Actual source code: vscatfce.c

petsc-master 2019-09-15
Report Typos and Errors
  1:  #include <petsc/private/vecscatterimpl.h>
  2: #if defined(PETSC_HAVE_CUDA)
  3:  #include <../src/vec/vec/impls/seq/seqcuda/cudavecimpl.h>
  4: #endif
  5: /* ------------------------------------------------------------------*/
  6: /*@
  7:    VecScatterGetMerged - Returns true if the scatter is completed in the VecScatterBegin()
  8:       and the VecScatterEnd() does nothing

 10:    Not Collective

 12:    Input Parameter:
 13: .   ctx - scatter context created with VecScatterCreate()

 15:    Output Parameter:
 16: .   flg - PETSC_TRUE if the VecScatterBegin/End() are all done during the VecScatterBegin()

 18:    Level: developer

 20: .seealso: VecScatterCreate(), VecScatterEnd(), VecScatterBegin()
 21: @*/
 22: PetscErrorCode  VecScatterGetMerged(VecScatter ctx,PetscBool  *flg)
 23: {
 26:   *flg = ctx->beginandendtogether;
 27:   return(0);
 28: }

 30: /*@
 31:    VecScatterBegin - Begins a generalized scatter from one vector to
 32:    another. Complete the scattering phase with VecScatterEnd().

 34:    Neighbor-wise Collective on VecScatter

 36:    Input Parameters:
 37: +  ctx - scatter context generated by VecScatterCreate()
 38: .  x - the vector from which we scatter
 39: .  y - the vector to which we scatter
 40: .  addv - either ADD_VALUES or INSERT_VALUES, with INSERT_VALUES mode any location
 41:           not scattered to retains its old value; i.e. the vector is NOT first zeroed.
 42: -  mode - the scattering mode, usually SCATTER_FORWARD.  The available modes are:
 43:     SCATTER_FORWARD or SCATTER_REVERSE


 46:    Level: intermediate

 48:    Options Database: See VecScatterCreate()

 50:    Notes:
 51:    The vectors x and y need not be the same vectors used in the call
 52:    to VecScatterCreate(), but x must have the same parallel data layout
 53:    as that passed in as the x to VecScatterCreate(), similarly for the y.
 54:    Most likely they have been obtained from VecDuplicate().

 56:    You cannot change the values in the input vector between the calls to VecScatterBegin()
 57:    and VecScatterEnd().

 59:    If you use SCATTER_REVERSE the two arguments x and y should be reversed, from
 60:    the SCATTER_FORWARD.

 62:    y[iy[i]] = x[ix[i]], for i=0,...,ni-1

 64:    This scatter is far more general than the conventional
 65:    scatter, since it can be a gather or a scatter or a combination,
 66:    depending on the indices ix and iy.  If x is a parallel vector and y
 67:    is sequential, VecScatterBegin() can serve to gather values to a
 68:    single processor.  Similarly, if y is parallel and x sequential, the
 69:    routine can scatter from one processor to many processors.


 72: .seealso: VecScatterCreate(), VecScatterEnd()
 73: @*/
 74: PetscErrorCode  VecScatterBegin(VecScatter ctx,Vec x,Vec y,InsertMode addv,ScatterMode mode)
 75: {
 77: #if defined(PETSC_USE_DEBUG)
 78:   PetscInt       to_n,from_n;
 79: #endif
 84:   if (ctx->inuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE," Scatter ctx already in use");

