149b5e25fSSatish Balay /* Using Modified Sparse Row (MSR) storage. 249b5e25fSSatish Balay See page 85, "Iterative Methods ..." by Saad. */ 349b5e25fSSatish Balay 4*671cb588SHong Zhang /*$Id: sbaijfact.c,v 1.47 2000/11/07 02:00:25 hzhang Exp hzhang $*/ 549b5e25fSSatish Balay /* 6effa298cSHong Zhang Symbolic U^T*D*U factorization for SBAIJ format. Modified from SSF of YSMP. 749b5e25fSSatish Balay */ 849b5e25fSSatish Balay #include "sbaij.h" 949b5e25fSSatish Balay #include "src/mat/impls/baij/seq/baij.h" 1049b5e25fSSatish Balay #include "src/vec/vecimpl.h" 1149b5e25fSSatish Balay #include "src/inline/ilu.h" 1249a6740bSHong Zhang #include "include/petscis.h" 1349b5e25fSSatish Balay 1449b5e25fSSatish Balay #undef __FUNC__ 156f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorSymbolic_SeqSBAIJ" 1677f73120SHong Zhang int MatCholeskyFactorSymbolic_SeqSBAIJ(Mat A,IS perm,PetscReal f,Mat *B) 1749b5e25fSSatish Balay { 1849b5e25fSSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b; 19cb718733SHong Zhang int *rip,ierr,i,mbs = a->mbs,*ai,*aj; 20066653e3SSatish Balay int *jutmp,bs = a->bs,bs2=a->bs2; 21066653e3SSatish Balay int m,nzi,realloc = 0; 22066653e3SSatish Balay int *jl,*q,jumin,jmin,jmax,juptr,nzk,qm,*iu,*ju,k,j,vj,umax,maxadd; 23*671cb588SHong Zhang PetscTruth perm_identity; 2449b5e25fSSatish Balay 2549b5e25fSSatish Balay PetscFunctionBegin; 26*671cb588SHong Zhang if (!perm) { 27*671cb588SHong Zhang ierr = ISCreateStride(PETSC_COMM_SELF,mbs,0,1,&perm);CHKERRQ(ierr); 28*671cb588SHong Zhang } 2977f73120SHong Zhang PetscValidHeaderSpecific(perm,IS_COOKIE); 302d007305SHong Zhang 31cb718733SHong Zhang /* check whether perm is the identity mapping */ 32*671cb588SHong Zhang ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 33*671cb588SHong Zhang if (!perm_identity) a->permute = PETSC_TRUE; 342d007305SHong Zhang 35*671cb588SHong Zhang ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 36*671cb588SHong Zhang 37*671cb588SHong Zhang if (perm_identity){ /* without permutation */ 382d007305SHong Zhang ai = a->i; aj = a->j; 392d007305SHong Zhang } else { /* non-trivial permutation */ 402d007305SHong Zhang ierr = MatReorderingSeqSBAIJ(A, perm);CHKERRA(ierr); 412d007305SHong Zhang ai = a->inew; aj = a->jnew; 422d007305SHong Zhang } 432d007305SHong Zhang 4449b5e25fSSatish Balay /* initialization */ 4549b5e25fSSatish Balay /* Don't know how many column pointers are needed so estimate. 4649b5e25fSSatish Balay Use Modified Sparse Row storage for u and ju, see Sasd pp.85 */ 479f9cb213SHong Zhang iu = (int*)PetscMalloc((mbs+1)*sizeof(int));CHKPTRQ(iu); 4849b5e25fSSatish Balay umax = (int)(f*ai[mbs] + 1); umax += mbs + 1; 499f9cb213SHong Zhang ju = (int*)PetscMalloc(umax*sizeof(int));CHKPTRQ(ju); 5049b5e25fSSatish Balay iu[0] = mbs+1; 5149b5e25fSSatish Balay juptr = mbs; 5249b5e25fSSatish Balay jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 5349b5e25fSSatish Balay q = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(q); 5449b5e25fSSatish Balay for (i=0; i<mbs; i++){ 5549b5e25fSSatish Balay jl[i] = mbs; q[i] = 0; 5649b5e25fSSatish Balay } 5749b5e25fSSatish Balay 5849b5e25fSSatish Balay /* for each row k */ 5949b5e25fSSatish Balay for (k=0; k<mbs; k++){ 6049b5e25fSSatish Balay nzk = 0; /* num. of nz blocks in k-th block row with diagonal block excluded */ 6149b5e25fSSatish Balay q[k] = mbs; 6249b5e25fSSatish Balay /* initialize nonzero structure of k-th row to row rip[k] of A */ 6349b5e25fSSatish Balay jmin = ai[rip[k]]; 6449b5e25fSSatish Balay jmax = ai[rip[k]+1]; 6549b5e25fSSatish Balay for (j=jmin; j<jmax; j++){ 66cb718733SHong Zhang vj = rip[aj[j]]; /* col. value */ 6749b5e25fSSatish Balay if(vj > k){ 6849b5e25fSSatish Balay qm = k; 6949b5e25fSSatish Balay do { 7049b5e25fSSatish Balay m = qm; qm = q[m]; 7149b5e25fSSatish Balay } while(qm < vj); 7249b5e25fSSatish Balay if (qm == vj) { 7349b5e25fSSatish Balay printf(" error: duplicate entry in A\n"); break; 7449b5e25fSSatish Balay } 7549b5e25fSSatish Balay nzk++; 7649b5e25fSSatish Balay q[m] = vj; 7749b5e25fSSatish Balay q[vj] = qm; 7849b5e25fSSatish Balay } /* if(vj > k) */ 7949b5e25fSSatish Balay } /* for (j=jmin; j<jmax; j++) */ 8049b5e25fSSatish Balay 8149b5e25fSSatish Balay /* modify nonzero structure of k-th row by computing fill-in 8249b5e25fSSatish Balay for each row i to be merged in */ 8349b5e25fSSatish Balay i = k; 8449b5e25fSSatish Balay i = jl[i]; /* next pivot row (== mbs for symbolic factorization) */ 8549b5e25fSSatish Balay /* printf(" next pivot row i=%d\n",i); */ 8649b5e25fSSatish Balay while (i < mbs){ 8749b5e25fSSatish Balay /* merge row i into k-th row */ 8849b5e25fSSatish Balay nzi = iu[i+1] - (iu[i]+1); 8949b5e25fSSatish Balay jmin = iu[i] + 1; jmax = iu[i] + nzi; 9049b5e25fSSatish Balay qm = k; 9149b5e25fSSatish Balay for (j=jmin; j<jmax+1; j++){ 9249b5e25fSSatish Balay vj = ju[j]; 9349b5e25fSSatish Balay do { 9449b5e25fSSatish Balay m = qm; qm = q[m]; 9549b5e25fSSatish Balay } while (qm < vj); 9649b5e25fSSatish Balay if (qm != vj){ 9749b5e25fSSatish Balay nzk++; q[m] = vj; q[vj] = qm; qm = vj; 9849b5e25fSSatish Balay } 9949b5e25fSSatish Balay } 10049b5e25fSSatish Balay i = jl[i]; /* next pivot row */ 10149b5e25fSSatish Balay } 10249b5e25fSSatish Balay 10349b5e25fSSatish Balay /* add k to row list for first nonzero element in k-th row */ 10449b5e25fSSatish Balay if (nzk > 0){ 10549b5e25fSSatish Balay i = q[k]; /* col value of first nonzero element in U(k, k+1:mbs-1) */ 10649b5e25fSSatish Balay jl[k] = jl[i]; jl[i] = k; 10749b5e25fSSatish Balay } 10849b5e25fSSatish Balay iu[k+1] = iu[k] + nzk; /* printf(" iu[%d]=%d, umax=%d\n", k+1, iu[k+1],umax);*/ 10949b5e25fSSatish Balay 11049b5e25fSSatish Balay /* allocate more space to ju if needed */ 11149b5e25fSSatish Balay if (iu[k+1] > umax) { printf("allocate more space, iu[%d]=%d > umax=%d\n",k+1, iu[k+1],umax); 11249b5e25fSSatish Balay /* estimate how much additional space we will need */ 11349b5e25fSSatish Balay /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */ 11449b5e25fSSatish