Actual source code: math2opus.cu

  1: #include <h2opusconf.h>
  2: /* skip compilation of this .cu file if H2OPUS is CPU only while PETSc has GPU support */
  4:   #include <h2opus.h>
  5:   #if defined(H2OPUS_USE_MPI)
  6:     #include <h2opus/distributed/distributed_h2opus_handle.h>
  7:     #include <h2opus/distributed/distributed_geometric_construction.h>
  8:     #include <h2opus/distributed/distributed_hgemv.h>
  9:     #include <h2opus/distributed/distributed_horthog.h>
 10:     #include <h2opus/distributed/distributed_hcompress.h>
 11:   #endif
 12:   #include <h2opus/util/boxentrygen.h>
 13: #include <petsc/private/matimpl.h>
 14: #include <petsc/private/vecimpl.h>
 15: #include <petsc/private/deviceimpl.h>
 16: #include <petscsf.h>

 18: /* math2opusutils */
 19: PETSC_INTERN PetscErrorCode MatDenseGetH2OpusStridedSF(Mat, PetscSF, PetscSF *);

 21:   #define MatH2OpusGetThrustPointer(v) thrust::raw_pointer_cast((v).data())

 23:   /* Use GPU only if H2OPUS is configured for GPU */
 24:   #if PetscDefined(HAVE_CUDA) && defined(H2OPUS_USE_GPU)
 25:     #define PETSC_H2OPUS_USE_GPU
 26:   #endif
 27:   #if PetscDefined(H2OPUS_USE_GPU)
 28:     #define MatH2OpusUpdateIfNeeded(A, B) MatBindToCPU(A, (PetscBool)((A)->boundtocpu || (B)))
 29:   #else
 30:     #define MatH2OpusUpdateIfNeeded(A, B) PETSC_SUCCESS
 31:   #endif

 33: // TODO H2OPUS:
 34: // DistributedHMatrix
 35: //   unsymmetric ?
 36: //   transpose for distributed_hgemv?
 37: //   clearData()
 38: // Unify interface for sequential and parallel?
 39: // Reuse geometric construction (almost possible, only the unsymmetric case is explicitly handled)
 40: //
 41: template <class T>
 42: class PetscPointCloud : public H2OpusDataSet<T> {
 43: private:
 44:   int            dimension;
 45:   size_t         num_points;
 46:   std::vector<T> pts;

 48: public:
 49:   PetscPointCloud(int dim, size_t num_pts, const T coords[])
 50:   {
 51:     dim              = dim > 0 ? dim : 1;
 52:     this->dimension  = dim;
 53:     this->num_points = num_pts;

 55:     pts.resize(num_pts * dim);
 56:     if (coords) {
 57:       for (size_t n = 0; n < num_pts; n++)
 58:         for (int i = 0; i < dim; i++) pts[n * dim + i] = coords[n * dim + i];
 59:     } else {
 60:       PetscReal h = 1.0; //num_pts > 1 ? 1./(num_pts - 1) : 0.0;
 61:       for (size_t n = 0; n < num_pts; n++) {
 62:         pts[n * dim] = n * h;
 63:         for (int i = 1; i < dim; i++) pts[n * dim + i] = 0.0;
 64:       }
 65:     }
 66:   }

 68:   PetscPointCloud(const PetscPointCloud<T> &other)
 69:   {
 70:     size_t N         = other.dimension * other.num_points;
 71:     this->dimension  = other.dimension;
 72:     this->num_points = other.num_points;
 73:     this->pts.resize(N);
 74:     for (size_t i = 0; i < N; i++) this->pts[i] = other.pts[i];
 75:   }

 77:   int getDimension() const { return dimension; }

 79:   size_t getDataSetSize() const { return num_points; }

 81:   T getDataPoint(size_t idx, int dim) const
 82:   {
 83:     assert(dim < dimension && idx < num_points);
 84:     return pts[idx * dimension + dim];
 85:   }

 87:   void Print(std::ostream &out = std::cout)
 88:   {
 89:     out << "Dimension: " << dimension << std::endl;
 90:     out << "NumPoints: " << num_points << std::endl;
 91:     for (size_t n = 0; n < num_points; n++) {
 92:       for (int d = 0; d < dimension; d++) out << pts[n * dimension + d] << " ";
 93:       out << std::endl;
 94:     }
 95:   }
 96: };

 98: template <class T>
 99: class PetscFunctionGenerator {
100: private:
101:   MatH2OpusKernelFn *k;
102:   int                dim;
103:   void              *ctx;

105: public:
106:   PetscFunctionGenerator(MatH2OpusKernelFn *k, int dim, PetscCtx ctx)
107:   {
108:     this->k   = k;
109:     this->dim = dim;
110:     this->ctx = ctx;
111:   }
112:   PetscFunctionGenerator(PetscFunctionGenerator &other)
113:   {
114:     this->k   = other.k;
115:     this->dim = other.dim;
116:     this->ctx = other.ctx;
117:   }
118:   T operator()(PetscReal *pt1, PetscReal *pt2) { return (T)((*this->k)(this->dim, pt1, pt2, this->ctx)); }
119: };

121: #include <../src/mat/impls/h2opus/math2opussampler.hpp>

123:   /* just to not clutter the code */
124:   #if !defined(H2OPUS_USE_GPU)
125: typedef HMatrix HMatrix_GPU;
126:     #if defined(H2OPUS_USE_MPI)
127: typedef DistributedHMatrix DistributedHMatrix_GPU;
128:     #endif
129:   #endif

131: typedef struct {
132:   #if defined(H2OPUS_USE_MPI)
133:   distributedH2OpusHandle_t handle;
134:   #else
135:   h2opusHandle_t handle;
136:   #endif
137:   /* Sequential and parallel matrices are two different classes at the moment */
138:   HMatrix *hmatrix;
139:   #if defined(H2OPUS_USE_MPI)
140:   DistributedHMatrix *dist_hmatrix;
141:   #else
142:   HMatrix *dist_hmatrix; /* just to not clutter the code */
143:   #endif
144:   /* May use permutations */
145:   PetscSF                           sf;
146:   PetscLayout                       h2opus_rmap, h2opus_cmap;
147:   IS                                h2opus_indexmap;
148:   thrust::host_vector<PetscScalar> *xx, *yy;
149:   PetscInt                          xxs, yys;
150:   PetscBool                         multsetup;

152:   /* GPU */
153:   HMatrix_GPU *hmatrix_gpu;
154:   #if defined(H2OPUS_USE_MPI)
155:   DistributedHMatrix_GPU *dist_hmatrix_gpu;
156:   #else
157:   HMatrix_GPU *dist_hmatrix_gpu; /* just to not clutter the code */
158:   #endif
159:   #if PetscDefined(H2OPUS_USE_GPU)
160:   thrust::device_vector<PetscScalar> *xx_gpu, *yy_gpu;
161:   PetscInt                            xxs_gpu, yys_gpu;
162:   #endif

164:   /* construction from matvecs */
165:   PetscMatrixSampler *sampler;
166:   PetscBool           nativemult;

168:   /* Admissibility */
169:   PetscReal eta;
170:   PetscInt  leafsize;

172:   /* for dof reordering */
173:   PetscPointCloud<PetscReal> *ptcloud;

175:   /* kernel for generating matrix entries */
176:   PetscFunctionGenerator<PetscScalar> *kernel;

178:   /* basis orthogonalized? */
179:   PetscBool orthogonal;

181:   /* customization */
182:   PetscInt  basisord;
183:   PetscInt  max_rank;
184:   PetscInt  bs;
185:   PetscReal rtol;
186:   PetscInt  norm_max_samples;
187:   PetscBool check_construction;
188:   PetscBool hara_verbose;
189:   PetscBool resize;

191:   /* keeps track of MatScale values */
192:   PetscScalar s;
193: } Mat_H2OPUS;

195: static PetscErrorCode MatDestroy_H2OPUS(Mat A)
196: {
197:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;

199:   PetscFunctionBegin;
200:   #if defined(H2OPUS_USE_MPI)
201:   h2opusDestroyDistributedHandle(a->handle);
202:   #else
203:   h2opusDestroyHandle(a->handle);
204:   #endif
205:   delete a->dist_hmatrix;
206:   delete a->hmatrix;
207:   PetscCall(PetscSFDestroy(&a->sf));
208:   PetscCall(PetscLayoutDestroy(&a->h2opus_rmap));
209:   PetscCall(PetscLayoutDestroy(&a->h2opus_cmap));
210:   PetscCall(ISDestroy(&a->h2opus_indexmap));
211:   delete a->xx;
212:   delete a->yy;
213:   delete a->hmatrix_gpu;
214:   delete a->dist_hmatrix_gpu;
215:   #if PetscDefined(H2OPUS_USE_GPU)
216:   delete a->xx_gpu;
217:   delete a->yy_gpu;
218:   #endif
219:   delete a->sampler;
220:   delete a->ptcloud;
221:   delete a->kernel;
222:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_seqdense_C", NULL));
223:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_seqdensecuda_C", NULL));
224:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_mpidense_C", NULL));
225:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_mpidensecuda_C", NULL));
226:   PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL));
227:   PetscCall(PetscFree(A->data));
228:   PetscFunctionReturn(PETSC_SUCCESS);
229: }

231: /*@
232:   MatH2OpusSetNativeMult - Enable or disable the native H2Opus matrix-vector multiplication path for a `MATH2OPUS`.

234:   Logically Collective

236:   Input Parameters:
237: + A  - the `MATH2OPUS` matrix
238: - nm - `PETSC_TRUE` to enable the native H2Opus multiply layout, `PETSC_FALSE` to use the PETSc layout

240:   Level: advanced

242:   Note:
243:   Switching this flag swaps the row and column `PetscLayout`s of `A` with those needed by H2Opus so that
244:   vectors created by `MatCreateVecs()` are compatible with the currently selected multiplication path.

246: .seealso: `Mat`, `MATH2OPUS`, `MatH2OpusGetNativeMult()`
247: @*/
248: PetscErrorCode MatH2OpusSetNativeMult(Mat A, PetscBool nm)
249: {
250:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;
251:   PetscBool   ish2opus;

253:   PetscFunctionBegin;
256:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
257:   if (ish2opus) {
258:     if (a->h2opus_rmap) { /* need to swap layouts for vector creation */
259:       if ((!a->nativemult && nm) || (a->nativemult && !nm)) {
260:         PetscLayout t;
261:         t              = A->rmap;
262:         A->rmap        = a->h2opus_rmap;
263:         a->h2opus_rmap = t;
264:         t              = A->cmap;
265:         A->cmap        = a->h2opus_cmap;
266:         a->h2opus_cmap = t;
267:       }
268:     }
269:     a->nativemult = nm;
270:   }
271:   PetscFunctionReturn(PETSC_SUCCESS);
272: }

274: /*@
275:   MatH2OpusGetNativeMult - Query whether the native H2Opus matrix-vector multiplication path is enabled for a `MATH2OPUS`.