 86: #if defined(PETSC_USE_DEBUG)
 87:   /*
 88:      Error checking to make sure these vectors match the vectors used
 89:    to create the vector scatter context. -1 in the from_n and to_n indicate the
 90:    vector lengths are unknown (for example with mapped scatters) and thus
 91:    no error checking is performed.
 92:   */
 93:   if (ctx->from_n >= 0 && ctx->to_n >= 0) {
 94:     VecGetLocalSize(x,&from_n);
 95:     VecGetLocalSize(y,&to_n);
 96:     if (mode & SCATTER_REVERSE) {
 97:       if (to_n != ctx->from_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Vector wrong size %D for scatter %D (scatter reverse and vector to != ctx from size)",to_n,ctx->from_n);
 98:       if (from_n != ctx->to_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Vector wrong size %D for scatter %D (scatter reverse and vector from != ctx to size)",from_n,ctx->to_n);
 99:     } else {
100:       if (to_n != ctx->to_n)     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Vector wrong size %D for scatter %D (scatter forward and vector to != ctx to size)",to_n,ctx->to_n);
101:       if (from_n != ctx->from_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Vector wrong size %D for scatter %D (scatter forward and vector from != ctx from size)",from_n,ctx->from_n);
102:     }
103:   }
104: #endif

106:   ctx->inuse = PETSC_TRUE;
107:   PetscLogEventBegin(VEC_ScatterBegin,ctx,x,y,0);
108:   (*ctx->ops->begin)(ctx,x,y,addv,mode);
109:   if (ctx->beginandendtogether && ctx->ops->end) {
110:     ctx->inuse = PETSC_FALSE;
111:     (*ctx->ops->end)(ctx,x,y,addv,mode);
112:   }
113:   PetscLogEventEnd(VEC_ScatterBegin,ctx,x,y,0);
114:   return(0);
115: }

117: /* --------------------------------------------------------------------*/
118: /*@
119:    VecScatterEnd - Ends a generalized scatter from one vector to another.  Call
120:    after first calling VecScatterBegin().

122:    Neighbor-wise Collective on VecScatter

124:    Input Parameters:
125: +  ctx - scatter context generated by VecScatterCreate()
126: .  x - the vector from which we scatter
127: .  y - the vector to which we scatter
128: .  addv - either ADD_VALUES or INSERT_VALUES.
129: -  mode - the scattering mode, usually SCATTER_FORWARD.  The available modes are:
130:      SCATTER_FORWARD, SCATTER_REVERSE

132:    Level: intermediate

134:    Notes:
135:    If you use SCATTER_REVERSE the arguments x and y should be reversed, from the SCATTER_FORWARD.

137:    y[iy[i]] = x[ix[i]], for i=0,...,ni-1

139: .seealso: VecScatterBegin(), VecScatterCreate()
140: @*/
141: PetscErrorCode  VecScatterEnd(VecScatter ctx,Vec x,Vec y,InsertMode addv,ScatterMode mode)
142: {

149:   ctx->inuse = PETSC_FALSE;
150:   if (!ctx->ops->end) return(0);
151:   if (!ctx->beginandendtogether) {
152:     PetscLogEventBegin(VEC_ScatterEnd,ctx,x,y,0);
153:     (*(ctx)->ops->end)(ctx,x,y,addv,mode);
154:     PetscLogEventEnd(VEC_ScatterEnd,ctx,x,y,0);
155:   }
156:   return(0);
157: }

159: /*@
160:    VecScatterDestroy - Destroys a scatter context created by VecScatterCreate()

162:    Collective on VecScatter

164:    Input Parameter:
165: .  ctx - the scatter context

167:    Level: intermediate

169: .seealso: VecScatterCreate(), VecScatterCopy()
170: @*/
171: PetscErrorCode VecScatterDestroy(VecScatter *ctx)
172: {

176:   if (!*ctx) return(0);
178:   if ((*ctx)->inuse && ((PetscObject)(*ctx))->refct == 1) SETERRQ(((PetscObject)(*ctx))->comm,PETSC_ERR_ARG_WRONGSTATE,"Scatter context is in use");
179:   if (--((PetscObject)(*ctx))->refct > 0) {*ctx = 0; return(0);}

181:   /* if memory was published with SAWs then destroy it */
182:   PetscObjectSAWsViewOff((PetscObject)(*ctx));
183:   if ((*ctx)->ops->destroy) {(*(*ctx)->ops->destroy)(*ctx);}
184: #if defined(PETSC_HAVE_CUDA)
185:   VecScatterCUDAIndicesDestroy((PetscCUDAIndices*)&((*ctx)->spptr));
186: #endif
187:   PetscHeaderDestroy(ctx);
188:   return(0);
189: }