Balay /* just double the memory each time */ 11549b5e25fSSatish Balay maxadd = umax; 11649b5e25fSSatish Balay if (maxadd < nzk) maxadd = (mbs-k)*(nzk+1)/2; 11749b5e25fSSatish Balay umax += maxadd; 11849b5e25fSSatish Balay 1199f9cb213SHong Zhang /* allocate a longer ju */ 12049b5e25fSSatish Balay jutmp = (int*)PetscMalloc(umax*sizeof(int));CHKPTRQ(jutmp); 12149b5e25fSSatish Balay ierr = PetscMemcpy(jutmp,ju,iu[k]*sizeof(int));CHKERRQ(ierr); 12249b5e25fSSatish Balay ierr = PetscFree(ju);CHKERRQ(ierr); 1239f9cb213SHong Zhang ju = jutmp; 12449b5e25fSSatish Balay realloc++; /* count how many times we realloc */ 12549b5e25fSSatish Balay } 12649b5e25fSSatish Balay 12749b5e25fSSatish Balay /* save nonzero structure of k-th row in ju */ 12849b5e25fSSatish Balay i=k; 12949b5e25fSSatish Balay jumin = juptr + 1; juptr += nzk; 13049b5e25fSSatish Balay for (j=jumin; j<juptr+1; j++){ 13149b5e25fSSatish Balay i=q[i]; 13249b5e25fSSatish Balay ju[j]=i; 13349b5e25fSSatish Balay } 13477f73120SHong Zhang } 13549b5e25fSSatish Balay 13649b5e25fSSatish Balay if (ai[mbs] != 0) { 13749b5e25fSSatish Balay PetscReal af = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]); 1386f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Reallocs %d Fill ratio:given %g needed %g\n",realloc,f,af); 1396f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Run with -pc_lu_fill %g or use \n",af); 1406f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:PCLUSetFill(pc,%g);\n",af); 1416f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:for best performance.\n"); 14249b5e25fSSatish Balay } else { 1436f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Empty matrix.\n"); 14449b5e25fSSatish Balay } 14549b5e25fSSatish Balay 14677f73120SHong Zhang ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 14749b5e25fSSatish Balay ierr = PetscFree(q);CHKERRQ(ierr); 14849b5e25fSSatish Balay ierr = PetscFree(jl);CHKERRQ(ierr); 14949b5e25fSSatish Balay 15049b5e25fSSatish Balay /* put together the new matrix */ 15149b5e25fSSatish Balay ierr = MatCreateSeqSBAIJ(A->comm,bs,bs*mbs,bs*mbs,0,PETSC_NULL,B);CHKERRQ(ierr); 1523d840bdaSHong Zhang /* PLogObjectParent(*B,iperm); */ 15349b5e25fSSatish Balay b = (Mat_SeqSBAIJ*)(*B)->data; 15449b5e25fSSatish Balay ierr = PetscFree(b->imax);CHKERRQ(ierr); 15549b5e25fSSatish Balay b->singlemalloc = PETSC_FALSE; 15649b5e25fSSatish Balay /* the next line frees the default space generated by the Create() */ 15749b5e25fSSatish Balay ierr = PetscFree(b->a);CHKERRQ(ierr); 15849b5e25fSSatish Balay ierr = PetscFree(b->ilen);CHKERRQ(ierr); 15949b5e25fSSatish Balay b->a = (MatScalar*)PetscMalloc((iu[mbs]+1)*sizeof(MatScalar)*bs2);CHKPTRQ(b->a); 16049b5e25fSSatish Balay b->j = ju; 16149b5e25fSSatish Balay b->i = iu; 16249b5e25fSSatish Balay b->diag = 0; 16349b5e25fSSatish Balay b->ilen = 0; 16449b5e25fSSatish Balay b->imax = 0; 16577f73120SHong Zhang b->row = perm; 16677f73120SHong Zhang ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 167cb718733SHong Zhang b->icol = perm; 168cb718733SHong Zhang ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 16949b5e25fSSatish Balay b->solve_work = (Scalar*)PetscMalloc((bs*mbs+bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 17049b5e25fSSatish Balay /* In b structure: Free imax, ilen, old a, old j. 17149b5e25fSSatish Balay Allocate idnew, solve_work, new a, new j */ 17249b5e25fSSatish Balay PLogObjectMemory(*B,(iu[mbs]-mbs)*(sizeof(int)+sizeof(MatScalar))); 17349b5e25fSSatish Balay b->s_maxnz = b->s_nz = iu[mbs]; 17449b5e25fSSatish Balay 1755c0bcdfcSHong Zhang (*B)->factor = FACTOR_CHOLESKY; 17649b5e25fSSatish Balay (*B)->info.factor_mallocs = realloc; 17749b5e25fSSatish Balay (*B)->info.fill_ratio_given = f; 17849b5e25fSSatish Balay if (ai[mbs] != 0) { 17949b5e25fSSatish Balay (*B)->info.fill_ratio_needed = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]); 18049b5e25fSSatish Balay } else { 18149b5e25fSSatish Balay (*B)->info.fill_ratio_needed = 0.0; 18249b5e25fSSatish Balay } 1830aa0ce06SHong Zhang 184*671cb588SHong Zhang if (perm_identity){ 185*671cb588SHong Zhang switch (bs) { 186*671cb588SHong Zhang case 1: 187*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering; 188*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering; 189*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=1\n"); 190*671cb588SHong Zhang break; 191*671cb588SHong Zhang case 2: 192*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering; 193*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_2_NaturalOrdering; 194*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=2\n"); 195*671cb588SHong Zhang break; 196*671cb588SHong Zhang case 3: 197*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering; 198*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_3_NaturalOrdering; 199*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:sing special in-place natural ordering factor and solve BS=3\n"); 200*671cb588SHong Zhang break; 201*671cb588SHong Zhang case 4: 202*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering; 203*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_4_NaturalOrdering; 204*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=4\n"); 205*671cb588SHong Zhang break; 206*671cb588SHong Zhang case 5: 207*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering; 208*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_5_NaturalOrdering; 209*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=5\n"); 210*671cb588SHong Zhang break; 211*671cb588SHong Zhang case 6: 212*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering; 213*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_6_NaturalOrdering; 214*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=6\n"); 215*671cb588SHong Zhang break; 216*671cb588SHong Zhang case 7: 217*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering; 218*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_7_NaturalOrdering; 219*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=7\n"); 220*671cb588SHong Zhang break; 