277:   Not Collective

279:   Input Parameter:
280: . A - the `MATH2OPUS` matrix

282:   Output Parameter:
283: . nm - `PETSC_TRUE` if the native H2Opus multiply is enabled, `PETSC_FALSE` otherwise

285:   Level: advanced

287: .seealso: `Mat`, `MATH2OPUS`, `MatH2OpusSetNativeMult()`
288: @*/
289: PetscErrorCode MatH2OpusGetNativeMult(Mat A, PetscBool *nm)
290: {
291:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;
292:   PetscBool   ish2opus;

294:   PetscFunctionBegin;
296:   PetscAssertPointer(nm, 2);
297:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
298:   PetscCheck(ish2opus, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
299:   *nm = a->nativemult;
300:   PetscFunctionReturn(PETSC_SUCCESS);
301: }

303: PETSC_EXTERN PetscErrorCode MatNorm_H2OPUS(Mat A, NormType normtype, PetscReal *n)
304: {
305:   PetscBool   ish2opus;
306:   PetscInt    nmax = PETSC_DECIDE;
307:   Mat_H2OPUS *a    = NULL;
308:   PetscBool   mult = PETSC_FALSE;

310:   PetscFunctionBegin;
311:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
312:   if (ish2opus) { /* set userdefine number of samples and fastpath for mult (norms are order independent) */
313:     a = (Mat_H2OPUS *)A->data;

315:     nmax = a->norm_max_samples;
316:     mult = a->nativemult;
317:     PetscCall(MatH2OpusSetNativeMult(A, PETSC_TRUE));
318:   } else {
319:     PetscCall(PetscOptionsGetInt(((PetscObject)A)->options, ((PetscObject)A)->prefix, "-mat_approximate_norm_samples", &nmax, NULL));
320:   }
321:   PetscCall(MatNormApproximate(A, normtype, nmax, n));
322:   if (a) PetscCall(MatH2OpusSetNativeMult(A, mult));
323:   PetscFunctionReturn(PETSC_SUCCESS);
324: }

326: static PetscErrorCode MatH2OpusResizeBuffers_Private(Mat A, PetscInt xN, PetscInt yN)
327: {
328:   Mat_H2OPUS *h2opus = (Mat_H2OPUS *)A->data;
329:   PetscInt    n;
330:   PetscBool   boundtocpu = PETSC_TRUE;

332:   PetscFunctionBegin;
333:   #if PetscDefined(H2OPUS_USE_GPU)
334:   boundtocpu = A->boundtocpu;
335:   #endif
336:   PetscCall(PetscSFGetGraph(h2opus->sf, NULL, &n, NULL, NULL));
337:   if (boundtocpu) {
338:     if (h2opus->xxs < xN) {
339:       h2opus->xx->resize(n * xN);
340:       h2opus->xxs = xN;
341:     }
342:     if (h2opus->yys < yN) {
343:       h2opus->yy->resize(n * yN);
344:       h2opus->yys = yN;
345:     }
346:   }
347:   #if PetscDefined(H2OPUS_USE_GPU)
348:   if (!boundtocpu) {
349:     if (h2opus->xxs_gpu < xN) {
350:       h2opus->xx_gpu->resize(n * xN);
351:       h2opus->xxs_gpu = xN;
352:     }
353:     if (h2opus->yys_gpu < yN) {
354:       h2opus->yy_gpu->resize(n * yN);
355:       h2opus->yys_gpu = yN;
356:     }
357:   }
358:   #endif
359:   PetscFunctionReturn(PETSC_SUCCESS);
360: }

362: static PetscErrorCode MatMultNKernel_H2OPUS(Mat A, PetscBool transA, Mat B, Mat C)
363: {
364:   Mat_H2OPUS *h2opus = (Mat_H2OPUS *)A->data;
365:   #if defined(H2OPUS_USE_MPI)
366:   h2opusHandle_t handle = h2opus->handle->handle;
367:   #else
368:   h2opusHandle_t handle = h2opus->handle;
369:   #endif
370:   PetscBool    boundtocpu = PETSC_TRUE;
371:   PetscScalar *xx, *yy, *uxx, *uyy;
372:   PetscInt     blda, clda;
373:   PetscMPIInt  size;
374:   PetscSF      bsf, csf;
375:   PetscBool    usesf = (PetscBool)(h2opus->sf && !h2opus->nativemult);

377:   PetscFunctionBegin;
378:   HLibProfile::clear();
379:   #if PetscDefined(H2OPUS_USE_GPU)
380:   boundtocpu = A->boundtocpu;
381:   #endif
382:   PetscCall(MatDenseGetLDA(B, &blda));
383:   PetscCall(MatDenseGetLDA(C, &clda));
384:   if (usesf) {
385:     PetscInt n;

387:     PetscCall(MatDenseGetH2OpusStridedSF(B, h2opus->sf, &bsf));
388:     PetscCall(MatDenseGetH2OpusStridedSF(C, h2opus->sf, &csf));

390:     PetscCall(MatH2OpusResizeBuffers_Private(A, B->cmap->N, C->cmap->N));
391:     PetscCall(PetscSFGetGraph(h2opus->sf, NULL, &n, NULL, NULL));
392:     blda = n;
393:     clda = n;
394:   }
395:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
396:   if (boundtocpu) {
397:     PetscCall(MatDenseGetArrayRead(B, (const PetscScalar **)&xx));
398:     PetscCall(MatDenseGetArrayWrite(C, &yy));
399:     if (usesf) {
400:       uxx = MatH2OpusGetThrustPointer(*h2opus->xx);
401:       uyy = MatH2OpusGetThrustPointer(*h2opus->yy);
402:       PetscCall(PetscSFBcastBegin(bsf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
403:       PetscCall(PetscSFBcastEnd(bsf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
404:     } else {
405:       uxx = xx;
406:       uyy = yy;
407:     }
408:     if (size > 1) {
409:       PetscCheck(h2opus->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing distributed CPU matrix");
410:       PetscCheck(!transA || A->symmetric, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MatMultTranspose not yet coded in parallel");
411:   #if defined(H2OPUS_USE_MPI)
412:       distributed_hgemv(/* transA ? H2Opus_Trans : H2Opus_NoTrans, */ h2opus->s, *h2opus->dist_hmatrix, uxx, blda, 0.0, uyy, clda, B->cmap->N, h2opus->handle);
413:   #endif
414:     } else {
415:       PetscCheck(h2opus->hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
416:       hgemv(transA ? H2Opus_Trans : H2Opus_NoTrans, h2opus->s, *h2opus->hmatrix, uxx, blda, 0.0, uyy, clda, B->cmap->N, handle);
417:     }
418:     PetscCall(MatDenseRestoreArrayRead(B, (const PetscScalar **)&xx));
419:     if (usesf) {
420:       PetscCall(PetscSFReduceBegin(csf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
421:       PetscCall(PetscSFReduceEnd(csf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
422:     }
423:     PetscCall(MatDenseRestoreArrayWrite(C, &yy));
424:   #if PetscDefined(H2OPUS_USE_GPU)
425:   } else {
426:     PetscBool ciscuda, biscuda;

428:     /* If not of type seqdensecuda, convert on the fly (i.e. allocate GPU memory) */
429:     PetscCall(PetscObjectTypeCompareAny((PetscObject)B, &biscuda, MATSEQDENSECUDA, MATMPIDENSECUDA, ""));
430:     if (!biscuda) PetscCall(MatConvert(B, MATDENSECUDA, MAT_INPLACE_MATRIX, &B));
431:     PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &ciscuda, MATSEQDENSECUDA, MATMPIDENSECUDA, ""));
432:     if (!ciscuda) {
433:       C->assembled = PETSC_TRUE;
434:       PetscCall(MatConvert(C, MATDENSECUDA, MAT_INPLACE_MATRIX, &C));
435:     }
436:     PetscCall(MatDenseCUDAGetArrayRead(B, (const PetscScalar **)&xx));
437:     PetscCall(MatDenseCUDAGetArrayWrite(C, &yy));
438:     if (usesf) {
439:       uxx = MatH2OpusGetThrustPointer(*h2opus->xx_gpu);
440:       uyy = MatH2OpusGetThrustPointer(*h2opus->yy_gpu);
441:       PetscCall(PetscSFBcastBegin(bsf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
442:       PetscCall(PetscSFBcastEnd(bsf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
443:     } else {
444:       uxx = xx;
445:       uyy = yy;
446:     }
447:     PetscCall(PetscLogGpuTimeBegin());
448:     if (size > 1) {
449:       PetscCheck(h2opus->dist_hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing distributed GPU matrix");
450:       PetscCheck(!transA || A->symmetric, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MatMultTranspose not yet coded in parallel");
451:     #if defined(H2OPUS_USE_MPI)
452:       distributed_hgemv(/* transA ? H2Opus_Trans : H2Opus_NoTrans, */ h2opus->s, *h2opus->dist_hmatrix_gpu, uxx, blda, 0.0, uyy, clda, B->cmap->N, h2opus->handle);
453:     #endif
454:     } else {
455:       PetscCheck(h2opus->hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
456:       hgemv(transA ? H2Opus_Trans : H2Opus_NoTrans, h2opus->s, *h2opus->hmatrix_gpu, uxx, blda, 0.0, uyy, clda, B->cmap->N, handle);
457:     }
458:     PetscCall(PetscLogGpuTimeEnd());
459:     PetscCall(MatDenseCUDARestoreArrayRead(B, (const PetscScalar **)&xx));
460:     if (usesf) {
461:       PetscCall(PetscSFReduceBegin(csf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
462:       PetscCall(PetscSFReduceEnd(csf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
463:     }
464:     PetscCall(MatDenseCUDARestoreArrayWrite(C, &yy));
465:     if (!biscuda) PetscCall(MatConvert(B, MATDENSE, MAT_INPLACE_MATRIX, &B));
466:     if (!ciscuda) PetscCall(MatConvert(C, MATDENSE, MAT_INPLACE_MATRIX, &C));
467:   #endif
468:   }
469:   { /* log flops */
470:     double gops, time, perf, dev;
471:     HLibProfile::getHgemvPerf(gops, time, perf, dev);
472:   #if PetscDefined(H2OPUS_USE_GPU)
473:     if (boundtocpu) PetscCall(PetscLogFlops(1e9 * gops));
474:     else PetscCall(PetscLogGpuFlops(1e9 * gops));
475:   #else
476:     PetscCall(PetscLogFlops(1e9 * gops));
477:   #endif
478:   }
479:   PetscFunctionReturn(PETSC_SUCCESS);
480: }

482: static PetscErrorCode MatProductNumeric_H2OPUS(Mat C)
483: {
484:   Mat_Product *product = C->product;

486:   PetscFunctionBegin;
487:   MatCheckProduct(C, 1);
488:   switch (product->type) {
489:   case MATPRODUCT_AB:
490:     PetscCall(MatMultNKernel_H2OPUS(product->A, PETSC_FALSE, product->B, C));
491:     break;
492:   case MATPRODUCT_AtB:
493:     PetscCall(MatMultNKernel_H2OPUS(product->A, PETSC_TRUE, product->B, C));
494:     break;
495:   default:
496:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatProduct type %s is not supported", MatProductTypes[product->type]);
497:   }
498:   PetscFunctionReturn(PETSC_SUCCESS);
499: }

501: static PetscErrorCode MatProductSymbolic_H2OPUS(Mat C)
502: {
503:   Mat_Product *product = C->product;
504:   PetscBool    cisdense;
505:   Mat          A, B;

507:   PetscFunctionBegin;
508:   MatCheckProduct(C, 1);
509:   A = product->A;
510:   B = product->B;
511:   switch (product->type) {
512:   case MATPRODUCT_AB:
513:     PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
514:     PetscCall(MatSetBlockSizesFromMats(C, product->A, product->B));
515:     PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATMPIDENSE, MATSEQDENSECUDA, MATMPIDENSECUDA, ""));
516:     if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)product->B)->type_name));
517:     PetscCall(MatSetUp(C));
518:     break;
519:   case MATPRODUCT_AtB:
520:     PetscCall(MatSetSizes(C, A->cmap->n, B->cmap->n, A->cmap->N, B->cmap->N));
521:     PetscCall(MatSetBlockSizesFromMats(C, product->A, product->B));
522:     PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATMPIDENSE, MATSEQDENSECUDA, MATMPIDENSECUDA, ""));
523:     if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)product->B)->type_name));
524:     PetscCall(MatSetUp(C));
525:     break;
526:   default:
527:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatProduct type %s is not supported", MatProductTypes[product->type]);
528:   }
529:   C->ops->productsymbolic = NULL;
530:   C->ops->productnumeric  = MatProductNumeric_H2OPUS;
531:   PetscFunctionReturn(PETSC_SUCCESS);
532: }

534: static PetscErrorCode MatProductSetFromOptions_H2OPUS(Mat C)
535: {
536:   PetscFunctionBegin;
537:   MatCheckProduct(C, 1);
538:   if (C->product->type == MATPRODUCT_AB || C->product->type == MATPRODUCT_AtB) C->ops->productsymbolic = MatProductSymbolic_H2OPUS;
539:   PetscFunctionReturn(PETSC_SUCCESS);
540: }

542: static PetscErrorCode MatMultKernel_H2OPUS(Mat A, Vec x, PetscScalar sy, Vec y, PetscBool trans)
543: {
544:   Mat_H2OPUS *h2opus = (Mat_H2OPUS *)A->data;
545:   #if defined(H2OPUS_USE_MPI)
546:   h2opusHandle_t handle = h2opus->handle->handle;
547:   #else
548:   h2opusHandle_t handle = h2opus->handle;
549:   #endif
550:   PetscBool    boundtocpu = PETSC_TRUE;
551:   PetscInt     n;
552:   PetscScalar *xx, *yy, *uxx, *uyy;
553:   PetscMPIInt  size;
554:   PetscBool    usesf = (PetscBool)(h2opus->sf && !h2opus->nativemult);

556:   PetscFunctionBegin;
557:   HLibProfile::clear();
558:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
559:   #if PetscDefined(H2OPUS_USE_GPU)
560:   boundtocpu = A->boundtocpu;
561:   #endif
562:   if (usesf) PetscCall(PetscSFGetGraph(h2opus->sf, NULL, &n, NULL, NULL));
563:   else n = A->rmap->n;
564:   if (boundtocpu) {
565:     PetscCall(VecGetArrayRead(x, (const PetscScalar **)&xx));
566:     if (sy == 0.0) {
567:       PetscCall(VecGetArrayWrite(y, &yy));
568:     } else {
569:       PetscCall(VecGetArray(y, &yy));
570:     }
571:     if (usesf) {
572:       uxx = MatH2OpusGetThrustPointer(*h2opus->xx);
573:       uyy = MatH2OpusGetThrustPointer(*h2opus->yy);

