static void ssolve_permuteA(IV ** oldToNewIV, IV ** newToOldIV, IVL ** symbfacIVL, ETree * frontETree, InpMtx * mtxA, FILE * msgFile, int symmetryflag) { int *oldToNew; *oldToNewIV = ETree_oldToNewVtxPerm(frontETree); oldToNew = IV_entries(*oldToNewIV); *newToOldIV = ETree_newToOldVtxPerm(frontETree); ETree_permuteVertices(frontETree, *oldToNewIV); InpMtx_permute(mtxA, oldToNew, oldToNew); if(symmetryflag!=2) InpMtx_mapToUpperTriangle(mtxA); InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS); InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS); *symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA); if (DEBUG_LVL > 1) { fprintf(msgFile, "\n\n old-to-new permutation vector"); IV_writeForHumanEye(*oldToNewIV, msgFile); fprintf(msgFile, "\n\n new-to-old permutation vector"); IV_writeForHumanEye(*newToOldIV, msgFile); fprintf(msgFile, "\n\n front tree after permutation"); ETree_writeForHumanEye(frontETree, msgFile); fprintf(msgFile, "\n\n input matrix after permutation"); InpMtx_writeForHumanEye(mtxA, msgFile); fprintf(msgFile, "\n\n symbolic factorization"); IVL_writeForHumanEye(*symbfacIVL, msgFile); fflush(msgFile); } }
/* ------------------------------------------------- purpose -- to write a Graph object for a human eye return value -- 1 if success, 0 otherwise created -- 95sep29, cca ------------------------------------------------- */ int Graph_writeForHumanEye ( Graph *graph, FILE *fp ) { int ierr, rc ; if ( graph == NULL || fp == NULL ) { fprintf(stderr, "\n fatal error in Graph_writeForHumanEye(%p,%p)" "\n bad input\n", graph, fp) ; spoolesFatal(); } /* ------------------------ write out the statistics ------------------------ */ if ( (rc = Graph_writeStats(graph, fp)) == 0 ) { fprintf(stderr, "\n fatal error in Graph_writeForHumanEye(%p,%p)" "\n rc = %d, return from Graph_writeStats(%p,%p)\n", graph, fp, rc, graph, fp) ; return(0) ; } if ( graph->adjIVL != NULL ) { /* ---------------------------------- write out the adjacency IVL object ---------------------------------- */ fprintf(fp, "\n\n adjacency IVL object") ; rc = IVL_writeForHumanEye(graph->adjIVL, fp) ; if ( rc < 0 ) { fprintf(stderr, "\n fatal error in Graph_writeForHumanEye(%p,%p)" "\n rc = %d, return from IVL_writeForHumanEye(%p,%p)" "\n while attempting to write out adjIVL\n", graph, fp, rc, graph->adjIVL, fp) ; return(0) ; } } /* ---------------------------------------------- write out the vertex weights vector if present ---------------------------------------------- */ if ( graph->type % 2 == 1 ) { if ( graph->vwghts == NULL ) { fprintf(stderr, "\n fatal error in Graph_writeForHumanEye(%p,%p)" "\n graph->type = %d, graph->vwghts == NULL\n", graph, fp, graph->type) ; return(0) ; } fprintf(fp, "\n\n vertex weights ") ; IVfp80(fp, graph->nvtx + graph->nvbnd, graph->vwghts, 80, &ierr) ; if ( ierr < 0 ) { fprintf(stderr, "\n fatal error in Graph_writeForHumanEye(%p,%p)" "\n ierr = %d, return from vwghts[] IVfp80\n", graph, fp, ierr) ; return(0) ; } } /* ------------------------------------------------ write out the edge weight IVL object, if present ------------------------------------------------ */ if ( graph->type >= 2 ) { if ( graph->ewghtIVL == NULL ) { fprintf(stderr, "\n fatal error in Graph_writeForHumanEye(%p,%p)" "\n graph->type = %d, graph->ewghtIVL == NULL\n", graph, fp, graph->type) ; return(0) ; } fprintf(fp, "\n\n edge weights IVL object") ; rc = IVL_writeForHumanEye(graph->ewghtIVL, fp) ; if ( rc < 0 ) { fprintf(stderr, "\n fatal error in Graph_writeForHumanEye(%p,%p)" "\n rc = %d, return from IVL_writeForHumanEye(%p,%p)" "\n while attempting to write out ewghtIVL\n", graph, fp, rc, graph->ewghtIVL, fp) ; return(0) ; } } return(1) ; }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) /* -------------------------------------------------------------------- this program tests the Graph_MPI_Bcast() method (1) process root generates a random Graph object and computes its checksum (2) process root broadcasts the Graph object to the other processors (3) each process computes the checksum of its Graph object (4) the checksums are compared on root created -- 98sep10, cca -------------------------------------------------------------------- */ { char *buffer ; double chksum, t1, t2 ; double *sums ; Drand drand ; int iproc, length, loc, msglvl, myid, nitem, nproc, nvtx, root, seed, size, type, v ; int *list ; FILE *msgFile ; Graph *graph ; /* --------------------------------------------------------------- find out the identity of this process and the number of process --------------------------------------------------------------- */ MPI_Init(&argc, &argv) ; MPI_Comm_rank(MPI_COMM_WORLD, &myid) ; MPI_Comm_size(MPI_COMM_WORLD, &nproc) ; fprintf(stdout, "\n process %d of %d, argc = %d", myid, nproc, argc) ; fflush(stdout) ; if ( argc != 8 ) { fprintf(stdout, "\n\n usage : %s msglvl msgFile type nvtx nitem root seed " "\n msglvl -- message level" "\n msgFile -- message file" "\n type -- type of graph" "\n nvtx -- # of vertices" "\n nitem -- # of items used to generate graph" "\n root -- root processor for broadcast" "\n seed -- random number seed" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else { length = strlen(argv[2]) + 1 + 4 ; buffer = CVinit(length, '\0') ; sprintf(buffer, "%s.%d", argv[2], myid) ; if ( (msgFile = fopen(buffer, "w")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } CVfree(buffer) ; } type = atoi(argv[3]) ; nvtx = atoi(argv[4]) ; nitem = atoi(argv[5]) ; root = atoi(argv[6]) ; seed = atoi(argv[7]) ; fprintf(msgFile, "\n %s " "\n msglvl -- %d" "\n msgFile -- %s" "\n type -- %d" "\n nvtx -- %d" "\n nitem -- %d" "\n root -- %d" "\n seed -- %d" "\n", argv[0], msglvl, argv[2], type, nvtx, nitem, root, seed) ; fflush(msgFile) ; /* ----------------------- set up the Graph object ----------------------- */ MARKTIME(t1) ; graph = Graph_new() ; if ( myid == root ) { InpMtx *inpmtx ; int nedges, totewght, totvwght, v ; int *adj, *vwghts ; IVL *adjIVL, *ewghtIVL ; /* ----------------------- generate a random graph ----------------------- */ inpmtx = InpMtx_new() ; InpMtx_init(inpmtx, INPMTX_BY_ROWS, INPMTX_INDICES_ONLY, nitem, 0) ; Drand_setDefaultFields(&drand) ; Drand_setSeed(&drand, seed) ; Drand_setUniform(&drand, 0, nvtx) ; Drand_fillIvector(&drand, nitem, InpMtx_ivec1(inpmtx)) ; Drand_fillIvector(&drand, nitem, InpMtx_ivec2(inpmtx)) ; InpMtx_setNent(inpmtx, nitem) ; InpMtx_changeStorageMode(inpmtx, INPMTX_BY_VECTORS) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n inpmtx mtx filled with raw entries") ; InpMtx_writeForHumanEye(inpmtx, msgFile) ; fflush(msgFile) ; } adjIVL = InpMtx_fullAdjacency(inpmtx) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n full adjacency structure") ; IVL_writeForHumanEye(adjIVL, msgFile) ; fflush(msgFile) ; } nedges = adjIVL->tsize ; if ( type == 1 || type == 3 ) { Drand_setUniform(&drand, 1, 10) ; vwghts = IVinit(nvtx, 0) ; Drand_fillIvector(&drand, nvtx, vwghts) ; totvwght = IVsum(nvtx, vwghts) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n vertex weights") ; IVfprintf(msgFile, nvtx, vwghts) ; fflush(msgFile) ; } } else { vwghts = NULL ; totvwght = nvtx ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n totvwght %d", totvwght) ; fflush(msgFile) ; } if ( type == 2 || type == 3 ) { ewghtIVL = IVL_new() ; IVL_init1(ewghtIVL, IVL_CHUNKED, nvtx) ; Drand_setUniform(&drand, 1, 100) ; totewght = 0 ; for ( v = 0 ; v < nvtx ; v++ ) { IVL_listAndSize(adjIVL, v, &size, &adj) ; IVL_setList(ewghtIVL, v, size, NULL) ; IVL_listAndSize(ewghtIVL, v, &size, &adj) ; Drand_fillIvector(&drand, size, adj) ; totewght += IVsum(size, adj) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n ewghtIVL") ; IVL_writeForHumanEye(ewghtIVL, msgFile) ; fflush(msgFile) ; } } else { ewghtIVL = NULL ; totewght = nedges ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n totewght %d", totewght) ; fflush(msgFile) ; } Graph_init2(graph, type, nvtx, 0, nedges, totvwght, totewght, adjIVL, vwghts, ewghtIVL) ; InpMtx_free(inpmtx) ; } MARKTIME(t2) ; fprintf(msgFile, "\n CPU %8.3f : initialize the Graph object", t2 - t1) ; fflush(msgFile) ; if ( msglvl > 2 ) { Graph_writeForHumanEye(graph, msgFile) ; } else { Graph_writeStats(graph, msgFile) ; } fflush(msgFile) ; if ( myid == root ) { /* ---------------------------------------- compute the checksum of the Graph object ---------------------------------------- */ chksum = graph->type + graph->nvtx + graph->nvbnd + graph->nedges + graph->totvwght + graph->totewght ; for ( v = 0 ; v < nvtx ; v++ ) { IVL_listAndSize(graph->adjIVL, v, &size, &list) ; chksum += 1 + v + size + IVsum(size, list) ; } if ( graph->vwghts != NULL ) { chksum += IVsum(nvtx, graph->vwghts) ; } if ( graph->ewghtIVL != NULL ) { for ( v = 0 ; v < nvtx ; v++ ) { IVL_listAndSize(graph->ewghtIVL, v, &size, &list) ; chksum += 1 + v + size + IVsum(size, list) ; } } fprintf(msgFile, "\n\n local chksum = %12.4e", chksum) ; fflush(msgFile) ; } /* -------------------------- broadcast the Graph object -------------------------- */ MARKTIME(t1) ; graph = Graph_MPI_Bcast(graph, root, msglvl, msgFile, MPI_COMM_WORLD) ; MARKTIME(t2) ; fprintf(msgFile, "\n CPU %8.3f : broadcast the Graph object", t2 - t1) ; if ( msglvl > 2 ) { Graph_writeForHumanEye(graph, msgFile) ; } else { Graph_writeStats(graph, msgFile) ; } /* ---------------------------------------- compute the checksum of the Graph object ---------------------------------------- */ chksum = graph->type + graph->nvtx + graph->nvbnd + graph->nedges + graph->totvwght + graph->totewght ; for ( v = 0 ; v < nvtx ; v++ ) { IVL_listAndSize(graph->adjIVL, v, &size, &list) ; chksum += 1 + v + size + IVsum(size, list) ; } if ( graph->vwghts != NULL ) { chksum += IVsum(nvtx, graph->vwghts) ; } if ( graph->ewghtIVL != NULL ) { for ( v = 0 ; v < nvtx ; v++ ) { IVL_listAndSize(graph->ewghtIVL, v, &size, &list) ; chksum += 1 + v + size + IVsum(size, list) ; } } fprintf(msgFile, "\n\n local chksum = %12.4e", chksum) ; fflush(msgFile) ; /* --------------------------------------- gather the checksums from the processes --------------------------------------- */ sums = DVinit(nproc, 0.0) ; MPI_Gather((void *) &chksum, 1, MPI_DOUBLE, (void *) sums, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD) ; if ( myid == 0 ) { fprintf(msgFile, "\n\n sums") ; DVfprintf(msgFile, nproc, sums) ; for ( iproc = 0 ; iproc < nproc ; iproc++ ) { sums[iproc] -= chksum ; } fprintf(msgFile, "\n\n errors") ; DVfprintf(msgFile, nproc, sums) ; fprintf(msgFile, "\n\n maxerror = %12.4e", DVmax(nproc, sums, &loc)); } /* ---------------- free the objects ---------------- */ DVfree(sums) ; Graph_free(graph) ; /* ------------------------ exit the MPI environment ------------------------ */ MPI_Finalize() ; fprintf(msgFile, "\n") ; fclose(msgFile) ; return(0) ; }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) /* ------------------------------------------------------ (1) read in an ETree object. (2) read in an Graph object. (3) find the optimal domain/schur complement partition for a semi-implicit factorization created -- 96oct03, cca ------------------------------------------------------ */ { char *inETreeFileName, *inGraphFileName, *outIVfileName ; double alpha, nA21, nfent1, nfops1, nL11, nL22, nPhi, nV, t1, t2 ; Graph *graph ; int ii, inside, J, K, msglvl, nfind1, nfront, nJ, nleaves1, nnode1, nvtx, rc, sizeJ, totalgain, vsize, v, w ; int *adjJ, *compids, *nodwghts, *vadj, *vtxToFront, *vwghts ; IV *compidsIV ; IVL *symbfacIVL ; ETree *etree ; FILE *msgFile ; Tree *tree ; if ( argc != 7 ) { fprintf(stdout, "\n\n usage : %s msglvl msgFile inETreeFile inGraphFile alpha" "\n outIVfile " "\n msglvl -- message level" "\n msgFile -- message file" "\n inETreeFile -- input file, must be *.etreef or *.etreeb" "\n inGraphFile -- input file, must be *.graphf or *.graphb" "\n alpha -- weight parameter" "\n alpha = 0 --> minimize storage" "\n alpha = 1 --> minimize solve ops" "\n outIVfile -- output file for oldToNew vector," "\n must be *.ivf or *.ivb" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else if ( (msgFile = fopen(argv[2], "a")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } inETreeFileName = argv[3] ; inGraphFileName = argv[4] ; alpha = atof(argv[5]) ; outIVfileName = argv[6] ; fprintf(msgFile, "\n %s " "\n msglvl -- %d" "\n msgFile -- %s" "\n inETreeFile -- %s" "\n inGraphFile -- %s" "\n alpha -- %f" "\n outIVfile -- %s" "\n", argv[0], msglvl, argv[2], inETreeFileName, inGraphFileName, alpha, outIVfileName) ; fflush(msgFile) ; /* ------------------------ read in the ETree object ------------------------ */ if ( strcmp(inETreeFileName, "none") == 0 ) { fprintf(msgFile, "\n no file to read from") ; spoolesFatal(); } etree = ETree_new() ; MARKTIME(t1) ; rc = ETree_readFromFile(etree, inETreeFileName) ; MARKTIME(t2) ; fprintf(msgFile, "\n CPU %9.5f : read in etree from file %s", t2 - t1, inETreeFileName) ; if ( rc != 1 ) { fprintf(msgFile, "\n return value %d from ETree_readFromFile(%p,%s)", rc, etree, inETreeFileName) ; spoolesFatal(); } ETree_leftJustify(etree) ; fprintf(msgFile, "\n\n after reading ETree object from file %s", inETreeFileName) ; if ( msglvl > 2 ) { ETree_writeForHumanEye(etree, msgFile) ; } else { ETree_writeStats(etree, msgFile) ; } fflush(msgFile) ; nfront = ETree_nfront(etree) ; tree = ETree_tree(etree) ; nodwghts = ETree_nodwghts(etree) ; vtxToFront = ETree_vtxToFront(etree) ; /* ------------------------ read in the Graph object ------------------------ */ if ( strcmp(inGraphFileName, "none") == 0 ) { fprintf(msgFile, "\n no file to read from") ; spoolesFatal(); } graph = Graph_new() ; MARKTIME(t1) ; rc = Graph_readFromFile(graph, inGraphFileName) ; nvtx = graph->nvtx ; vwghts = graph->vwghts ; MARKTIME(t2) ; fprintf(msgFile, "\n CPU %9.5f : read in graph from file %s", t2 - t1, inGraphFileName) ; if ( rc != 1 ) { fprintf(msgFile, "\n return value %d from Graph_readFromFile(%p,%s)", rc, graph, inGraphFileName) ; spoolesFatal(); } fprintf(msgFile, "\n\n after reading Graph object from file %s", inGraphFileName) ; if ( msglvl > 2 ) { Graph_writeForHumanEye(graph, msgFile) ; } else { Graph_writeStats(graph, msgFile) ; } fflush(msgFile) ; /* ---------------------- compute the statistics ---------------------- */ nnode1 = etree->tree->n ; nfind1 = ETree_nFactorIndices(etree) ; nfent1 = ETree_nFactorEntries(etree, SPOOLES_SYMMETRIC) ; nfops1 = ETree_nFactorOps(etree, SPOOLES_REAL, SPOOLES_SYMMETRIC) ; nleaves1 = Tree_nleaves(etree->tree) ; fprintf(stdout, "\n root front %d has %d vertices", etree->tree->root, etree->nodwghtsIV->vec[etree->tree->root]) ; /* --------------------------------- create the symbolic factorization --------------------------------- */ symbfacIVL = SymbFac_initFromGraph(etree, graph) ; if ( msglvl > 2 ) { IVL_writeForHumanEye(symbfacIVL, msgFile) ; } else { IVL_writeStats(symbfacIVL, msgFile) ; } fflush(msgFile) ; /* -------------------------- find the optimal partition -------------------------- */ compidsIV = ETree_optPart(etree, graph, symbfacIVL, alpha, &totalgain, msglvl, msgFile) ; if ( msglvl > 2 ) { IV_writeForHumanEye(compidsIV, msgFile) ; } else { IV_writeStats(compidsIV, msgFile) ; } fflush(msgFile) ; compids = IV_entries(compidsIV) ; /* ------------------------------------------------------ compute the number of vertices in the schur complement ------------------------------------------------------ */ for ( J = 0, nPhi = nV = 0. ; J < nfront ; J++ ) { if ( compids[J] == 0 ) { nPhi += nodwghts[J] ; } nV += nodwghts[J] ; } /* -------------------------------------------- compute the number of entries in L11 and L22 -------------------------------------------- */ nL11 = nL22 = 0 ; for ( J = Tree_postOTfirst(tree) ; J != -1 ; J = Tree_postOTnext(tree, J) ) { nJ = nodwghts[J] ; if ( msglvl > 3 ) { fprintf(msgFile, "\n\n front %d, nJ = %d", J, nJ) ; } IVL_listAndSize(symbfacIVL, J, &sizeJ, &adjJ) ; for ( ii = 0, inside = 0 ; ii < sizeJ ; ii++ ) { w = adjJ[ii] ; K = vtxToFront[w] ; if ( msglvl > 3 ) { fprintf(msgFile, "\n w = %d, K = %d", w, K) ; } if ( K > J && compids[K] == compids[J] ) { inside += (vwghts == NULL) ? 