221*671cb588SHong Zhang default: 222*671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering; 223*671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_N_NaturalOrdering; 224*671cb588SHong Zhang PLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS>7\n"); 225*671cb588SHong Zhang break; 226*671cb588SHong Zhang } 227*671cb588SHong Zhang } 228*671cb588SHong Zhang 22949b5e25fSSatish Balay PetscFunctionReturn(0); 23049b5e25fSSatish Balay } 23149b5e25fSSatish Balay 23249b5e25fSSatish Balay #undef __FUNC__ 2336f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_N" 2346f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_N(Mat A,Mat *B) 23549b5e25fSSatish Balay { 23649b5e25fSSatish Balay Mat C = *B; 2374c16a6a6SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2384c16a6a6SHong Zhang IS perm = b->row; 2394c16a6a6SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2404c16a6a6SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2414c16a6a6SHong Zhang int bs=a->bs,bs2 = a->bs2; 2424c16a6a6SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2434c16a6a6SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 2444c16a6a6SHong Zhang MatScalar *W,*work; 2454c16a6a6SHong Zhang int *pivots; 2464c16a6a6SHong Zhang 2474c16a6a6SHong Zhang PetscFunctionBegin; 2489706f043SHong Zhang 2494c16a6a6SHong Zhang /* initialization */ 2504c16a6a6SHong Zhang rtmp = (MatScalar*)PetscMalloc(bs2*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2514c16a6a6SHong Zhang ierr = PetscMemzero(rtmp,bs2*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2524c16a6a6SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2534c16a6a6SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2544c16a6a6SHong Zhang for (i=0; i<mbs; i++) { 2554c16a6a6SHong Zhang jl[i] = mbs; il[0] = 0; 2564c16a6a6SHong Zhang } 2574c16a6a6SHong Zhang dk = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(dk); 2584c16a6a6SHong Zhang uik = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(uik); 2594c16a6a6SHong Zhang W = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(W); 2604c16a6a6SHong Zhang work = (MatScalar*)PetscMalloc(bs*sizeof(MatScalar));CHKPTRQ(work); 2614c16a6a6SHong Zhang pivots= (int*)PetscMalloc(bs*sizeof(int));CHKPTRQ(pivots); 2624c16a6a6SHong Zhang 2634c16a6a6SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2644c16a6a6SHong Zhang 2654c16a6a6SHong Zhang /* check permutation */ 2664c16a6a6SHong Zhang if (!a->permute){ 2674c16a6a6SHong Zhang ai = a->i; aj = a->j; aa = a->a; 2684c16a6a6SHong Zhang } else { 2694c16a6a6SHong Zhang ai = a->inew; aj = a->jnew; 2704c16a6a6SHong Zhang aa = (MatScalar*)PetscMalloc(bs2*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2714c16a6a6SHong Zhang ierr = PetscMemcpy(aa,a->a,bs2*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2724c16a6a6SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2734c16a6a6SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2744c16a6a6SHong Zhang 2754c16a6a6SHong Zhang for (i=0; i<mbs; i++){ 2764c16a6a6SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2774c16a6a6SHong Zhang for (j=jmin; j<jmax; j++){ 2784c16a6a6SHong Zhang while (a2anew[j] != j){ 2794c16a6a6SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2804c16a6a6SHong Zhang for (k1=0; k1<bs2; k1++){ 2814c16a6a6SHong Zhang dk[k1] = aa[k*bs2+k1]; 2824c16a6a6SHong Zhang aa[k*bs2+k1] = aa[j*bs2+k1]; 2834c16a6a6SHong Zhang aa[j*bs2+k1] = dk[k1]; 2844c16a6a6SHong Zhang } 2854c16a6a6SHong Zhang } 2864c16a6a6SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2874c16a6a6SHong Zhang if (i > aj[j]){ 2884c16a6a6SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2894c16a6a6SHong Zhang ap = aa + j*bs2; /* ptr to the beginning of j-th block of aa */ 2904c16a6a6SHong Zhang for (k=0; k<bs2; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2914c16a6a6SHong Zhang for (k=0; k<bs; k++){ /* j-th block of aa <- dk^T */ 2924c16a6a6SHong Zhang for (k1=0; k1<bs; k1++) *ap++ = dk[k + bs*k1]; 2934c16a6a6SHong Zhang } 2944c16a6a6SHong Zhang } 2954c16a6a6SHong Zhang } 2964c16a6a6SHong Zhang } 2974c16a6a6SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2984c16a6a6SHong Zhang } 2994c16a6a6SHong Zhang 3004c16a6a6SHong Zhang /* for each row k */ 3014c16a6a6SHong Zhang for (k = 0; k<mbs; k++){ 3024c16a6a6SHong Zhang 3034c16a6a6SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 3044c16a6a6SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 3054c16a6a6SHong Zhang if (jmin < jmax) { 3064c16a6a6SHong Zhang ap = aa + jmin*bs2; 3074c16a6a6SHong Zhang for (j = jmin; j < jmax; j++){ 3084c16a6a6SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 3094c16a6a6SHong Zhang rtmp_ptr = rtmp + vj*bs2; 3104c16a6a6SHong Zhang for (i=0; i<bs2; i++) *rtmp_ptr++ = *ap++; 3114c16a6a6SHong Zhang } 3124c16a6a6SHong Zhang } 3134c16a6a6SHong Zhang 3144c16a6a6SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 3154c16a6a6SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 3164c16a6a6SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 3174c16a6a6SHong Zhang 3184c16a6a6SHong Zhang while (i < mbs){ 3194c16a6a6SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 3204c16a6a6SHong Zhang 3214c16a6a6SHong Zhang /* compute multiplier */ 3224c16a6a6SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 3234c16a6a6SHong Zhang 3244c16a6a6SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 3254c16a6a6SHong Zhang diag = ba + i*bs2; 3264c16a6a6SHong Zhang u = ba + ili*bs2; 3274c16a6a6SHong Zhang ierr = PetscMemzero(uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 3284c16a6a6SHong Zhang Kernel_A_gets_A_minus_B_times_C(bs,uik,diag,u); 3294c16a6a6SHong Zhang 3304c16a6a6SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 3314c16a6a6SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,dk,uik,u); 3324c16a6a6SHong Zhang 3334c16a6a6SHong Zhang /* update -U(i,k) */ 3344c16a6a6SHong Zhang ierr = PetscMemcpy(ba+ili*bs2,uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 3354c16a6a6SHong Zhang 3364c16a6a6SHong Zhang /* add multiple