575:       PetscCall(PetscSFBcastBegin(h2opus->sf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
576:       PetscCall(PetscSFBcastEnd(h2opus->sf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
577:       if (sy != 0.0) {
578:         PetscCall(PetscSFBcastBegin(h2opus->sf, MPIU_SCALAR, yy, uyy, MPI_REPLACE));
579:         PetscCall(PetscSFBcastEnd(h2opus->sf, MPIU_SCALAR, yy, uyy, MPI_REPLACE));
580:       }
581:     } else {
582:       uxx = xx;
583:       uyy = yy;
584:     }
585:     if (size > 1) {
586:       PetscCheck(h2opus->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing distributed CPU matrix");
587:       PetscCheck(!trans || A->symmetric, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MatMultTranspose not yet coded in parallel");
588:   #if defined(H2OPUS_USE_MPI)
589:       distributed_hgemv(/*trans ? H2Opus_Trans : H2Opus_NoTrans, */ h2opus->s, *h2opus->dist_hmatrix, uxx, n, sy, uyy, n, 1, h2opus->handle);
590:   #endif
591:     } else {
592:       PetscCheck(h2opus->hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
593:       hgemv(trans ? H2Opus_Trans : H2Opus_NoTrans, h2opus->s, *h2opus->hmatrix, uxx, n, sy, uyy, n, 1, handle);
594:     }
595:     PetscCall(VecRestoreArrayRead(x, (const PetscScalar **)&xx));
596:     if (usesf) {
597:       PetscCall(PetscSFReduceBegin(h2opus->sf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
598:       PetscCall(PetscSFReduceEnd(h2opus->sf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
599:     }
600:     if (sy == 0.0) {
601:       PetscCall(VecRestoreArrayWrite(y, &yy));
602:     } else {
603:       PetscCall(VecRestoreArray(y, &yy));
604:     }
605:   #if PetscDefined(H2OPUS_USE_GPU)
606:   } else {
607:     PetscCall(VecCUDAGetArrayRead(x, (const PetscScalar **)&xx));
608:     if (sy == 0.0) {
609:       PetscCall(VecCUDAGetArrayWrite(y, &yy));
610:     } else {
611:       PetscCall(VecCUDAGetArray(y, &yy));
612:     }
613:     if (usesf) {
614:       uxx = MatH2OpusGetThrustPointer(*h2opus->xx_gpu);
615:       uyy = MatH2OpusGetThrustPointer(*h2opus->yy_gpu);

617:       PetscCall(PetscSFBcastBegin(h2opus->sf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
618:       PetscCall(PetscSFBcastEnd(h2opus->sf, MPIU_SCALAR, xx, uxx, MPI_REPLACE));
619:       if (sy != 0.0) {
620:         PetscCall(PetscSFBcastBegin(h2opus->sf, MPIU_SCALAR, yy, uyy, MPI_REPLACE));
621:         PetscCall(PetscSFBcastEnd(h2opus->sf, MPIU_SCALAR, yy, uyy, MPI_REPLACE));
622:       }
623:     } else {
624:       uxx = xx;
625:       uyy = yy;
626:     }
627:     PetscCall(PetscLogGpuTimeBegin());
628:     if (size > 1) {
629:       PetscCheck(h2opus->dist_hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing distributed GPU matrix");
630:       PetscCheck(!trans || A->symmetric, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "MatMultTranspose not yet coded in parallel");
631:     #if defined(H2OPUS_USE_MPI)
632:       distributed_hgemv(/*trans ? H2Opus_Trans : H2Opus_NoTrans, */ h2opus->s, *h2opus->dist_hmatrix_gpu, uxx, n, sy, uyy, n, 1, h2opus->handle);
633:     #endif
634:     } else {
635:       PetscCheck(h2opus->hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
636:       hgemv(trans ? H2Opus_Trans : H2Opus_NoTrans, h2opus->s, *h2opus->hmatrix_gpu, uxx, n, sy, uyy, n, 1, handle);
637:     }
638:     PetscCall(PetscLogGpuTimeEnd());
639:     PetscCall(VecCUDARestoreArrayRead(x, (const PetscScalar **)&xx));
640:     if (usesf) {
641:       PetscCall(PetscSFReduceBegin(h2opus->sf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
642:       PetscCall(PetscSFReduceEnd(h2opus->sf, MPIU_SCALAR, uyy, yy, MPI_REPLACE));
643:     }
644:     if (sy == 0.0) {
645:       PetscCall(VecCUDARestoreArrayWrite(y, &yy));
646:     } else {
647:       PetscCall(VecCUDARestoreArray(y, &yy));
648:     }
649:   #endif
650:   }
651:   { /* log flops */
652:     double gops, time, perf, dev;
653:     HLibProfile::getHgemvPerf(gops, time, perf, dev);
654:   #if PetscDefined(H2OPUS_USE_GPU)
655:     if (boundtocpu) PetscCall(PetscLogFlops(1e9 * gops));
656:     else PetscCall(PetscLogGpuFlops(1e9 * gops));
657:   #else
658:     PetscCall(PetscLogFlops(1e9 * gops));
659:   #endif
660:   }
661:   PetscFunctionReturn(PETSC_SUCCESS);
662: }

664: static PetscErrorCode MatMultTranspose_H2OPUS(Mat A, Vec x, Vec y)
665: {
666:   PetscBool xiscuda, yiscuda;

668:   PetscFunctionBegin;
669:   PetscCall(PetscObjectTypeCompareAny((PetscObject)x, &xiscuda, VECSEQCUDA, VECMPICUDA, ""));
670:   PetscCall(PetscObjectTypeCompareAny((PetscObject)y, &yiscuda, VECSEQCUDA, VECMPICUDA, ""));
671:   PetscCall(MatH2OpusUpdateIfNeeded(A, !xiscuda || !yiscuda));
672:   PetscCall(MatMultKernel_H2OPUS(A, x, 0.0, y, PETSC_TRUE));
673:   PetscFunctionReturn(PETSC_SUCCESS);
674: }

676: static PetscErrorCode MatMult_H2OPUS(Mat A, Vec x, Vec y)
677: {
678:   PetscBool xiscuda, yiscuda;

680:   PetscFunctionBegin;
681:   PetscCall(PetscObjectTypeCompareAny((PetscObject)x, &xiscuda, VECSEQCUDA, VECMPICUDA, ""));
682:   PetscCall(PetscObjectTypeCompareAny((PetscObject)y, &yiscuda, VECSEQCUDA, VECMPICUDA, ""));
683:   PetscCall(MatH2OpusUpdateIfNeeded(A, !xiscuda || !yiscuda));
684:   PetscCall(MatMultKernel_H2OPUS(A, x, 0.0, y, PETSC_FALSE));
685:   PetscFunctionReturn(PETSC_SUCCESS);
686: }

688: static PetscErrorCode MatMultTransposeAdd_H2OPUS(Mat A, Vec x, Vec y, Vec z)
689: {
690:   PetscBool xiscuda, ziscuda;

692:   PetscFunctionBegin;
693:   PetscCall(VecCopy(y, z));
694:   PetscCall(PetscObjectTypeCompareAny((PetscObject)x, &xiscuda, VECSEQCUDA, VECMPICUDA, ""));
695:   PetscCall(PetscObjectTypeCompareAny((PetscObject)z, &ziscuda, VECSEQCUDA, VECMPICUDA, ""));
696:   PetscCall(MatH2OpusUpdateIfNeeded(A, !xiscuda || !ziscuda));
697:   PetscCall(MatMultKernel_H2OPUS(A, x, 1.0, z, PETSC_TRUE));
698:   PetscFunctionReturn(PETSC_SUCCESS);
699: }

701: static PetscErrorCode MatMultAdd_H2OPUS(Mat A, Vec x, Vec y, Vec z)
702: {
703:   PetscBool xiscuda, ziscuda;

705:   PetscFunctionBegin;
706:   PetscCall(VecCopy(y, z));
707:   PetscCall(PetscObjectTypeCompareAny((PetscObject)x, &xiscuda, VECSEQCUDA, VECMPICUDA, ""));
708:   PetscCall(PetscObjectTypeCompareAny((PetscObject)z, &ziscuda, VECSEQCUDA, VECMPICUDA, ""));
709:   PetscCall(MatH2OpusUpdateIfNeeded(A, !xiscuda || !ziscuda));
710:   PetscCall(MatMultKernel_H2OPUS(A, x, 1.0, z, PETSC_FALSE));
711:   PetscFunctionReturn(PETSC_SUCCESS);
712: }

714: static PetscErrorCode MatScale_H2OPUS(Mat A, PetscScalar s)
715: {
716:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;

718:   PetscFunctionBegin;
719:   a->s *= s;
720:   PetscFunctionReturn(PETSC_SUCCESS);
721: }

723: static PetscErrorCode MatSetFromOptions_H2OPUS(Mat A, PetscOptionItems PetscOptionsObject)
724: {
725:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;

727:   PetscFunctionBegin;
728:   PetscOptionsHeadBegin(PetscOptionsObject, "H2OPUS options");
729:   PetscCall(PetscOptionsInt("-mat_h2opus_leafsize", "Leaf size of cluster tree", NULL, a->leafsize, &a->leafsize, NULL));
730:   PetscCall(PetscOptionsReal("-mat_h2opus_eta", "Admissibility condition tolerance", NULL, a->eta, &a->eta, NULL));
731:   PetscCall(PetscOptionsInt("-mat_h2opus_order", "Basis order for off-diagonal sampling when constructed from kernel", NULL, a->basisord, &a->basisord, NULL));
732:   PetscCall(PetscOptionsInt("-mat_h2opus_maxrank", "Maximum rank when constructed from matvecs", NULL, a->max_rank, &a->max_rank, NULL));
733:   PetscCall(PetscOptionsInt("-mat_h2opus_samples", "Maximum number of samples to be taken concurrently when constructing from matvecs", NULL, a->bs, &a->bs, NULL));
734:   PetscCall(PetscOptionsInt("-mat_h2opus_normsamples", "Maximum number of samples to be when estimating norms", NULL, a->norm_max_samples, &a->norm_max_samples, NULL));
735:   PetscCall(PetscOptionsReal("-mat_h2opus_rtol", "Relative tolerance for construction from sampling", NULL, a->rtol, &a->rtol, NULL));
736:   PetscCall(PetscOptionsBool("-mat_h2opus_check", "Check error when constructing from sampling during MatAssemblyEnd()", NULL, a->check_construction, &a->check_construction, NULL));
737:   PetscCall(PetscOptionsBool("-mat_h2opus_hara_verbose", "Verbose output from hara construction", NULL, a->hara_verbose, &a->hara_verbose, NULL));
738:   PetscCall(PetscOptionsBool("-mat_h2opus_resize", "Resize after compression", NULL, a->resize, &a->resize, NULL));
739:   PetscOptionsHeadEnd();
740:   PetscFunctionReturn(PETSC_SUCCESS);
741: }

743: static PetscErrorCode MatH2OpusSetCoords_H2OPUS(Mat, PetscInt, const PetscReal[], PetscBool, MatH2OpusKernelFn *, void *);

745: static PetscErrorCode MatH2OpusInferCoordinates_Private(Mat A)
746: {
747:   Mat_H2OPUS        *a = (Mat_H2OPUS *)A->data;
748:   Vec                c;
749:   PetscInt           spacedim;
750:   const PetscScalar *coords;

752:   PetscFunctionBegin;
753:   if (a->ptcloud) PetscFunctionReturn(PETSC_SUCCESS);
754:   PetscCall(PetscObjectQuery((PetscObject)A, "__math2opus_coords", (PetscObject *)&c));
755:   if (!c && a->sampler) {
756:     Mat S = a->sampler->GetSamplingMat();

758:     PetscCall(PetscObjectQuery((PetscObject)S, "__math2opus_coords", (PetscObject *)&c));
759:   }
760:   if (!c) {
761:     PetscCall(MatH2OpusSetCoords_H2OPUS(A, -1, NULL, PETSC_FALSE, NULL, NULL));
762:   } else {
763:     PetscCall(VecGetArrayRead(c, &coords));
764:     PetscCall(VecGetBlockSize(c, &spacedim));
765:     PetscCall(MatH2OpusSetCoords_H2OPUS(A, spacedim, coords, PETSC_FALSE, NULL, NULL));
766:     PetscCall(VecRestoreArrayRead(c, &coords));
767:   }
768:   PetscFunctionReturn(PETSC_SUCCESS);
769: }