191: /*@
192:    VecScatterSetUp - Sets up the VecScatter to be able to actually scatter information between vectors

194:    Collective on VecScatter

196:    Input Parameter:
197: .  ctx - the scatter context

199:    Level: intermediate

201: .seealso: VecScatterCreate(), VecScatterCopy()
202: @*/
203: PetscErrorCode VecScatterSetUp(VecScatter ctx)
204: {

209:   (*ctx->ops->setup)(ctx);
210:   return(0);
211: }

213: /*@
214:    VecScatterCopy - Makes a copy of a scatter context.

216:    Collective on VecScatter

218:    Input Parameter:
219: .  sctx - the scatter context

221:    Output Parameter:
222: .  ctx - the context copy

224:    Level: advanced

226: .seealso: VecScatterCreate(), VecScatterDestroy()
227: @*/
228: PetscErrorCode  VecScatterCopy(VecScatter sctx,VecScatter *ctx)
229: {
231:   VecScatterType type;

236:   if (!sctx->ops->copy) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot copy this type");
237:   PetscHeaderCreate(*ctx,VEC_SCATTER_CLASSID,"VecScatter","VecScatter","Vec",PetscObjectComm((PetscObject)sctx),VecScatterDestroy,VecScatterView);
238:   (*ctx)->to_n   = sctx->to_n;
239:   (*ctx)->from_n = sctx->from_n;
240:   (*sctx->ops->copy)(sctx,*ctx);

242:   VecScatterGetType(sctx,&type);
243:   PetscObjectChangeTypeName((PetscObject)(*ctx),type);
244:   return(0);
245: }

247: /* ------------------------------------------------------------------*/
248: /*@C
249:    VecScatterView - Views a vector scatter context.

251:    Collective on VecScatter

253:    Input Parameters:
254: +  ctx - the scatter context
255: -  viewer - the viewer for displaying the context

257:    Level: intermediate

259: @*/
260: PetscErrorCode  VecScatterView(VecScatter ctx,PetscViewer viewer)
261: {

266:   if (!viewer) {
267:     PetscViewerASCIIGetStdout(PetscObjectComm((PetscObject)ctx),&viewer);
268:   }
270:   if (ctx->ops->view) {
271:     (*ctx->ops->view)(ctx,viewer);
272:   }
273:   return(0);
274: }

276: /*@C
277:    VecScatterRemap - Remaps the "from" and "to" indices in a
278:    vector scatter context. FOR EXPERTS ONLY!

280:    Collective on VecScatter

282:    Input Parameters:
283: +  scat    - vector scatter context
284: .  tomap   - remapping plan for "to" indices (may be NULL).
285: -  frommap - remapping plan for "from" indices (may be NULL)

287:    Level: developer

289:    Notes:
290:      In the parallel case the todata contains indices from where the data is taken
291:      (and then sent to others)! The fromdata contains indices from where the received
292:      data is finally put locally.

294:      In the sequential case the todata contains indices from where the data is put
295:      and the fromdata contains indices from where the data is taken from.
296:      This is backwards from the paralllel case!

298: @*/
299: PetscErrorCode  VecScatterRemap(VecScatter scat,PetscInt tomap[],PetscInt frommap[])
300: {
301:   VecScatter_MPI_General *to,*from;
302:   VecScatter_Seq_General *sgto,*sgfrom;
303:   VecScatter_Seq_Stride  *ssto;
304:   PetscInt               i,ierr;


311:   if (scat->ops->remap) {
312:     (*scat->ops->remap)(scat,tomap,frommap);
313:   } else {
314:     to     = (VecScatter_MPI_General*)scat->todata;
315:     from   = (VecScatter_MPI_General*)scat->fromdata;
316:     ssto   = (VecScatter_Seq_Stride*)scat->todata;
317:     sgto   = (VecScatter_Seq_General*)scat->todata;
318:     sgfrom = (VecScatter_Seq_General*)scat->fromdata;