of row i to k-th row ... */ 3374c16a6a6SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 3384c16a6a6SHong Zhang if (jmin < jmax){ 3394c16a6a6SHong Zhang for (j=jmin; j<jmax; j++) { 3404c16a6a6SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 3414c16a6a6SHong Zhang rtmp_ptr = rtmp + bj[j]*bs2; 3424c16a6a6SHong Zhang u = ba + j*bs2; 3434c16a6a6SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,rtmp_ptr,uik,u); 3444c16a6a6SHong Zhang } 3454c16a6a6SHong Zhang 3464c16a6a6SHong Zhang /* ... add i to row list for next nonzero entry */ 3474c16a6a6SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 3484c16a6a6SHong Zhang j = bj[jmin]; 3494c16a6a6SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 3504c16a6a6SHong Zhang } 3514c16a6a6SHong Zhang i = nexti; 3524c16a6a6SHong Zhang } 3534c16a6a6SHong Zhang 3544c16a6a6SHong Zhang /* save nonzero entries in k-th row of U ... */ 3554c16a6a6SHong Zhang 3564c16a6a6SHong Zhang /* invert diagonal block */ 3574c16a6a6SHong Zhang diag = ba+k*bs2; 3584c16a6a6SHong Zhang ierr = PetscMemcpy(diag,dk,bs2*sizeof(MatScalar));CHKERRQ(ierr); 3594c16a6a6SHong Zhang Kernel_A_gets_inverse_A(bs,diag,pivots,work); 3604c16a6a6SHong Zhang 3614c16a6a6SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 3624c16a6a6SHong Zhang if (jmin < jmax) { 3634c16a6a6SHong Zhang for (j=jmin; j<jmax; j++){ 3644c16a6a6SHong Zhang vj = bj[j]; /* block col. index of U */ 3654c16a6a6SHong Zhang u = ba + j*bs2; 3664c16a6a6SHong Zhang rtmp_ptr = rtmp + vj*bs2; 3674c16a6a6SHong Zhang for (k1=0; k1<bs2; k1++){ 3684c16a6a6SHong Zhang *u++ = *rtmp_ptr; 3694c16a6a6SHong Zhang *rtmp_ptr++ = 0.0; 3704c16a6a6SHong Zhang } 3714c16a6a6SHong Zhang } 3724c16a6a6SHong Zhang 3734c16a6a6SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 3744c16a6a6SHong Zhang il[k] = jmin; 3754c16a6a6SHong Zhang i = bj[jmin]; 3764c16a6a6SHong Zhang jl[k] = jl[i]; jl[i] = k; 3774c16a6a6SHong Zhang } 3784c16a6a6SHong Zhang } 3794c16a6a6SHong Zhang 3804c16a6a6SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 3814c16a6a6SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 3824c16a6a6SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 3834c16a6a6SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 3844c16a6a6SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 3854c16a6a6SHong Zhang ierr = PetscFree(W);CHKERRQ(ierr); 3864c16a6a6SHong Zhang ierr = PetscFree(work);CHKERRQ(ierr); 3874c16a6a6SHong Zhang ierr = PetscFree(pivots);CHKERRQ(ierr); 3884c16a6a6SHong Zhang if (a->permute){ 3894c16a6a6SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 3904c16a6a6SHong Zhang } 3914c16a6a6SHong Zhang 3924c16a6a6SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 3934c16a6a6SHong Zhang C->factor = FACTOR_CHOLESKY; 3944c16a6a6SHong Zhang C->assembled = PETSC_TRUE; 3954c16a6a6SHong Zhang C->preallocated = PETSC_TRUE; 3964c16a6a6SHong Zhang PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 3974c16a6a6SHong Zhang PetscFunctionReturn(0); 3984c16a6a6SHong Zhang } 399d4edadd4SHong Zhang 400*671cb588SHong Zhang #undef __FUNC__ 401*671cb588SHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering" 402*671cb588SHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering(Mat A,Mat *B) 403*671cb588SHong Zhang { 404*671cb588SHong Zhang Mat C = *B; 405*671cb588SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 406*671cb588SHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 407*671cb588SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 408*671cb588SHong Zhang int bs=a->bs,bs2 = a->bs2; 409*671cb588SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 410*671cb588SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 411*671cb588SHong Zhang MatScalar *W,*work; 412*671cb588SHong Zhang int *pivots; 413*671cb588SHong Zhang 414*671cb588SHong Zhang PetscFunctionBegin; 415*671cb588SHong Zhang 416*671cb588SHong Zhang /* initialization */ 417*671cb588SHong Zhang 418*671cb588SHong Zhang rtmp = (MatScalar*)PetscMalloc(bs2*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 419*671cb588SHong Zhang ierr = PetscMemzero(rtmp,bs2*mbs*sizeof(MatScalar));CHKERRQ(ierr); 420*671cb588SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 421*671cb588SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 422*671cb588SHong Zhang for (i=0; i<mbs; i++) { 423*671cb588SHong Zhang jl[i] = mbs; il[0] = 0; 424*671cb588SHong Zhang } 425*671cb588SHong Zhang dk = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(dk); 426*671cb588SHong Zhang uik = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(uik); 427*671cb588SHong Zhang W = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(W); 428*671cb588SHong Zhang work = (MatScalar*)PetscMalloc(bs*sizeof(MatScalar));CHKPTRQ(work); 429*671cb588SHong Zhang pivots= (int*)PetscMalloc(bs*sizeof(int));CHKPTRQ(pivots); 430*671cb588SHong Zhang 431*671cb588SHong Zhang ai = a->i; aj = a->j; aa = a->a; 432*671cb588SHong Zhang 433*671cb588SHong Zhang /* for each row k */ 434*671cb588SHong Zhang for (k = 0; k<mbs; k++){ 435*671cb588SHong Zhang 436*671cb588SHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 437*671cb588SHong Zhang jmin = ai[k]; jmax = ai[k+1]; 438*671cb588SHong Zhang if (jmin < jmax) { 439*671cb588SHong Zhang ap = aa + jmin*bs2; 440*671cb588SHong Zhang for (j = jmin; j < jmax; j++){ 441*671cb588SHong Zhang vj = aj[j]; /* block col. index */ 442*671cb588SHong Zhang rtmp_ptr = rtmp + vj*bs2; 443*671cb588SHong Zhang for (i=0; i<bs2; i++) *rtmp_ptr++ = *ap++; 444*671cb588SHong Zhang } 445*671cb588SHong Zhang } 446*671cb588SHong Zhang 447*671cb588SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 448*671cb588SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 449*671cb588SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 450*671cb588SHong Zhang 451*671cb588SHong Zhang while (i < mbs){ 452*671cb588SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 453*671cb588SHong Zhang 454*671cb588SHong Zhang /* compute multiplier */ 455*671cb588SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 