771: static PetscErrorCode MatSetUpMultiply_H2OPUS(Mat A)
772: {
773:   MPI_Comm      comm;
774:   PetscMPIInt   size;
775:   Mat_H2OPUS   *a = (Mat_H2OPUS *)A->data;
776:   PetscInt      n = 0, *idx = NULL;
777:   int          *iidx = NULL;
778:   PetscCopyMode own;
779:   PetscBool     rid;

781:   PetscFunctionBegin;
782:   if (a->multsetup) PetscFunctionReturn(PETSC_SUCCESS);
783:   if (a->sf) { /* MatDuplicate_H2OPUS takes reference to the SF */
784:     PetscCall(PetscSFGetGraph(a->sf, NULL, &n, NULL, NULL));
785:   #if PetscDefined(H2OPUS_USE_GPU)
786:     a->xx_gpu  = new thrust::device_vector<PetscScalar>(n);
787:     a->yy_gpu  = new thrust::device_vector<PetscScalar>(n);
788:     a->xxs_gpu = 1;
789:     a->yys_gpu = 1;
790:   #endif
791:     a->xx  = new thrust::host_vector<PetscScalar>(n);
792:     a->yy  = new thrust::host_vector<PetscScalar>(n);
793:     a->xxs = 1;
794:     a->yys = 1;
795:   } else {
796:     IS is;
797:     PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
798:     PetscCallMPI(MPI_Comm_size(comm, &size));
799:     if (!a->h2opus_indexmap) {
800:       if (size > 1) {
801:         PetscCheck(a->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing distributed CPU matrix");
802:   #if defined(H2OPUS_USE_MPI)
803:         iidx = MatH2OpusGetThrustPointer(a->dist_hmatrix->basis_tree.basis_branch.index_map);
804:         n    = a->dist_hmatrix->basis_tree.basis_branch.index_map.size();
805:   #endif
806:       } else {
807:         iidx = MatH2OpusGetThrustPointer(a->hmatrix->u_basis_tree.index_map);
808:         n    = a->hmatrix->u_basis_tree.index_map.size();
809:       }

811:       if (PetscDefined(USE_64BIT_INDICES)) {
812:         PetscInt i;

814:         own = PETSC_OWN_POINTER;
815:         PetscCall(PetscMalloc1(n, &idx));
816:         for (i = 0; i < n; i++) idx[i] = iidx[i];
817:       } else {
818:         own = PETSC_COPY_VALUES;
819:         idx = (PetscInt *)iidx;
820:       }
821:       PetscCall(ISCreateGeneral(comm, n, idx, own, &is));
822:       PetscCall(ISSetPermutation(is));
823:       PetscCall(ISViewFromOptions(is, (PetscObject)A, "-mat_h2opus_indexmap_view"));
824:       a->h2opus_indexmap = is;
825:     }
826:     PetscCall(ISGetLocalSize(a->h2opus_indexmap, &n));
827:     PetscCall(ISGetIndices(a->h2opus_indexmap, (const PetscInt **)&idx));
828:     rid = (PetscBool)(n == A->rmap->n);
829:     PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &rid, 1, MPI_C_BOOL, MPI_LAND, comm));
830:     if (rid) PetscCall(ISIdentity(a->h2opus_indexmap, &rid));
831:     if (!rid) {
832:       if (size > 1) { /* Parallel distribution may be different, save it here for fast path in MatMult (see MatH2OpusSetNativeMult) */
833:         PetscCall(PetscLayoutCreate(comm, &a->h2opus_rmap));
834:         PetscCall(PetscLayoutSetLocalSize(a->h2opus_rmap, n));
835:         PetscCall(PetscLayoutSetUp(a->h2opus_rmap));
836:         PetscCall(PetscLayoutReference(a->h2opus_rmap, &a->h2opus_cmap));
837:       }
838:       PetscCall(PetscSFCreate(comm, &a->sf));
839:       PetscCall(PetscSFSetGraphLayout(a->sf, A->rmap, n, NULL, PETSC_OWN_POINTER, idx));
840:       PetscCall(PetscSFSetUp(a->sf));
841:       PetscCall(PetscSFViewFromOptions(a->sf, (PetscObject)A, "-mat_h2opus_sf_view"));
842:   #if PetscDefined(H2OPUS_USE_GPU)
843:       a->xx_gpu  = new thrust::device_vector<PetscScalar>(n);
844:       a->yy_gpu  = new thrust::device_vector<PetscScalar>(n);
845:       a->xxs_gpu = 1;
846:       a->yys_gpu = 1;
847:   #endif
848:       a->xx  = new thrust::host_vector<PetscScalar>(n);
849:       a->yy  = new thrust::host_vector<PetscScalar>(n);
850:       a->xxs = 1;
851:       a->yys = 1;
852:     }
853:     PetscCall(ISRestoreIndices(a->h2opus_indexmap, (const PetscInt **)&idx));
854:   }
855:   a->multsetup = PETSC_TRUE;
856:   PetscFunctionReturn(PETSC_SUCCESS);
857: }

859: static PetscErrorCode MatAssemblyEnd_H2OPUS(Mat A, MatAssemblyType assemblytype)
860: {
861:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;
862:   #if defined(H2OPUS_USE_MPI)
863:   h2opusHandle_t handle = a->handle->handle;
864:   #else
865:   h2opusHandle_t handle = a->handle;
866:   #endif
867:   PetscBool   kernel       = PETSC_FALSE;
868:   PetscBool   boundtocpu   = PETSC_TRUE;
869:   PetscBool   samplingdone = PETSC_FALSE;
870:   MPI_Comm    comm;
871:   PetscMPIInt size;

873:   PetscFunctionBegin;
874:   PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
875:   PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Different row and column local sizes are not supported");
876:   PetscCheck(A->rmap->N == A->cmap->N, comm, PETSC_ERR_SUP, "Rectangular matrices are not supported");

878:   /* XXX */
879:   a->leafsize = PetscMin(a->leafsize, PetscMin(A->rmap->N, A->cmap->N));

881:   PetscCallMPI(MPI_Comm_size(comm, &size));
882:   /* TODO REUSABILITY of geometric construction */
883:   delete a->hmatrix;
884:   delete a->dist_hmatrix;
885:   #if PetscDefined(H2OPUS_USE_GPU)
886:   delete a->hmatrix_gpu;
887:   delete a->dist_hmatrix_gpu;
888:   #endif
889:   a->orthogonal = PETSC_FALSE;

891:   /* TODO: other? */
892:   H2OpusBoxCenterAdmissibility adm(a->eta);

894:   PetscCall(PetscLogEventBegin(MAT_H2Opus_Build, A, 0, 0, 0));
895:   if (size > 1) {
896:   #if defined(H2OPUS_USE_MPI)
897:     a->dist_hmatrix = new DistributedHMatrix(A->rmap->n /* ,A->symmetric */);
898:   #else
899:     a->dist_hmatrix = NULL;
900:   #endif
901:   } else a->hmatrix = new HMatrix(A->rmap->n, A->symmetric == PETSC_BOOL3_TRUE);
902:   PetscCall(MatH2OpusInferCoordinates_Private(A));
903:   PetscCheck(a->ptcloud, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing pointcloud");
904:   if (a->kernel) {
905:     BoxEntryGen<PetscScalar, H2OPUS_HWTYPE_CPU, PetscFunctionGenerator<PetscScalar>> entry_gen(*a->kernel);
906:     if (size > 1) {
907:       PetscCheck(a->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing distributed CPU matrix");
908:   #if defined(H2OPUS_USE_MPI)
909:       buildDistributedHMatrix(*a->dist_hmatrix, a->ptcloud, adm, entry_gen, a->leafsize, a->basisord, a->handle);
910:   #endif
911:     } else {
912:       buildHMatrix(*a->hmatrix, a->ptcloud, adm, entry_gen, a->leafsize, a->basisord);
913:     }
914:     kernel = PETSC_TRUE;
915:   } else {
916:     PetscCheck(size <= 1, comm, PETSC_ERR_SUP, "Construction from sampling not supported in parallel");
917:     buildHMatrixStructure(*a->hmatrix, a->ptcloud, a->leafsize, adm);
918:   }
919:   PetscCall(MatSetUpMultiply_H2OPUS(A));

921:   #if PetscDefined(H2OPUS_USE_GPU)
922:   boundtocpu = A->boundtocpu;
923:   if (!boundtocpu) {
924:     if (size > 1) {
925:       PetscCheck(a->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing distributed CPU matrix");
926:     #if defined(H2OPUS_USE_MPI)
927:       a->dist_hmatrix_gpu = new DistributedHMatrix_GPU(*a->dist_hmatrix);
928:     #endif
929:     } else {
930:       a->hmatrix_gpu = new HMatrix_GPU(*a->hmatrix);
931:     }
932:   }
933:   #endif
934:   if (size == 1) {
935:     if (!kernel && a->sampler && a->sampler->GetSamplingMat()) {
936:       PetscReal Anorm;
937:       bool      verbose;

939:       PetscCall(PetscOptionsGetBool(((PetscObject)A)->options, ((PetscObject)A)->prefix, "-mat_h2opus_hara_verbose", &a->hara_verbose, NULL));
940:       verbose = a->hara_verbose;
941:       PetscCall(MatNormApproximate(a->sampler->GetSamplingMat(), NORM_2, a->norm_max_samples, &Anorm));
942:       if (a->hara_verbose) PetscCall(PetscPrintf(PETSC_COMM_SELF, "Sampling uses max rank %d, tol %g (%g*%g), %s samples %d\n", a->max_rank, a->rtol * Anorm, a->rtol, Anorm, boundtocpu ? "CPU" : "GPU", a->bs));
943:       if (a->sf && !a->nativemult) a->sampler->SetIndexMap(a->hmatrix->u_basis_tree.index_map.size(), a->hmatrix->u_basis_tree.index_map.data());
944:       a->sampler->SetStream(handle->getMainStream());
945:       if (boundtocpu) {
946:         a->sampler->SetGPUSampling(false);
947:         hara(a->sampler, *a->hmatrix, a->max_rank, 10 /* TODO */, a->rtol * Anorm, a->bs, handle, verbose);
948:   #if PetscDefined(H2OPUS_USE_GPU)
949:       } else {
950:         a->sampler->SetGPUSampling(true);
951:         hara(a->sampler, *a->hmatrix_gpu, a->max_rank, 10 /* TODO */, a->rtol * Anorm, a->bs, handle, verbose);
952:   #endif
953:       }
954:       samplingdone = PETSC_TRUE;
955:     }
956:   }
957:   #if PetscDefined(H2OPUS_USE_GPU)
958:   if (!boundtocpu) {
959:     delete a->hmatrix;
960:     delete a->dist_hmatrix;
961:     a->hmatrix      = NULL;
962:     a->dist_hmatrix = NULL;
963:   }
964:   A->offloadmask = boundtocpu ? PETSC_OFFLOAD_CPU : PETSC_OFFLOAD_GPU;
965:   #endif
966:   PetscCall(PetscLogEventEnd(MAT_H2Opus_Build, A, 0, 0, 0));

968:   if (!a->s) a->s = 1.0;
969:   A->assembled = PETSC_TRUE;

971:   if (samplingdone) {
972:     PetscBool check  = a->check_construction;
973:     PetscBool checke = PETSC_FALSE;

975:     PetscCall(PetscOptionsGetBool(((PetscObject)A)->options, ((PetscObject)A)->prefix, "-mat_h2opus_check", &check, NULL));
976:     PetscCall(PetscOptionsGetBool(((PetscObject)A)->options, ((PetscObject)A)->prefix, "-mat_h2opus_check_explicit", &checke, NULL));
977:     if (check) {
978:       Mat               E, Ae;
979:       PetscReal         n1, ni, n2;
980:       PetscReal         n1A, niA, n2A;
981:       PetscErrorCodeFn *normfunc;

983:       Ae = a->sampler->GetSamplingMat();
984:       PetscCall(MatConvert(A, MATSHELL, MAT_INITIAL_MATRIX, &E));
985:       PetscCall(MatShellSetOperation(E, MATOP_NORM, (PetscErrorCodeFn *)MatNorm_H2OPUS));
986:       PetscCall(MatAXPY(E, -1.0, Ae, DIFFERENT_NONZERO_PATTERN));
987:       PetscCall(MatNorm(E, NORM_1, &n1));
988:       PetscCall(MatNorm(E, NORM_INFINITY, &ni));
989:       PetscCall(MatNorm(E, NORM_2, &n2));
990:       if (checke) {
991:         Mat eA, eE, eAe;

