1df826632SBarry Smith #define PETSCKSP_DLL 2df826632SBarry Smith 3df826632SBarry Smith /* 4df826632SBarry Smith This file defines a "solve the problem redistributely on each subgroup of processor" preconditioner. 5df826632SBarry Smith */ 6df826632SBarry Smith #include "private/pcimpl.h" /*I "petscpc.h" I*/ 7df826632SBarry Smith #include "petscksp.h" 8df826632SBarry Smith 9df826632SBarry Smith typedef struct { 10df826632SBarry Smith KSP ksp; 11df826632SBarry Smith Vec x,b; 12df826632SBarry Smith Mat mat; 13df826632SBarry Smith VecScatter scatter; 14911f9fe8SBarry Smith IS is; 15df826632SBarry Smith } PC_Redistribute; 16df826632SBarry Smith 17df826632SBarry Smith #undef __FUNCT__ 18df826632SBarry Smith #define __FUNCT__ "PCView_Redistribute" 19df826632SBarry Smith static PetscErrorCode PCView_Redistribute(PC pc,PetscViewer viewer) 20df826632SBarry Smith { 21df826632SBarry Smith PC_Redistribute *red = (PC_Redistribute*)pc->data; 22df826632SBarry Smith PetscErrorCode ierr; 23df826632SBarry Smith PetscTruth iascii,isstring; 24df826632SBarry Smith 25df826632SBarry Smith PetscFunctionBegin; 26df826632SBarry Smith ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);CHKERRQ(ierr); 27df826632SBarry Smith ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_STRING,&isstring);CHKERRQ(ierr); 28df826632SBarry Smith if (iascii) { 29df826632SBarry Smith ierr = PetscViewerASCIIPrintf(viewer," Redistribute preconditioner: \n");CHKERRQ(ierr); 30df826632SBarry Smith ierr = KSPView(red->ksp,viewer);CHKERRQ(ierr); 31df826632SBarry Smith } else if (isstring) { 32df826632SBarry Smith ierr = PetscViewerStringSPrintf(viewer," Redistribute preconditioner");CHKERRQ(ierr); 33df826632SBarry Smith ierr = KSPView(red->ksp,viewer);CHKERRQ(ierr); 34df826632SBarry Smith } else { 35df826632SBarry Smith SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported for PC redistribute",((PetscObject)viewer)->type_name); 36df826632SBarry Smith } 37df826632SBarry Smith PetscFunctionReturn(0); 38df826632SBarry Smith } 39df826632SBarry Smith 40df826632SBarry Smith #include "private/matimpl.h" /*I "petscmat.h" I*/ 41df826632SBarry Smith #undef __FUNCT__ 42df826632SBarry Smith #define __FUNCT__ "PCSetUp_Redistribute" 43df826632SBarry Smith static PetscErrorCode PCSetUp_Redistribute(PC pc) 44df826632SBarry Smith { 45df826632SBarry Smith PC_Redistribute *red = (PC_Redistribute*)pc->data; 46df826632SBarry Smith PetscErrorCode ierr; 47df826632SBarry Smith MPI_Comm comm; 48df826632SBarry Smith PetscInt rstart,rend,i,nz,cnt,*rows,ncnt; 49df826632SBarry Smith PetscMap *map,*nmap; 50df826632SBarry Smith PetscMPIInt size,rank,imdex,tag,n; 51e8dd6687SHong Zhang PetscInt *source = PETSC_NULL; 52df826632SBarry Smith PetscMPIInt *nprocs = PETSC_NULL,nrecvs; 53df826632SBarry Smith PetscInt j,nsends; 54df826632SBarry Smith PetscInt *owner = PETSC_NULL,*starts = PETSC_NULL,count,slen; 55e8dd6687SHong Zhang PetscInt *rvalues,*svalues,recvtotal; 56df826632SBarry Smith PetscMPIInt *onodes1,*olengths1; 57df826632SBarry Smith MPI_Request *send_waits = PETSC_NULL,*recv_waits = PETSC_NULL; 58df826632SBarry Smith MPI_Status recv_status,*send_status; 59*ca320bd4SBarry Smith Vec tvec; 60*ca320bd4SBarry Smith Mat tmat; 61df826632SBarry Smith 62df826632SBarry Smith PetscFunctionBegin; 63df826632SBarry Smith ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 64df826632SBarry Smith ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 