1 : vwghts[w] ; if ( msglvl > 3 ) { fprintf(msgFile, ", inside") ; } } } if ( compids[J] != 0 ) { if ( msglvl > 3 ) { fprintf(msgFile, "\n inside = %d, adding %d to L11", inside, nJ*nJ + 2*nJ*inside) ; } nL11 += (nJ*(nJ+1))/2 + nJ*inside ; } else { if ( msglvl > 3 ) { fprintf(msgFile, "\n inside = %d, adding %d to L22", inside, (nJ*(nJ+1))/2 + nJ*inside) ; } nL22 += (nJ*(nJ+1))/2 + nJ*inside ; } } if ( msglvl > 0 ) { fprintf(msgFile, "\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f", nfent1, nL11, nL22) ; } /* ------------------------------------ compute the number of entries in A21 ------------------------------------ */ nA21 = 0 ; if ( vwghts != NULL ) { for ( v = 0 ; v < nvtx ; v++ ) { J = vtxToFront[v] ; if ( compids[J] != 0 ) { Graph_adjAndSize(graph, v, &vsize, &vadj) ; for ( ii = 0 ; ii < vsize ; ii++ ) { w = vadj[ii] ; K = vtxToFront[w] ; if ( compids[K] == 0 ) { if ( msglvl > 3 ) { fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ; } nA21 += vwghts[v] * vwghts[w] ; } } } } } else { for ( v = 0 ; v < nvtx ; v++ ) { J = vtxToFront[v] ; if ( compids[J] != 0 ) { Graph_adjAndSize(graph, v, &vsize, &vadj) ; for ( ii = 0 ; ii < vsize ; ii++ ) { w = vadj[ii] ; K = vtxToFront[w] ; if ( compids[K] == 0 ) { if ( msglvl > 3 ) { fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ; } nA21++ ; } } } } } if ( msglvl > 0 ) { fprintf(msgFile, "\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f, |A21| = %.0f", nfent1, nL11, nL22, nA21) ; fprintf(msgFile, "\n storage: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f" "\n opcount: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f", nfent1, nL11 + nA21 + nL22, nfent1/(nL11 + nA21 + nL22), 2*nfent1, 4*nL11 + 2*nA21 + 2*nL22, 2*nfent1/(4*nL11 + 2*nA21 + 2*nL22)) ; fprintf(msgFile, "\n ratios %8.3f %8.3f %8.3f", nPhi/nV, nfent1/(nL11 + nA21 + nL22), 2*nfent1/(4*nL11 + 2*nA21 + 2*nL22)) ; } /* ---------------- free the objects ---------------- */ ETree_free(etree) ; Graph_free(graph) ; IVL_free(symbfacIVL) ; fprintf(msgFile, "\n") ; fclose(msgFile) ; return(1) ; }
/* ------------------------------------------------------------- purpose -- the IVL object ivl and IV object ownersIV are both found on each process. the ownersIV object is identical over all the processes, and owners[ii] tells which processes owns list ii of the ivl object. on return from this method, the ivl object is replicated over all the processes. each process sends the lists that it owns to all the other processes. created -- 98apr03, cca ------------------------------------------------------------- */ void IVL_MPI_allgather ( IVL *ivl, IV *ownersIV, int stats[], int msglvl, FILE *msgFile, int firsttag, MPI_Comm comm ) { int count, destination, ii, ilist, incount, jlist, jproc, left, maxcount, myid, nlist, nmylists, notherlists, nowners, nproc, offset, outcount, right, size, source, tag ; int *counts, *inbuffer, *list, *outbuffer, *owners ; MPI_Status status ; /* --------------- check the input --------------- */ if ( ivl == NULL || ownersIV == NULL ) { fprintf(stderr, "\n fatal error in IVL_MPI_allgather()" "\n ivl = %p, ownersIV = %p\n", ivl, ownersIV) ; exit(-1) ; } /* ---------------------------------------------- get id of self, # of processes and # of fronts ---------------------------------------------- */ MPI_Comm_rank(comm, &myid) ; MPI_Comm_size(comm, &nproc) ; nlist = ivl->nlist ; IV_sizeAndEntries(ownersIV, &nowners, &owners) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n inside IVL_MPI_allgather()" "\n nproc = %d, myid = %d, nlist = %d, nowners = %d", nproc, myid, nlist, nowners) ; fflush(msgFile) ; } if ( nlist != nowners || owners == NULL ) { fprintf(stderr, "\n fatal error in IVL_MPI_allgather()" "\n nlist = %d, nowners = %d, owners = %p\n", nlist, nowners, owners) ; exit(-1) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n ivl") ; IVL_writeForHumanEye(ivl, msgFile) ; fprintf(msgFile, "\n\n ownersIV") ; IV_writeForHumanEye(ownersIV, msgFile) ; fflush(msgFile) ; } /* ----------------------------------------------- step 1 : determine the size of the message that this process will send to the others ----------------------------------------------- */ for ( ilist = 0, outcount = 1 ; ilist < nlist ; ilist++ ) { if ( owners[ilist] < 0 || owners[ilist] >= nproc ) { fprintf(stderr, "\n owners[%d] = %d", ilist, owners[ilist]) ; exit(-1) ; } if ( owners[ilist] == myid ) { outcount += 2 ; IVL_listAndSize(ivl, ilist, &size, &list) ; outcount += size ; } } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n outcount = %d", outcount) ; fflush(msgFile) ; } /* ---------------------------------------------------- do an all-to-all gather/scatter counts[jproc] = # of int's in the message from jproc ---------------------------------------------------- */ counts = IVinit(nproc, 0) ; counts[myid] = outcount ; MPI_Allgather((void *) &counts[myid], 1, MPI_INT, (void *) counts, 1, MPI_INT, comm) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n counts") ; IVfprintf(msgFile, nproc, counts) ; fflush(msgFile) ; } /* ----------------------------- set up the in and out buffers ----------------------------- */ if ( outcount > 0 ) { outbuffer = IVinit(outcount, -1) ; for ( ilist = nmylists = 0, ii = 1 ; ilist < nlist ; ilist++ ) { if ( owners[ilist] == myid ) { nmylists++ ; IVL_listAndSize(ivl, ilist, &size, &list) ; outbuffer[ii++] = ilist ; outbuffer[ii++] = size ; if ( size > 0 ) { IVcopy(size, &outbuffer[ii], list) ; ii += size ; } } } outbuffer[0] = nmylists ; if ( ii != outcount ) { fprintf(stderr, "\n myid = %d, ii = %d, outcount = %d", myid, ii, outcount) ; fprintf(msgFile, "\n myid = %d, ii = %d, outcount = %d", myid, ii, outcount) ; exit(-1) ; } } else { outbuffer = NULL ; } maxcount = IVmax(nproc, counts, &jproc) ; if ( maxcount > 0 ) { inbuffer = IVinit(maxcount, -1) ; } else { inbuffer = NULL ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n outbuffer %p, maxcount %d, inbuffer %p", outbuffer, maxcount, inbuffer) ; fflush(msgFile) ; } /* ------------------------------------- step 2: loop over the other processes send and receive information ------------------------------------- */ for ( offset = 1, tag = firsttag ; offset < nproc ; offset++, tag++ ) { right = (myid + offset) % nproc ; if ( offset <= myid ) { left = myid - offset ; } else { left = nproc + myid - offset ; } if ( outcount > 0 ) { destination = right ; stats[0]++ ; stats[2] += outcount*sizeof(int) ; } else { destination = MPI_PROC_NULL ; } incount = counts[left] ; if ( incount > 0 ) { source = left ; stats[1]++ ; stats[3] += incount*sizeof(int) ; } else { source = MPI_PROC_NULL ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n offset %d, source %d, destination %d", offset, source, destination) ; fflush(msgFile) ; } /* ----------------- do a send/receive ----------------- */ MPI_Sendrecv(outbuffer, outcount, MPI_INT, destination, tag, inbuffer, incount, MPI_INT, source, tag, comm, &status) ; if ( source != MPI_PROC_NULL ) { MPI_Get_count(&status, MPI_INT, &count) ; if ( count != incount ) { fprintf(stderr, "\n 1. fatal error in IVL_MPI_allgather()" "\n proc %d : source = %d, count = %d, incount = %d\n", myid, source, count, incount) ; exit(-1) ; } } /* ---------------------------- set the values in the vector ---------------------------- */ notherlists = inbuffer[0] ; for ( ilist = 0, ii = 1 ; ilist < notherlists ; ilist++ ) { jlist = inbuffer[ii++] ; size = inbuffer[ii++] ; if ( size > 0 ) { IVL_setList(ivl, jlist, size, &inbuffer[ii]) ; ii += size ; } } if ( ii != incount ) { fprintf(msgFile, "\n ii = %d, incount = %d", ii, incount) ; fprintf(stderr, "\n ii = %d, incount = %d", ii, incount) ; exit(-1) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n after setting values") ; IVL_writeForHumanEye(ivl, msgFile) ; fflush(msgFile) ; } } /* ------------------------ free the working storage ------------------------ */ IVfree(counts) ; if ( outbuffer != NULL ) { IVfree(outbuffer) ; } if ( inbuffer != NULL ) { IVfree(inbuffer) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n leaving IVL_MPI_gatherall()") ; fflush(msgFile) ; } return ; }
/* ------------------------------------------- purpose -- to write an IVL object to a file input -- fn -- filename *.ivlb -- binary *.ivlf -- formatted anything else -- for human eye return value -- 1 if success, 0 otherwise created -- 95sep29, cca ------------------------------------------- */ int IVL_writeToFile ( IVL *ivl, char *fn ) { FILE *fp ; int fnlength, rc, sulength ; /* --------------- check the input --------------- */ if ( ivl == NULL || fn == NULL ) { fprintf(stderr, "\n fatal error in IVL_writeToFile(%p,%s)" "\n bad input\n", ivl, fn) ; } switch ( ivl->type ) { case IVL_CHUNKED : case IVL_SOLO : case IVL_UNKNOWN : break ; default : fprintf(stderr, "\n fatal error in IVL_writeToFile(%p,%s)" "\n bad type = %d", ivl, fn, ivl->type) ; return(0) ; } /* ------------------ write out the file ------------------ */ fnlength = strlen(fn) ; sulength = strlen(suffixb) ; if ( fnlength > sulength ) { if ( strcmp(&fn[fnlength-sulength], suffixb) == 0 ) { if ( (fp = fopen(fn, "wb")) == NULL ) { fprintf(stderr, "\n error in IVL_writeToFile(%p,%s)" "\n unable to open file %s", ivl, fn, fn) ; rc = 0 ; } else { rc = IVL_writeToBinaryFile(ivl, fp) ; fclose(fp) ; } } else if ( strcmp(&fn[fnlength-sulength], suffixf) == 0 ) { if ( (fp = fopen(fn, "w")) == NULL ) { fprintf(stderr, "\n error in IVL_writeToFile(%p,%s)" "\n unable to open file %s", ivl, fn, fn) ; rc = 0 ; } else { rc = IVL_writeToFormattedFile(ivl, fp) ; fclose(fp) ; } } else { if ( (fp = fopen(fn, "a")) == NULL ) { fprintf(stderr, "\n error in IVL_writeToFile(%p,%s)" "\n unable to open file %s", ivl, fn, fn) ; rc = 0 ; } else { rc = IVL_writeForHumanEye(ivl, fp) ; fclose(fp) ; } } } else { if ( (fp = fopen(fn, "a")) == NULL ) { fprintf(stderr, "\n error in IVL_writeToFile(%p,%s)" "\n unable to open file %s", ivl, fn, fn) ; rc = 0 ; } else { rc = IVL_writeForHumanEye(ivl, fp) ; fclose(fp) ; } } return(rc) ; }
PetscErrorCode MatFactorNumeric_SeqSpooles(Mat F,Mat A,const MatFactorInfo *info) { Mat_Spooles *lu = (Mat_Spooles*)(F)->spptr; ChvManager *chvmanager ; Chv *rootchv ; IVL *adjIVL; PetscErrorCode ierr; PetscInt nz,nrow=A->rmap->n,irow,nedges,neqns=A->cmap->n,*ai,*aj,i,*diag=0,fierr; PetscScalar *av; double cputotal,facops; #if defined(PETSC_USE_COMPLEX) PetscInt nz_row,*aj_tmp; PetscScalar *av_tmp; #else PetscInt *ivec1,*ivec2,j; double *dvec; #endif PetscBool isSeqAIJ,isMPIAIJ; PetscFunctionBegin; if (lu->flg == DIFFERENT_NONZERO_PATTERN) { /* first numeric factorization */ (F)->ops->solve = MatSolve_SeqSpooles; (F)->assembled = PETSC_TRUE; /* set Spooles options */ ierr = SetSpoolesOptions(A, &lu->options);CHKERRQ(ierr); lu->mtxA = InpMtx_new(); } /* copy A to Spooles' InpMtx object */ ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isSeqAIJ);CHKERRQ(ierr); ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isMPIAIJ);CHKERRQ(ierr); if (isSeqAIJ){ Mat_SeqAIJ *mat = (Mat_SeqAIJ*)A->data; ai=mat->i; aj=mat->j; av=mat->a; if (lu->options.symflag == SPOOLES_NONSYMMETRIC) { nz=mat->nz; } else { /* SPOOLES_SYMMETRIC || SPOOLES_HERMITIAN */ nz=(mat->nz + A->rmap->n)/2; diag=mat->diag; } } else { /* A is SBAIJ */ Mat_SeqSBAIJ *mat = (Mat_SeqSBAIJ*)A->data; ai=mat->i; aj=mat->j; av=mat->a; nz=mat->nz; } InpMtx_init(lu->mtxA, INPMTX_BY_ROWS, lu->options.typeflag, nz, 0); #if defined(PETSC_USE_COMPLEX) for (irow=0; irow<nrow; irow++) { if ( lu->options.symflag == SPOOLES_NONSYMMETRIC || !