993:         PetscCall(MatComputeOperator(A, MATAIJ, &eA));
994:         PetscCall(MatComputeOperator(E, MATAIJ, &eE));
995:         PetscCall(MatComputeOperator(Ae, MATAIJ, &eAe));
996:         PetscCall(MatFilter(eA, PETSC_SMALL, PETSC_FALSE, PETSC_FALSE));
997:         PetscCall(MatFilter(eE, PETSC_SMALL, PETSC_FALSE, PETSC_FALSE));
998:         PetscCall(MatFilter(eAe, PETSC_SMALL, PETSC_FALSE, PETSC_FALSE));
999:         PetscCall(PetscObjectSetName((PetscObject)eA, "H2Mat"));
1000:         PetscCall(MatView(eA, NULL));
1001:         PetscCall(PetscObjectSetName((PetscObject)eAe, "S"));
1002:         PetscCall(MatView(eAe, NULL));
1003:         PetscCall(PetscObjectSetName((PetscObject)eE, "H2Mat - S"));
1004:         PetscCall(MatView(eE, NULL));
1005:         PetscCall(MatDestroy(&eA));
1006:         PetscCall(MatDestroy(&eE));
1007:         PetscCall(MatDestroy(&eAe));
1008:       }

1010:       PetscCall(MatGetOperation(Ae, MATOP_NORM, &normfunc));
1011:       PetscCall(MatSetOperation(Ae, MATOP_NORM, (PetscErrorCodeFn *)MatNorm_H2OPUS));
1012:       PetscCall(MatNorm(Ae, NORM_1, &n1A));
1013:       PetscCall(MatNorm(Ae, NORM_INFINITY, &niA));
1014:       PetscCall(MatNorm(Ae, NORM_2, &n2A));
1015:       n1A = PetscMax(n1A, PETSC_SMALL);
1016:       n2A = PetscMax(n2A, PETSC_SMALL);
1017:       niA = PetscMax(niA, PETSC_SMALL);
1018:       PetscCall(MatSetOperation(Ae, MATOP_NORM, normfunc));
1019:       PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "MATH2OPUS construction errors: NORM_1 %g, NORM_INFINITY %g, NORM_2 %g (%g %g %g)\n", (double)n1, (double)ni, (double)n2, (double)(n1 / n1A), (double)(ni / niA), (double)(n2 / n2A)));
1020:       PetscCall(MatDestroy(&E));
1021:     }
1022:     a->sampler->SetSamplingMat(NULL);
1023:   }
1024:   PetscFunctionReturn(PETSC_SUCCESS);
1025: }

1027: static PetscErrorCode MatZeroEntries_H2OPUS(Mat A)
1028: {
1029:   PetscMPIInt size;
1030:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;

1032:   PetscFunctionBegin;
1033:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1034:   PetscCheck(size <= 1, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not yet supported");
1035:   a->hmatrix->clearData();
1036:   #if PetscDefined(H2OPUS_USE_GPU)
1037:   if (a->hmatrix_gpu) a->hmatrix_gpu->clearData();
1038:   #endif
1039:   PetscFunctionReturn(PETSC_SUCCESS);
1040: }

1042: static PetscErrorCode MatDuplicate_H2OPUS(Mat B, MatDuplicateOption op, Mat *nA)
1043: {
1044:   Mat         A;
1045:   Mat_H2OPUS *a, *b = (Mat_H2OPUS *)B->data;
1046:   PetscBool   iscpu = PetscDefined(H2OPUS_USE_GPU) ? PETSC_FALSE : PETSC_TRUE;
1047:   MPI_Comm    comm;

1049:   PetscFunctionBegin;
1050:   PetscCall(PetscObjectGetComm((PetscObject)B, &comm));
1051:   PetscCall(MatCreate(comm, &A));
1052:   PetscCall(MatSetSizes(A, B->rmap->n, B->cmap->n, B->rmap->N, B->cmap->N));
1053:   PetscCall(MatSetType(A, MATH2OPUS));
1054:   PetscCall(MatPropagateSymmetryOptions(B, A));
1055:   a = (Mat_H2OPUS *)A->data;

1057:   a->eta              = b->eta;
1058:   a->leafsize         = b->leafsize;
1059:   a->basisord         = b->basisord;
1060:   a->max_rank         = b->max_rank;
1061:   a->bs               = b->bs;
1062:   a->rtol             = b->rtol;
1063:   a->norm_max_samples = b->norm_max_samples;
1064:   if (op == MAT_COPY_VALUES) a->s = b->s;

1066:   a->ptcloud = new PetscPointCloud<PetscReal>(*b->ptcloud);
1067:   if (op == MAT_COPY_VALUES && b->kernel) a->kernel = new PetscFunctionGenerator<PetscScalar>(*b->kernel);

1069:   #if defined(H2OPUS_USE_MPI)
1070:   if (b->dist_hmatrix) a->dist_hmatrix = new DistributedHMatrix(*b->dist_hmatrix);
1071:     #if PetscDefined(H2OPUS_USE_GPU)
1072:   if (b->dist_hmatrix_gpu) a->dist_hmatrix_gpu = new DistributedHMatrix_GPU(*b->dist_hmatrix_gpu);
1073:     #endif
1074:   #endif
1075:   if (b->hmatrix) {
1076:     a->hmatrix = new HMatrix(*b->hmatrix);
1077:     if (op == MAT_DO_NOT_COPY_VALUES) a->hmatrix->clearData();
1078:   }
1079:   #if PetscDefined(H2OPUS_USE_GPU)
1080:   if (b->hmatrix_gpu) {
1081:     a->hmatrix_gpu = new HMatrix_GPU(*b->hmatrix_gpu);
1082:     if (op == MAT_DO_NOT_COPY_VALUES) a->hmatrix_gpu->clearData();
1083:   }
1084:   #endif
1085:   if (b->sf) {
1086:     PetscCall(PetscObjectReference((PetscObject)b->sf));
1087:     a->sf = b->sf;
1088:   }
1089:   if (b->h2opus_indexmap) {
1090:     PetscCall(PetscObjectReference((PetscObject)b->h2opus_indexmap));
1091:     a->h2opus_indexmap = b->h2opus_indexmap;
1092:   }

1094:   PetscCall(MatSetUp(A));
1095:   PetscCall(MatSetUpMultiply_H2OPUS(A));
1096:   if (op == MAT_COPY_VALUES) {
1097:     A->assembled  = PETSC_TRUE;
1098:     a->orthogonal = b->orthogonal;
1099:   #if PetscDefined(H2OPUS_USE_GPU)
1100:     A->offloadmask = B->offloadmask;
1101:   #endif
1102:   }
1103:   #if PetscDefined(H2OPUS_USE_GPU)
1104:   iscpu = B->boundtocpu;
1105:   #endif
1106:   PetscCall(MatBindToCPU(A, iscpu));

1108:   *nA = A;
1109:   PetscFunctionReturn(PETSC_SUCCESS);
1110: }

1112: static PetscErrorCode MatView_H2OPUS(Mat A, PetscViewer view)
1113: {
1114:   Mat_H2OPUS       *h2opus = (Mat_H2OPUS *)A->data;
1115:   PetscBool         isascii, vieweps;
1116:   PetscMPIInt       size;
1117:   PetscViewerFormat format;

1119:   PetscFunctionBegin;
1120:   PetscCall(PetscObjectTypeCompare((PetscObject)view, PETSCVIEWERASCII, &isascii));
1121:   PetscCall(PetscViewerGetFormat(view, &format));
1122:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1123:   if (isascii) {
1124:     if (format == PETSC_VIEWER_ASCII_MATLAB) {
1125:       if (size == 1) {
1126:         FILE *fp;
1127:         PetscCall(PetscViewerASCIIGetPointer(view, &fp));
1128:         dumpHMatrix(*h2opus->hmatrix, 6, fp);
1129:       }
1130:     } else {
1131:       PetscCall(PetscViewerASCIIPrintf(view, "  H-Matrix constructed from %s\n", h2opus->kernel ? "Kernel" : "Mat"));
1132:       PetscCall(PetscViewerASCIIPrintf(view, "  PointCloud dim %" PetscInt_FMT "\n", h2opus->ptcloud ? h2opus->ptcloud->getDimension() : 0));
1133:       PetscCall(PetscViewerASCIIPrintf(view, "  Admissibility parameters: leaf size %" PetscInt_FMT ", eta %g\n", h2opus->leafsize, (double)h2opus->eta));
1134:       if (!h2opus->kernel) {
1135:         PetscCall(PetscViewerASCIIPrintf(view, "  Sampling parameters: max_rank %" PetscInt_FMT ", samples %" PetscInt_FMT ", tolerance %g\n", h2opus->max_rank, h2opus->bs, (double)h2opus->rtol));
1136:       } else {
1137:         PetscCall(PetscViewerASCIIPrintf(view, "  Off-diagonal blocks approximation order %" PetscInt_FMT "\n", h2opus->basisord));
1138:       }
1139:       PetscCall(PetscViewerASCIIPrintf(view, "  Number of samples for norms %" PetscInt_FMT "\n", h2opus->norm_max_samples));
1140:       if (size == 1) {
1141:         double dense_mem_cpu = h2opus->hmatrix ? h2opus->hmatrix->getDenseMemoryUsage() : 0;
1142:         double low_rank_cpu  = h2opus->hmatrix ? h2opus->hmatrix->getLowRankMemoryUsage() : 0;
1143:   #if PetscDefined(HAVE_CUDA)
1144:         double dense_mem_gpu = h2opus->hmatrix_gpu ? h2opus->hmatrix_gpu->getDenseMemoryUsage() : 0;
1145:         double low_rank_gpu  = h2opus->hmatrix_gpu ? h2opus->hmatrix_gpu->getLowRankMemoryUsage() : 0;
1146:   #endif
1147:         PetscCall(PetscViewerASCIIPrintf(view, "  Memory consumption GB (CPU): %g (dense) %g (low rank) %g (total)\n", dense_mem_cpu, low_rank_cpu, low_rank_cpu + dense_mem_cpu));
1148:   #if PetscDefined(HAVE_CUDA)
1149:         PetscCall(PetscViewerASCIIPrintf(view, "  Memory consumption GB (GPU): %g (dense) %g (low rank) %g (total)\n", dense_mem_gpu, low_rank_gpu, low_rank_gpu + dense_mem_gpu));
1150:   #endif
1151:       } else {
1152:   #if PetscDefined(HAVE_CUDA)
1153:         double      matrix_mem[4] = {0., 0., 0., 0.};
1154:         PetscMPIInt rsize         = 4;
1155:   #else
1156:         double      matrix_mem[2] = {0., 0.};
1157:         PetscMPIInt rsize         = 2;
1158:   #endif
1159:   #if defined(H2OPUS_USE_MPI)
1160:         matrix_mem[0] = h2opus->dist_hmatrix ? h2opus->dist_hmatrix->getLocalDenseMemoryUsage() : 0;
1161:         matrix_mem[1] = h2opus->dist_hmatrix ? h2opus->dist_hmatrix->getLocalLowRankMemoryUsage() : 0;
1162:     #if PetscDefined(HAVE_CUDA)
1163:         matrix_mem[2] = h2opus->dist_hmatrix_gpu ? h2opus->dist_hmatrix_gpu->getLocalDenseMemoryUsage() : 0;
1164:         matrix_mem[3] = h2opus->dist_hmatrix_gpu ? h2opus->dist_hmatrix_gpu->getLocalLowRankMemoryUsage() : 0;
1165:     #endif
1166:   #endif
1167:         PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, matrix_mem, rsize, MPI_DOUBLE_PRECISION, MPI_SUM, PetscObjectComm((PetscObject)A)));
1168:         PetscCall(PetscViewerASCIIPrintf(view, "  Memory consumption GB (CPU): %g (dense) %g (low rank) %g (total)\n", matrix_mem[0], matrix_mem[1], matrix_mem[0] + matrix_mem[1]));
1169:   #if PetscDefined(HAVE_CUDA)
1170:         PetscCall(PetscViewerASCIIPrintf(view, "  Memory consumption GB (GPU): %g (dense) %g (low rank) %g (total)\n", matrix_mem[2], matrix_mem[3], matrix_mem[2] + matrix_mem[3]));
1171:   #endif
1172:       }
1173:     }
1174:   }
1175:   vieweps = PETSC_FALSE;
1176:   PetscCall(PetscOptionsGetBool(((PetscObject)A)->options, ((PetscObject)A)->prefix, "-mat_h2opus_vieweps", &vieweps, NULL));
1177:   if (vieweps) {
1178:     char        filename[256];
1179:     const char *name;

1181:     PetscCall(PetscObjectGetName((PetscObject)A, &name));
1182:     PetscCall(PetscSNPrintf(filename, sizeof(filename), "%s_structure.eps", name));
1183:     PetscCall(PetscOptionsGetString(((PetscObject)A)->options, ((PetscObject)A)->prefix, "-mat_h2opus_vieweps_filename", filename, sizeof(filename), NULL));
1184:     outputEps(*h2opus->hmatrix, filename);
1185:   }
1186:   PetscFunctionReturn(PETSC_SUCCESS);
1187: }