65df826632SBarry Smith ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 66361c1e09SMatthew Knepley ierr = PetscObjectGetNewTag((PetscObject)pc,&tag);CHKERRQ(ierr); 67df826632SBarry Smith 68df826632SBarry Smith if (!pc->setupcalled) { 69df826632SBarry Smith } else SETERRQ(PETSC_ERR_SUP,"Cannot yet re-setup a redistribute KSP"); 70df826632SBarry Smith 71*ca320bd4SBarry Smith /* count non-diagonal rows on process */ 72df826632SBarry Smith ierr = MatGetOwnershipRange(pc->mat,&rstart,&rend);CHKERRQ(ierr); 73*ca320bd4SBarry Smith cnt = 0; 74df826632SBarry Smith for (i=rstart; i<rend; i++) { 75df826632SBarry Smith ierr = MatGetRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr); 76df826632SBarry Smith if (nz > 1) cnt++; 77911f9fe8SBarry Smith ierr = MatRestoreRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr); 78df826632SBarry Smith } 79df826632SBarry Smith ierr = PetscMalloc(cnt*sizeof(PetscInt),&rows);CHKERRQ(ierr); 80*ca320bd4SBarry Smith 81*ca320bd4SBarry Smith /* list non-diagonal rows on process */ 82df826632SBarry Smith cnt = 0; 83df826632SBarry Smith for (i=rstart; i<rend; i++) { 84df826632SBarry Smith ierr = MatGetRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr); 85df826632SBarry Smith if (nz > 1) rows[cnt++] = i; 86911f9fe8SBarry Smith ierr = MatRestoreRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr); 87df826632SBarry Smith } 88*ca320bd4SBarry Smith 89*ca320bd4SBarry Smith /* create PetscMap for non-diagonal rows on each process */ 90df826632SBarry Smith ierr = PetscMalloc(sizeof(PetscMap),&map);CHKERRQ(ierr); 91df826632SBarry Smith ierr = PetscMapInitialize(comm,map);CHKERRQ(ierr); 92df826632SBarry Smith ierr = PetscMapSetLocalSize(map,cnt);CHKERRQ(ierr); 93df826632SBarry Smith ierr = PetscMapSetBlockSize(map,1);CHKERRQ(ierr); 94df826632SBarry Smith ierr = PetscMapSetUp(map);CHKERRQ(ierr); 95df826632SBarry Smith rstart = map->rstart; 96df826632SBarry Smith rend = map->rend; 97df826632SBarry Smith 98*ca320bd4SBarry Smith /* create PetscMap for load-balanced non-diagonal rows on each process */ 99df826632SBarry Smith ierr = PetscMalloc(sizeof(PetscMap),&nmap);CHKERRQ(ierr); 100df826632SBarry Smith ierr = PetscMapInitialize(comm,nmap);CHKERRQ(ierr); 101df826632SBarry Smith ierr = MPI_Allreduce(&cnt,&ncnt,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 102df826632SBarry Smith ierr = PetscMapSetSize(nmap,ncnt);CHKERRQ(ierr); 103df826632SBarry Smith ierr = PetscMapSetBlockSize(nmap,1);CHKERRQ(ierr); 104df826632SBarry Smith ierr = PetscMapSetUp(nmap);CHKERRQ(ierr); 105df826632SBarry Smith 106*ca320bd4SBarry Smith /* 107*ca320bd4SBarry Smith this code is taken from VecScatterCreate_PtoS() 108*ca320bd4SBarry Smith Determines what rows need to be moved where to 109*ca320bd4SBarry Smith load balance the non-diagonal rows 110*ca320bd4SBarry Smith */ 111df826632SBarry Smith /* count number of contributors to each processor */ 112df826632SBarry Smith ierr = PetscMalloc2(size,PetscMPIInt,&nprocs,cnt,PetscInt,&owner);CHKERRQ(ierr); 113df826632SBarry Smith ierr = PetscMemzero(nprocs,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 114df826632SBarry Smith j = 0; 115df826632SBarry Smith nsends = 0; 116df826632SBarry Smith for (i=rstart; i<rend; i++) { 117df826632SBarry Smith if (i < nmap->range[j]) j = 0; 118df826632SBarry Smith for (; j<size; j++) { 119df826632SBarry Smith if (i < nmap->range[j+1]) { 120df826632SBarry