#endif
				/*
309 310 311 312 313 314 315
				// applying weights ////////////////////////////////////////
				FReal weights[nnodes];
				FChebTensor<ORDER>::setRootOfWeights(weights);
				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
				}
Matthias Messner's avatar
Matthias Messner committed
316 317 318 319 320 321 322 323 324 325
				*/

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

326

Matthias Messner's avatar
Matthias Messner committed
327 328

				//////////////////////////////////////////////////////////////
Matthias Messner's avatar
Matthias Messner committed
329
#if (defined FULLY_PIVOTED_ACASVD || defined PARTIALLY_PIVOTED_ACASVD) ////////////
Matthias Messner's avatar
Matthias Messner committed
330 331
				FReal *UU, *VV;
				unsigned int rank;
Matthias Messner's avatar
Matthias Messner committed
332 333

#ifdef FULLY_PIVOTED_ACASVD
334
				fACA(U,        nnodes, nnodes, Epsilon, UU, VV, rank);
Matthias Messner's avatar
Matthias Messner committed
335
#else
336
				pACA(Computer, nnodes, nnodes, Epsilon, UU, VV, rank);
Matthias Messner's avatar
Matthias Messner committed
337
#endif 
Matthias Messner's avatar
Matthias Messner committed
338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414

				// 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) throw std::runtime_error("SVD did not converge with " + INFO);
					rank = getRank(S, aca_rank, Epsilon);
				}					
				
				const unsigned int idx = (i+3)*7*7 + (j+3)*7 + (k+3);

				// store
				{
					// allocate
					assert(K[idx]==NULL);
					K[idx] = new FReal [2*rank*nnodes];
					
					// set low rank
					LowRank[idx] = 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;

Matthias Messner's avatar
Matthias Messner committed
415 416
				elapsed_time = time.tacAndElapsed(); 
				overall_time += elapsed_time;
417
				overall_rank += rank;
Matthias Messner's avatar
Matthias Messner committed
418
				std::cout << "(" << i << "," << j << "," << k << ") " << idx <<
Matthias Messner's avatar
Matthias Messner committed
419
					", low rank = " << rank << " (" << aca_rank << ") in " << elapsed_time << "s" << std::endl;
Matthias Messner's avatar
Matthias Messner committed
420

Matthias Messner's avatar
Matthias Messner committed
421 422 423 424 425 426 427 428 429
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////
				// ALL PREPROC FLAGS ARE SET ON TOP OF THIS FILE !!! /////////
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////

#elif defined ONLY_SVD
430
				// truncated singular value decomposition of matrix
Matthias Messner's avatar
Matthias Messner committed
431 432
				INFO = FBlas::gesvd(nnodes, nnodes, U, S, VT, nnodes, LWORK, WORK);
				if (INFO!=0) throw std::runtime_error("SVD did not converge with " + INFO);
433
				const unsigned int rank = getRank<ORDER>(S, Epsilon);
Matthias Messner's avatar
Matthias Messner committed
434
				
435 436 437 438 439 440 441 442 443 444 445
				// store 
				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;
				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);

Matthias Messner's avatar
Matthias Messner committed
446 447
				elapsed_time = time.tacAndElapsed(); 
				overall_time += elapsed_time;
448
				overall_rank += rank;
Matthias Messner's avatar
Matthias Messner committed
449
				std::cout << "(" << i << "," << j << "," << k << ") " << idx <<
Matthias Messner's avatar
Matthias Messner committed
450 451 452
					", low rank = " << rank << " in " << elapsed_time << "s" << std::endl;
#else
#error Either fully-, partially pivoted ACA or only SVD must be defined!
Matthias Messner's avatar
Matthias Messner committed
453 454 455
#endif ///////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////

Matthias Messner's avatar
Matthias Messner committed
456 457 458 459 460 461 462 463
				
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////
				// ALL PREPROC FLAGS ARE SET ON TOP OF THIS FILE !!! /////////
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////
				//////////////////////////////////////////////////////////////
Matthias Messner's avatar
Matthias Messner committed
464 465