1189: static PetscErrorCode MatH2OpusSetCoords_H2OPUS(Mat A, PetscInt spacedim, const PetscReal coords[], PetscBool cdist, MatH2OpusKernelFn *kernel, void *kernelctx)
1190: {
1191:   Mat_H2OPUS *h2opus = (Mat_H2OPUS *)A->data;
1192:   PetscReal  *gcoords;
1193:   PetscInt    N;
1194:   MPI_Comm    comm;
1195:   PetscMPIInt size;
1196:   PetscBool   cong;

1198:   PetscFunctionBegin;
1199:   PetscCall(PetscLayoutSetUp(A->rmap));
1200:   PetscCall(PetscLayoutSetUp(A->cmap));
1201:   PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
1202:   PetscCall(MatHasCongruentLayouts(A, &cong));
1203:   PetscCheck(cong, comm, PETSC_ERR_SUP, "Only for square matrices with congruent layouts");
1204:   N = A->rmap->N;
1205:   PetscCallMPI(MPI_Comm_size(comm, &size));
1206:   if (spacedim > 0 && size > 1 && cdist) {
1207:     PetscSF      sf;
1208:     MPI_Datatype dtype;

1210:     PetscCallMPI(MPI_Type_contiguous(spacedim, MPIU_REAL, &dtype));
1211:     PetscCallMPI(MPI_Type_commit(&dtype));

1213:     PetscCall(PetscSFCreate(comm, &sf));
1214:     PetscCall(PetscSFSetGraphWithPattern(sf, A->rmap, PETSCSF_PATTERN_ALLGATHER));
1215:     PetscCall(PetscMalloc1(spacedim * N, &gcoords));
1216:     PetscCall(PetscSFBcastBegin(sf, dtype, coords, gcoords, MPI_REPLACE));
1217:     PetscCall(PetscSFBcastEnd(sf, dtype, coords, gcoords, MPI_REPLACE));
1218:     PetscCall(PetscSFDestroy(&sf));
1219:     PetscCallMPI(MPI_Type_free(&dtype));
1220:   } else gcoords = (PetscReal *)coords;

1222:   delete h2opus->ptcloud;
1223:   delete h2opus->kernel;
1224:   h2opus->ptcloud = new PetscPointCloud<PetscReal>(spacedim, N, gcoords);
1225:   if (kernel) h2opus->kernel = new PetscFunctionGenerator<PetscScalar>(kernel, spacedim, kernelctx);
1226:   if (gcoords != coords) PetscCall(PetscFree(gcoords));
1227:   A->preallocated = PETSC_TRUE;
1228:   PetscFunctionReturn(PETSC_SUCCESS);
1229: }

1231:   #if PetscDefined(H2OPUS_USE_GPU)
1232: static PetscErrorCode MatBindToCPU_H2OPUS(Mat A, PetscBool flg)
1233: {
1234:   PetscMPIInt size;
1235:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;

1237:   PetscFunctionBegin;
1238:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1239:   if (flg && A->offloadmask == PETSC_OFFLOAD_GPU) {
1240:     if (size > 1) {
1241:       PetscCheck(a->dist_hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
1242:     #if defined(H2OPUS_USE_MPI)
1243:       if (!a->dist_hmatrix) a->dist_hmatrix = new DistributedHMatrix(*a->dist_hmatrix_gpu);
1244:       else *a->dist_hmatrix = *a->dist_hmatrix_gpu;
1245:     #endif
1246:     } else {
1247:       PetscCheck(a->hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
1248:       if (!a->hmatrix) a->hmatrix = new HMatrix(*a->hmatrix_gpu);
1249:       else *a->hmatrix = *a->hmatrix_gpu;
1250:     }
1251:     delete a->hmatrix_gpu;
1252:     delete a->dist_hmatrix_gpu;
1253:     a->hmatrix_gpu      = NULL;
1254:     a->dist_hmatrix_gpu = NULL;
1255:     A->offloadmask      = PETSC_OFFLOAD_CPU;
1256:   } else if (!flg && A->offloadmask == PETSC_OFFLOAD_CPU) {
1257:     if (size > 1) {
1258:       PetscCheck(a->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
1259:     #if defined(H2OPUS_USE_MPI)
1260:       if (!a->dist_hmatrix_gpu) a->dist_hmatrix_gpu = new DistributedHMatrix_GPU(*a->dist_hmatrix);
1261:       else *a->dist_hmatrix_gpu = *a->dist_hmatrix;
1262:     #endif
1263:     } else {
1264:       PetscCheck(a->hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
1265:       if (!a->hmatrix_gpu) a->hmatrix_gpu = new HMatrix_GPU(*a->hmatrix);
1266:       else *a->hmatrix_gpu = *a->hmatrix;
1267:     }
1268:     delete a->hmatrix;
1269:     delete a->dist_hmatrix;
1270:     a->hmatrix      = NULL;
1271:     a->dist_hmatrix = NULL;
1272:     A->offloadmask  = PETSC_OFFLOAD_GPU;
1273:   }
1274:   PetscCall(PetscFree(A->defaultvectype));
1275:   if (!flg) {
1276:     PetscCall(PetscStrallocpy(VECCUDA, &A->defaultvectype));
1277:   } else {
1278:     PetscCall(PetscStrallocpy(VECSTANDARD, &A->defaultvectype));
1279:   }
1280:   A->boundtocpu = flg;
1281:   PetscFunctionReturn(PETSC_SUCCESS);
1282: }
1283:   #endif

1285: /*MC
1286:    MATH2OPUS = "h2opus" - A matrix type for hierarchical matrices using the H2Opus package {cite}`zampinibouakaramturkiyyahkniokeyes2022`.

1288:    Options Database Key:
1289: .  -mat_type h2opus - matrix type to "h2opus"

1291:    Level: beginner

1293:    Notes:
1294:    H2Opus implements hierarchical matrices in the $H^2$ flavor. It supports CPU or NVIDIA GPUs.

1296:    For CPU only builds, use `./configure --download-h2opus --download-thrust` to install PETSc to use H2Opus.
1297:    In order to run on NVIDIA GPUs, use `./configure --download-h2opus --download-magma --download-kblas`.

1299: .seealso: [](ch_matrices), `Mat`, `MATH2OPUS`, `MATHTOOL`, `MATDENSE`, `MatCreateH2OpusFromKernel()`, `MatCreateH2OpusFromMat()`
1300: M*/
1301: PETSC_EXTERN PetscErrorCode MatCreate_H2OPUS(Mat A)
1302: {
1303:   Mat_H2OPUS *a;
1304:   PetscMPIInt size;

1306:   PetscFunctionBegin;
1307:   #if PetscDefined(H2OPUS_USE_GPU)
1308:   PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA));
1309:   #endif
1310:   PetscCall(PetscNew(&a));
1311:   A->data = (void *)a;

1313:   a->eta              = 0.9;
1314:   a->leafsize         = 32;
1315:   a->basisord         = 4;
1316:   a->max_rank         = 64;
1317:   a->bs               = 32;
1318:   a->rtol             = 1.e-4;
1319:   a->s                = 1.0;
1320:   a->norm_max_samples = 10;
1321:   a->resize           = PETSC_TRUE; /* reallocate after compression */
1322:   #if defined(H2OPUS_USE_MPI)
1323:   h2opusCreateDistributedHandleComm(&a->handle, PetscObjectComm((PetscObject)A));
1324:   #else
1325:   h2opusCreateHandle(&a->handle);
1326:   #endif
1327:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1328:   PetscCall(PetscObjectChangeTypeName((PetscObject)A, MATH2OPUS));
1329:   PetscCall(PetscMemzero(A->ops, sizeof(struct _MatOps)));

1331:   A->ops->destroy          = MatDestroy_H2OPUS;
1332:   A->ops->view             = MatView_H2OPUS;
1333:   A->ops->assemblyend      = MatAssemblyEnd_H2OPUS;
1334:   A->ops->mult             = MatMult_H2OPUS;
1335:   A->ops->multtranspose    = MatMultTranspose_H2OPUS;
1336:   A->ops->multadd          = MatMultAdd_H2OPUS;
1337:   A->ops->multtransposeadd = MatMultTransposeAdd_H2OPUS;
1338:   A->ops->scale            = MatScale_H2OPUS;
1339:   A->ops->duplicate        = MatDuplicate_H2OPUS;
1340:   A->ops->setfromoptions   = MatSetFromOptions_H2OPUS;
1341:   A->ops->norm             = MatNorm_H2OPUS;
1342:   A->ops->zeroentries      = MatZeroEntries_H2OPUS;
1343:   #if PetscDefined(H2OPUS_USE_GPU)
1344:   A->ops->bindtocpu = MatBindToCPU_H2OPUS;
1345:   #endif

1347:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_seqdense_C", MatProductSetFromOptions_H2OPUS));
1348:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_seqdensecuda_C", MatProductSetFromOptions_H2OPUS));
1349:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_mpidense_C", MatProductSetFromOptions_H2OPUS));
1350:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_h2opus_mpidensecuda_C", MatProductSetFromOptions_H2OPUS));
1351:   #if PetscDefined(H2OPUS_USE_GPU)
1352:   PetscCall(PetscFree(A->defaultvectype));
1353:   PetscCall(PetscStrallocpy(VECCUDA, &A->defaultvectype));
1354:   #endif
1355:   PetscFunctionReturn(PETSC_SUCCESS);
1356: }

1358: /*@
1359:   MatH2OpusOrthogonalize - Orthogonalize the basis tree of a hierarchical matrix.

1361:   Input Parameter:
1362: . A - the matrix

1364:   Level: intermediate

1366: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromMat()`, `MatCreateH2OpusFromKernel()`, `MatH2OpusCompress()`
1367: @*/
1368: PetscErrorCode MatH2OpusOrthogonalize(Mat A)
1369: {
1370:   PetscBool   ish2opus;
1371:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;
1372:   PetscMPIInt size;
1373:   PetscBool   boundtocpu = PETSC_TRUE;

1375:   PetscFunctionBegin;
1378:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
1379:   if (!ish2opus) PetscFunctionReturn(PETSC_SUCCESS);
1380:   if (a->orthogonal) PetscFunctionReturn(PETSC_SUCCESS);
1381:   HLibProfile::clear();
1382:   PetscCall(PetscLogEventBegin(MAT_H2Opus_Orthog, A, 0, 0, 0));
1383:   #if PetscDefined(H2OPUS_USE_GPU)
1384:   boundtocpu = A->boundtocpu;
1385:   #endif
1386:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1387:   if (size > 1) {
1388:     if (boundtocpu) {
1389:       PetscCheck(a->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
1390:   #if defined(H2OPUS_USE_MPI)
1391:       distributed_horthog(*a->dist_hmatrix, a->handle);
1392:   #endif
1393:   #if PetscDefined(H2OPUS_USE_GPU)
1394:       A->offloadmask = PETSC_OFFLOAD_CPU;
1395:     } else {
1396:       PetscCheck(a->dist_hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
1397:       PetscCall(PetscLogGpuTimeBegin());
1398:     #if defined(H2OPUS_USE_MPI)
1399:       distributed_horthog(*a->dist_hmatrix_gpu, a->handle);
1400:     #endif
1401:       PetscCall(PetscLogGpuTimeEnd());
1402:   #endif
1403:     }
1404:   } else {
1405:   #if defined(H2OPUS_USE_MPI)
1406:     h2opusHandle_t handle = a->handle->handle;
1407:   #else
1408:     h2opusHandle_t handle = a->handle;
1409:   #endif
1410:     if (boundtocpu) {
1411:       PetscCheck(a->hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
1412:       horthog(*a->hmatrix, handle);
1413:   #if PetscDefined(H2OPUS_USE_GPU)
1414:       A->offloadmask = PETSC_OFFLOAD_CPU;
1415:     } else {
1416:       PetscCheck(a->hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
1417:       PetscCall(PetscLogGpuTimeBegin());
1418:       horthog(*a->hmatrix_gpu, handle);
1419:       PetscCall(PetscLogGpuTimeEnd());
1420:   #endif
1421:     }
1422:   }
1423:   a->orthogonal = PETSC_TRUE;
1424:   { /* log flops */
1425:     double gops, time, perf, dev;
1426:     HLibProfile::getHorthogPerf(gops, time, perf, dev);
1427:   #if PetscDefined(H2OPUS_USE_GPU)
1428:     if (boundtocpu) PetscCall(PetscLogFlops(1e9 * gops));
1429:     else PetscCall(PetscLogGpuFlops(1e9 * gops));
1430:   #else
1431:     PetscCall(PetscLogFlops(1e9 * gops));
1432:   #endif
1433:   }
1434:   PetscCall(PetscLogEventEnd(MAT_H2Opus_Orthog, A, 0, 0, 0));
1435:   PetscFunctionReturn(PETSC_SUCCESS);
1436: }

1438: /*@
1439:   MatH2OpusCompress - Compress a hierarchical matrix.