320:     /* remap indices from where we take/read data */
321:     if (tomap) {
322:       if (to->format == VEC_SCATTER_MPI_TOALL) {
323:         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Not for to all scatter");
324:       } else if (to->format == VEC_SCATTER_MPI_GENERAL) {
325:         /* handle off processor parts */
326:         for (i=0; i<to->starts[to->n]; i++) to->indices[i] = tomap[to->indices[i]];

328:         /* handle local part */
329:         for (i=0; i<to->local.n; i++) to->local.vslots[i] = tomap[to->local.vslots[i]];

331:         /* the memcpy optimizations in vecscatter was based on index patterns it has.
332:            They need to be recalculated when indices are changed (remapped).
333:          */
334:         VecScatterMemcpyPlanDestroy_PtoP(to,from);
335:         VecScatterMemcpyPlanCreate_PtoP(to,from);
336:       } else if (sgfrom->format == VEC_SCATTER_SEQ_GENERAL) {
337:         /* remap indices*/
338:         for (i=0; i<sgfrom->n; i++) sgfrom->vslots[i] = tomap[sgfrom->vslots[i]];
339:         /* update optimizations, which happen when it is a Stride1toSG, SGtoStride1 or SGToSG vecscatter */
340:         if (ssto->format == VEC_SCATTER_SEQ_STRIDE && ssto->step == 1) {
341:           PetscInt tmp[2];
342:           tmp[0] = 0; tmp[1] = sgfrom->n;
343:           VecScatterMemcpyPlanDestroy(&sgfrom->memcpy_plan);
344:           VecScatterMemcpyPlanCreate_Index(1,tmp,sgfrom->vslots,1/*bs*/,&sgfrom->memcpy_plan);
345:         } else if (sgto->format == VEC_SCATTER_SEQ_GENERAL) {
346:           VecScatterMemcpyPlanDestroy(&sgto->memcpy_plan);;
347:           VecScatterMemcpyPlanDestroy(&sgfrom->memcpy_plan);
348:           VecScatterMemcpyPlanCreate_SGToSG(1/*bs*/,sgto,sgfrom);
349:         }
350:       } else if (sgfrom->format == VEC_SCATTER_SEQ_STRIDE) {
351:         VecScatter_Seq_Stride *ssto = (VecScatter_Seq_Stride*)sgfrom;

353:         /* if the remapping is the identity and stride is identity then skip remap */
354:         if (ssto->step == 1 && ssto->first == 0) {
355:           for (i=0; i<ssto->n; i++) {
356:             if (tomap[i] != i) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Unable to remap such scatters");
357:           }
358:         } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Unable to remap such scatters");
359:       } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Unable to remap such scatters");
360:     }
361:   }
362:   if (frommap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unable to remap the FROM in scatters yet");

364:   /*
365:     Mark then vector lengths as unknown because we do not know the
366:     lengths of the remapped vectors
367:   */
368:   scat->from_n = -1;
369:   scat->to_n   = -1;
370:   return(0);
371: }

373: /* Given a parallel VecScatter context, return number of procs and vector entries involved in remote (i.e., off-process) communication

375:   Input Parameters:
376: + ctx   - the context (must be a parallel vecscatter)
377: - send  - true to select the send info (i.e., todata), otherwise to select the recv info (i.e., fromdata)

379:   Output parameters:
380: + num_procs   - number of remote processors
381: - num_entries - number of vector entries to send or recv


384:   .seealso: VecScatterGetRemote_Private(), VecScatterGetRemoteOrdered_Private()

386:   Notes:
387:   Sometimes PETSc internally needs to use the matrix-vector-multiply vecscatter context for other purposes. The client code
388:   usually only uses MPI_Send/Recv. This group of subroutines provides info needed for such uses.
389:  */
390: PetscErrorCode VecScatterGetRemoteCount_Private(VecScatter ctx,PetscBool send,PetscInt *num_procs,PetscInt *num_entries)
391: {
392:   VecScatter_MPI_General *vs;
393:   PetscBool              par;
394:   PetscErrorCode         ierr;