456*671cb588SHong Zhang 457*671cb588SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 458*671cb588SHong Zhang diag = ba + i*bs2; 459*671cb588SHong Zhang u = ba + ili*bs2; 460*671cb588SHong Zhang ierr = PetscMemzero(uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 461*671cb588SHong Zhang Kernel_A_gets_A_minus_B_times_C(bs,uik,diag,u); 462*671cb588SHong Zhang 463*671cb588SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 464*671cb588SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,dk,uik,u); 465*671cb588SHong Zhang 466*671cb588SHong Zhang /* update -U(i,k) */ 467*671cb588SHong Zhang ierr = PetscMemcpy(ba+ili*bs2,uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 468*671cb588SHong Zhang 469*671cb588SHong Zhang /* add multiple of row i to k-th row ... */ 470*671cb588SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 471*671cb588SHong Zhang if (jmin < jmax){ 472*671cb588SHong Zhang for (j=jmin; j<jmax; j++) { 473*671cb588SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 474*671cb588SHong Zhang rtmp_ptr = rtmp + bj[j]*bs2; 475*671cb588SHong Zhang u = ba + j*bs2; 476*671cb588SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,rtmp_ptr,uik,u); 477*671cb588SHong Zhang } 478*671cb588SHong Zhang 479*671cb588SHong Zhang /* ... add i to row list for next nonzero entry */ 480*671cb588SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 481*671cb588SHong Zhang j = bj[jmin]; 482*671cb588SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 483*671cb588SHong Zhang } 484*671cb588SHong Zhang i = nexti; 485*671cb588SHong Zhang } 486*671cb588SHong Zhang 487*671cb588SHong Zhang /* save nonzero entries in k-th row of U ... */ 488*671cb588SHong Zhang 489*671cb588SHong Zhang /* invert diagonal block */ 490*671cb588SHong Zhang diag = ba+k*bs2; 491*671cb588SHong Zhang ierr = PetscMemcpy(diag,dk,bs2*sizeof(MatScalar));CHKERRQ(ierr); 492*671cb588SHong Zhang Kernel_A_gets_inverse_A(bs,diag,pivots,work); 493*671cb588SHong Zhang 494*671cb588SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 495*671cb588SHong Zhang if (jmin < jmax) { 496*671cb588SHong Zhang for (j=jmin; j<jmax; j++){ 497*671cb588SHong Zhang vj = bj[j]; /* block col. index of U */ 498*671cb588SHong Zhang u = ba + j*bs2; 499*671cb588SHong Zhang rtmp_ptr = rtmp + vj*bs2; 500*671cb588SHong Zhang for (k1=0; k1<bs2; k1++){ 501*671cb588SHong Zhang *u++ = *rtmp_ptr; 502*671cb588SHong Zhang *rtmp_ptr++ = 0.0; 503*671cb588SHong Zhang } 504*671cb588SHong Zhang } 505*671cb588SHong Zhang 506*671cb588SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 507*671cb588SHong Zhang il[k] = jmin; 508*671cb588SHong Zhang i = bj[jmin]; 509*671cb588SHong Zhang jl[k] = jl[i]; jl[i] = k; 510*671cb588SHong Zhang } 511*671cb588SHong Zhang } 512*671cb588SHong Zhang 513*671cb588SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 514*671cb588SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 515*671cb588SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 516*671cb588SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 517*671cb588SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 518*671cb588SHong Zhang ierr = PetscFree(W);CHKERRQ(ierr); 519*671cb588SHong Zhang ierr = PetscFree(work);CHKERRQ(ierr); 520*671cb588SHong Zhang ierr = PetscFree(pivots);CHKERRQ(ierr); 521*671cb588SHong Zhang 522*671cb588SHong Zhang C->factor = FACTOR_CHOLESKY; 523*671cb588SHong Zhang C->assembled = PETSC_TRUE; 524*671cb588SHong Zhang C->preallocated = PETSC_TRUE; 525*671cb588SHong Zhang PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 526*671cb588SHong Zhang PetscFunctionReturn(0); 527*671cb588SHong Zhang } 528*671cb588SHong Zhang 5290aa0ce06SHong Zhang /* Version for when blocks are 7 by 7 */ 53049b5e25fSSatish Balay #undef __FUNC__ 5316f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_7" 5326f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_7(Mat A,Mat *B) 53349b5e25fSSatish Balay { 53449b5e25fSSatish Balay Mat C = *B; 5350aa0ce06SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 5360aa0ce06SHong Zhang IS perm = b->row; 5370aa0ce06SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 5380aa0ce06SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 5390aa0ce06SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 5400aa0ce06SHong Zhang MatScalar *u,*d,*w,*wp; 54149b5e25fSSatish Balay 54249b5e25fSSatish Balay PetscFunctionBegin; 5430aa0ce06SHong Zhang /* initialization */ 5440aa0ce06SHong Zhang w = (MatScalar*)PetscMalloc(49*mbs*sizeof(MatScalar));CHKPTRQ(w); 5450aa0ce06SHong Zhang ierr = PetscMemzero(w,49*mbs*sizeof(MatScalar));CHKERRQ(ierr); 5460aa0ce06SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 5470aa0ce06SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 5480aa0ce06SHong Zhang for (i=0; i<mbs; i++) { 5490aa0ce06SHong Zhang jl[i] = mbs; il[0] = 0; 55049b5e25fSSatish Balay } 5510aa0ce06SHong Zhang dk = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(dk); 5520aa0ce06SHong Zhang uik = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(uik); 5530aa0ce06SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 5540aa0ce06SHong Zhang 5550aa0ce06SHong Zhang /* check permutation */ 5560aa0ce06SHong Zhang if (!a->permute){ 5570aa0ce06SHong Zhang ai = a->i; aj = a->j; aa = a->a; 5580aa0ce06SHong Zhang } else { 5590aa0ce06SHong Zhang ai = a->inew; aj = a->jnew; 5600aa0ce06SHong Zhang aa = (MatScalar*)PetscMalloc(49*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 5610aa0ce06SHong Zhang ierr = PetscMemcpy(aa,a->a,49*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 5620aa0ce06SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 5630aa0ce06SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 5640aa0ce06SHong Zhang 5650aa0ce06SHong Zhang for (i=0; i<mbs; i++){ 5660aa0ce06SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 5670aa0ce06SHong Zhang for (j=jmin; j<jmax; j++){ 5680aa0ce06SHong Zhang while (a2anew[j] != j){ 5690aa0ce06SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 5700aa0ce06SHong Zhang for (k1=0; k1<49; k1++){ 5710aa0ce06SHong Zhang dk[k1] = aa[k*49+k1]; 5720aa0ce06SHong Zhang aa[k*49+k1] = aa[j*49+k1]; 5730aa0ce06SHong Zhang aa[j*49+k1] = dk[k1]; 57449b5e25fSSatish Balay } 57549b5e25fSSatish Balay } 5760aa0ce06SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 5770aa0ce06SHong Zhang if (i > aj[j]){ 5780aa0ce06SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 5790aa0ce06SHong Zhang ap = aa + j*49; /* ptr to the beginning of j-th block of aa */ 5800aa0ce06SHong Zhang for (k=0; k<49; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 5810aa0ce06SHong Zhang for (k=0; k<7; k++){ /* j-th block of aa <- dk^T */ 5820aa0ce06SHong Zhang for (k1=0; k1<7; k1++) *ap++ = dk[k + 7*k1]; 58349b5e25fSSatish Balay } 58449b5e25fSSatish Balay } 58549b5e25fSSatish Balay } 5860aa0ce06SHong Zhang } 5870aa0ce06SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 5880aa0ce06SHong Zhang } 5890aa0ce06SHong Zhang 5900aa0ce06SHong Zhang /* for each row k */ 5910aa0ce06SHong Zhang for (k = 0; k<mbs; k++){ 5920aa0ce06SHong Zhang 5930aa0ce06SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 5940aa0ce06SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 5950aa0ce06SHong Zhang if (jmin < jmax) { 5960aa0ce06SHong Zhang ap = aa + jmin*49; 5970aa0ce06SHong Zhang for (j = jmin; j < jmax; j++){ 5980aa0ce06SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 5990aa0ce06SHong Zhang wp = w + vj*49; 6000aa0ce06SHong Zhang for (i=0; i<49; i++) *wp++ = *ap++; 6010aa0ce06SHong Zhang } 6020aa0ce06SHong Zhang } 6030aa0ce06SHong Zhang 6040aa0ce06SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 6050aa0ce06SHong Zhang ierr = PetscMemcpy(dk,w+k*49,49*sizeof(MatScalar));CHKERRQ(ierr); 6060aa0ce06SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 6070aa0ce06SHong Zhang 6080aa0ce06SHong Zhang while (i < mbs){ 6090aa0ce06SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 6100aa0ce06SHong Zhang 6110aa0ce06SHong Zhang /* compute multiplier */ 6120aa0ce06SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 6130aa0ce06SHong Zhang 6140aa0ce06SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 6150aa0ce06SHong Zhang d = ba + i*49; 6160aa0ce06SHong Zhang u = ba + ili*49; 6170aa0ce06SHong Zhang 6180aa0ce06SHong Zhang uik[0] = -(d[0]*u[0] + d[7]*u[1]+ d[14]*u[2]+ d[21]*u[3]+ d[28]*u[4]+ d[35]*u[5]+ d[42]*u[6]); 6190aa0ce06SHong Zhang uik[1] = -(d[1]*u[0] + d[8]*u[1]+ d[15]*u[2]+ d[22]*u[3]+ d[29]*u[4]+ d[36]*u[5]+ d[43]*u[6]); 6200aa0ce06SHong Zhang uik[2] = -(d[2]*u[0] + d[9]*u[1]+ d[16]*u[2]+ d[23]*u[3]+ d[30]*u[4]+ d[37]*u[5]+ d[44]*u[6]); 6210aa0ce06SHong Zhang uik[3] = -(d[3]*u[0]+ d[10]*u[1]+ d[17]*u[2]+ d[24]*u[3]+ d[31]*u[4]+ d[38]*u[5]+ d[45]*u[6]); 6220aa0ce06SHong Zhang uik[4] = -(d[4]*u[0]+ d[11]*u[1]+ d[18]*u[2]+ d[25]*u[3]+ d[32]*u[4]+ d[39]*u[5]+ d[46]*u[6]); 6230aa0ce06SHong Zhang uik[5] = -(d[5]*u[0]+ d[12]*u[1]+ d[19]*u[2]+ d[26]*u[3]+ d[33]*u[4]+ d[40]*u[5]+ d[47]*u[6]); 6240aa0ce06SHong Zhang uik[6] = -(d[6]*u[0]+ d[13]*u[1]+ d[20]*u[2]+ d[27]*u[3]+ d[34]*u[4]+ d[41]*u[5]+ d[48]*u[6]); 6250aa0ce06SHong Zhang 6260aa0ce06SHong Zhang uik[7] = -(d[0]*u[7] + d[7]*u[8]+ d[14]*u[9]+ d[21]*u[10]+ d[28]*u[11]+ d[35]*u[12]+ d[42]*u[13]); 6270aa0ce06SHong Zhang uik[8] = -(d[1]*u[7] + d[8]*u[8]+ d[15]*u[9]+ d[22]*u[10]+ d[29]*u[11]+ d[36]*u[12]+ d[43]*u[13]); 6280aa0ce06SHong Zhang uik[9] = -(d[2]*u[7] + d[9]*u[8]+ d[16]*u[9]+ d[23]*u[10]+ d[30]*u[11]+ d[37]*u[12]+ d[44]*u[13]); 6290aa0ce06SHong Zhang uik[10]= -(d[3]*u[7]+ d[10]*u[8]+ d[17]*u[9]+ d[24]*u[10]+ d[31]*u[11]+ d[38]*u[12]+ d[45]*u[13]); 6300aa0ce06SHong Zhang uik[11]= -(d[4]*u[7]+ d[11]*u[8]+ d[18]*u[9]+ d[25]*u[10]+ d[32]*u[11]+ d[39]*u[12]+ d[46]*u[13]); 6310aa0ce06SHong Zhang uik[12]= -(d[5]*u[7]+ d[12]*u[8]+ d[19]*u[9]+ d[26]*u[10]+ d[33]*u[11]+ d[40]*u[12]+ d[47]*u[13]); 6320aa0ce06SHong Zhang uik[13]= -(d[6]*u[7]+ d[13]*u[8]+ d[20]*u[9]+ d[27]*u[10]+ d[34]*u[11]+ d[41]*u[12]+ d[48]*u[13]); 6330aa0ce06SHong Zhang 6340aa0ce06SHong Zhang uik[14]= -(d[0]*u[14] + d[7]*u[15]+ d[14]*u[16]+ d[21]*u[17]+ d[28]*u[18]+ d[35]*u[19]+ d[42]*u[20]); 6350aa0ce06SHong Zhang uik[15]= -(d[1]*u[14] + d[8]*u[15]+ d[15]*u[16]+ d[22]*u[17]+ d[29]*u[18]+ d[36]*u[19]+ d[43]*u[20]); 6360aa0ce06SHong Zhang uik[16]= -(d[2]*u[14] + d[9]*u[15]+ d[16]*u[16]+ d[23]*u[17]+ d[30]*u[18]+ d[37]*u[19]+ d[44]*u[20]); 6370aa0ce06SHong Zhang uik[17]= -(d[3]*u[14]+ d[10]*u[15]+ d[17]*u[16]+ d[24]*u[17]+ d[31]*u[18]+ d[38]*u[19]+ d[45]*u[20]); 6380aa0ce06SHong Zhang uik[18]= -(d[4]*u[14]+ d[11]*u[15]+ d[18]*u[16]+ d[25]*u[17]+ d[32]*u[18]+ d[39]*u[19]+ d[46]*u[20]); 6390aa0ce06SHong Zhang uik[19]= -(d[5]*u[14]+ d[12]*u[15]+ d[19]*u[16]+ d[26]*u[17]+ d[33]*u[18]+ d[40]*u[19]+ d[47]*u[20]); 6400aa0ce06SHong Zhang uik[20]= -(d[6]*u[14]+ d[13]*u[15]+ d[20]*u[16]+ d[27]*u[17]+ d[34]*u[18]+ d[41]*u[19]+ d[48]*u[20]); 6410aa0ce06SHong Zhang 6420aa0ce06SHong Zhang uik[21]= -(d[0]*u[21] + d[7]*u[22]+ d[14]*u[23]+ d[21]*u[24]+ d[28]*u[25]+ d[35]*u[26]+ d[42]*u[27]); 6430aa0ce06SHong Zhang uik[22]= -(d[1]*u[21] + d[8]*u[22]+ d[15]*u[23]+ d[22]*u[24]+ d[29]*u[25]+ d[36]*u[26]+ d[43]*u[27]); 6440aa0ce06SHong Zhang uik[23]= -(d[2]*u[21] + d[9]*u[22]+ d[16]*u[23]+ d[23]*u[24]+ d[30]*u[25]+ d[37]*u[26]+ d[44]*u[27]); 6450aa0ce06SHong Zhang uik[24]= -(d[3]*u[21]+ d[10]*u[22]+ d[17]*u[23]+ d[24]*u[24]+ d[31]*u[25]+ d[38]*u[26]+ d[45]*u[27]); 6460aa0ce06SHong Zhang uik[25]= -(d[4]*u[21]+ d[11]*u[22]+ d[18]*u[23]+ d[25]*u[24]+ d[32]*u[25]+ d[39]*u[26]+ d[46]*u[27]); 6470aa0ce06SHong Zhang uik[26]= -(d[5]*u[21]+ d[12]*u[22]+ d[19]*u[23]+ d[26]*u[24]+ d[33]*u[25]+ d[40]*u[26]+ d[47]*u[27]); 6480aa0ce06SHong Zhang uik[27]= -(d[6]*u[21]+ d[13]*u[22]+ d[20]*u[23]+ d[27]*u[24]+ d[34]*u[25]+ d[41]*u[26]+ d[48]*u[27]); 6490aa0ce06SHong Zhang 6500aa0ce06SHong Zhang uik[28]= -(d[0]*u[28] + d[7]*u[29]+ d[14]*u[30]+ d[21]*u[31]+ d[28]*u[32]+ d[35]*u[33]+ d[42]*u[34]); 6510aa0ce06SHong Zhang uik[29]= -(d[1]*u[28] + d[8]*u[29]+ d[15]*u[30]+ d[22]*u[31]+ d[29]*u[32]+ d[36]*u[33]+ d[43]*u[34]); 6520aa0ce06SHong Zhang