(isSeqAIJ || isMPIAIJ)){ nz_row = ai[irow+1] - ai[irow]; aj_tmp = aj + ai[irow]; av_tmp = av + ai[irow]; } else { nz_row = ai[irow+1] - diag[irow]; aj_tmp = aj + diag[irow]; av_tmp = av + diag[irow]; } for (i=0; i<nz_row; i++){ InpMtx_inputComplexEntry(lu->mtxA, irow, *aj_tmp++,PetscRealPart(*av_tmp),PetscImaginaryPart(*av_tmp)); av_tmp++; } } #else ivec1 = InpMtx_ivec1(lu->mtxA); ivec2 = InpMtx_ivec2(lu->mtxA); dvec = InpMtx_dvec(lu->mtxA); if ( lu->options.symflag == SPOOLES_NONSYMMETRIC || !isSeqAIJ){ for (irow = 0; irow < nrow; irow++){ for (i = ai[irow]; i<ai[irow+1]; i++) ivec1[i] = irow; } IVcopy(nz, ivec2, aj); DVcopy(nz, dvec, av); } else { nz = 0; for (irow = 0; irow < nrow; irow++){ for (j = diag[irow]; j<ai[irow+1]; j++) { ivec1[nz] = irow; ivec2[nz] = aj[j]; dvec[nz] = av[j]; nz++; } } } InpMtx_inputRealTriples(lu->mtxA, nz, ivec1, ivec2, dvec); #endif InpMtx_changeStorageMode(lu->mtxA, INPMTX_BY_VECTORS); if ( lu->options.msglvl > 0 ) { int err; printf("\n\n input matrix"); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n input matrix");CHKERRQ(ierr); InpMtx_writeForHumanEye(lu->mtxA, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } if ( lu->flg == DIFFERENT_NONZERO_PATTERN){ /* first numeric factorization */ /*--------------------------------------------------- find a low-fill ordering (1) create the Graph object (2) order the graph -------------------------------------------------------*/ if (lu->options.useQR){ adjIVL = InpMtx_adjForATA(lu->mtxA); } else { adjIVL = InpMtx_fullAdjacency(lu->mtxA); } nedges = IVL_tsize(adjIVL); lu->graph = Graph_new(); Graph_init2(lu->graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL); if ( lu->options.msglvl > 2 ) { int err; if (lu->options.useQR){ ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n graph of A^T A");CHKERRQ(ierr); } else { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n graph of the input matrix");CHKERRQ(ierr); } Graph_writeForHumanEye(lu->graph, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } switch (lu->options.ordering) { case 0: lu->frontETree = orderViaBestOfNDandMS(lu->graph, lu->options.maxdomainsize, lu->options.maxzeros, lu->options.maxsize, lu->options.seed, lu->options.msglvl, lu->options.msgFile); break; case 1: lu->frontETree = orderViaMMD(lu->graph,lu->options.seed,lu->options.msglvl,lu->options.msgFile); break; case 2: lu->frontETree = orderViaMS(lu->graph, lu->options.maxdomainsize, lu->options.seed,lu->options.msglvl,lu->options.msgFile); break; case 3: lu->frontETree = orderViaND(lu->graph, lu->options.maxdomainsize, lu->options.seed,lu->options.msglvl,lu->options.msgFile); break; default: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown Spooles's ordering"); } if ( lu->options.msglvl > 0 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n front tree from ordering");CHKERRQ(ierr); ETree_writeForHumanEye(lu->frontETree, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } /* get the permutation, permute the front tree */ lu->oldToNewIV = ETree_oldToNewVtxPerm(lu->frontETree); lu->oldToNew = IV_entries(lu->oldToNewIV); lu->newToOldIV = ETree_newToOldVtxPerm(lu->frontETree); if (!lu->options.useQR) ETree_permuteVertices(lu->frontETree, lu->oldToNewIV); /* permute the matrix */ if (lu->options.useQR){ InpMtx_permute(lu->mtxA, NULL, lu->oldToNew); } else { InpMtx_permute(lu->mtxA, lu->oldToNew, lu->oldToNew); if ( lu->options.symflag == SPOOLES_SYMMETRIC) { InpMtx_mapToUpperTriangle(lu->mtxA); } #if defined(PETSC_USE_COMPLEX) if ( lu->options.symflag == SPOOLES_HERMITIAN ) { InpMtx_mapToUpperTriangleH(lu->mtxA); } #endif InpMtx_changeCoordType(lu->mtxA, INPMTX_BY_CHEVRONS); } InpMtx_changeStorageMode(lu->mtxA, INPMTX_BY_VECTORS); /* get symbolic factorization */ if (lu->options.useQR){ lu->symbfacIVL = SymbFac_initFromGraph(lu->frontETree, lu->graph); IVL_overwrite(lu->symbfacIVL, lu->oldToNewIV); IVL_sortUp(lu->symbfacIVL); ETree_permuteVertices(lu->frontETree, lu->oldToNewIV); } else { lu->symbfacIVL = SymbFac_initFromInpMtx(lu->frontETree, lu->mtxA); } if ( lu->options.msglvl > 2 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n old-to-new permutation vector");CHKERRQ(ierr); IV_writeForHumanEye(lu->oldToNewIV, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n new-to-old permutation vector");CHKERRQ(ierr); IV_writeForHumanEye(lu->newToOldIV, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n front tree after permutation");CHKERRQ(ierr); ETree_writeForHumanEye(lu->frontETree, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n input matrix after permutation");CHKERRQ(ierr); InpMtx_writeForHumanEye(lu->mtxA, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n symbolic factorization");CHKERRQ(ierr); IVL_writeForHumanEye(lu->symbfacIVL, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } lu->frontmtx = FrontMtx_new(); lu->mtxmanager = SubMtxManager_new(); SubMtxManager_init(lu->mtxmanager, NO_LOCK, 0); } else { /* new num factorization using previously computed symbolic factor */ if (lu->options.pivotingflag) { /* different FrontMtx is required */ FrontMtx_free(lu->frontmtx); lu->frontmtx = FrontMtx_new(); } else { FrontMtx_clearData (lu->frontmtx); } SubMtxManager_free(lu->mtxmanager); lu->mtxmanager = SubMtxManager_new(); SubMtxManager_init(lu->mtxmanager, NO_LOCK, 0); /* permute mtxA */ if (lu->options.useQR){ InpMtx_permute(lu->mtxA, NULL, lu->oldToNew); } else { InpMtx_permute(lu->mtxA, lu->oldToNew, lu->oldToNew); if ( lu->options.symflag == SPOOLES_SYMMETRIC ) { InpMtx_mapToUpperTriangle(lu->mtxA); } InpMtx_changeCoordType(lu->mtxA, INPMTX_BY_CHEVRONS); } InpMtx_changeStorageMode(lu->mtxA, INPMTX_BY_VECTORS); if ( lu->options.msglvl > 2 ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n input matrix after permutation");CHKERRQ(ierr); InpMtx_writeForHumanEye(lu->mtxA, lu->options.msgFile); } } /* end of if( lu->flg == DIFFERENT_NONZERO_PATTERN) */ if (lu->options.useQR){ FrontMtx_init(lu->frontmtx, lu->frontETree, lu->symbfacIVL, lu->options.typeflag, SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL, lu->mtxmanager, lu->options.msglvl, lu->options.msgFile); } else { FrontMtx_init(lu->frontmtx, lu->frontETree, lu->symbfacIVL, lu->options.typeflag, lu->options.symflag, FRONTMTX_DENSE_FRONTS, lu->options.pivotingflag, NO_LOCK, 0, NULL, lu->mtxmanager, lu->options.msglvl, lu->options.msgFile); } if ( lu->options.symflag == SPOOLES_SYMMETRIC ) { /* || SPOOLES_HERMITIAN ? */ if ( lu->options.patchAndGoFlag == 1 ) { lu->frontmtx->patchinfo = PatchAndGoInfo_new(); PatchAndGoInfo_init(lu->frontmtx->patchinfo, 1, lu->options.toosmall, lu->options.fudge, lu->options.storeids, lu->options.storevalues); } else if ( lu->options.patchAndGoFlag == 2 ) { lu->frontmtx->patchinfo = PatchAndGoInfo_new(); PatchAndGoInfo_init(lu->frontmtx->patchinfo, 2, lu->options.toosmall, lu->options.fudge, lu->options.storeids, lu->options.storevalues); } } /* numerical factorization */ chvmanager = ChvManager_new(); ChvManager_init(chvmanager, NO_LOCK, 1); DVfill(10, lu->cpus, 0.0); if (lu->options.useQR){ facops = 0.0 ; FrontMtx_QR_factor(lu->frontmtx, lu->mtxA, chvmanager, lu->cpus, &facops, lu->options.msglvl, lu->options.msgFile); if ( lu->options.msglvl > 1 ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n factor matrix");CHKERRQ(ierr); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n facops = %9.2f", facops);CHKERRQ(ierr); } } else { IVfill(20, lu->stats, 0); rootchv = FrontMtx_factorInpMtx(lu->frontmtx, lu->mtxA, lu->options.tau, 0.0, chvmanager, &fierr, lu->cpus,lu->stats,lu->options.msglvl,lu->options.msgFile); if (rootchv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"\n matrix found to be singular"); if (fierr >= 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"\n error encountered at front %D", fierr); if(lu->options.FrontMtxInfo){ ierr = PetscPrintf(PETSC_COMM_SELF,"\n %8d pivots, %8d pivot tests, %8d delayed rows and columns\n",lu->stats[0], lu->stats[1], lu->stats[2]);CHKERRQ(ierr); cputotal = lu->cpus[8] ; if ( cputotal > 0.0 ) { ierr = PetscPrintf(PETSC_COMM_SELF, "\n cpus cpus/totaltime" "\n initialize fronts %8.3f %6.2f" "\n load original entries %8.3f %6.2f" "\n update fronts %8.3f %6.2f" "\n assemble postponed data %8.3f %6.2f" "\n factor fronts %8.3f %6.2f" "\n extract postponed data %8.3f %6.2f" "\n store factor entries %8.3f %6.2f" "\n miscellaneous %8.3f %6.2f" "\n total time %8.3f \n", lu->cpus[0], 100.*lu->cpus[0]/cputotal, lu->cpus[1], 100.*lu->cpus[1]/cputotal, lu->cpus[2], 100.*lu->cpus[2]/cputotal, lu->cpus[3], 100.*lu->cpus[3]/cputotal, lu->cpus[4], 100.*lu->cpus[4]/cputotal, lu->cpus[5], 100.*lu->cpus[5]/cputotal, lu->cpus[6], 100.*lu->cpus[6]/cputotal, lu->cpus[7], 100.*lu->cpus[7]/cputotal, cputotal);CHKERRQ(ierr); } } } ChvManager_free(chvmanager); if ( lu->options.msglvl > 0 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n factor matrix");CHKERRQ(ierr); FrontMtx_writeForHumanEye(lu->frontmtx, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } if ( lu->options.symflag == SPOOLES_SYMMETRIC ) { /* || SPOOLES_HERMITIAN ? */ if ( lu->options.patchAndGoFlag == 1 ) { if ( lu->frontmtx->patchinfo->fudgeIV != NULL ) { if (lu->options.msglvl > 0 ){ ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n small pivots found at these locations");CHKERRQ(ierr); IV_writeForHumanEye(lu->frontmtx->patchinfo->fudgeIV, lu->options.msgFile); } } PatchAndGoInfo_free(lu->frontmtx->patchinfo); } else if ( lu->options.patchAndGoFlag == 2 ) { if (lu->options.msglvl > 0 ){ if ( lu->frontmtx->patchinfo->fudgeIV != NULL ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n small pivots found at these locations");CHKERRQ(ierr); IV_writeForHumanEye(lu->frontmtx->patchinfo->fudgeIV, lu->options.msgFile); } if ( lu->frontmtx->patchinfo->fudgeDV != NULL ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n perturbations");CHKERRQ(ierr); DV_writeForHumanEye(lu->frontmtx->patchinfo->fudgeDV, lu->options.msgFile); } } PatchAndGoInfo_free(lu->frontmtx->patchinfo); } } /* post-process the factorization */ FrontMtx_postProcess(lu->frontmtx, lu->options.msglvl, lu->options.msgFile); if ( lu->options.msglvl > 2 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n factor matrix after post-processing");CHKERRQ(ierr); FrontMtx_writeForHumanEye(lu->frontmtx, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } lu->flg = SAME_NONZERO_PATTERN; lu->CleanUpSpooles = PETSC_TRUE; PetscFunctionReturn(0); }
NM_Status SpoolesSolver :: solve(SparseMtrx *A, FloatArray *b, FloatArray *x) { int errorValue, mtxType, symmetryflag; int seed = 30145, pivotingflag = 0; int *oldToNew, *newToOld; double droptol = 0.0, tau = 1.e300; double cpus [ 10 ]; int stats [ 20 ]; ChvManager *chvmanager; Chv *rootchv; InpMtx *mtxA; DenseMtx *mtxY, *mtxX; // first check whether Lhs is defined if ( !A ) { _error("solveYourselfAt: unknown Lhs"); } // and whether Rhs if ( !b ) { _error("solveYourselfAt: unknown Rhs"); } // and whether previous Solution exist if ( !x ) { _error("solveYourselfAt: unknown solution array"); } if ( x->giveSize() != b->giveSize() ) { _error("solveYourselfAt: size mismatch"); } Timer timer; timer.startTimer(); if ( A->giveType() != SMT_SpoolesMtrx ) { _error("solveYourselfAt: SpoolesSparseMtrx Expected"); } mtxA = ( ( SpoolesSparseMtrx * ) A )->giveInpMtrx(); mtxType = ( ( SpoolesSparseMtrx * ) A )->giveValueType(); symmetryflag = ( ( SpoolesSparseMtrx * ) A )->giveSymmetryFlag(); int i; int neqns = A->giveNumberOfRows(); int nrhs = 1; /* convert right-hand side to DenseMtx */ mtxY = DenseMtx_new(); DenseMtx_init(mtxY, mtxType, 0, 0, neqns, nrhs, 1, neqns); DenseMtx_zero(mtxY); for ( i = 0; i < neqns; i++ ) { DenseMtx_setRealEntry( mtxY, i, 0, b->at(i + 1) ); } if ( ( Lhs != A ) || ( this->lhsVersion != A->giveVersion() ) ) { // // lhs has been changed -> new factorization // Lhs = A; this->lhsVersion = A->giveVersion(); if ( frontmtx ) { FrontMtx_free(frontmtx); } if ( newToOldIV ) { IV_free(newToOldIV); } if ( oldToNewIV ) { IV_free(oldToNewIV); } if ( frontETree ) { ETree_free(frontETree); } if ( symbfacIVL ) { IVL_free(symbfacIVL); } if ( mtxmanager ) { SubMtxManager_free(mtxmanager); } if ( graph ) { Graph_free(graph); } /* * ------------------------------------------------- * STEP 3 : find a low-fill ordering * (1) create the Graph object * (2) order the graph using multiple minimum degree * ------------------------------------------------- */ int nedges; graph = Graph_new(); adjIVL = InpMtx_fullAdjacency(mtxA); nedges = IVL_tsize(adjIVL); Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n graph of the input matrix"); Graph_writeForHumanEye(graph, msgFile); fflush(msgFile); } frontETree = orderViaMMD(graph, seed, msglvl, msgFile); if ( msglvl > 0 ) { fprintf(msgFile, "\n\n front tree from ordering"); ETree_writeForHumanEye(frontETree, msgFile); fflush(msgFile); } /* * ---------------------------------------------------- * STEP 4: get the permutation, permute the front tree, * permute the matrix and right hand side, and * get the symbolic factorization * ---------------------------------------------------- */ oldToNewIV = ETree_oldToNewVtxPerm(frontETree); oldToNew = IV_entries(oldToNewIV); newToOldIV = ETree_newToOldVtxPerm(frontETree); newToOld = IV_entries(newToOldIV); ETree_permuteVertices(frontETree, oldToNewIV); InpMtx_permute(mtxA, oldToNew, oldToNew); if ( symmetryflag == SPOOLES_SYMMETRIC || symmetryflag == SPOOLES_HERMITIAN ) { InpMtx_mapToUpperTriangle(mtxA); } InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS); InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS); symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n old-to-new permutation vector"); IV_writeForHumanEye(oldToNewIV, msgFile); fprintf(msgFile, "\n\n new-to-old permutation vector"); IV_writeForHumanEye(newToOldIV, msgFile); fprintf(msgFile, "\n\n front tree after permutation"); ETree_writeForHumanEye(frontETree, msgFile); fprintf(msgFile, "\n\n input matrix after permutation"); InpMtx_writeForHumanEye(mtxA, msgFile); fprintf(msgFile, "\n\n symbolic factorization"); IVL_writeForHumanEye(symbfacIVL, msgFile); fflush(msgFile); } Tree_writeToFile(frontETree->tree, (char*)"haggar.treef"); /*--------------------------------------------------------------------*/ /* * ------------------------------------------ * STEP 5: initialize the front matrix object * ------------------------------------------ */ frontmtx = FrontMtx_new(); mtxmanager = SubMtxManager_new(); SubMtxManager_init(mtxmanager, NO_LOCK, 0); FrontMtx_init(frontmtx, frontETree, symbfacIVL, mtxType, symmetryflag, FRONTMTX_DENSE_FRONTS, pivotingflag, NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile); /*--------------------------------------------------------------------*/ /* * ----------------------------------------- * STEP 6: compute the numeric factorization * ----------------------------------------- */ chvmanager = ChvManager_new(); ChvManager_init(chvmanager, NO_LOCK, 1); DVfill(10, cpus, 0.0); IVfill(20, stats, 0); rootchv = FrontMtx_factorInpMtx(frontmtx, mtxA, tau, droptol, chvmanager, & errorValue, cpus, stats, msglvl, msgFile); ChvManager_free(chvmanager); if ( msglvl > 0 ) { fprintf(msgFile, "\n\n factor matrix"); FrontMtx_writeForHumanEye(frontmtx, msgFile); fflush(msgFile); } if ( rootchv != NULL ) { fprintf(msgFile, "\n\n matrix found to be singular\n"); exit(-1); } if ( errorValue >= 0 ) { fprintf(msgFile, "\n\n error encountered at front %d", errorValue); exit(-1); } /*--------------------------------------------------------------------*/ /* * -------------------------------------- * STEP 7: post-process the factorization * -------------------------------------- */ FrontMtx_postProcess(frontmtx, msglvl, msgFile); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n factor matrix after post-processing"); FrontMtx_writeForHumanEye(frontmtx, msgFile); fflush(msgFile); } /*--------------------------------------------------------------------*/ } /* * ---------------------------------------------------- * STEP 4: permute the right hand side * ---------------------------------------------------- */ DenseMtx_permuteRows(mtxY, oldToNewIV); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n right hand side matrix after permutation"); DenseMtx_writeForHumanEye(mtxY, msgFile); } /* * ------------------------------- * STEP 8: solve the linear system * ------------------------------- */ mtxX = DenseMtx_new(); DenseMtx_init(mtxX, mtxType, 0, 0, neqns, nrhs, 1, neqns); DenseMtx_zero(mtxX); FrontMtx_solve(frontmtx, mtxX, mtxY, mtxmanager, cpus, msglvl, msgFile); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n solution matrix in new ordering"); DenseMtx_writeForHumanEye(mtxX, msgFile); fflush(msgFile); } /*--------------------------------------------------------------------*/ /* * ------------------------------------------------------- * STEP 9: permute the solution into the original ordering * ------------------------------------------------------- */ DenseMtx_permuteRows(mtxX, newToOldIV); if ( msglvl > 0 ) { fprintf(msgFile, "\n\n solution matrix in original ordering"); DenseMtx_writeForHumanEye(mtxX, msgFile); fflush(msgFile); } // DenseMtx_writeForMatlab(mtxX, "x", msgFile) ; /*--------------------------------------------------------------------*/ /* fetch data to oofem vectors */ double *xptr = x->givePointer(); for ( i = 0; i < neqns; i++ ) { DenseMtx_realEntry(mtxX, i, 0, xptr + i); // printf ("x(%d) = %e\n", i+1, *(xptr+i)); } // DenseMtx_copyRowIntoVector(mtxX, 0, x->givePointer()); timer.stopTimer(); OOFEM_LOG_DEBUG( "SpoolesSolver info: user time consumed by solution: %.2fs\n", timer.getUtime() ); /* * ----------- * free memory * ----------- */ DenseMtx_free(mtxX); DenseMtx_free(mtxY); /*--------------------------------------------------------------------*/ return ( 1 ); }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) /* ------------------------------------------------- this program tests the IVL_MPI_allgather() method (1) each process generates the same owners[n] map (2) each process creates an IVL object and fills its owned lists with random numbers (3) the processes gather-all's the lists of ivl created -- 98apr03, cca ------------------------------------------------- */ { char *buffer ; double chksum, globalsum, t1, t2 ; Drand drand ; int ilist, length, myid, msglvl, nlist, nproc, rc, seed, size, tag ; int *list, *owners, *vec ; int stats[4], tstats[4] ; IV *ownersIV ; IVL *ivl ; FILE *msgFile ; /* --------------------------------------------------------------- find out the identity of this process and the number of process --------------------------------------------------------------- */ MPI_Init(&argc, &argv) ; MPI_Comm_rank(MPI_COMM_WORLD, &myid) ; MPI_Comm_size(MPI_COMM_WORLD, &nproc) ; fprintf(stdout, "\n process %d of %d, argc = %d", myid, nproc, argc) ; fflush(stdout) ; if ( argc != 5 ) { fprintf(stdout, "\n\n usage : %s msglvl msgFile n seed " "\n msglvl -- message level" "\n msgFile -- message file" "\n nlist -- number of lists in the IVL object" "\n seed -- random number seed" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else { length = strlen(argv[2]) + 1 + 4 ; buffer = CVinit(length, '\0') ; sprintf(buffer, "%s.%d", argv[2], myid) ; if ( (msgFile = fopen(buffer, "w")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } CVfree(buffer) ; } nlist = atoi(argv[3]) ; seed = atoi(argv[4]) ; fprintf(msgFile, "\n %s " "\n msglvl -- %d" "\n msgFile -- %s" "\n nlist -- %d" "\n seed -- %d" "\n", argv[0], msglvl, argv[2], nlist, seed) ; fflush(msgFile) ; /* ---------------------------- generate the ownersIV object ---------------------------- */ MARKTIME(t1) ; ownersIV = IV_new() ; IV_init(ownersIV, nlist, NULL) ; owners = IV_entries(ownersIV) ; Drand_setDefaultFields(&drand) ; Drand_setSeed(&drand, seed) ; Drand_setUniform(&drand, 0, nproc) ; Drand_fillIvector(&drand, nlist, owners) ; MARKTIME(t2) ; fprintf(msgFile, "\n CPU %8.3f : initialize the ownersIV object", t2 - t1) ; fflush(msgFile) ; fprintf(msgFile, "\n\n ownersIV generated") ; if ( msglvl > 2 ) { IV_writeForHumanEye(ownersIV, msgFile) ; } else { IV_writeStats(ownersIV, msgFile) ; } fflush(msgFile) ; /* -------------------------------------------- set up the IVL object and fill owned entries -------------------------------------------- */ MARKTIME(t1) ; ivl = IVL_new() ; IVL_init1(ivl, IVL_CHUNKED, nlist) ; vec = IVinit(nlist, -1) ; Drand_setSeed(&drand, seed + myid) ; Drand_setUniform(&drand, 0, nlist) ; for ( ilist = 0 ; ilist < nlist ; ilist++ ) { if ( owners[ilist] == myid ) { size = (int) Drand_value(&drand) ; Drand_fillIvector(&drand, size, vec) ; IVL_setList(ivl, ilist, size, vec) ; } } MARKTIME(t2) ; fprintf(msgFile, "\n CPU %8.3f : initialize the IVL object", t2 - t1) ; fflush(msgFile) ; if ( msglvl > 2 ) { IVL_writeForHumanEye(ivl, msgFile) ; } else { IVL_writeStats(ivl, msgFile) ; } fflush(msgFile) ; /* -------------------------------------------- compute the local checksum of the ivl object -------------------------------------------- */ for ( ilist = 0, chksum = 0.0 ; ilist < nlist ; ilist++ ) { if ( owners[ilist] == myid ) { IVL_listAndSize(ivl, ilist, &size, &list) ; chksum += 1 + ilist + size + IVsum(size, list) ; } } fprintf(msgFile, "\n\n local partial chksum = %12.4e", chksum) ; fflush(msgFile) ; /* ----------------------- get the global checksum ----------------------- */ rc = MPI_Allreduce((void *) &chksum, (void *) &globalsum, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD) ; /* -------------------------------- execute the all-gather operation -------------------------------- */ tag = 47 ; IVzero(4, stats) ; IVL_MPI_allgather(ivl, ownersIV, stats, msglvl, msgFile, tag, MPI_COMM_WORLD) ; if ( msglvl > 0 ) { fprintf(msgFile, "\n\n return from IVL_MPI_allgather()") ; fprintf(msgFile, "\n local send stats : %10d messages with %10d bytes" "\n local recv stats : %10d messages with %10d bytes", stats[0], stats[2], stats[1], stats[3]) ; fflush(msgFile) ; } MPI_Reduce((void *) stats, (void *) tstats, 4, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD) ; if ( myid == 0 ) { fprintf(msgFile, "\n total send stats : %10d messages with %10d bytes" "\n total recv stats : %10d messages with %10d bytes", tstats[0], tstats[2], tstats[1], tstats[3]) ; fflush(msgFile) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n ivl") ; IVL_writeForHumanEye(ivl, msgFile) ; fflush(msgFile) ; } /* ----------------------------------------- compute the checksum of the entire object ----------------------------------------- */ for ( ilist = 0, chksum = 0.0 ; ilist < nlist ; ilist++ ) { IVL_listAndSize(ivl, ilist, &size, &list) ; chksum += 1 + ilist + size + IVsum(size, list) ; } fprintf(msgFile, "\n globalsum = %12.4e, chksum = %12.4e, error = %12.4e", globalsum, chksum, fabs(globalsum - chksum)) ; fflush(msgFile) ; /* ---------------- free the objects ---------------- */ IV_free(ownersIV) ; IVL_free(ivl) ; /* ------------------------ exit the MPI environment ------------------------ */ MPI_Finalize() ; fprintf(msgFile, "\n") ; fclose(msgFile) ; return(0) ; }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) { /* -------------------------------------------------- QR all-in-one program (1) read in matrix entries and form InpMtx object of A and A^TA (2) form Graph object of A^TA (3) order matrix and form front tree (4) get the permutation, permute the matrix and front tree and get the symbolic factorization (5) compute the numeric factorization (6) read in right hand side entries (7) compute the solution created -- 98jun11, cca -------------------------------------------------- */ /*--------------------------------------------------------------------*/ char *matrixFileName, *rhsFileName ; ChvManager *chvmanager ; DenseMtx *mtxB, *mtxX ; double facops, imag, real, value ; double cpus[10] ; ETree *frontETree ; FILE *inputFile, *msgFile ; FrontMtx *frontmtx ; Graph *graph ; int ient, irow, jcol, jrhs, jrow, msglvl, neqns, nedges, nent, nrhs, nrow, seed, type ; InpMtx *mtxA ; IV *newToOldIV, *oldToNewIV ; IVL *adjIVL, *symbfacIVL ; SubMtxManager *mtxmanager ; /*--------------------------------------------------------------------*/ /* -------------------- get input parameters -------------------- */ if ( argc != 7 ) { fprintf(stdout, "\n usage: %s msglvl msgFile type matrixFileName rhsFileName seed" "\n msglvl -- message level" "\n msgFile -- message file" "\n type -- type of entries" "\n 1 (SPOOLES_REAL) -- real entries" "\n 2 (SPOOLES_COMPLEX) -- complex entries" "\n matrixFileName -- matrix file name, format" "\n nrow ncol nent" "\n irow jcol entry" "\n ..." "\n note: indices are zero based" "\n rhsFileName -- right hand side file name, format" "\n nrow " "\n entry[0]" "\n ..." "\n entry[nrow-1]" "\n seed -- random number seed, used for ordering" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else if ( (msgFile = fopen(argv[2], "a")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } type = atoi(argv[3]) ; matrixFileName = argv[4] ; rhsFileName = argv[5] ; seed = atoi(argv[6]) ; /*--------------------------------------------------------------------*/ /* -------------------------------------------- STEP 1: read the entries from the input file and create the InpMtx object of A -------------------------------------------- */ inputFile = fopen(matrixFileName, "r") ; fscanf(inputFile, "%d %d %d", &nrow, &neqns, &nent) ; mtxA = InpMtx_new() ; InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ; if ( type == SPOOLES_REAL ) { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ; InpMtx_inputRealEntry(mtxA, irow, jcol, value) ; } } else { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ; InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ; } } fclose(inputFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n input matrix") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ---------------------------------------- STEP 2: read the right hand side entries ---------------------------------------- */ inputFile = fopen(rhsFileName, "r") ; fscanf(inputFile, "%d %d", &nrow, &nrhs) ; mtxB = DenseMtx_new() ; DenseMtx_init(mtxB, type, 0, 0, nrow, nrhs, 1, nrow) ; DenseMtx_zero(mtxB) ; if ( type == SPOOLES_REAL ) { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le", &value) ; DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ; } } } else { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le %le", &real, &imag) ; DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ; } } } fclose(inputFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n rhs matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxB, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------------- STEP 3 : find a low-fill ordering (1) create the Graph object for A^TA or A^HA (2) order the graph using multiple minimum degree ------------------------------------------------- */ graph = Graph_new() ; adjIVL = InpMtx_adjForATA(mtxA) ; nedges = IVL_tsize(adjIVL) ; Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n graph of A^T A") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------------------- STEP 4: get the permutation, permute the matrix and front tree and get the symbolic factorization ----------------------------------------------------- */ oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ; newToOldIV = ETree_newToOldVtxPerm(frontETree) ; InpMtx_permute(mtxA, NULL, IV_entries(oldToNewIV)) ; InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ; IVL_overwrite(symbfacIVL, oldToNewIV) ; IVL_sortUp(symbfacIVL) ; ETree_permuteVertices(frontETree, oldToNewIV) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n old-to-new permutation vector") ; IV_writeForHumanEye(oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation vector") ; IV_writeForHumanEye(newToOldIV, msgFile) ; fprintf(msgFile, "\n\n front tree after permutation") ; ETree_writeForHumanEye(frontETree, msgFile) ; fprintf(msgFile, "\n\n input matrix after permutation") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(symbfacIVL, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------ STEP 5: initialize the front matrix object ------------------------------------------ */ frontmtx = FrontMtx_new() ; mtxmanager = SubMtxManager_new() ; SubMtxManager_init(mtxmanager, NO_LOCK, 0) ; if ( type == SPOOLES_REAL ) { FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile) ; } else { FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, SPOOLES_HERMITIAN, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------- STEP 6: compute the numeric factorization ----------------------------------------- */ chvmanager = ChvManager_new() ; ChvManager_init(chvmanager, NO_LOCK, 1) ; DVzero(10, cpus) ; facops = 0.0 ; FrontMtx_QR_factor(frontmtx, mtxA, chvmanager, cpus, &facops, msglvl, msgFile) ; ChvManager_free(chvmanager) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix") ; fprintf(msgFile, "\n facops = %9.2f", facops) ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------- STEP 7: post-process the factorization -------------------------------------- */ FrontMtx_postProcess(frontmtx, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix after post-processing") ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------- STEP 8: solve the linear system ------------------------------- */ mtxX = DenseMtx_new() ; DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ; FrontMtx_QR_solve(frontmtx, mtxA, mtxX, mtxB, mtxmanager, cpus, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n solution matrix in new ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------------------- STEP 9: permute the solution into the original ordering ------------------------------------------------------- */ DenseMtx_permuteRows(mtxX, newToOldIV) ; if ( msglvl > 0 ) { fprintf(msgFile, "\n\n solution matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------ free the working storage ------------------------ */ InpMtx_free(mtxA) ; FrontMtx_free(frontmtx) ; Graph_free(graph) ; DenseMtx_free(mtxX) ; DenseMtx_free(mtxB) ; ETree_free(frontETree) ; IV_free(newToOldIV) ; IV_free(oldToNewIV) ; IVL_free(symbfacIVL) ; SubMtxManager_free(mtxmanager) ; /*--------------------------------------------------------------------*/ return(1) ; }