397:   if (ctx->ops->getremotecount) {
398:     (*ctx->ops->getremotecount)(ctx,send,num_procs,num_entries);
399:   } else {
400:     vs = (VecScatter_MPI_General*)(send ? ctx->todata : ctx->fromdata);
401:     par = (vs->format == VEC_SCATTER_MPI_GENERAL)? PETSC_TRUE : PETSC_FALSE;
402:     if (num_procs)   *num_procs   = par ? vs->n             : 0;
403:     if (num_entries) *num_entries = par ? vs->starts[vs->n] : 0;
404:   }
405:   return(0);
406: }

408: /* Given a parallel VecScatter context, return a plan that represents the remote communication.
409:    Any output parameter can be NULL.

411:   Input Parameters:
412: + ctx   - the context
413: - send  - true to select the send info (i.e., todata), otherwise to select the recv info (i.e., fromdata)

415:   Output parameters:
416: + n        - number of remote processors
417: . starts   - starting point in indices for each proc. ATTENTION: starts[0] is not necessarily zero.
418:              Therefore, expressions like starts[i+1]-starts[i] and indices[starts[i]+j] work as
419:              expected for a CSR structure but buf[starts[i]+j] may be out of range if buf was allocated
420:              with length starts[n]-starts[0]. One should use buf[starts[i]-starts[0]+j] instead.
421: . indices  - indices of entries to send/recv
422: . procs    - ranks of remote processors
423: - bs       - block size

425:   .seealso: VecScatterRestoreRemote_Private(), VecScatterGetRemoteOrdered_Private()
426:  */
427: PetscErrorCode VecScatterGetRemote_Private(VecScatter ctx,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
428: {
429:   VecScatter_MPI_General *vs;
430:   PetscBool              par;
431:   PetscErrorCode         ierr;

434:   if (ctx->ops->getremote) {
435:     (*ctx->ops->getremote)(ctx,send,n,starts,indices,procs,bs);
436:   } else {
437:     vs = (VecScatter_MPI_General*)(send ? ctx->todata : ctx->fromdata);
438:     par = (vs->format == VEC_SCATTER_MPI_GENERAL)? PETSC_TRUE : PETSC_FALSE;
439:     if (n)       *n       = par ? vs->n       : 0;
440:     if (indices) *indices = par ? vs->indices : NULL;
441:     if (starts)  *starts  = par ? vs->starts  : NULL;
442:     if (procs)   *procs   = par ? vs->procs   : NULL;
443:     if (bs)      *bs      = par ? vs->bs      : 0;
444:   }
445:   return(0);
446: }


449: /* Given a parallel VecScatter context, return a plan that represents the remote communication. Ranks of remote
450:    processors returned in procs must be sorted in ascending order. Any output parameter can be NULL.

452:   Input Parameters:
453: + ctx   - the context
454: - send  - true to select the send info (i.e., todata), otherwise to select the recv info (i.e., fromdata)

456:   Output parameters:
457: + n        - number of remote processors
458: . starts   - starting point in indices for each proc. ATTENTION: starts[0] is not necessarily zero.
459:              Therefore, expressions like starts[i+1]-starts[i] and indices[starts[i]+j] work as
460:              expected for a CSR structure but buf[starts[i]+j] may be out of range if buf was allocated
461:              with length starts[n]-starts[0]. One should use buf[starts[i]-starts[0]+j] instead.
462: . indices  - indices of entries to send/recv
463: . procs    - ranks of remote processors
464: - bs       - block size

466:   .seealso: VecScatterRestoreRemoteOrdered_Private(), VecScatterGetRemote_Private()

468:   Notes:
469:   Output parameters like starts, indices must also be adapted according to the sorted ranks.
470:  */
471: PetscErrorCode VecScatterGetRemoteOrdered_Private(VecScatter ctx,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
472: {
473:   VecScatter_MPI_General *vs;
474:   PetscBool              par;
475:   PetscErrorCode         ierr;