1441:   Input Parameters:
1442: + A   - the matrix
1443: - tol - the absolute truncation threshold

1445:   Level: intermediate

1447: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromMat()`, `MatCreateH2OpusFromKernel()`, `MatH2OpusOrthogonalize()`
1448: @*/
1449: PetscErrorCode MatH2OpusCompress(Mat A, PetscReal tol)
1450: {
1451:   PetscBool   ish2opus;
1452:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;
1453:   PetscMPIInt size;
1454:   PetscBool   boundtocpu = PETSC_TRUE;

1456:   PetscFunctionBegin;
1460:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
1461:   if (!ish2opus || tol <= 0.0) PetscFunctionReturn(PETSC_SUCCESS);
1462:   PetscCall(MatH2OpusOrthogonalize(A));
1463:   HLibProfile::clear();
1464:   PetscCall(PetscLogEventBegin(MAT_H2Opus_Compress, A, 0, 0, 0));
1465:   #if PetscDefined(H2OPUS_USE_GPU)
1466:   boundtocpu = A->boundtocpu;
1467:   #endif
1468:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1469:   if (size > 1) {
1470:     if (boundtocpu) {
1471:       PetscCheck(a->dist_hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
1472:   #if defined(H2OPUS_USE_MPI)
1473:       distributed_hcompress(*a->dist_hmatrix, tol, a->handle);
1474:       if (a->resize) {
1475:         DistributedHMatrix *dist_hmatrix = new DistributedHMatrix(*a->dist_hmatrix);
1476:         delete a->dist_hmatrix;
1477:         a->dist_hmatrix = dist_hmatrix;
1478:       }
1479:   #endif
1480:   #if PetscDefined(H2OPUS_USE_GPU)
1481:       A->offloadmask = PETSC_OFFLOAD_CPU;
1482:     } else {
1483:       PetscCheck(a->dist_hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
1484:       PetscCall(PetscLogGpuTimeBegin());
1485:     #if defined(H2OPUS_USE_MPI)
1486:       distributed_hcompress(*a->dist_hmatrix_gpu, tol, a->handle);

1488:       if (a->resize) {
1489:         DistributedHMatrix_GPU *dist_hmatrix_gpu = new DistributedHMatrix_GPU(*a->dist_hmatrix_gpu);
1490:         delete a->dist_hmatrix_gpu;
1491:         a->dist_hmatrix_gpu = dist_hmatrix_gpu;
1492:       }
1493:     #endif
1494:       PetscCall(PetscLogGpuTimeEnd());
1495:   #endif
1496:     }
1497:   } else {
1498:   #if defined(H2OPUS_USE_MPI)
1499:     h2opusHandle_t handle = a->handle->handle;
1500:   #else
1501:     h2opusHandle_t handle = a->handle;
1502:   #endif
1503:     if (boundtocpu) {
1504:       PetscCheck(a->hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
1505:       hcompress(*a->hmatrix, tol, handle);

1507:       if (a->resize) {
1508:         HMatrix *hmatrix = new HMatrix(*a->hmatrix);
1509:         delete a->hmatrix;
1510:         a->hmatrix = hmatrix;
1511:       }
1512:   #if PetscDefined(H2OPUS_USE_GPU)
1513:       A->offloadmask = PETSC_OFFLOAD_CPU;
1514:     } else {
1515:       PetscCheck(a->hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
1516:       PetscCall(PetscLogGpuTimeBegin());
1517:       hcompress(*a->hmatrix_gpu, tol, handle);
1518:       PetscCall(PetscLogGpuTimeEnd());

1520:       if (a->resize) {
1521:         HMatrix_GPU *hmatrix_gpu = new HMatrix_GPU(*a->hmatrix_gpu);
1522:         delete a->hmatrix_gpu;
1523:         a->hmatrix_gpu = hmatrix_gpu;
1524:       }
1525:   #endif
1526:     }
1527:   }
1528:   { /* log flops */
1529:     double gops, time, perf, dev;
1530:     HLibProfile::getHcompressPerf(gops, time, perf, dev);
1531:   #if PetscDefined(H2OPUS_USE_GPU)
1532:     if (boundtocpu) PetscCall(PetscLogFlops(1e9 * gops));
1533:     else PetscCall(PetscLogGpuFlops(1e9 * gops));
1534:   #else
1535:     PetscCall(PetscLogFlops(1e9 * gops));
1536:   #endif
1537:   }
1538:   PetscCall(PetscLogEventEnd(MAT_H2Opus_Compress, A, 0, 0, 0));
1539:   PetscFunctionReturn(PETSC_SUCCESS);
1540: }

1542: /*@
1543:   MatH2OpusSetSamplingMat - Set a matrix to be sampled from matrix-vector products on another matrix to construct a hierarchical matrix.

1545:   Input Parameters:
1546: + A   - the hierarchical matrix
1547: . B   - the matrix to be sampled
1548: . bs  - maximum number of samples to be taken concurrently
1549: - tol - relative tolerance for construction

1551:   Level: intermediate

1553:   Notes:
1554:   You need to call `MatAssemblyBegin()` and `MatAssemblyEnd()` to update the hierarchical matrix.

1556: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromMat()`, `MatCreateH2OpusFromKernel()`, `MatH2OpusCompress()`, `MatH2OpusOrthogonalize()`
1557: @*/
1558: PetscErrorCode MatH2OpusSetSamplingMat(Mat A, Mat B, PetscInt bs, PetscReal tol)
1559: {
1560:   PetscBool ish2opus;

1562:   PetscFunctionBegin;
1568:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
1569:   if (ish2opus) {
1570:     Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;

1572:     if (!a->sampler) a->sampler = new PetscMatrixSampler();
1573:     a->sampler->SetSamplingMat(B);
1574:     if (bs > 0) a->bs = bs;
1575:     if (tol > 0.) a->rtol = tol;
1576:     delete a->kernel;
1577:   }
1578:   PetscFunctionReturn(PETSC_SUCCESS);
1579: }

1581: /*@C
1582:   MatCreateH2OpusFromKernel - Creates a `MATH2OPUS` from a user-supplied kernel.

1584:   Input Parameters:
1585: + comm      - MPI communicator
1586: . m         - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
1587: . n         - number of local columns (or `PETSC_DECIDE` to have calculated if `N` is given)
1588: . M         - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
1589: . N         - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
1590: . spacedim  - dimension of the space coordinates
1591: . coords    - coordinates of the points
1592: . cdist     - whether or not coordinates are distributed
1593: . kernel    - computational kernel (or `NULL`)
1594: . kernelctx - kernel context
1595: . eta       - admissibility condition tolerance
1596: . leafsize  - leaf size in cluster tree
1597: - basisord  - approximation order for Chebychev interpolation of low-rank blocks

1599:   Output Parameter:
1600: . nA - matrix

1602:   Options Database Keys:
1603: + -mat_h2opus_leafsize <`PetscInt`>    - Leaf size of cluster tree
1604: . -mat_h2opus_eta <`PetscReal`>        - Admissibility condition tolerance
1605: . -mat_h2opus_order <`PetscInt`>       - Chebychev approximation order
1606: - -mat_h2opus_normsamples <`PetscInt`> - Maximum number of samples to be used when estimating norms

1608:   Level: intermediate

1610: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromMat()`
1611: @*/
1612: PetscErrorCode MatCreateH2OpusFromKernel(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscInt spacedim, const PetscReal coords[], PetscBool cdist, MatH2OpusKernelFn *kernel, void *kernelctx, PetscReal eta, PetscInt leafsize, PetscInt basisord, Mat *nA)
1613: {
1614:   Mat         A;
1615:   Mat_H2OPUS *h2opus;
1616:   PetscBool   iscpu = PetscDefined(H2OPUS_USE_GPU) ? PETSC_FALSE : PETSC_TRUE;

1618:   PetscFunctionBegin;
1619:   PetscCheck(m == n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Different row and column local sizes are not supported");
1620:   PetscCall(MatCreate(comm, &A));
1621:   PetscCall(MatSetSizes(A, m, n, M, N));
1622:   PetscCheck(M == N, comm, PETSC_ERR_SUP, "Rectangular matrices are not supported");
1623:   PetscCall(MatSetType(A, MATH2OPUS));
1624:   PetscCall(MatBindToCPU(A, iscpu));
1625:   PetscCall(MatH2OpusSetCoords_H2OPUS(A, spacedim, coords, cdist, kernel, kernelctx));

1627:   h2opus = (Mat_H2OPUS *)A->data;
1628:   if (eta > 0.) h2opus->eta = eta;
1629:   if (leafsize > 0) h2opus->leafsize = leafsize;
1630:   if (basisord > 0) h2opus->basisord = basisord;

1632:   *nA = A;
1633:   PetscFunctionReturn(PETSC_SUCCESS);
1634: }

1636: /*@
1637:   MatCreateH2OpusFromMat - Creates a `MATH2OPUS` sampling from a user-supplied operator.

1639:   Input Parameters:
1640: + B        - the matrix to be sampled
1641: . spacedim - dimension of the space coordinates
1642: . coords   - coordinates of the points
1643: . cdist    - whether or not coordinates are distributed
1644: . eta      - admissibility condition tolerance
1645: . leafsize - leaf size in cluster tree
1646: . maxrank  - maximum rank allowed
1647: . bs       - maximum number of samples to be taken concurrently
1648: - rtol     - relative tolerance for construction

1650:   Output Parameter:
1651: . nA - matrix

1653:   Options Database Keys:
1654: + -mat_h2opus_leafsize <`PetscInt`>      - Leaf size of cluster tree
1655: . -mat_h2opus_eta <`PetscReal`>          - Admissibility condition tolerance
1656: . -mat_h2opus_maxrank <`PetscInt`>       - Maximum rank when constructed from matvecs
1657: . -mat_h2opus_samples <`PetscInt`>       - Maximum number of samples to be taken concurrently when constructing from matvecs
1658: . -mat_h2opus_rtol <`PetscReal`>         - Relative tolerance for construction from sampling
1659: . -mat_h2opus_check <`PetscBool`>        - Check error when constructing from sampling during MatAssemblyEnd()
1660: . -mat_h2opus_hara_verbose <`PetscBool`> - Verbose output from hara construction
1661: - -mat_h2opus_normsamples <`PetscInt`>   - Maximum number of samples to be when estimating norms

1663:   Level: intermediate

1665:   Note:
1666:   Not available in parallel

1668: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromKernel()`
1669: @*/
1670: PetscErrorCode MatCreateH2OpusFromMat(Mat B, PetscInt spacedim, const PetscReal coords[], PetscBool cdist, PetscReal eta, PetscInt leafsize, PetscInt maxrank, PetscInt bs, PetscReal rtol, Mat *nA)
1671: {
1672:   Mat         A;
1673:   Mat_H2OPUS *h2opus;
1674:   MPI_Comm    comm;
1675:   PetscBool   boundtocpu = PETSC_TRUE;

1677:   PetscFunctionBegin;
1686:   PetscAssertPointer(nA, 10);
1687:   PetscCall(PetscObjectGetComm((PetscObject)B, &comm));
1688:   PetscCheck(B->rmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Different row and column local sizes are not supported");
1689:   PetscCheck(B->rmap->N == B->cmap->N, comm, PETSC_ERR_SUP, "Rectangular matrices are not supported");
1690:   PetscCall(MatCreate(comm, &A));
1691:   PetscCall(MatSetSizes(A, B->rmap->n, B->cmap->n, B->rmap->N, B->cmap->N));
1692:   #if PetscDefined(H2OPUS_USE_GPU)
1693:   {
1694:     VecType   vtype;
1695:     PetscBool isstd, iscuda, iskok;

1697:     PetscCall(MatGetVecType(B, &vtype));
1698:     PetscCall(PetscStrcmpAny(vtype, &isstd, VECSTANDARD, VECSEQ, VECMPI, ""));
1699:     PetscCall(PetscStrcmpAny(vtype, &iscuda, VECCUDA, VECSEQCUDA, VECMPICUDA, ""));
1700:     PetscCall(PetscStrcmpAny(vtype, &iskok, VECKOKKOS, VECSEQKOKKOS, VECMPIKOKKOS, ""));
1701:     PetscCheck(isstd || iscuda || iskok, comm, PETSC_ERR_SUP, "Not for type %s", vtype);
1702:     if (iscuda && !B->boundtocpu) boundtocpu = PETSC_FALSE;
1703:     if (iskok && PetscDefined(HAVE_MACRO_KOKKOS_ENABLE_CUDA)) boundtocpu = PETSC_FALSE;
1704:   }
1705:   #endif
1706:   PetscCall(MatSetType(A, MATH2OPUS));
1707:   PetscCall(MatBindToCPU(A, boundtocpu));
1708:   if (spacedim) PetscCall(MatH2OpusSetCoords_H2OPUS(A, spacedim, coords, cdist, NULL, NULL));
1709:   PetscCall(MatPropagateSymmetryOptions(B, A));
1710:   /* PetscCheck(A->symmetric,comm,PETSC_ERR_SUP,"Unsymmetric sampling does not work"); */

1712:   h2opus          = (Mat_H2OPUS *)A->data;
1713:   h2opus->sampler = new PetscMatrixSampler(B);
1714:   if (eta > 0.) h2opus->eta = eta;
1715:   if (leafsize > 0) h2opus->leafsize = leafsize;
1716:   if (maxrank > 0) h2opus->max_rank = maxrank;
1717:   if (bs > 0) h2opus->bs = bs;
1718:   if (rtol > 0.) h2opus->rtol = rtol;
1719:   *nA             = A;
1720:   A->preallocated = PETSC_TRUE;
1721:   PetscFunctionReturn(PETSC_SUCCESS);
1722: }

1724: /*@
1725:   MatH2OpusGetIndexMap - Access reordering index set.