/* -------------------------------------------------------------------- fill *pndom with ndom, the number of domains. fill *pnseg with nseg, the number of segments. domains are numbered in [0, ndom), segments in [ndom,ndom+nseg). return -- an IV object that contains the map from vertices to segments created -- 99feb29, cca -------------------------------------------------------------------- */ IV * GPart_domSegMap ( GPart *gpart, int *pndom, int *pnseg ) { FILE *msgFile ; Graph *g ; int adjdom, count, d, first, ierr, ii, jj1, jj2, last, ndom, msglvl, nextphi, nPhi, nPsi, nV, phi, phi0, phi1, phi2, phi3, psi, sigma, size, size0, size1, size2, v, vsize, w ; int *adj, *adj0, *adj1, *adj2, *compids, *dmark, *dsmap, *head, *link, *list, *offsets, *PhiToPsi, *PhiToV, *PsiToSigma, *vadj, *VtoPhi ; IV *dsmapIV ; IVL *PhiByPhi, *PhiByPowD, *PsiByPowD ; /* -------------------- set the initial time -------------------- */ icputimes = 0 ; MARKTIME(cputimes[icputimes]) ; /* --------------- check the input --------------- */ if ( gpart == NULL || (g = gpart->g) == NULL || pndom == NULL || pnseg == NULL ) { fprintf(stderr, "\n fatal error in GPart_domSegMap(%p,%p,%p)" "\n bad input\n", gpart, pndom, pnseg) ; exit(-1) ; } compids = IV_entries(&gpart->compidsIV) ; msglvl = gpart->msglvl ; msgFile = gpart->msgFile ; /* ------------------------ create the map IV object ------------------------ */ nV = g->nvtx ; dsmapIV = IV_new() ; IV_init(dsmapIV, nV, NULL) ; dsmap = IV_entries(dsmapIV) ; /* ---------------------------------- check compids[] and get the number of domains and interface vertices ---------------------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; ndom = nPhi = 0 ; for ( v = 0 ; v < nV ; v++ ) { if ( (d = compids[v]) < 0 ) { fprintf(stderr, "\n fatal error in GPart_domSegMap(%p,%p,%p)" "\n compids[%d] = %d\n", gpart, pndom, pnseg, v, compids[v]) ; exit(-1) ; } else if ( d == 0 ) { nPhi++ ; } else if ( ndom < d ) { ndom = d ; } } *pndom = ndom ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n Inside GPart_domSegMap") ; fprintf(msgFile, "\n %d domains, %d Phi vertices", ndom, nPhi) ; } if ( ndom == 1 ) { IVfill(nV, dsmap, 0) ; *pndom = 1 ; *pnseg = 0 ; return(dsmapIV) ; } /* -------------------------------- get the maps PhiToV : [0,nPhi) |---> [0,nV) VtoPhi : [0,nV) |---> [0,nPhi) -------------------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; PhiToV = IVinit(nPhi, -1) ; VtoPhi = IVinit(nV, -1) ; for ( v = 0, phi = 0 ; v < nV ; v++ ) { if ( (d = compids[v]) == 0 ) { PhiToV[phi] = v ; VtoPhi[v] = phi++ ; } } if ( phi != nPhi ) { fprintf(stderr, "\n fatal error in GPart_domSegMap(%p,%p,%p)" "\n phi = %d != %d = nPhi\n", gpart, pndom, pnseg, phi, nPhi) ; exit(-1) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n PhiToV(%d) :", nPhi) ; IVfp80(msgFile, nPhi, PhiToV, 15, &ierr) ; fflush(msgFile) ; } if ( msglvl > 3 ) { fprintf(msgFile, "\n VtoPhi(%d) :", nV) ; IVfp80(msgFile, nV, VtoPhi, 15, &ierr) ; fflush(msgFile) ; } /* --------------------------------------------------- create an IVL object, PhiByPowD, to hold lists from the interface vertices to their adjacent domains --------------------------------------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; dmark = IVinit(ndom+1, -1) ; if ( nPhi >= ndom ) { list = IVinit(nPhi, -1) ; } else { list = IVinit(ndom, -1) ; } PhiByPowD = IVL_new() ; IVL_init1(PhiByPowD, IVL_CHUNKED, nPhi) ; for ( phi = 0 ; phi < nPhi ; phi++ ) { v = PhiToV[phi] ; Graph_adjAndSize(g, v, &vsize, &vadj) ; /* if ( phi == 0 ) { int ierr ; fprintf(msgFile, "\n adj(%d,%d) = ", v, phi) ; IVfp80(msgFile, vsize, vadj, 15, &ierr) ; fflush(msgFile) ; } */ count = 0 ; for ( ii = 0 ; ii < vsize ; ii++ ) { if ( (w = vadj[ii]) < nV && (d = compids[w]) > 0 && dmark[d] != phi ) { dmark[d] = phi ; list[count++] = d ; } } if ( count > 0 ) { IVqsortUp(count, list) ; IVL_setList(PhiByPowD, phi, count, list) ; } } if ( msglvl > 2 ) { fprintf(msgFile, "\n PhiByPowD : interface x adjacent domains") ; IVL_writeForHumanEye(PhiByPowD, msgFile) ; fflush(msgFile) ; } /* ------------------------------------------------------- create an IVL object, PhiByPhi to hold lists from the interface vertices to interface vertices. (s,t) are in the list if (s,t) is an edge in the graph and s and t do not share an adjacent domain ------------------------------------------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; PhiByPhi = IVL_new() ; IVL_init1(PhiByPhi, IVL_CHUNKED, nPhi) ; offsets = IVinit(nPhi, 0) ; head = IVinit(nPhi, -1) ; link = IVinit(nPhi, -1) ; for ( phi1 = 0 ; phi1 < nPhi ; phi1++ ) { count = 0 ; if ( msglvl > 2 ) { v = PhiToV[phi1] ; Graph_adjAndSize(g, v, &vsize, &vadj) ; fprintf(msgFile, "\n checking out phi = %d, v = %d", phi1, v) ; fprintf(msgFile, "\n adj(%d) : ", v) ; IVfp80(msgFile, vsize, vadj, 10, &ierr) ; } /* ------------------------------------------------------------- get (phi1, phi0) edges that were previously put into the list ------------------------------------------------------------- */ if ( msglvl > 3 ) { if ( head[phi1] == -1 ) { fprintf(msgFile, "\n no previous edges") ; } else { fprintf(msgFile, "\n previous edges :") ; } } for ( phi0 = head[phi1] ; phi0 != -1 ; phi0 = nextphi ) { if ( msglvl > 3 ) { fprintf(msgFile, " %d", phi0) ; } nextphi = link[phi0] ; list[count++] = phi0 ; IVL_listAndSize(PhiByPhi, phi0, &size0, &adj0) ; if ( (ii = ++offsets[phi0]) < size0 ) { /* ---------------------------- link phi0 into the next list ---------------------------- */ phi2 = adj0[ii] ; link[phi0] = head[phi2] ; head[phi2] = phi0 ; } } /* -------------------------- get new edges (phi1, phi2) -------------------------- */ IVL_listAndSize(PhiByPowD, phi1, &size1, &adj1) ; v = PhiToV[phi1] ; Graph_adjAndSize(g, v, &vsize, &vadj) ; for ( ii = 0 ; ii < vsize ; ii++ ) { if ( (w = vadj[ii]) < nV && compids[w] == 0 && (phi2 = VtoPhi[w]) > phi1 ) { if ( msglvl > 3 ) { fprintf(msgFile, "\n checking out phi2 = %d", phi2) ; } /* -------------------------------------------------- see if phi1 and phi2 have a common adjacent domain -------------------------------------------------- */ IVL_listAndSize(PhiByPowD, phi2, &size2, &adj2) ; adjdom = 0 ; jj1 = jj2 = 0 ; while ( jj1 < size1 && jj2 < size2 ) { if ( adj1[jj1] < adj2[jj2] ) { jj1++ ; } else if ( adj1[jj1] > adj2[jj2] ) { jj2++ ; } else { if ( msglvl > 3 ) { fprintf(msgFile, ", common adj domain %d", adj1[jj1]) ; } adjdom = 1 ; break ; } } if ( adjdom == 0 ) { if ( msglvl > 3 ) { fprintf(msgFile, ", no adjacent domain") ; } list[count++] = phi2 ; } } } if ( count > 0 ) { /* --------------------- set the list for phi1 --------------------- */ IVqsortUp(count, list) ; IVL_setList(PhiByPhi, phi1, count, list) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n edge list for %d :", phi1) ; IVfp80(msgFile, count, list, 15, &ierr) ; } for ( ii = 0, phi2 = -1 ; ii < count ; ii++ ) { if ( list[ii] > phi1 ) { offsets[phi1] = ii ; phi2 = list[ii] ; break ; } } if ( phi2 != -1 ) { if ( msglvl > 2 ) { fprintf(msgFile, "\n linking %d into list for %d", phi1, phi2) ; } link[phi1] = head[phi2] ; head[phi2] = phi1 ; } /* phi2 = list[0] ; link[phi1] = head[phi2] ; head[phi2] = phi1 ; */ } } if ( msglvl > 2 ) { fprintf(msgFile, "\n PhiByPhi : interface x interface") ; IVL_writeForHumanEye(PhiByPhi, msgFile) ; fflush(msgFile) ; } /* -------------------- get the PhiToPsi map -------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; PhiToPsi = IVinit(nPhi, -1) ; nPsi = 0 ; for ( phi = 0 ; phi < nPhi ; phi++ ) { if ( PhiToPsi[phi] == -1 ) { /* --------------------------- phi not yet mapped to a psi --------------------------- */ first = last = 0 ; list[0] = phi ; PhiToPsi[phi] = nPsi ; while ( first <= last ) { phi2 = list[first++] ; IVL_listAndSize(PhiByPhi, phi2, &size, &adj) ; for ( ii = 0 ; ii < size ; ii++ ) { phi3 = adj[ii] ; if ( PhiToPsi[phi3] == -1 ) { PhiToPsi[phi3] = nPsi ; list[++last] = phi3 ; } } } nPsi++ ; } } if ( msglvl > 1 ) { fprintf(msgFile, "\n nPsi = %d", nPsi) ; fflush(msgFile) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n PhiToPsi(%d) :", nPhi) ; IVfp80(msgFile, nPhi, PhiToPsi, 15, &ierr) ; fflush(msgFile) ; } /* --------------------------------- create an IVL object, Psi --> 2^D --------------------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; IVfill(nPsi, head, -1) ; IVfill(nPhi, link, -1) ; for ( phi = 0 ; phi < nPhi ; phi++ ) { psi = PhiToPsi[phi] ; link[phi] = head[psi] ; head[psi] = phi ; } PsiByPowD = IVL_new() ; IVL_init1(PsiByPowD, IVL_CHUNKED, nPsi) ; IVfill(ndom+1, dmark, -1) ; for ( psi = 0 ; psi < nPsi ; psi++ ) { count = 0 ; for ( phi = head[psi] ; phi != -1 ; phi = link[phi] ) { v = PhiToV[phi] ; Graph_adjAndSize(g, v, &size, &adj) ; for ( ii = 0 ; ii < size ; ii++ ) { if ( (w = adj[ii]) < nV && (d = compids[w]) > 0 && dmark[d] != psi ) { dmark[d] = psi ; list[count++] = d ; } } } if ( count > 0 ) { IVqsortUp(count, list) ; IVL_setList(PsiByPowD, psi, count, list) ; } } if ( msglvl > 2 ) { fprintf(msgFile, "\n PsiByPowD(%d) :", nPhi) ; IVL_writeForHumanEye(PsiByPowD, msgFile) ; fflush(msgFile) ; } icputimes++ ; MARKTIME(cputimes[icputimes]) ; /* ------------------------------------- now get the map Psi |---> Sigma that is the equivalence map over PhiByPowD ------------------------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; PsiToSigma = IVL_equivMap1(PsiByPowD) ; *pnseg = 1 + IVmax(nPsi, PsiToSigma, &ii) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n nSigma = %d", *pnseg) ; fprintf(msgFile, "\n PsiToSigma(%d) :", nPhi) ; IVfp80(msgFile, nPsi, PsiToSigma, 15, &ierr) ; fflush(msgFile) ; } /* -------------------------------------------------------------- now fill the map from the vertices to the domains and segments -------------------------------------------------------------- */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; for ( v = 0 ; v < nV ; v++ ) { if ( (d = compids[v]) > 0 ) { dsmap[v] = d - 1 ; } else { phi = VtoPhi[v] ; psi = PhiToPsi[phi] ; sigma = PsiToSigma[psi] ; dsmap[v] = ndom + sigma ; } } /* ------------------------ free the working storage ------------------------ */ icputimes++ ; MARKTIME(cputimes[icputimes]) ; IVL_free(PhiByPhi) ; IVL_free(PhiByPowD) ; IVL_free(PsiByPowD) ; IVfree(PhiToV) ; IVfree(VtoPhi) ; IVfree(dmark) ; IVfree(list) ; IVfree(PhiToPsi) ; IVfree(head) ; IVfree(link) ; IVfree(offsets) ; IVfree(PsiToSigma) ; icputimes++ ; MARKTIME(cputimes[icputimes]) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n domain/segment map timings split") ; fprintf(msgFile, "\n %9.5f : create the DSmap object" "\n %9.5f : get numbers of domain and interface vertices" "\n %9.5f : generate PhiToV and VtoPhi" "\n %9.5f : generate PhiByPowD" "\n %9.5f : generate PhiByPhi" "\n %9.5f : generate PhiToPsi" "\n %9.5f : generate PsiByPowD" "\n %9.5f : generate PsiToSigma" "\n %9.5f : generate dsmap" "\n %9.5f : free working storage" "\n %9.5f : total time", cputimes[1] - cputimes[0], cputimes[2] - cputimes[1], cputimes[3] - cputimes[2], cputimes[4] - cputimes[3], cputimes[5] - cputimes[4], cputimes[6] - cputimes[5], cputimes[7] - cputimes[6], cputimes[8] - cputimes[7], cputimes[9] - cputimes[8], cputimes[10] - cputimes[9], cputimes[11] - cputimes[0]) ; } return(dsmapIV) ; }
/* -------------------------------------------------------------------- purpose -- to fill submtx with a submatrix of the front matrix. the fronts that form the submatrix are found in frontidsIV. all information in submtx is local, front #'s are from 0 to one less than the number of fronts in the submatrix, equation #'s are from 0 to one less than the number of rows and columns in the submatrix. the global row and column ids for the submatrix are stored in rowsIV and colsIV on return. return values --- 1 -- normal return -1 -- submtx is NULL -2 -- frontmtx is NULL -3 -- frontmtx is not in 2-D mode -4 -- frontidsIV is NULL -5 -- frontidsIV is invalid -6 -- rowsIV is NULL -7 -- colsIV is NULL -8 -- unable to create front tree -9 -- unable to create symbfacIVL -10 -- unable to create coladjIVL -11 -- unable to create rowadjIVL -12 -- unable to create upperblockIVL -13 -- unable to create lowerblockIVL created -- 98oct17, cca -------------------------------------------------------------------- */ int FrontMtx_initFromSubmatrix ( FrontMtx *submtx, FrontMtx *frontmtx, IV *frontidsIV, IV *rowsIV, IV *colsIV, int msglvl, FILE *msgFile ) { ETree *etreeSub ; int ii, J, Jsub, K, Ksub, ncol, nfront, nfrontSub, neqnSub, nJ, nrow, offset, rc, size, vSub ; int *bndwghts, *colind, *colmap, *cols, *frontSubIds, *list, *nodwghts, *rowind, *rowmap, *rows ; IV *frontsizesIVsub, *vtxIV ; IVL *coladjIVLsub, *lowerblockIVLsub, *rowadjIVLsub, *symbfacIVLsub, *upperblockIVLsub ; SubMtx *mtx ; /* --------------- check the input --------------- */ if ( submtx == NULL ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n submtx is NULL\n") ; return(-1) ; } if ( frontmtx == NULL ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n frontmtx is NULL\n") ; return(-2) ; } if ( ! FRONTMTX_IS_2D_MODE(frontmtx) ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n frontmtx mode is not 2D\n") ; return(-3) ; } if ( frontidsIV == NULL ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n frontidsIV is NULL\n") ; return(-4) ; } nfront = FrontMtx_nfront(frontmtx) ; IV_sizeAndEntries(frontidsIV, &nfrontSub, &frontSubIds) ; if ( nfrontSub < 0 || nfrontSub > nfront ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n invalid frontidsIV" "\n nfrontSub = %d, nfront %d\n", nfrontSub, nfront) ; return(-5) ; } for ( ii = 0 ; ii < nfrontSub ; ii++ ) { if ( (J = frontSubIds[ii]) < 0 || J >= nfront ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n invalid frontidsIV" "\n frontSubIds[%d] = %d, nfront = %d\n", ii, J, nfront) ; return(-5) ; } } if ( rowsIV == NULL ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n rowsIV is NULL\n") ; return(-6) ; } if ( colsIV == NULL ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n colsIV is NULL\n") ; return(-7) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------------------- clear the data for the submatrix