Smith if (!nprocs[j]++) nsends++; 121*ca320bd4SBarry Smith owner[i-rstart] = j; 122df826632SBarry Smith break; 123df826632SBarry Smith } 124df826632SBarry Smith } 125df826632SBarry Smith } 126df826632SBarry Smith /* inform other processors of number of messages and max length*/ 127df826632SBarry Smith ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,nprocs,&nrecvs);CHKERRQ(ierr); 128df826632SBarry Smith ierr = PetscGatherMessageLengths(comm,nsends,nrecvs,nprocs,&onodes1,&olengths1);CHKERRQ(ierr); 129df826632SBarry Smith ierr = PetscSortMPIIntWithArray(nrecvs,onodes1,olengths1);CHKERRQ(ierr); 130df826632SBarry Smith recvtotal = 0; for (i=0; i<nrecvs; i++) recvtotal += olengths1[i]; 131df826632SBarry Smith 132df826632SBarry Smith /* post receives: rvalues - rows I will own; count - nu */ 133df826632SBarry Smith ierr = PetscMalloc3(recvtotal,PetscInt,&rvalues,nrecvs,PetscInt,&source,nrecvs,MPI_Request,&recv_waits);CHKERRQ(ierr); 134df826632SBarry Smith count = 0; 135df826632SBarry Smith for (i=0; i<nrecvs; i++) { 136df826632SBarry Smith ierr = MPI_Irecv((rvalues+count),olengths1[i],MPIU_INT,onodes1[i],tag,comm,recv_waits+i);CHKERRQ(ierr); 137df826632SBarry Smith count += olengths1[i]; 138df826632SBarry Smith } 139df826632SBarry Smith 140df826632SBarry Smith /* do sends: 141df826632SBarry Smith 1) starts[i] gives the starting index in svalues for stuff going to 142df826632SBarry Smith the ith processor 143df826632SBarry Smith */ 144911f9fe8SBarry Smith ierr = PetscMalloc3(cnt,PetscInt,&svalues,nsends,MPI_Request,&send_waits,size,PetscInt,&starts);CHKERRQ(ierr); 145df826632SBarry Smith starts[0] = 0; 146df826632SBarry Smith for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 147df826632SBarry Smith for (i=0; i<cnt; i++) { 148df826632SBarry Smith svalues[starts[owner[i]]++] = rows[i]; 149df826632SBarry Smith } 150*ca320bd4SBarry Smith ierr = PetscFree(rows);CHKERRQ(ierr); 151df826632SBarry Smith starts[0] = 0; 152911f9fe8SBarry Smith for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 153df826632SBarry Smith count = 0; 154df826632SBarry Smith for (i=0; i<size; i++) { 155df826632SBarry Smith if (nprocs[i]) { 156df826632SBarry Smith ierr = MPI_Isend(svalues+starts[i],nprocs[i],MPIU_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr); 157df826632SBarry Smith } 158df826632SBarry Smith } 159df826632SBarry Smith 160df826632SBarry Smith /* wait on receives */ 161df826632SBarry Smith count = nrecvs; 162df826632SBarry Smith slen = 0; 163df826632SBarry Smith while (count) { 164df826632SBarry Smith ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr); 165df826632SBarry Smith /* unpack receives into our local space */ 166df826632SBarry Smith ierr = MPI_Get_count(&recv_status,MPIU_INT,&n);CHKERRQ(ierr); 167df826632SBarry Smith slen += n; 168df826632SBarry Smith count--; 169df826632SBarry Smith } 170df826632SBarry Smith if (slen != recvtotal) SETERRQ2(PETSC_ERR_PLIB,"Total message lengths %D not expected %D",slen,recvtotal); 171df826632SBarry Smith 172911f9fe8SBarry Smith ierr = ISCreateGeneral(comm,slen,rvalues,&red->is);CHKERRQ(ierr); 173911f9fe8SBarry Smith 174*ca320bd4SBarry Smith /* free up all work space */ 175df826632SBarry Smith ierr = PetscFree(olengths1);CHKERRQ(ierr); 176df826632SBarry Smith ierr = PetscFree(onodes1);CHKERRQ(ierr); 177df826632SBarry Smith ierr = PetscFree3(rvalues,source,recv_waits);CHKERRQ(ierr); 178*ca320bd4SBarry Smith ierr = PetscFree2(nprocs,owner);CHKERRQ(ierr); 179*ca320bd4SBarry Smith if (nsends) { /* wait on sends */ 180df826632SBarry Smith ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr); 181df826632SBarry Smith ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr); 182df826632SBarry Smith ierr = PetscFree(send_status);CHKERRQ(ierr); 183df826632SBarry Smith } 184df826632SBarry Smith ierr = PetscFree3(svalues,send_waits,starts);CHKERRQ(ierr); 185*ca320bd4SBarry Smith ierr = PetscFree(map->range);CHKERRQ(ierr); 186df826632SBarry Smith ierr = PetscFree(map);CHKERRQ(ierr); 187*ca320bd4SBarry Smith ierr = PetscFree(nmap->range);CHKERRQ(ierr); 188df826632SBarry Smith ierr = PetscFree(nmap);CHKERRQ(ierr); 189df826632SBarry Smith 190*ca320bd4SBarry Smith ierr = ISView(red->is,PETSC_VIEWER_STDOUT_(comm));CHKERRQ(ierr); 191*ca320bd4SBarry Smith 192*ca320bd4SBarry Smith ierr = VecCreateMPI(comm,slen,PETSC_DETERMINE,&red->b);CHKERRQ(ierr); 193*ca320bd4SBarry Smith ierr = VecDuplicate(red->b,&red->x);CHKERRQ(ierr); 194*ca320bd4SBarry Smith ierr = MatGetVecs(pc->pmat,&tvec,PETSC_NULL);CHKERRQ(ierr); 195*ca320bd4SBarry Smith ierr = VecScatterCreate(tvec,red->is,red->b,PETSC_NULL,&red->scatter);CHKERRQ(ierr); 196*ca320bd4SBarry Smith ierr = VecDestroy(tvec);CHKERRQ(ierr); 197*ca320bd4SBarry Smith ierr = MatGetSubMatrix(pc->pmat,red->is,red->is,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 198*ca320bd4SBarry Smith ierr = KSPSetOperators(red->ksp,tmat,tmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 199*ca320bd4SBarry Smith ierr = MatDestroy(tmat);CHKERRQ(ierr); 200*ca320bd4SBarry Smith 201df826632SBarry Smith ierr = KSPSetUp(red->ksp);CHKERRQ(ierr); 202df826632SBarry Smith PetscFunctionReturn(0); 203df826632SBarry Smith } 204df826632SBarry Smith 205df826632SBarry Smith #undef __FUNCT__ 206df826632SBarry Smith #define __FUNCT__ "PCApply_Redistribute" 207df826632SBarry Smith static PetscErrorCode PCApply_Redistribute(PC pc,Vec b,Vec x) 208df826632SBarry Smith { 209df826632SBarry Smith PC_Redistribute *red = (PC_Redistribute*)pc->data; 210df826632SBarry Smith PetscErrorCode ierr; 211df826632SBarry Smith 212df826632SBarry Smith PetscFunctionBegin; 213df826632SBarry Smith ierr = VecScatterBegin(red->scatter,b,red->b,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 214df826632SBarry Smith ierr = VecScatterEnd(red->scatter,b,red->b,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 215df826632SBarry Smith ierr = KSPSolve(red->ksp,red->b,red->x);CHKERRQ(ierr); 216df826632SBarry Smith ierr = VecScatterBegin(red->scatter,red->x,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 217df826632SBarry Smith ierr = VecScatterEnd(red->scatter,red->x,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 218df826632SBarry Smith PetscFunctionReturn(0); 219df826632SBarry Smith } 220df826632SBarry Smith 221df826632SBarry Smith #undef __FUNCT__ 222df826632SBarry Smith #define __FUNCT__ "PCDestroy_Redistribute" 223df826632SBarry Smith static PetscErrorCode PCDestroy_Redistribute(PC pc) 224df826632SBarry Smith { 225df826632SBarry Smith PC_Redistribute *red = (PC_Redistribute*)pc->data; 226df826632SBarry Smith PetscErrorCode ierr; 227df826632SBarry Smith 228df826632SBarry Smith PetscFunctionBegin; 229df826632SBarry Smith if (red->scatter) {ierr = VecScatterDestroy(red->scatter);CHKERRQ(ierr);} 230*ca320bd4SBarry Smith if (red->is) {ierr = ISDestroy(red->is);CHKERRQ(ierr);} 231df826632SBarry