478:   if (ctx->ops->getremoteordered) {
479:     (*ctx->ops->getremoteordered)(ctx,send,n,starts,indices,procs,bs);
480:   } else {
481:     vs = (VecScatter_MPI_General*)(send ? ctx->todata : ctx->fromdata);
482:     par = (vs->format == VEC_SCATTER_MPI_GENERAL)? PETSC_TRUE : PETSC_FALSE;
483:     if (n)       *n       = par ? vs->n       : 0;
484:     if (indices) *indices = par ? vs->indices : NULL;
485:     if (starts)  *starts  = par ? vs->starts  : NULL;
486:     if (procs)   *procs   = par ? vs->procs   : NULL;
487:     if (bs)      *bs      = par ? vs->bs      : 0;
488:   }
489: #if defined(PETSC_USE_DEBUG)
490:   if (n && procs) {
491:     PetscInt i;
492:     /* from back to front to also handle cases *n=0 */
493:     for (i=*n-1; i>0; i--) { if ((*procs)[i-1] > (*procs)[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"procs[] are not ordered"); }
494:   }
495: #endif
496:   return(0);
497: }

499: /* Given a parallel VecScatter context, restore the plan returned by VecScatterGetRemote_Private. This gives a chance for
500:    an implementation to free memory allocated in the VecScatterGetRemote_Private call.

502:   Input Parameters:
503: + ctx   - the context
504: - send  - true to select the send info (i.e., todata), otherwise to select the recv info (i.e., fromdata)

506:   Output parameters:
507: + n        - number of remote processors
508: . starts   - starting point in indices for each proc
509: . indices  - indices of entries to send/recv
510: . procs    - ranks of remote processors
511: - bs       - block size

513:   .seealso: VecScatterGetRemote_Private()
514:  */
515: PetscErrorCode VecScatterRestoreRemote_Private(VecScatter ctx,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
516: {
517:   PetscErrorCode         ierr;

520:   if (ctx->ops->restoreremote) {
521:     (*ctx->ops->restoreremote)(ctx,send,n,starts,indices,procs,bs);
522:   } else {
523:     if (starts)  *starts  = NULL;
524:     if (indices) *indices = NULL;
525:     if (procs)   *procs   = NULL;
526:   }
527:   return(0);
528: }

530: /* Given a parallel VecScatter context, restore the plan returned by VecScatterGetRemoteOrdered_Private. This gives a chance for
531:    an implementation to free memory allocated in the VecScatterGetRemoteOrdered_Private call.

533:   Input Parameters:
534: + ctx   - the context
535: - send  - true to select the send info (i.e., todata), otherwise to select the recv info (i.e., fromdata)

537:   Output parameters:
538: + n        - number of remote processors
539: . starts   - starting point in indices for each proc
540: . indices  - indices of entries to send/recv
541: . procs    - ranks of remote processors
542: - bs       - block size

544:   .seealso: VecScatterGetRemoteOrdered_Private()
545:  */
546: PetscErrorCode VecScatterRestoreRemoteOrdered_Private(VecScatter ctx,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
547: {
550:   if (ctx->ops->restoreremoteordered) {
551:     (*ctx->ops->restoreremoteordered)(ctx,send,n,starts,indices,procs,bs);
552:   } else {
553:     VecScatterRestoreRemote_Private(ctx,send,n,starts,indices,procs,bs);
554:   }
555:   return(0);
556: }

558: #if defined(PETSC_HAVE_CUDA)

560: /*@C
561:    VecScatterInitializeForGPU - Initializes a generalized scatter from one vector
562:    to another for GPU based computation.

564:    Input Parameters:
565: +  inctx - scatter context generated by VecScatterCreate()
566: -  x - the vector from which we scatter

568:   Level: intermediate

570:   Notes:
571:    Effectively, this function creates all the necessary indexing buffers and work
572:    vectors needed to move data only those data points in a vector which need to
573:    be communicated across ranks. This is done at the first time this function is
574:    called. Currently, this only used in the context of the parallel SpMV call in
575:    MatMult_MPIAIJCUSPARSE.