and set the scalar values (some inherited from the global matrix) ----------------------------------------------------- */ FrontMtx_clearData(submtx) ; submtx->nfront = nfrontSub ; submtx->type = frontmtx->type ; submtx->symmetryflag = frontmtx->symmetryflag ; submtx->sparsityflag = frontmtx->sparsityflag ; submtx->pivotingflag = frontmtx->pivotingflag ; submtx->dataMode = FRONTMTX_2D_MODE ; /* --------------------------------------------------------------- initialize the front tree for the submatrix. note: on return, vtxIV is filled with the vertices originally in the submatrix, (pivoting may change this), needed to find symbolic factorization IVL object note: at return, the boundary weights are likely to be invalid, since we have no way of knowing what boundary indices for a front are really in the domain. this will be changed after we have the symbolic factorization. --------------------------------------------------------------- */ etreeSub = submtx->frontETree = ETree_new() ; vtxIV = IV_new() ; rc = ETree_initFromSubtree(etreeSub, frontidsIV, frontmtx->frontETree, vtxIV) ; if ( rc != 1 ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n unable to create submatrix's front ETree, rc = %d\n", rc) ; return(-8) ; } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n submatrix ETree") ; ETree_writeForHumanEye(etreeSub, msgFile) ; fprintf(msgFile, "\n\n submatrix original equations") ; IV_writeForHumanEye(vtxIV, msgFile) ; fflush(msgFile) ; } /* ------------------------------------------------------ set the # of equations (perhap temporarily if pivoting has delayed some rows and columns), and the tree. ------------------------------------------------------ */ submtx->neqns = neqnSub = IV_size(vtxIV) ; submtx->tree = etreeSub->tree ; /* ----------------------------------------------------- initialize the symbolic factorization for the subtree ----------------------------------------------------- */ symbfacIVLsub = submtx->symbfacIVL = IVL_new() ; rc = IVL_initFromSubIVL(symbfacIVLsub, frontmtx->symbfacIVL, frontidsIV, vtxIV) ; if ( rc != 1 ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n unable to create submatrix's symbfac, rc = %d\n", rc) ; return(-9) ; } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n submatrix symbolic factorizatio") ; IVL_writeForHumanEye(symbfacIVLsub, msgFile) ; fflush(msgFile) ; } /* --------------------------------------------- adjust the boundary weights of the front tree --------------------------------------------- */ nodwghts = ETree_nodwghts(etreeSub) ; bndwghts = ETree_bndwghts(etreeSub) ; for ( J = 0 ; J < nfrontSub ; J++ ) { IVL_listAndSize(symbfacIVLsub, J, &size, &list) ; bndwghts[J] = size - nodwghts[J] ; } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n submatrix ETree after bndweight adjustment") ; ETree_writeForHumanEye(etreeSub, msgFile) ; fflush(msgFile) ; } /* ------------------------------------- set the front sizes for the submatrix ------------------------------------- */ frontsizesIVsub = submtx->frontsizesIV = IV_new() ; IV_init(frontsizesIVsub, nfrontSub, NULL) ; IVgather(nfrontSub, IV_entries(frontsizesIVsub), IV_entries(frontmtx->frontsizesIV), IV_entries(frontidsIV)) ; neqnSub = submtx->neqns = IV_sum(frontsizesIVsub) ; if ( msglvl > 4 ) { fprintf(msgFile, "\n\n %d equations in submatrix", neqnSub) ; fprintf(msgFile, "\n\n front sizes for submatrix") ; IV_writeForHumanEye(frontsizesIVsub, msgFile) ; fflush(msgFile) ; } /* ------------------------------------------------------------------- fill rowsIV and colsIV with the row and column ids of the submatrix ------------------------------------------------------------------- */ IV_setSize(rowsIV, neqnSub) ; IV_setSize(colsIV, neqnSub) ; rows = IV_entries(rowsIV) ; cols = IV_entries(colsIV) ; for ( Jsub = offset = 0 ; Jsub < nfrontSub ; Jsub++ ) { if ( (nJ = FrontMtx_frontSize(submtx, Jsub)) > 0 ) { J = frontSubIds[Jsub] ; FrontMtx_columnIndices(frontmtx, J, &size, &list) ; IVcopy(nJ, cols + offset, list) ; FrontMtx_rowIndices(frontmtx, J, &size, &list) ; IVcopy(nJ, rows + offset, list) ; offset += nJ ; } } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n row ids for submatrix") ; IV_writeForHumanEye(rowsIV, msgFile) ; fprintf(msgFile, "\n\n column ids for submatrix") ; IV_writeForHumanEye(colsIV, msgFile) ; fflush(msgFile) ; } /* ---------------------------------- get the row and column adjacencies ---------------------------------- */ if ( FRONTMTX_IS_PIVOTING(frontmtx) ) { submtx->neqns = neqnSub ; coladjIVLsub = submtx->coladjIVL = IVL_new() ; rc = IVL_initFromSubIVL(coladjIVLsub, frontmtx->coladjIVL, frontidsIV, colsIV) ; if ( rc != 1 ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n unable to create submatrix's coladjIVL, rc = %d\n", rc) ; return(-10) ; } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n submatrix col adjacency") ; IVL_writeForHumanEye(coladjIVLsub, msgFile) ; fflush(msgFile) ; } if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) { rowadjIVLsub = submtx->rowadjIVL = IVL_new() ; rc = IVL_initFromSubIVL(rowadjIVLsub, frontmtx->rowadjIVL, frontidsIV, rowsIV) ; if ( rc != 1 ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n unable to create submatrix's rowadjIVL, rc = %d\n", rc) ; return(-11) ; } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n submatrix row adjacency") ; IVL_writeForHumanEye(rowadjIVLsub, msgFile) ; fflush(msgFile) ; } } } IV_free(vtxIV) ; /* ---------------------------------------------- get the rowmap[] and colmap[] vectors, needed to translate indices in the submatrices ---------------------------------------------- */ colmap = IVinit(frontmtx->neqns, -1) ; for ( ii = 0 ; ii < neqnSub ; ii++ ) { colmap[cols[ii]] = ii ; } if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) { rowmap = IVinit(frontmtx->neqns, -1) ; for ( ii = 0 ; ii < neqnSub ; ii++ ) { rowmap[rows[ii]] = ii ; } } else { rowmap = colmap ; } /* ----------------------------------------------------------- get the upper and lower block IVL objects for the submatrix ----------------------------------------------------------- */ upperblockIVLsub = submtx->upperblockIVL = IVL_new() ; rc = IVL_initFromSubIVL(upperblockIVLsub, frontmtx->upperblockIVL, frontidsIV, frontidsIV) ; if ( rc != 1 ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n unable to create upperblockIVL, rc = %d\n", rc) ; return(-12) ; } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n upper block adjacency IVL object") ; IVL_writeForHumanEye(upperblockIVLsub, msgFile) ; fflush(msgFile) ; } if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) { lowerblockIVLsub = submtx->lowerblockIVL = IVL_new() ; rc = IVL_initFromSubIVL(lowerblockIVLsub, frontmtx->lowerblockIVL, frontidsIV, frontidsIV) ; if ( rc != 1 ) { fprintf(stderr, "\n error in FrontMtx_initFromSubmatrix()" "\n unable to create lowerblockIVL, rc = %d\n", rc) ; return(-13) ; } if ( msglvl > 4 ) { fprintf(msgFile, "\n\n lower block adjacency IVL object") ; IVL_writeForHumanEye(lowerblockIVLsub, msgFile) ; fflush(msgFile) ; } } /* ---------------------------------------------------------------- allocate the vector and hash table(s) for the factor submatrices ---------------------------------------------------------------- */ ALLOCATE(submtx->p_mtxDJJ, struct _SubMtx *, nfrontSub) ; for ( J = 0 ; J < nfrontSub ; J++ ) { submtx->p_mtxDJJ[J] = NULL ; } submtx->upperhash = I2Ohash_new() ; I2Ohash_init(submtx->upperhash, nfrontSub, nfrontSub, nfrontSub) ; if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) { submtx->lowerhash = I2Ohash_new() ; I2Ohash_init(submtx->lowerhash, nfrontSub, nfrontSub, nfrontSub) ; } /* ----------------------------------------------------------------- remove the diagonal submatrices from the factor matrix and insert into the submatrix object. note: front row and column ids must be changed to their local values, and the row and column indices must be mapped to local indices. ----------------------------------------------------------------- */ for ( Jsub = 0 ; Jsub < nfrontSub ; Jsub++ ) { J = frontSubIds[Jsub] ; if ( (mtx = frontmtx->p_mtxDJJ[J]) != NULL ) { SubMtx_setIds(mtx, Jsub, Jsub) ; SubMtx_columnIndices(mtx, &ncol, &colind) ; IVgather(ncol, colind, colmap, colind) ; SubMtx_rowIndices(mtx, &nrow, &rowind) ; IVgather(nrow, rowind, rowmap, rowind) ; submtx->p_mtxDJJ[Jsub] = mtx ; frontmtx->p_mtxDJJ[J] = NULL ; submtx->nentD += mtx->nent ; } } /* ---------------------------------------------------------------- remove the upper triangular submatrices from the factor matrix and insert into the submatrix object. note: front row and column ids must be changed to their local values. if the matrix is on the diagonal, i.e., U(J,J), its row and column indices must be mapped to local indices. ---------------------------------------------------------------- */ for ( Jsub = 0 ; Jsub < nfrontSub ; Jsub++ ) { J = frontSubIds[Jsub] ; FrontMtx_upperAdjFronts(submtx, Jsub, &size, &list) ; for ( ii = 0 ; ii < size ; ii++ ) { Ksub = list[ii] ; K = frontSubIds[Ksub] ; if ( 1 == I2Ohash_remove(frontmtx->upperhash, J, K, (void *) &mtx) ) { SubMtx_setIds(mtx, Jsub, Ksub) ; if ( K == J ) { SubMtx_columnIndices(mtx, &ncol, &colind) ; IVgather(ncol, colind, colmap, colind) ; SubMtx_rowIndices(mtx, &nrow, &rowind) ; IVgather(nrow, rowind, rowmap, rowind) ; } I2Ohash_insert(submtx->upperhash, Jsub, Ksub, (void *) mtx) ; submtx->nentU += mtx->nent ; } } } if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) { /* ---------------------------------------------------------------- remove the lower triangular submatrices from the factor matrix and insert into the submatrix object. note: front row and column ids must be changed to their local values. if the matrix is on the diagonal, i.e., L(J,J), its row and column indices must be mapped to local indices. ---------------------------------------------------------------- */ for ( Jsub = 0 ; Jsub < nfrontSub ; Jsub++ ) { J = frontSubIds[Jsub] ; FrontMtx_lowerAdjFronts(submtx, Jsub, &size, &list) ; for ( ii = 0 ; ii < size ; ii++ ) { Ksub = list[ii] ; K = frontSubIds[Ksub] ; if ( 1 == I2Ohash_remove(frontmtx->lowerhash, K, J, (void *) &mtx) ) { SubMtx_setIds(mtx, Ksub, Jsub) ; if ( K == J ) { SubMtx_columnIndices(mtx, &ncol, &colind) ; IVgather(ncol, colind, colmap, colind) ; SubMtx_rowIndices(mtx, &nrow, &rowind) ; IVgather(nrow, rowind, rowmap, rowind) ; } I2Ohash_insert(submtx->lowerhash, Ksub, Jsub, (void *) mtx); submtx->nentL += mtx->nent ; } } } } /* ------------------------ free the working storage ------------------------ */ IVfree(colmap) ; if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) { IVfree(rowmap) ; } return(1) ; }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) { /* -------------------------------------------------- all-in-one program to solve A X = B using a multithreaded factorization and solve We use a patch-and-go strategy for the factorization without pivoting (1) read in matrix entries and form DInpMtx object (2) form Graph object (3) order matrix and form front tree (4) get the permutation, permute the matrix and front tree and get the symbolic factorization (5) compute the numeric factorization (6) read in right hand side entries (7) compute the solution created -- 98jun04, cca -------------------------------------------------- */ /*--------------------------------------------------------------------*/ char *matrixFileName, *rhsFileName ; DenseMtx *mtxB, *mtxX ; Chv *rootchv ; ChvManager *chvmanager ; double fudge, imag, real, tau = 100., toosmall, value ; double cpus[10] ; DV *cumopsDV ; ETree *frontETree ; FrontMtx *frontmtx ; FILE *inputFile, *msgFile ; Graph *graph ; InpMtx *mtxA ; int error, ient, irow, jcol, jrhs, jrow, lookahead, msglvl, ncol, nedges, nent, neqns, nfront, nrhs, nrow, nthread, patchAndGoFlag, seed, storeids, storevalues, symmetryflag, type ; int *newToOld, *oldToNew ; int stats[20] ; IV *newToOldIV, *oldToNewIV, *ownersIV ; IVL *adjIVL, *symbfacIVL ; SolveMap *solvemap ; SubMtxManager *mtxmanager ; /*--------------------------------------------------------------------*/ /* -------------------- get input parameters -------------------- */ if ( argc != 14 ) { fprintf(stdout, "\n" "\n usage: %s msglvl msgFile type symmetryflag patchAndGoFlag" "\n fudge toosmall storeids storevalues" "\n matrixFileName rhsFileName seed" "\n msglvl -- message level" "\n msgFile -- message file" "\n type -- type of entries" "\n 1 (SPOOLES_REAL) -- real entries" "\n 2 (SPOOLES_COMPLEX) -- complex entries" "\n symmetryflag -- type of matrix" "\n 0 (SPOOLES_SYMMETRIC) -- symmetric entries" "\n 1 (SPOOLES_HERMITIAN) -- Hermitian entries" "\n 2 (SPOOLES_NONSYMMETRIC) -- nonsymmetric entries" "\n patchAndGoFlag -- flag for the patch-and-go strategy" "\n 0 -- none, stop factorization" "\n 1 -- optimization strategy" "\n 2 -- structural analysis strategy" "\n fudge -- perturbation parameter" "\n toosmall -- upper bound on a small pivot" "\n storeids -- flag to store ids of small pivots" "\n storevalues -- flag to store perturbations" "\n matrixFileName -- matrix file name, format" "\n nrow ncol nent" "\n irow jcol entry" "\n ..." "\n note: indices are zero based" "\n rhsFileName -- right hand side file name, format" "\n nrow nrhs " "\n ..." "\n jrow entry(jrow,0) ... entry(jrow,nrhs-1)" "\n ..." "\n seed -- random number seed, used for ordering" "\n nthread -- number of threads" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else if ( (msgFile = fopen(argv[2], "a")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } type = atoi(argv[3]) ; symmetryflag = atoi(argv[4]) ; patchAndGoFlag = atoi(argv[5]) ; fudge = atof(argv[6]) ; toosmall = atof(argv[7]) ; storeids = atoi(argv[8]) ; storevalues = atoi(argv[9]) ; matrixFileName = argv[10] ; rhsFileName = argv[11] ; seed = atoi(argv[12]) ; nthread = atoi(argv[13]) ; /*--------------------------------------------------------------------*/ /* -------------------------------------------- STEP 1: read the entries from the input file and create the InpMtx object -------------------------------------------- */ if ( (inputFile = fopen(matrixFileName, "r")) == NULL ) { fprintf(stderr, "\n unable to open file %s", matrixFileName) ; spoolesFatal(); } fscanf(inputFile, "%d %d %d", &nrow, &ncol, &nent) ; neqns = nrow ; mtxA = InpMtx_new() ; InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ; if ( type == SPOOLES_REAL ) { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ; InpMtx_inputRealEntry(mtxA, irow, jcol, value) ; } } else { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ; InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ; } } fclose(inputFile) ; InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n input matrix") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------------- STEP 2 : find a low-fill ordering (1) create the Graph object (2) order the graph using multiple minimum degree ------------------------------------------------- */ graph = Graph_new() ; adjIVL = InpMtx_fullAdjacency(mtxA) ; nedges = IVL_tsize(adjIVL) ; Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n graph of the input matrix") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------------------- STEP 3: get the permutation, permute the matrix and front tree and get the symbolic factorization ----------------------------------------------------- */ oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ; oldToNew = IV_entries(oldToNewIV) ; newToOldIV = ETree_newToOldVtxPerm(frontETree) ; newToOld = IV_entries(newToOldIV) ; ETree_permuteVertices(frontETree, oldToNewIV) ; InpMtx_permute(mtxA, oldToNew, oldToNew) ; InpMtx_mapToUpperTriangle(mtxA) ; InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS) ; InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n old-to-new permutation vector") ; IV_writeForHumanEye(oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation vector") ; IV_writeForHumanEye(newToOldIV, msgFile) ; fprintf(msgFile, "\n\n front tree after permutation") ; ETree_writeForHumanEye(frontETree, msgFile) ; fprintf(msgFile, "\n\n input matrix after permutation") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(symbfacIVL, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------ STEP 4: initialize the front matrix object and the PatchAndGoInfo object to handle small pivots ------------------------------------------ */ frontmtx = FrontMtx_new() ; mtxmanager = SubMtxManager_new() ; SubMtxManager_init(mtxmanager, LOCK_IN_PROCESS, 0) ; FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, symmetryflag, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, LOCK_IN_PROCESS, 0, NULL, mtxmanager, msglvl, msgFile) ; if ( patchAndGoFlag == 1 ) { frontmtx->patchinfo = PatchAndGoInfo_new() ; PatchAndGoInfo_init(frontmtx->patchinfo, 1, toosmall, fudge, storeids, storevalues) ; } else if ( patchAndGoFlag == 2 ) { frontmtx->patchinfo = PatchAndGoInfo_new() ; PatchAndGoInfo_init(frontmtx->patchinfo, 2, toosmall, fudge, storeids, storevalues) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------ STEP 5: setup the domain decomposition map ------------------------------------------ */ if ( nthread > (nfront = FrontMtx_nfront(frontmtx)) ) { nthread = nfront ; } cumopsDV = DV_new() ; DV_init(cumopsDV, nthread, NULL) ; ownersIV = ETree_ddMap(frontETree, type, symmetryflag, cumopsDV, 1./(2.*nthread)) ; DV_free(cumopsDV) ; /*--------------------------------------------------------------------*/ /* ----------------------------------------------------- STEP 6: compute the numeric factorization in parallel ----------------------------------------------------- */ chvmanager = ChvManager_new() ; ChvManager_init(chvmanager, LOCK_IN_PROCESS, 1) ; DVfill(10, cpus, 0.0) ; IVfill(20, stats, 0) ; lookahead = 0 ; rootchv = FrontMtx_MT_factorInpMtx(frontmtx, mtxA, tau, 0.0, chvmanager, ownersIV, lookahead, &error, cpus, stats, msglvl, msgFile) ; if ( patchAndGoFlag == 1 ) { if ( frontmtx->patchinfo->fudgeIV != NULL ) { fprintf(msgFile, "\n small pivots found at these locations") ; IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ; } PatchAndGoInfo_free(frontmtx->patchinfo) ; } else if ( patchAndGoFlag == 2 ) { if ( frontmtx->patchinfo->fudgeIV != NULL ) { fprintf(msgFile, "\n small pivots found at these locations") ; IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ; } if ( frontmtx->patchinfo->fudgeDV != NULL ) { fprintf(msgFile, "\n perturbations") ; DV_writeForHumanEye(frontmtx->patchinfo->fudgeDV, msgFile) ; } PatchAndGoInfo_free(frontmtx->patchinfo) ; } ChvManager_free(chvmanager) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix") ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } if ( rootchv != NULL ) { fprintf(msgFile, "\n\n matrix found to be singular\n") ; spoolesFatal(); } if ( error >= 0 ) { fprintf(msgFile, "\n\n fatal error at front %d\n", error) ; spoolesFatal(); } /* -------------------------------------- STEP 7: post-process the factorization -------------------------------------- */ FrontMtx_postProcess(frontmtx, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix after post-processing") ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------- STEP 8: read the right hand side matrix B ----------------------------------------- */ if ( (inputFile = fopen(rhsFileName, "r")) == NULL ) { fprintf(stderr, "\n unable to open file %s", rhsFileName) ; spoolesFatal(); } fscanf(inputFile, "%d %d", &nrow, &nrhs) ; mtxB = DenseMtx_new() ; DenseMtx_init(mtxB, type, 0, 0, neqns, nrhs, 1, neqns) ; DenseMtx_zero(mtxB) ; if ( type == SPOOLES_REAL ) { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le", &value) ; DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ; } } } else { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le %le", &real, &imag) ; DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ; } } } fclose(inputFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n rhs matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxB, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------------------------------- STEP 9: permute the right hand side into the original ordering -------------------------------------------------------------- */ DenseMtx_permuteRows(mtxB, oldToNewIV) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n right hand side matrix in new ordering") ; DenseMtx_writeForHumanEye(mtxB, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------------------------- STEP 10: get the solve map object for the parallel solve -------------------------------------------------------- */ solvemap = SolveMap_new() ; SolveMap_ddMap(solvemap, type, FrontMtx_upperBlockIVL(frontmtx), FrontMtx_lowerBlockIVL(frontmtx), nthread, ownersIV, FrontMtx_frontTree(frontmtx), seed, msglvl, msgFile) ; /*--------------------------------------------------------------------*/ /* -------------------------------------------- STEP 11: solve the linear system in parallel -------------------------------------------- */ mtxX = DenseMtx_new() ; DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ; DenseMtx_zero(mtxX) ; FrontMtx_MT_solve(frontmtx, mtxX, mtxB, mtxmanager, solvemap, cpus, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n solution matrix in new ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------------------------- STEP 12: permute the solution into the original ordering -------------------------------------------------------- */ DenseMtx_permuteRows(mtxX, newToOldIV) ; if ( msglvl > 0 ) { fprintf(msgFile, "\n\n solution matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------- free memory ----------- */ FrontMtx_free(frontmtx) ; DenseMtx_free(mtxX) ; DenseMtx_free(mtxB) ; IV_free(newToOldIV) ; IV_free(oldToNewIV) ; InpMtx_free(mtxA) ; ETree_free(frontETree) ; IVL_free(symbfacIVL) ; SubMtxManager_free(mtxmanager) ; Graph_free(graph) ; SolveMap_free(solvemap) ; IV_free(ownersIV) ; /*--------------------------------------------------------------------*/ return(1) ; }
/* ---------------------------------------------------------------- purpose -- given InpMtx objects that contain A and B, initialize the bridge data structure for the serial factor's, solve's and mvm's. data -- pointer to a Bridge object pprbtype -- pointer to value containing problem type *prbtype = 1 --> A X = B X Lambda, vibration problem *prbtype = 2 --> A X = B X Lambda, buckling problem *prbtype = 3 --> A X = X Lambda, simple eigenvalue problem pneqns -- pointer to value containing number of equations pmxbsz -- pointer to value containing blocksize A -- pointer to InpMtx object containing A B -- pointer to InpMtx object containing B pseed -- pointer to value containing a random number seed pmsglvl -- pointer to value containing a message level msgFile -- message file pointer return value -- 1 -- normal return -1 -- data is NULL -2 -- pprbtype is NULL -3 -- *pprbtype is invalid -4 -- pneqns is NULL -5 -- *pneqns is invalid -6 -- pmxbsz is NULL -7 -- *pmxbsz is invalid -8 -- A and B are NULL -9 -- pseed is NULL -10 -- pmsglvl is NULL -11 -- *pmsglvl > 0 and msgFile is NULL created -- 98aug10, cca ---------------------------------------------------------------- */ int Setup ( void *data, int *pprbtype, int *pneqns, int *pmxbsz, InpMtx *A, InpMtx *B, int *pseed, int *pmsglvl, FILE *msgFile ) { Bridge *bridge = (Bridge *) data ; double sigma[2] ; Graph *graph ; int maxdomainsize, maxsize, maxzeros, msglvl, mxbsz, nedges, neqns, prbtype, seed ; IVL *adjIVL ; #if MYDEBUG > 0 double t1, t2 ; MARKTIME(t1) ; count_Setup++ ; fprintf(stdout, "\n (%d) Setup()", count_Setup) ; fflush(stdout) ; #endif /* -------------------- check the input data -------------------- */ if ( data == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n data is NULL\n") ; return(-1) ; } if ( pprbtype == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n prbtype is NULL\n") ; return(-2) ; } prbtype = *pprbtype ; if ( prbtype < 1 || prbtype > 3 ) { fprintf(stderr, "\n fatal error in Setup()" "\n prbtype = %d, is invalid\n", prbtype) ; return(-3) ; } if ( pneqns == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pneqns is NULL\n") ; return(-4) ; } neqns = *pneqns ; if ( neqns <= 0 ) { fprintf(stderr, "\n fatal error in Setup()" "\n neqns = %d, is invalid\n", neqns) ; return(-5) ; } if ( pmxbsz == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pmxbsz is NULL\n") ; return(-6) ; } mxbsz = *pmxbsz ; if ( mxbsz <= 0 ) { fprintf(stderr, "\n fatal error in Setup()" "\n *pmxbsz = %d, is invalid\n", mxbsz) ; return(-7) ; } if ( A == NULL && B == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n A and B are NULL\n") ; return(-8) ; } if ( pseed == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pseed is NULL\n") ; return(-9) ; } seed = *pseed ; if ( pmsglvl == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pmsglvl is NULL\n") ; return(-10) ; } msglvl = *pmsglvl ; if ( msglvl > 0 && msgFile == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n msglvl = %d, msgFile = NULL\n", msglvl) ; return(-11) ; } bridge->msglvl = msglvl ; bridge->msgFile = msgFile ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n inside Setup()" "\n neqns = %d, prbtype = %d, mxbsz = %d, seed = %d", neqns, prbtype, mxbsz, seed) ; if ( A != NULL ) { fprintf(msgFile, "\n\n matrix A") ; InpMtx_writeForHumanEye(A, msgFile) ; } if ( B != NULL ) { fprintf(msgFile, "\n\n matrix B") ; InpMtx_writeForHumanEye(B, msgFile) ; } fflush(msgFile) ; } bridge->prbtype = prbtype ; bridge->neqns = neqns ; bridge->mxbsz = mxbsz ; bridge->A = A ; bridge->B = B ; bridge->seed = seed ; /* ---------------------------- create and initialize pencil ---------------------------- */ sigma[0] = 1.0; sigma[1] = 0.0; bridge->pencil = Pencil_new() ; Pencil_setDefaultFields(bridge->pencil) ; Pencil_init(bridge->pencil, SPOOLES_REAL, SPOOLES_SYMMETRIC, A, sigma, B) ; /* -------------------------------- convert to row or column vectors -------------------------------- */ if ( A != NULL ) { if ( ! INPMTX_IS_BY_ROWS(A) && ! INPMTX_IS_BY_COLUMNS(A) ) { InpMtx_changeCoordType(A, INPMTX_BY_ROWS) ; } if ( ! INPMTX_IS_BY_VECTORS(A) ) { InpMtx_changeStorageMode(A, INPMTX_BY_VECTORS) ; } } if ( B != NULL ) { if ( ! INPMTX_IS_BY_ROWS(B) && ! INPMTX_IS_BY_COLUMNS(B) ) { InpMtx_changeCoordType(B, INPMTX_BY_ROWS) ; } if ( ! INPMTX_IS_BY_VECTORS(B) ) { InpMtx_changeStorageMode(B, INPMTX_BY_VECTORS) ; } } /* ------------------------------- create a Graph object for A + B ------------------------------- */ graph = Graph_new() ; adjIVL = Pencil_fullAdjacency(bridge->pencil) ; nedges = IVL_tsize(adjIVL), Graph_init2(graph, 0, bridge->neqns, 0, nedges, bridge->neqns, nedges, adjIVL, NULL, NULL) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n graph of the input matrix") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } /* --------------- order the graph --------------- */ maxdomainsize = neqns / 64 ; if ( maxdomainsize == 0 ) { maxdomainsize = 1 ; } maxzeros = (int) (0.01*neqns) ; maxsize = 64 ; bridge->frontETree = orderViaBestOfNDandMS(graph, maxdomainsize, maxzeros, maxsize, bridge->seed, msglvl, msgFile) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(bridge->frontETree, msgFile) ; fflush(msgFile) ; } /* ---------------------------------------------- get the old-to-new and new-to-old permutations ---------------------------------------------- */ bridge->oldToNewIV = ETree_oldToNewVtxPerm(bridge->frontETree) ; bridge->newToOldIV = ETree_newToOldVtxPerm(bridge->frontETree) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n old-to-new permutation") ; IV_writeForHumanEye(bridge->oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation") ; IV_writeForHumanEye(bridge->newToOldIV, msgFile) ; fflush(msgFile) ; } /* -------------------------------------- permute the vertices in the front tree -------------------------------------- */ ETree_permuteVertices(bridge->frontETree, bridge->oldToNewIV) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n permuted front etree") ; ETree_writeForHumanEye(bridge->frontETree, msgFile) ; fflush(msgFile) ; } /* ------------------------------------------- permute the entries in the pencil. note, after the permutation the entries are mapped into the upper triangle. ------------------------------------------- */ Pencil_permute(bridge->pencil, bridge->oldToNewIV, bridge->oldToNewIV) ; Pencil_mapToUpperTriangle(bridge->pencil) ; Pencil_changeCoordType(bridge->pencil, INPMTX_BY_CHEVRONS) ; Pencil_changeStorageMode(bridge->pencil, INPMTX_BY_VECTORS) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n permuted pencil") ; Pencil_writeForHumanEye(bridge->pencil, msgFile) ; fflush(msgFile) ; } /* ---------------------------------- compute the symbolic factorization ---------------------------------- */ bridge->symbfacIVL = SymbFac_initFromPencil(bridge->frontETree, bridge->pencil) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(bridge->symbfacIVL, msgFile) ; fflush(msgFile) ; } /* -------------------------------------------------- create a FrontMtx object to hold the factorization -------------------------------------------------- */ bridge->frontmtx = FrontMtx_new() ; /* ------------------------------------------------------------ create a SubMtxManager object to hold the factor submatrices ------------------------------------------------------------ */ bridge->mtxmanager = SubMtxManager_new() ; SubMtxManager_init(bridge->mtxmanager, NO_LOCK, 0) ; /* ------------------------------------------------------------ allocate the working objects X and Y for the matrix multiply ------------------------------------------------------------ */ bridge->X = DenseMtx_new() ; DenseMtx_init(bridge->X, SPOOLES_REAL, 0, 0, neqns, mxbsz, 1, neqns) ; bridge->Y = DenseMtx_new() ; DenseMtx_init(bridge->Y, SPOOLES_REAL, 0, 0, neqns, mxbsz, 1, neqns) ; /* ------------------------ free the working storage ------------------------ */ Graph_free(graph) ; #if MYDEBUG > 0 MARKTIME(t2) ; time_Setup += t2 - t1 ; fprintf(stdout, ", %8.3f seconds, %8.3f total time", t2 - t1, time_Setup) ; fflush(stdout) ; #endif return(1) ; }