1727:   Input Parameter:
1728: . A - the matrix

1730:   Output Parameter:
1731: . indexmap - the index set for the reordering

1733:   Level: intermediate

1735: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromMat()`, `MatCreateH2OpusFromKernel()`
1736: @*/
1737: PetscErrorCode MatH2OpusGetIndexMap(Mat A, IS *indexmap)
1738: {
1739:   PetscBool   ish2opus;
1740:   Mat_H2OPUS *a = (Mat_H2OPUS *)A->data;

1742:   PetscFunctionBegin;
1745:   PetscAssertPointer(indexmap, 2);
1746:   PetscCheck(A->assembled, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
1747:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
1748:   PetscCheck(ish2opus, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
1749:   *indexmap = a->h2opus_indexmap;
1750:   PetscFunctionReturn(PETSC_SUCCESS);
1751: }

1753: /*@
1754:   MatH2OpusMapVec - Maps a vector between PETSc and H2Opus ordering

1756:   Input Parameters:
1757: + A             - the matrix
1758: . nativetopetsc - if true, maps from H2Opus ordering to PETSc ordering. If false, applies the reverse map
1759: - in            - the vector to be mapped

1761:   Output Parameter:
1762: . out - the newly created mapped vector

1764:   Level: intermediate

1766: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromMat()`, `MatCreateH2OpusFromKernel()`
1767: @*/
1768: PetscErrorCode MatH2OpusMapVec(Mat A, PetscBool nativetopetsc, Vec in, Vec *out)
1769: {
1770:   PetscBool    ish2opus;
1771:   Mat_H2OPUS  *a = (Mat_H2OPUS *)A->data;
1772:   PetscScalar *xin, *xout;
1773:   PetscBool    nm;

1775:   PetscFunctionBegin;
1780:   PetscAssertPointer(out, 4);
1781:   PetscCheck(A->assembled, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
1782:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &ish2opus));
1783:   PetscCheck(ish2opus, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
1784:   nm = a->nativemult;
1785:   PetscCall(MatH2OpusSetNativeMult(A, (PetscBool)!nativetopetsc));
1786:   PetscCall(MatCreateVecs(A, out, NULL));
1787:   PetscCall(MatH2OpusSetNativeMult(A, nm));
1788:   if (!a->sf) { /* same ordering */
1789:     PetscCall(VecCopy(in, *out));
1790:     PetscFunctionReturn(PETSC_SUCCESS);
1791:   }
1792:   PetscCall(VecGetArrayRead(in, (const PetscScalar **)&xin));
1793:   PetscCall(VecGetArrayWrite(*out, &xout));
1794:   if (nativetopetsc) {
1795:     PetscCall(PetscSFReduceBegin(a->sf, MPIU_SCALAR, xin, xout, MPI_REPLACE));
1796:     PetscCall(PetscSFReduceEnd(a->sf, MPIU_SCALAR, xin, xout, MPI_REPLACE));
1797:   } else {
1798:     PetscCall(PetscSFBcastBegin(a->sf, MPIU_SCALAR, xin, xout, MPI_REPLACE));
1799:     PetscCall(PetscSFBcastEnd(a->sf, MPIU_SCALAR, xin, xout, MPI_REPLACE));
1800:   }
1801:   PetscCall(VecRestoreArrayRead(in, (const PetscScalar **)&xin));
1802:   PetscCall(VecRestoreArrayWrite(*out, &xout));
1803:   PetscFunctionReturn(PETSC_SUCCESS);
1804: }

1806: /*@
1807:   MatH2OpusLowRankUpdate - Perform a low-rank update of the form $ A = A + s * U * V^T $

1809:   Input Parameters:
1810: + A - the hierarchical `MATH2OPUS` matrix
1811: . s - the scaling factor
1812: . U - the dense low-rank update matrix
1813: - V - (optional) the dense low-rank update matrix (if `NULL`, then `V` = `U` is assumed)

1815:   Note:
1816:   The `U` and `V` matrices must be in `MATDENSE` dense format

1818:   Level: intermediate

1820: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MATH2OPUS`, `MatCreateH2OpusFromMat()`, `MatCreateH2OpusFromKernel()`, `MatH2OpusCompress()`, `MatH2OpusOrthogonalize()`, `MATDENSE`
1821: @*/
1822: PetscErrorCode MatH2OpusLowRankUpdate(Mat A, Mat U, Mat V, PetscScalar s)
1823: {
1824:   PetscBool flg;

1826:   PetscFunctionBegin;
1829:   PetscCheck(A->assembled, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
1831:   PetscCheckSameComm(A, 1, U, 2);
1832:   if (V) {
1834:     PetscCheckSameComm(A, 1, V, 3);
1835:   }

1838:   if (!V) V = U;
1839:   PetscCheck(U->cmap->N == V->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Non matching rank update %" PetscInt_FMT " != %" PetscInt_FMT, U->cmap->N, V->cmap->N);
1840:   if (!U->cmap->N) PetscFunctionReturn(PETSC_SUCCESS);
1841:   PetscCall(PetscLayoutCompare(U->rmap, A->rmap, &flg));
1842:   PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "A and U must have the same row layout");
1843:   PetscCall(PetscLayoutCompare(V->rmap, A->cmap, &flg));
1844:   PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "A column layout must match V row column layout");
1845:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATH2OPUS, &flg));
1846:   if (flg) {
1847:     Mat_H2OPUS        *a = (Mat_H2OPUS *)A->data;
1848:     const PetscScalar *u, *v, *uu, *vv;
1849:     PetscInt           ldu, ldv;
1850:     PetscMPIInt        size;
1851:   #if defined(H2OPUS_USE_MPI)
1852:     h2opusHandle_t handle = a->handle->handle;
1853:   #else
1854:     h2opusHandle_t handle = a->handle;
1855:   #endif
1856:     PetscBool usesf = (PetscBool)(a->sf && !a->nativemult);
1857:     PetscSF   usf, vsf;

1859:     PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1860:     PetscCheck(size <= 1, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not yet implemented in parallel");
1861:     PetscCall(PetscLogEventBegin(MAT_H2Opus_LR, A, 0, 0, 0));
1862:     PetscCall(PetscObjectBaseTypeCompareAny((PetscObject)U, &flg, MATSEQDENSE, MATMPIDENSE, ""));
1863:     PetscCheck(flg, PetscObjectComm((PetscObject)U), PETSC_ERR_SUP, "Not for U of type %s", ((PetscObject)U)->type_name);
1864:     PetscCall(PetscObjectBaseTypeCompareAny((PetscObject)V, &flg, MATSEQDENSE, MATMPIDENSE, ""));
1865:     PetscCheck(flg, PetscObjectComm((PetscObject)V), PETSC_ERR_SUP, "Not for V of type %s", ((PetscObject)V)->type_name);
1866:     PetscCall(MatDenseGetLDA(U, &ldu));
1867:     PetscCall(MatDenseGetLDA(V, &ldv));
1868:     PetscCall(MatBoundToCPU(A, &flg));
1869:     if (usesf) {
1870:       PetscInt n;

1872:       PetscCall(MatDenseGetH2OpusStridedSF(U, a->sf, &usf));
1873:       PetscCall(MatDenseGetH2OpusStridedSF(V, a->sf, &vsf));
1874:       PetscCall(MatH2OpusResizeBuffers_Private(A, U->cmap->N, V->cmap->N));
1875:       PetscCall(PetscSFGetGraph(a->sf, NULL, &n, NULL, NULL));
1876:       ldu = n;
1877:       ldv = n;
1878:     }
1879:     if (flg) {
1880:       PetscCheck(a->hmatrix, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing CPU matrix");
1881:       PetscCall(MatDenseGetArrayRead(U, &u));
1882:       PetscCall(MatDenseGetArrayRead(V, &v));
1883:       if (usesf) {
1884:         vv = MatH2OpusGetThrustPointer(*a->yy);
1885:         PetscCall(PetscSFBcastBegin(vsf, MPIU_SCALAR, v, (PetscScalar *)vv, MPI_REPLACE));
1886:         PetscCall(PetscSFBcastEnd(vsf, MPIU_SCALAR, v, (PetscScalar *)vv, MPI_REPLACE));
1887:         if (U != V) {
1888:           uu = MatH2OpusGetThrustPointer(*a->xx);
1889:           PetscCall(PetscSFBcastBegin(usf, MPIU_SCALAR, u, (PetscScalar *)uu, MPI_REPLACE));
1890:           PetscCall(PetscSFBcastEnd(usf, MPIU_SCALAR, u, (PetscScalar *)uu, MPI_REPLACE));
1891:         } else uu = vv;
1892:       } else {
1893:         uu = u;
1894:         vv = v;
1895:       }
1896:       hlru_global(*a->hmatrix, uu, ldu, vv, ldv, U->cmap->N, s, handle);
1897:       PetscCall(MatDenseRestoreArrayRead(U, &u));
1898:       PetscCall(MatDenseRestoreArrayRead(V, &v));
1899:     } else {
1900:   #if PetscDefined(H2OPUS_USE_GPU)
1901:       PetscBool flgU, flgV;

1903:       PetscCheck(a->hmatrix_gpu, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing GPU matrix");
1904:       PetscCall(PetscObjectTypeCompareAny((PetscObject)U, &flgU, MATSEQDENSE, MATMPIDENSE, ""));
1905:       if (flgU) PetscCall(MatConvert(U, MATDENSECUDA, MAT_INPLACE_MATRIX, &U));
1906:       PetscCall(PetscObjectTypeCompareAny((PetscObject)V, &flgV, MATSEQDENSE, MATMPIDENSE, ""));
1907:       if (flgV) PetscCall(MatConvert(V, MATDENSECUDA, MAT_INPLACE_MATRIX, &V));
1908:       PetscCall(MatDenseCUDAGetArrayRead(U, &u));
1909:       PetscCall(MatDenseCUDAGetArrayRead(V, &v));
1910:       if (usesf) {
1911:         vv = MatH2OpusGetThrustPointer(*a->yy_gpu);
1912:         PetscCall(PetscSFBcastBegin(vsf, MPIU_SCALAR, v, (PetscScalar *)vv, MPI_REPLACE));
1913:         PetscCall(PetscSFBcastEnd(vsf, MPIU_SCALAR, v, (PetscScalar *)vv, MPI_REPLACE));
1914:         if (U != V) {
1915:           uu = MatH2OpusGetThrustPointer(*a->xx_gpu);
1916:           PetscCall(PetscSFBcastBegin(usf, MPIU_SCALAR, u, (PetscScalar *)uu, MPI_REPLACE));
1917:           PetscCall(PetscSFBcastEnd(usf, MPIU_SCALAR, u, (PetscScalar *)uu, MPI_REPLACE));
1918:         } else uu = vv;
1919:       } else {
1920:         uu = u;
1921:         vv = v;
1922:       }
1923:   #else
1924:       SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "This should not happen");
1925:   #endif
1926:       hlru_global(*a->hmatrix_gpu, uu, ldu, vv, ldv, U->cmap->N, s, handle);
1927:   #if PetscDefined(H2OPUS_USE_GPU)
1928:       PetscCall(MatDenseCUDARestoreArrayRead(U, &u));
1929:       PetscCall(MatDenseCUDARestoreArrayRead(V, &v));
1930:       if (flgU) PetscCall(MatConvert(U, MATDENSE, MAT_INPLACE_MATRIX, &U));
1931:       if (flgV) PetscCall(MatConvert(V, MATDENSE, MAT_INPLACE_MATRIX, &V));
1932:   #endif
1933:     }
1934:     PetscCall(PetscLogEventEnd(MAT_H2Opus_LR, A, 0, 0, 0));
1935:     a->orthogonal = PETSC_FALSE;
1936:   }
1937:   PetscFunctionReturn(PETSC_SUCCESS);
1938: }
1939: #endif