577:    This function is executed before the call to MatMult. This enables the memory
578:    transfers to be overlapped with the MatMult SpMV kernel call.

580: .seealso: VecScatterFinalizeForGPU(), VecScatterCreate(), VecScatterEnd()
581: @*/
582: PETSC_EXTERN PetscErrorCode VecScatterInitializeForGPU(VecScatter inctx,Vec x)
583: {
585:   PetscInt       i,nrecvs,nsends,sbs,rbs,ns,nr;
586:   const PetscInt *sstarts,*rstarts,*sindices,*rindices;

589:   VecScatterGetRemote_Private(inctx,PETSC_TRUE/*send*/, &nsends,&sstarts,&sindices,NULL/*procs*/,&sbs);
590:   VecScatterGetRemote_Private(inctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,&rindices,NULL/*procs*/,&rbs);
591:   ns   = nsends ? sstarts[nsends]-sstarts[0] : 0; /* s/rstarts[0] is not necessarily zero */
592:   nr   = nrecvs ? rstarts[nrecvs]-rstarts[0] : 0;

594:   if (x->valid_GPU_array != PETSC_OFFLOAD_UNALLOCATED && (nsends>0 || nrecvs>0)) {
595:     if (!inctx->spptr) {
596:       PetscInt k,*tindicesSends,*sindicesSends,*tindicesRecvs,*sindicesRecvs;
597:       /* Here we create indices for both the senders and receivers. */
598:       PetscMalloc1(ns,&tindicesSends);
599:       PetscMalloc1(nr,&tindicesRecvs);

601:       /* s/rindices and s/rstarts could be NULL when ns or nr is zero */
602:       if (ns) {PetscArraycpy(tindicesSends,&sindices[sstarts[0]],ns);}
603:       if (nr) {PetscArraycpy(tindicesRecvs,&rindices[rstarts[0]],nr);}

605:       PetscSortRemoveDupsInt(&ns,tindicesSends);
606:       PetscSortRemoveDupsInt(&nr,tindicesRecvs);

608:       PetscMalloc1(sbs*ns,&sindicesSends);
609:       PetscMalloc1(rbs*nr,&sindicesRecvs);

611:       /* sender indices */
612:       for (i=0; i<ns; i++) {
613:         for (k=0; k<sbs; k++) sindicesSends[i*sbs+k] = tindicesSends[i]+k;
614:       }
615:       PetscFree(tindicesSends);

617:       /* receiver indices */
618:       for (i=0; i<nr; i++) {
619:         for (k=0; k<rbs; k++) sindicesRecvs[i*rbs+k] = tindicesRecvs[i]+k;
620:       }
621:       PetscFree(tindicesRecvs);

623:       /* create GPU indices, work vectors, ... */
624:       VecScatterCUDAIndicesCreate_PtoP(ns*sbs,sindicesSends,nr*rbs,sindicesRecvs,(PetscCUDAIndices*)&inctx->spptr);
625:       PetscFree(sindicesSends);
626:       PetscFree(sindicesRecvs);
627:     }
628:   }
629:   return(0);
630: }

632: /*@C
633:    VecScatterFinalizeForGPU - Finalizes a generalized scatter from one vector to
634:    another for GPU based computation.

636:    Input Parameter:
637: .  inctx - scatter context generated by VecScatterCreate()

639:   Level: intermediate

641:   Notes:
642:    Effectively, this function resets the temporary buffer flags. Currently, this
643:    only used in the context of the parallel SpMV call in in MatMult_MPIAIJCUDA
644:    or MatMult_MPIAIJCUDAARSE. Once the MatMultAdd is finished, the GPU temporary
645:    buffers used for messaging are no longer valid.

647: .seealso: VecScatterInitializeForGPU(), VecScatterCreate(), VecScatterEnd()
648: @*/
649: PETSC_EXTERN PetscErrorCode VecScatterFinalizeForGPU(VecScatter inctx)
650: {
652:   return(0);
653: }

655: #endif