/* ------------------------------------------------------------------- purpose -- given an InpMtx object that contains the structure of A, initialize the bridge data structure for the serial factor's and solve's. note: all parameters are pointers to be compatible with fortran's call by reference. return value -- 1 -- normal return -1 -- bridge is NULL -2 -- mtxA is NULL created -- 98sep17, cca ------------------------------------------------------------------- */ int BridgeMT_setup ( BridgeMT *bridge, InpMtx *mtxA ) { double t0, t1, t2 ; ETree *frontETree ; FILE *msgFile ; Graph *graph ; int compressed, msglvl, nedges, neqns, Neqns ; IV *eqmapIV ; IVL *adjIVL, *symbfacIVL ; MARKTIME(t0) ; /* -------------------- check the input data -------------------- */ if ( bridge == NULL ) { fprintf(stderr, "\n fatal error in BridgeMT_setup()" "\n data is NULL\n") ; return(-1) ; } if ( mtxA == NULL ) { fprintf(stderr, "\n fatal error in BridgeMT_setup()" "\n A is NULL\n") ; return(-2) ; } msglvl = bridge->msglvl ; msgFile = bridge->msgFile ; neqns = bridge->neqns ; if ( ! (INPMTX_IS_BY_ROWS(mtxA) || INPMTX_IS_BY_COLUMNS(mtxA)) ) { /* ------------------------------ change coordinate type to rows ------------------------------ */ InpMtx_changeCoordType(mtxA, INPMTX_BY_ROWS) ; } if ( ! INPMTX_IS_BY_VECTORS(mtxA) ) { /* ------------------------------ change storage mode to vectors ------------------------------ */ InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; } /* --------------------------- create a Graph object for A --------------------------- */ MARKTIME(t1) ; graph = Graph_new() ; adjIVL = InpMtx_fullAdjacency(mtxA); nedges = bridge->nedges = IVL_tsize(adjIVL), Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL) ; MARKTIME(t2) ; bridge->cpus[0] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time to create Graph", t2 - t1) ; fflush(msgFile) ; } if ( msglvl > 3 ) { fprintf(msgFile, "\n\n graph of the input matrix") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } /* ------------------ compress the graph ------------------ */ MARKTIME(t1) ; eqmapIV = Graph_equivMap(graph) ; Neqns = bridge->Neqns = 1 + IV_max(eqmapIV) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n graph's equivalence map") ; IV_writeForHumanEye(eqmapIV, msgFile) ; fflush(msgFile) ; } if ( Neqns < bridge->compressCutoff * neqns ) { Graph *cgraph ; /* ------------------ compress the graph ------------------ */ cgraph = Graph_compress2(graph, eqmapIV, 1) ; Graph_free(graph) ; graph = cgraph ; compressed = 1 ; bridge->Nedges = graph->nedges ; } else { compressed = 0 ; } MARKTIME(t2) ; bridge->cpus[1] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time to create compressed graph", t2 - t1) ; fflush(msgFile) ; } if ( msglvl > 3 ) { fprintf(msgFile, "\n\n graph to order") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } /* --------------- order the graph --------------- */ MARKTIME(t1) ; if ( bridge->maxdomainsize <= 0 ) { bridge->maxdomainsize = neqns/32 ; } if ( bridge->maxdomainsize <= 0 ) { bridge->maxdomainsize = 1 ; } if ( bridge->maxnzeros < 0 ) { bridge->maxnzeros = 0.01*neqns ; } if ( bridge->maxsize < 0 ) { bridge->maxsize = neqns ; } frontETree = orderViaBestOfNDandMS(graph, bridge->maxdomainsize, bridge->maxnzeros, bridge->maxsize, bridge->seed, msglvl, msgFile) ; MARKTIME(t2) ; bridge->cpus[2] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time to order graph", t2 - t1) ; fflush(msgFile) ; } if ( msglvl > 3 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } MARKTIME(t1) ; if ( compressed == 1 ) { ETree *etree ; IVL *tempIVL ; /* ---------------------------------------------------------- compute the symbolic factorization of the compressed graph ---------------------------------------------------------- */ tempIVL = SymbFac_initFromGraph(frontETree, graph) ; /* ------------------------------------------------------- expand the symbolic factorization to the original graph ------------------------------------------------------- */ symbfacIVL = IVL_expand(tempIVL, eqmapIV) ; IVL_free(tempIVL) ; /* --------------------- expand the front tree --------------------- */ etree = ETree_expand(frontETree, eqmapIV) ; ETree_free(frontETree) ; frontETree = etree ; } else { /* -------------------------------------------------------- compute the symbolic factorization of the original graph -------------------------------------------------------- */ symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ; } MARKTIME(t2) ; bridge->frontETree = frontETree ; bridge->symbfacIVL = symbfacIVL ; /* ---------------------------------------------- get the old-to-new and new-to-old permutations ---------------------------------------------- */ bridge->oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ; bridge->newToOldIV = ETree_newToOldVtxPerm(frontETree) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n old-to-new permutation") ; IV_writeForHumanEye(bridge->oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation") ; IV_writeForHumanEye(bridge->newToOldIV, msgFile) ; fflush(msgFile) ; } /* ------------------------------------------------------ overwrite the symbolic factorization with the permuted indices and sort the lists into ascending order ------------------------------------------------------ */ IVL_overwrite(symbfacIVL, bridge->oldToNewIV) ; IVL_sortUp(symbfacIVL) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(symbfacIVL, msgFile) ; fflush(msgFile) ; } /* -------------------------------------- permute the vertices in the front tree -------------------------------------- */ ETree_permuteVertices(frontETree, bridge->oldToNewIV) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n permuted front etree") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } MARKTIME(t2) ; bridge->cpus[3] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time for symbolic factorization", t2 - t1) ; fflush(msgFile) ; } /* ------------------------ free the working storage ------------------------ */ Graph_free(graph) ; IV_free(eqmapIV) ; MARKTIME(t2) ; bridge->cpus[4] += t2 - t0 ; return(1) ; }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) /* ---------------------------------------- get statistics for a semi-implicit solve created -- 97dec11, cca ---------------------------------------- */ { char *inGraphFileName, *inETreeFileName, *inMapFileName ; double nA21, nL, nL11, nL22, nPhi, nV, t1, t2 ; ETree *etree ; int ii, inside, J, K, msglvl, nfront, nJ, nvtx, rc, sizeJ, v, vsize, w ; int *adjJ, *frontmap, *map, *nodwghts, *vadj, *vtxToFront, *vwghts ; IV *mapIV ; IVL *symbfacIVL ; Graph *graph ; FILE *msgFile ; Tree *tree ; if ( argc != 6 ) { fprintf(stdout, "\n\n usage : %s msglvl msgFile GraphFile ETreeFile mapFile " "\n msglvl -- message level" "\n msgFile -- message file" "\n GraphFile -- input graph file, must be *.graphf or *.graphb" "\n ETreeFile -- input ETree file, must be *.etreef or *.etreeb" "\n mapFile -- input map IV file, must be *.ivf or *.ivb" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else if ( (msgFile = fopen(argv[2], "a")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } inGraphFileName = argv[3] ; inETreeFileName = argv[4] ; inMapFileName = argv[5] ; fprintf(msgFile, "\n %s " "\n msglvl -- %d" "\n msgFile -- %s" "\n GraphFile -- %s" "\n ETreeFile -- %s" "\n mapFile -- %s" "\n", argv[0], msglvl, argv[2], inGraphFileName, inETreeFileName, inMapFileName) ; fflush(msgFile) ; /* ------------------------ read in the Graph object ------------------------ */ graph = Graph_new() ; if ( strcmp(inGraphFileName, "none") == 0 ) { fprintf(msgFile, "\n no file to read from") ; exit(0) ; } MARKTIME(t1) ; rc = Graph_readFromFile(graph, inGraphFileName) ; MARKTIME(t2) ; fprintf(msgFile, "\n CPU %9.5f : read in graph from file %s", t2 - t1, inGraphFileName) ; nvtx = graph->nvtx ; vwghts = graph->vwghts ; if ( rc != 1 ) { fprintf(msgFile, "\n return value %d from Graph_readFromFile(%p,%s)", rc, graph, inGraphFileName) ; exit(-1) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n after reading Graph object from file %s", inGraphFileName) ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } /* ------------------------ read in the ETree object ------------------------ */ etree = ETree_new() ; if ( strcmp(inETreeFileName, "none") == 0 ) { fprintf(msgFile, "\n no file to read from") ; exit(0) ; } MARKTIME(t1) ; rc = ETree_readFromFile(etree, inETreeFileName) ; MARKTIME(t2) ; fprintf(msgFile, "\n CPU %9.5f : read in etree from file %s", t2 - t1, inETreeFileName) ; nfront = ETree_nfront(etree) ; tree = ETree_tree(etree) ; vtxToFront = ETree_vtxToFront(etree) ; nodwghts = ETree_nodwghts(etree) ; nL = ETree_nFactorEntries(etree, 2) ; if ( rc != 1 ) { fprintf(msgFile, "\n return value %d from ETree_readFromFile(%p,%s)", rc, etree, inETreeFileName) ; exit(-1) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n after reading ETree object from file %s", inETreeFileName) ; ETree_writeForHumanEye(etree, msgFile) ; fflush(msgFile) ; } /* ------------------------- read in the map IV object ------------------------- */ mapIV = IV_new() ; if ( strcmp(inMapFileName, "none") == 0 ) { fprintf(msgFile, "\n no file to read from") ; exit(0) ; } MARKTIME(t1) ; rc = IV_readFromFile(mapIV, inMapFileName) ; MARKTIME(t2) ; fprintf(msgFile, "\n CPU %9.5f : read in mapIV from file %s", t2 - t1, inMapFileName) ; map = IV_entries(mapIV) ; if ( rc != 1 ) { fprintf(msgFile, "\n return value %d from IV_readFromFile(%p,%s)", rc, mapIV, inMapFileName) ; exit(-1) ; } if ( msglvl > 2 ) { fprintf(msgFile, "\n\n after reading IV object from file %s", inMapFileName) ; IV_writeForHumanEye(mapIV, msgFile) ; fflush(msgFile) ; } nV = nPhi = 0 ; if ( vwghts == NULL ) { for ( v = 0 ; v < nvtx ; v++ ) { nV++ ; if ( map[v] == 0 ) { nPhi++ ; } } } else { for ( v = 0 ; v < nvtx ; v++ ) { nV += vwghts[v] ; if ( map[v] == 0 ) { nPhi += vwghts[v] ; } } } fprintf(msgFile, "\n nPhi = %.0f, nV = %.0f", nPhi, nV) ; /* ------------------------- get the frontmap[] vector ------------------------- */ frontmap = IVinit(nfront, -1) ; for ( v = 0 ; v < nvtx ; v++ ) { J = vtxToFront[v] ; if ( frontmap[J] == -1 ) { frontmap[J] = map[v] ; } else if ( frontmap[J] != map[v] ) { fprintf(msgFile, "\n\n error, frontmap[%d] = %d, map[%d] = %d", J, frontmap[J], v, map[v]) ; } } /* ---------------------------------- compute the symbolic factorization ---------------------------------- */ symbfacIVL = SymbFac_initFromGraph(etree, graph) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(symbfacIVL, msgFile) ; fflush(msgFile) ; } /* -------------------------------------------- compute the number of entries in L11 and L22 -------------------------------------------- */ nL11 = nL22 = 0 ; for ( J = Tree_postOTfirst(tree) ; J != -1 ; J = Tree_postOTnext(tree, J) ) { nJ = nodwghts[J] ; if ( msglvl > 3 ) { fprintf(msgFile, "\n\n front %d, nJ = %d", J, nJ) ; } IVL_listAndSize(symbfacIVL, J, &sizeJ, &adjJ) ; for ( ii = 0, inside = 0 ; ii < sizeJ ; ii++ ) { w = adjJ[ii] ; K = vtxToFront[w] ; if ( msglvl > 3 ) { fprintf(msgFile, "\n w = %d, K = %d", w, K) ; } if ( K > J && frontmap[K] == frontmap[J] ) { inside += (vwghts == NULL) ? 1 : vwghts[w] ; if ( msglvl > 3 ) { fprintf(msgFile, ", inside") ; } } } if ( frontmap[J] != 0 ) { if ( msglvl > 3 ) { fprintf(msgFile, "\n inside = %d, adding %d to L11", inside, nJ*nJ + 2*nJ*inside) ; } nL11 += nJ*nJ + 2*nJ*inside ; } else { if ( msglvl > 3 ) { fprintf(msgFile, "\n inside = %d, adding %d to L22", inside, nJ*nJ + 2*nJ*inside) ; } nL22 += nJ*nJ + 2*nJ*inside ; } } if ( msglvl > 0 ) { fprintf(msgFile, "\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f", nL, nL11, nL22) ; } /* ------------------------------------ compute the number of entries in A21 ------------------------------------ */ nA21 = 0 ; if ( vwghts != NULL ) { for ( v = 0 ; v < nvtx ; v++ ) { if ( map[v] == 0 ) { Graph_adjAndSize(graph, v, &vsize, &vadj) ; for ( ii = 0 ; ii < vsize ; ii++ ) { w = vadj[ii] ; if ( map[v] != map[w] ) { if ( msglvl > 3 ) { fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ; } nA21 += vwghts[v] * vwghts[w] ; } } } } } else { for ( v = 0 ; v < nvtx ; v++ ) { if ( map[v] == 0 ) { Graph_adjAndSize(graph, v, &vsize, &vadj) ; for ( ii = 0 ; ii < vsize ; ii++ ) { w = vadj[ii] ; if ( map[v] != map[w] ) { if ( msglvl > 3 ) { fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ; } nA21++ ; } } } } } if ( msglvl > 0 ) { fprintf(msgFile, "\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f, |A21| = %.0f", nL, nL11, nL22, nA21) ; fprintf(msgFile, "\n storage: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f" "\n opcount: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f", nL, nL11 + nA21 + nL22, nL/(nL11 + nA21 + nL22), 2*nL, 4*nL11 + 2*nA21 + 2*nL22, 2*nL/(4*nL11 + 2*nA21 + 2*nL22)) ; fprintf(msgFile, "\n ratios %8.3f %8.3f %8.3f", nPhi/nV, nL/(nL11 + nA21 + nL22), 2*nL/(4*nL11 + 2*nA21 + 2*nL22)) ; } /* ------------------------ free the working storage ------------------------ */ Graph_free(graph) ; ETree_free(etree) ; IV_free(mapIV) ; IVL_free(symbfacIVL) ; IVfree(frontmap) ; fprintf(msgFile, "\n") ; fclose(msgFile) ; return(1) ; }