Smith if (red->b) {ierr = VecDestroy(red->b);CHKERRQ(ierr);} 232df826632SBarry Smith if (red->x) {ierr = VecDestroy(red->x);CHKERRQ(ierr);} 233df826632SBarry Smith if (red->mat) {ierr = MatDestroy(red->mat);CHKERRQ(ierr);} 234df826632SBarry Smith if (red->ksp) {ierr = KSPDestroy(red->ksp);CHKERRQ(ierr);} 235df826632SBarry Smith ierr = PetscFree(red);CHKERRQ(ierr); 236df826632SBarry Smith PetscFunctionReturn(0); 237df826632SBarry Smith } 238df826632SBarry Smith 239df826632SBarry Smith #undef __FUNCT__ 240df826632SBarry Smith #define __FUNCT__ "PCSetFromOptions_Redistribute" 241df826632SBarry Smith static PetscErrorCode PCSetFromOptions_Redistribute(PC pc) 242df826632SBarry Smith { 243df826632SBarry Smith PetscErrorCode ierr; 244df826632SBarry Smith PC_Redistribute *red = (PC_Redistribute*)pc->data; 245df826632SBarry Smith 246df826632SBarry Smith PetscFunctionBegin; 247df826632SBarry Smith ierr = KSPSetFromOptions(red->ksp);CHKERRQ(ierr); 248df826632SBarry Smith PetscFunctionReturn(0); 249df826632SBarry Smith } 250df826632SBarry Smith 251df826632SBarry Smith /* -------------------------------------------------------------------------------------*/ 252df826632SBarry Smith /*MC 253df826632SBarry Smith PCREDISTRIBUTE - Redistributes a matrix for load balancing and then applys a KSP to that new matrix 254df826632SBarry Smith 255df826632SBarry Smith Options for the redistribute preconditioners can be set with -redistribute_ksp_xxx <values> and -redistribute_pc_xxx <values> 256df826632SBarry Smith 257df826632SBarry Smith Level: intermediate 258df826632SBarry Smith 259df826632SBarry Smith .seealso: PCCreate(), PCSetType(), PCType (for list of available types) 260df826632SBarry Smith M*/ 261df826632SBarry Smith 262df826632SBarry Smith EXTERN_C_BEGIN 263df826632SBarry Smith #undef __FUNCT__ 264df826632SBarry Smith #define __FUNCT__ "PCCreate_Redistribute" 265df826632SBarry Smith PetscErrorCode PETSCKSP_DLLEXPORT PCCreate_Redistribute(PC pc) 266df826632SBarry Smith { 267df826632SBarry Smith PetscErrorCode ierr; 268df826632SBarry Smith PC_Redistribute *red; 269911f9fe8SBarry Smith const char *prefix; 270df826632SBarry Smith 271df826632SBarry Smith PetscFunctionBegin; 272df826632SBarry Smith ierr = PetscNewLog(pc,PC_Redistribute,&red);CHKERRQ(ierr); 273df826632SBarry Smith pc->data = (void*)red; 274df826632SBarry Smith 275df826632SBarry Smith pc->ops->apply = PCApply_Redistribute; 276df826632SBarry Smith pc->ops->applytranspose = 0; 277df826632SBarry Smith pc->ops->setup = PCSetUp_Redistribute; 278df826632SBarry Smith pc->ops->destroy = PCDestroy_Redistribute; 279df826632SBarry Smith pc->ops->setfromoptions = PCSetFromOptions_Redistribute; 280df826632SBarry Smith pc->ops->view = PCView_Redistribute; 281911f9fe8SBarry Smith 282911f9fe8SBarry Smith ierr = KSPCreate(((PetscObject)pc)->comm,&red->ksp);CHKERRQ(ierr); 283911f9fe8SBarry Smith ierr = PetscObjectIncrementTabLevel((PetscObject)red->ksp,(PetscObject)pc,1);CHKERRQ(ierr); 284911f9fe8SBarry Smith ierr = PetscLogObjectParent(pc,red->ksp);CHKERRQ(ierr); 285911f9fe8SBarry Smith ierr = PCGetOptionsPrefix(pc,&prefix);CHKERRQ(ierr); 286911f9fe8SBarry Smith ierr = KSPSetOptionsPrefix(red->ksp,prefix);CHKERRQ(ierr); 287911f9fe8SBarry Smith ierr = KSPAppendOptionsPrefix(red->ksp,"redistribute_");CHKERRQ(ierr); 288df826632SBarry Smith PetscFunctionReturn(0); 289df826632SBarry Smith } 290df826632SBarry Smith EXTERN_C_END 291