466 467 468 469 470 471 472 473 474 475 476
				// 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++;
			}
		}
	}
477 478
	std::cout << "The approximation of the " << counter
						<< " far-field interactions (overall rank " << overall_rank << ") took "
479
						<< overall_time << "s\n" << std::endl;
480 481 482 483 484 485 486 487 488 489 490 491 492 493
	delete [] U;
	delete [] WORK;
	delete [] VT;
	delete [] S;
}









Matthias Messner's avatar
Matthias Messner committed
494
/*!  \class SymmetryHandler 
495

Matthias Messner's avatar
Matthias Messner committed
496 497 498 499 500 501 502
	\brief Deals with all the symmetries in the arrangement of the far-field interactions

	Stores permutation indices and permutation vectors to reduce 316 (7^3-3^3)
  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. */
503 504 505
template <int ORDER, KERNEL_FUNCTION_TYPE TYPE> class SymmetryHandler;

/*! Specialization for homogeneous kernel functions */
506
template <int ORDER>
507
class SymmetryHandler<ORDER, HOMOGENEOUS>
508 509 510 511 512 513
{
  static const unsigned int nnodes = ORDER*ORDER*ORDER;

	// M2L operators
	FReal*    K[343];
	int LowRank[343];
514 515

public:
516 517 518 519 520 521 522 523
	
	// permutation vectors and permutated indices
	unsigned int pvectors[343][nnodes];
	unsigned int pindices[343];


	/** Constructor: with 16 small SVDs */
	template <typename MatrixKernelClass>
524
	SymmetryHandler(const MatrixKernelClass *const MatrixKernel, const FReal Epsilon,
525
									const FReal, const unsigned int)
526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544
	{
		// init all 343 item to zero, because effectively only 16 exist
		for (unsigned int t=0; t<343; ++t) {
			K[t] = NULL;
			LowRank[t] = 0;
		}
			
		// set permutation vector and indices
		const FChebSymmetries<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
545 546
		const FReal ReferenceCellWidth = FReal(2.);
		precompute<ORDER>(MatrixKernel, ReferenceCellWidth, Epsilon, K, LowRank);
547 548 549 550 551 552 553 554 555 556
	}



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

557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657

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

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

};






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

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

	// M2L operators for all levels in the octree
	FReal***    K;
	int** LowRank;

public:
	
	// 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]       = NULL; K[1]       = NULL;
		LowRank[0] = NULL; LowRank[1] = NULL;
		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]       = NULL;
				LowRank[l][t] = 0;
			}
		}
		

		// set permutation vector and indices
		const FChebSymmetries<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]!=NULL) {
				for (unsigned int t=0; t<343; ++t) if (K[l][t]!=NULL) delete [] K[l][t];
				delete [] K[l];
			}
			if (LowRank[l]!=NULL)	delete [] LowRank[l];
		}
		delete [] K;
		delete [] LowRank;
	}

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

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

658 659 660 661 662 663 664 665 666 667 668 669 670
};








#include <fstream>
#include <sstream>


Matthias Messner's avatar
Matthias Messner committed
671 672 673
/**
 * Computes, compresses and stores the 16 M2L kernels in a binary file.
 */
674 675 676 677 678 679 680 681 682
template <int ORDER, typename MatrixKernelClass>
static void ComputeAndCompressAndStoreInBinaryFile(const MatrixKernelClass *const MatrixKernel, const FReal Epsilon)
{
	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] = NULL; LowRank[idx] = 0;	}
683
	precompute<ORDER>(MatrixKernel, FReal(2.), Epsilon, K, LowRank);
684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719

	// 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]!=NULL) {
				// 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]!=NULL) delete [] K[t];
}


Matthias Messner's avatar
Matthias Messner committed
720 721 722 723
/**
 * Reads the 16 compressed M2L kernels from the binary files and writes them
 * in K and the respective low-rank in LowRank.
 */
724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776
template <int ORDER>
void ReadFromBinaryFile(const FReal Epsilon, FReal* K[343], int LowRank[343])
{
	// 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] = NULL;
			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();
}





#endif