uik[30]= -(d[2]*u[28] + d[9]*u[29]+ d[16]*u[30]+ d[23]*u[31]+ d[30]*u[32]+ d[37]*u[33]+ d[44]*u[34]); 6530aa0ce06SHong Zhang uik[31]= -(d[3]*u[28]+ d[10]*u[29]+ d[17]*u[30]+ d[24]*u[31]+ d[31]*u[32]+ d[38]*u[33]+ d[45]*u[34]); 6540aa0ce06SHong Zhang uik[32]= -(d[4]*u[28]+ d[11]*u[29]+ d[18]*u[30]+ d[25]*u[31]+ d[32]*u[32]+ d[39]*u[33]+ d[46]*u[34]); 6550aa0ce06SHong Zhang uik[33]= -(d[5]*u[28]+ d[12]*u[29]+ d[19]*u[30]+ d[26]*u[31]+ d[33]*u[32]+ d[40]*u[33]+ d[47]*u[34]); 6560aa0ce06SHong Zhang uik[34]= -(d[6]*u[28]+ d[13]*u[29]+ d[20]*u[30]+ d[27]*u[31]+ d[34]*u[32]+ d[41]*u[33]+ d[48]*u[34]); 6570aa0ce06SHong Zhang 6580aa0ce06SHong Zhang uik[35]= -(d[0]*u[35] + d[7]*u[36]+ d[14]*u[37]+ d[21]*u[38]+ d[28]*u[39]+ d[35]*u[40]+ d[42]*u[41]); 6590aa0ce06SHong Zhang uik[36]= -(d[1]*u[35] + d[8]*u[36]+ d[15]*u[37]+ d[22]*u[38]+ d[29]*u[39]+ d[36]*u[40]+ d[43]*u[41]); 6600aa0ce06SHong Zhang uik[37]= -(d[2]*u[35] + d[9]*u[36]+ d[16]*u[37]+ d[23]*u[38]+ d[30]*u[39]+ d[37]*u[40]+ d[44]*u[41]); 6610aa0ce06SHong Zhang uik[38]= -(d[3]*u[35]+ d[10]*u[36]+ d[17]*u[37]+ d[24]*u[38]+ d[31]*u[39]+ d[38]*u[40]+ d[45]*u[41]); 6620aa0ce06SHong Zhang uik[39]= -(d[4]*u[35]+ d[11]*u[36]+ d[18]*u[37]+ d[25]*u[38]+ d[32]*u[39]+ d[39]*u[40]+ d[46]*u[41]); 6630aa0ce06SHong Zhang uik[40]= -(d[5]*u[35]+ d[12]*u[36]+ d[19]*u[37]+ d[26]*u[38]+ d[33]*u[39]+ d[40]*u[40]+ d[47]*u[41]); 6640aa0ce06SHong Zhang uik[41]= -(d[6]*u[35]+ d[13]*u[36]+ d[20]*u[37]+ d[27]*u[38]+ d[34]*u[39]+ d[41]*u[40]+ d[48]*u[41]); 6650aa0ce06SHong Zhang 6660aa0ce06SHong Zhang uik[42]= -(d[0]*u[42] + d[7]*u[43]+ d[14]*u[44]+ d[21]*u[45]+ d[28]*u[46]+ d[35]*u[47]+ d[42]*u[48]); 6670aa0ce06SHong Zhang uik[43]= -(d[1]*u[42] + d[8]*u[43]+ d[15]*u[44]+ d[22]*u[45]+ d[29]*u[46]+ d[36]*u[47]+ d[43]*u[48]); 6680aa0ce06SHong Zhang uik[44]= -(d[2]*u[42] + d[9]*u[43]+ d[16]*u[44]+ d[23]*u[45]+ d[30]*u[46]+ d[37]*u[47]+ d[44]*u[48]); 6690aa0ce06SHong Zhang uik[45]= -(d[3]*u[42]+ d[10]*u[43]+ d[17]*u[44]+ d[24]*u[45]+ d[31]*u[46]+ d[38]*u[47]+ d[45]*u[48]); 6700aa0ce06SHong Zhang uik[46]= -(d[4]*u[42]+ d[11]*u[43]+ d[18]*u[44]+ d[25]*u[45]+ d[32]*u[46]+ d[39]*u[47]+ d[46]*u[48]); 6710aa0ce06SHong Zhang uik[47]= -(d[5]*u[42]+ d[12]*u[43]+ d[19]*u[44]+ d[26]*u[45]+ d[33]*u[46]+ d[40]*u[47]+ d[47]*u[48]); 6720aa0ce06SHong Zhang uik[48]= -(d[6]*u[42]+ d[13]*u[43]+ d[20]*u[44]+ d[27]*u[45]+ d[34]*u[46]+ d[41]*u[47]+ d[48]*u[48]); 6730aa0ce06SHong Zhang 6740aa0ce06SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 6750aa0ce06SHong Zhang dk[0]+= uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6]; 6760aa0ce06SHong Zhang dk[1]+= uik[7]*u[0] + uik[8]*u[1] + uik[9]*u[2]+ uik[10]*u[3]+ uik[11]*u[4]+ uik[12]*u[5]+ uik[13]*u[6]; 6770aa0ce06SHong Zhang dk[2]+= uik[14]*u[0]+ uik[15]*u[1]+ uik[16]*u[2]+ uik[17]*u[3]+ uik[18]*u[4]+ uik[19]*u[5]+ uik[20]*u[6]; 6780aa0ce06SHong Zhang dk[3]+= uik[21]*u[0]+ uik[22]*u[1]+ uik[23]*u[2]+ uik[24]*u[3]+ uik[25]*u[4]+ uik[26]*u[5]+ uik[27]*u[6]; 6790aa0ce06SHong Zhang dk[4]+= uik[28]*u[0]+ uik[29]*u[1]+ uik[30]*u[2]+ uik[31]*u[3]+ uik[32]*u[4]+ uik[33]*u[5]+ uik[34]*u[6]; 6800aa0ce06SHong Zhang dk[5]+= uik[35]*u[0]+ uik[36]*u[1]+ uik[37]*u[2]+ uik[38]*u[3]+ uik[39]*u[4]+ uik[40]*u[5]+ uik[41]*u[6]; 6810aa0ce06SHong Zhang dk[6]+= uik[42]*u[0]+ uik[43]*u[1]+ uik[44]*u[2]+ uik[45]*u[3]+ uik[46]*u[4]+ uik[47]*u[5]+ uik[48]*u[6]; 6820aa0ce06SHong Zhang 6830aa0ce06SHong Zhang dk[7]+= uik[0]*u[7] + uik[1]*u[8] + uik[2]*u[9] + uik[3]*u[10] + uik[4]*u[11] + uik[5]*u[12] + uik[6]*u[13]; 6840aa0ce06SHong Zhang dk[8]+= uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]; 6850aa0ce06SHong Zhang dk[9]+= uik[14]*u[7]+ uik[15]*u[8]+ uik[16]*u[9]+ uik[17]*u[10]+ uik[18]*u[11]+ uik[19]*u[12]+ uik[20]*u[13]; 6860aa0ce06SHong Zhang dk[10]+=uik[21]*u[7]+ uik[22]*u[8]+ uik[23]*u[9]+ uik[24]*u[10]+ uik[25]*u[11]+ uik[26]*u[12]+ uik[27]*u[13]; 6870aa0ce06SHong Zhang dk[11]+=uik[28]*u[7]+ uik[29]*u[8]+ uik[30]*u[9]+ uik[31]*u[10]+ uik[32]*u[11]+ uik[33]*u[12]+ uik[34]*u[13]; 6880aa0ce06SHong Zhang dk[12]+=uik[35]*u[7]+ uik[36]*u[8]+ uik[37]*u[9]+ uik[38]*u[10]+ uik[39]*u[11]+ uik[40]*u[12]+ uik[41]*u[13]; 6890aa0ce06SHong Zhang dk[13]+=uik[42]*u[7]+ uik[43]*u[8]+ uik[44]*u[9]+ uik[45]*u[10]+ uik[46]*u[11]+ uik[47]*u[12]+ uik[48]*u[13]; 6900aa0ce06SHong Zhang 6910aa0ce06SHong Zhang dk[14]+= uik[0]*u[14] + uik[1]*u[15] + uik[2]*u[16] + uik[3]*u[17] + uik[4]*u[18] + uik[5]*u[19] + uik[6]*u[20]; 6920aa0ce06SHong Zhang dk[15]+= uik[7]*u[14] + uik[8]*u[15] + uik[9]*u[16]+ uik[10]*u[17]+ uik[11]*u[18]+ uik[12]*u[19]+ uik[13]*u[20]; 6930aa0ce06SHong Zhang dk[16]+= uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]; 6940aa0ce06SHong Zhang dk[17]+= uik[21]*u[14]+ uik[22]*u[15]+ uik[23]*u[16]+ uik[24]*u[17]+ uik[25]*u[18]+ uik[26]*u[19]+ uik[27]*u[20]; 6950aa0ce06SHong Zhang dk[18]+= uik[28]*u[14]+ uik[29]*u[15]+ uik[30]*u[16]+ uik[31]*u[17]+ uik[32]*u[18]+ uik[33]*u[19]+ uik[34]*u[20]; 6960aa0ce06SHong Zhang dk[19]+= uik[35]*u[14]+ uik[36]*u[15]+ uik[37]*u[16]+ uik[38]*u[17]+ uik[39]*u[18]+ uik[40]*u[19]+ uik[41]*u[20]; 6970aa0ce06SHong Zhang dk[20]+= uik[42]*u[14]+ uik[43]*u[15]+ uik[44]*u[16]+ uik[45]*u[17]+ uik[46]*u[18]+ uik[47]*u[19]+ uik[48]*u[20]; 6980aa0ce06SHong Zhang 6990aa0ce06SHong Zhang dk[21]+= uik[0]*u[21] + uik[1]*u[22] + uik[2]*u[23] + uik[3]*u[24] + uik[4]*u[25] + uik[5]*u[26] + uik[6]*u[27]; 7000aa0ce06SHong Zhang dk[22]+= uik[7]*u[21] + uik[8]*u[22] + uik[9]*u[23]+ uik[10]*u[24]+ uik[11]*u[25]+ uik[12]*u[26]+ uik[13]*u[27]; 7010aa0ce06SHong Zhang dk[23]+= uik[14]*u[21]+ uik[15]*u[22]+ uik[16]*u[23]+ uik[17]*u[24]+ uik[18]*u[25]+ uik[19]*u[26]+ uik[20]*u[27]; 7020aa0ce06SHong Zhang dk[24]+= uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]; 7030aa0ce06SHong Zhang dk[25]+= uik[28]*u[21]+ uik[29]*u[22]+ uik[30]*u[23]+ uik[31]*u[24]+ uik[32]*u[25]+ uik[33]*u[26]+ uik[34]*u[27]; 7040aa0ce06SHong Zhang dk[26]+= uik[35]*u[21]+ uik[36]*u[22]+ uik[37]*u[23]+ uik[38]*u[24]+ uik[39]*u[25]+ uik[40]*u[26]+ uik[41]*u[27]; 7050aa0ce06SHong Zhang dk[27]+= uik[42]*u[21]+ uik[43]*u[22]+ uik[44]*u[23]+ uik[45]*u[24]+ uik[46]*u[25]+ uik[47]*u[26]+ uik[48]*u[27]; 7060aa0ce06SHong Zhang 7070aa0ce06SHong Zhang dk[28]+= uik[0]*u[28] + uik[1]*u[29] + uik[2]*u[30] + uik[3]*u[31] + uik[4]*u[32] + uik[5]*u[33] + uik[6]*u[34]; 7080aa0ce06SHong Zhang dk[29]+= uik[7]*u[28] + uik[8]*u[29] + uik[9]*u[30]+ uik[10]*u[31]+ uik[11]*u[32]+ uik[12]*u[33]+ uik[13]*u[34]; 7090aa0ce06SHong Zhang dk[30]+= uik[14]*u[28]+ uik[15]*u[29]+ uik[16]*u[30]+ uik[17]*u[31]+ uik[18]*u[32]+ uik[19]*u[33]+ uik[20]*u[34]; 7100aa0ce06SHong Zhang dk[31]+= uik[21]*u[28]+ uik[22]*u[29]+ uik[23]*u[30]+ uik[24]*u[31]+ uik[25]*u[32]+ uik[26]*u[33]+ uik[27]*u[34]; 7110aa0ce06SHong Zhang dk[32]+= uik[28]*u[28]+ uik[29]*u[29]+ uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]; 7120aa0ce06SHong Zhang dk[33]+= uik[35]*u[28]+ uik[36]*u[29]+ uik[37]*u[30]+ uik[38]*u[31]+ uik[39]*u[32]+ uik[40]*u[33]+ uik[41]*u[34]; 7130aa0ce06SHong Zhang dk[34]+= uik[42]*u[28]+ uik[43]*u[29]+ uik[44]*u[30]+ uik[45]*u[31]+ uik[46]*u[32]+ uik[47]*u[33]+ uik[48]*u[34]; 7140aa0ce06SHong Zhang 7150aa0ce06SHong Zhang dk[35]+= uik[0]*u[35] + uik[1]*u[36] + uik[2]*u[37] + uik[3]*u[38] + uik[4]*u[39] + uik[5]*u[40] + uik[6]*u[41]; 7160aa0ce06SHong Zhang dk[36]+= uik[7]*u[35] + uik[8]*u[36] + uik[9]*u[37]+ uik[10]*u[38]+ uik[11]*u[39]+ uik[12]*u[40]+ uik[13]*u[41]; 7170aa0ce06SHong Zhang dk[37]+= uik[14]*u[35]+ uik[15]*u[36]+ uik[16]*u[37]+ uik[17]*u[38]+ uik[18]*u[39]+ uik[19]*u[40]+ uik[20]*u[41]; 7180aa0ce06SHong Zhang dk[38]+= uik[21]*u[35]+ uik[22]*u[36]+ uik[23]*u[37]+ uik[24]*u[38]+ uik[25]*u[39]+ uik[26]*u[40]+ uik[27]*u[41]; 7190aa0ce06SHong Zhang dk[39]+= uik[28]*u[35]+ uik[29]*u[36]+ uik[30]*u[37]+ uik[31]*u[38]+ uik[32]*u[39]+ uik[33]*u[40]+ uik[34]*u[41]; 7200aa0ce06SHong Zhang dk[40]+= uik[35]*u[35]+ uik[36]*u[36]+ uik[37]*u[37]+ uik[38]*u[38]+ uik[39]*u[39]+ uik[40]*u[40]+ uik[41]*u[41]; 7210aa0ce06SHong Zhang dk[41]+= uik[42]*u[35]+ uik[43]*u[36]+ uik[44]*u[37]+ uik[45]*u[38]+ uik[46]*u[39]+ uik[47]*u[40]+ uik[48]*u[41]; 7220aa0ce06SHong Zhang 7230aa0ce06SHong Zhang dk[42]+= uik[0]*u[42] + uik[1]*u[43] + uik[2]*u[44] + uik[3]*u[45] + uik[4]*u[46] + uik[5]*u[47] + uik[6]*u[48]; 7240aa0ce06SHong Zhang dk[43]+= uik[7]*u[42] + uik[8]*u[43] + uik[9]*u[44]+ uik[10]*u[45]+ uik[11]*u[46]+ uik[12]*u[47]+ uik[13]*u[48]; 7250aa0ce06SHong Zhang dk[44]+= uik[14]*u[42]+ uik[15]*u[43]+ uik[16]*u[44]+ uik[17]*u[45]+ uik[18]*u[46]+ uik[19]*u[47]+ uik[20]*u[48]; 7260aa0ce06SHong Zhang dk[45]+= uik[21]*u[42]+ uik[22]*u[43]+ uik[23]*u[44]+ uik[24]*u[45]+ uik[25]*u[46]+ uik[26]*u[47]+ uik[27]*u[48]; 7270aa0ce06SHong Zhang dk[46]+= uik[28]*u[42]+ uik[29]*u[43]+ uik[30]*u[44]+ uik[31]*u[45]+ uik[32]*u[46]+ uik[33]*u[47]+ uik[34]*u[48]; 7280aa0ce06SHong Zhang dk[47]+= uik[35]*u[42]+ uik[36]*u[43]+ uik[37]*u[44]+ uik[38]*u[45]+ uik[39]*u[46]+ uik[40]*u[47]+ uik[41]*u[48]; 7290aa0ce06SHong Zhang dk[48]+= uik[42]*u[42]+ uik[43]*u[43]+ uik[44]*u[44]+ uik[45]*u[45]+ uik[46]*u[46]+ uik[47]*u[47]+ uik[48]*u[48]; 7300aa0ce06SHong Zhang 7310aa0ce06SHong Zhang /* update -U(i,k) */ 7320aa0ce06SHong Zhang ierr = PetscMemcpy(ba+ili*49,uik,49*sizeof(MatScalar));CHKERRQ(ierr); 7330aa0ce06SHong Zhang 7340aa0ce06SHong Zhang /* add multiple of row i to k-th row ... */ 7350aa0ce06SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 7360aa0ce06SHong Zhang if (jmin < jmax){ 7370aa0ce06SHong Zhang for (j=jmin; j<jmax; j++) { 7380aa0ce06SHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 7390aa0ce06SHong Zhang wp = w + bj[j]*49; 7400aa0ce06SHong Zhang u = ba + j*49; 7410aa0ce06SHong Zhang 7420aa0ce06SHong Zhang wp[0]+= uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6]; 7430aa0ce06SHong Zhang wp[1]+= uik[7]*u[0] + uik[8]*u[1] + uik[9]*u[2]+ uik[10]*u[3]+ uik[11]*u[4]+ uik[12]*u[5]+ uik[13]*u[6]; 7440aa0ce06SHong Zhang wp[2]+= uik[14]*u[0]+ uik[15]*u[1]+ uik[16]*u[2]+ uik[17]*u[3]+ uik[18]*u[4]+ uik[19]*u[5]+ uik[20]*u[6]; 7450aa0ce06SHong Zhang wp[3]+= uik[21]*u[0]+ uik[22]*u[1]+ uik[23]*u[2]+ uik[24]*u[3]+ uik[25]*u[4]+ uik[26]*u[5]+ uik[27]*u[6]; 7460aa0ce06SHong Zhang wp[4]+= uik[28]*u[0]+ uik[29]*u[1]+ uik[30]*u[2]+ uik[31]*u[3]+ uik[32]*u[4]+ uik[33]*u[5]+ uik[34]*u[6]; 7470aa0ce06SHong Zhang wp[5]+= uik[35]*u[0]+ uik[36]*u[1]+ uik[37]*u[2]+ uik[38]*u[3]+ uik[39]*u[4]+ uik[40]*u[5]+ uik[41]*u[6]; 7480aa0ce06SHong Zhang wp[6]+= uik[42]*u[0]+ uik[43]*u[1]+ uik[44]*u[2]+ uik[45]*u[3]+ uik[46]*u[4]+ uik[47]*u[5]+ uik[48]*u[6]; 7490aa0ce06SHong Zhang 7500aa0ce06SHong Zhang wp[7]+= uik[0]*u[7] + uik[1]*u[8] + uik[2]*u[9] + uik[3]*u[10] + uik[4]*u[11] + uik[5]*u[12] + uik[6]*u[13]; 7510aa0ce06SHong Zhang wp[8]+= uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]; 7520aa0ce06SHong Zhang wp[9]+= uik[14]*u[7]+ uik[15]*u[8]+ uik[16]*u[9]+ uik[17]*u[10]+ uik[18]*u[11]+ uik[19]*u[12]+ uik[20]*u[13]; 7530aa0ce06SHong Zhang wp[10]+=uik[21]*u[7]+ uik[22]*u[8]+ uik[23]*u[9]+ uik[24]*u[10]+ uik[25]*u[11]+ uik[26]*u[12]+ uik[27]*u[13]; 7540aa0ce06SHong Zhang wp[11]+=uik[28]*u[7]+ uik[29]*u[8]+ uik[30]*u[9]+ uik[31]*u[10]+ uik[32]*u[11]+ uik[33]*u[12]+ uik[34]*u[13]; 7550aa0ce06SHong Zhang wp[12]+=uik[35]*u[7]+ uik[36]*u[8]+ uik[37]*u[9]+ uik[38]*u[10]+ uik[39]*u[11]+ uik[40]*u[12]+ uik[41]*u[13]; 7560aa0ce06SHong Zhang wp[13]+=uik[42]*u[7]+ uik[43]*u[8]+ uik[44]*u[9]+ uik[45]*u[10]+ uik[46]*u[11]+ uik[47]*u[12]+ uik[48]*u[13]; 7570aa0ce06SHong Zhang 7580aa0ce06SHong Zhang wp[14]+= uik[0]*u[14] + uik[1]*u[15] + uik[2]*u[16] + uik[3]*u[17] + uik[4]*u[18] + uik[5]*u[19] + uik[6]*u[20]; 7590aa0ce06SHong Zhang wp[15]+= uik[7]*u[14] + uik[8]*u[15] + uik[9]*u[16]+ uik[10]*u[17]+ uik[11]*u[18]+ uik[12]*u[19]+ uik[13]*u[20]; 7600aa0ce06SHong Zhang wp[16]+= uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]; 7610aa0ce06SHong Zhang wp[17]+= uik[21]*u[14]+ uik[22]*u[15]+ uik[23]*u[16]+ uik[24]*u[17]+ uik[25]*u[18]+ uik[26]*u[19]+ uik[27]*u[20]; 7620aa0ce06SHong Zhang wp[18]+= uik[28]*u[14]+ uik[29]*u[15]+ uik[30]*u[16]+ uik[31]*u[17]+ uik[32]*u[18]+ uik[33]*u[19]+ uik[34]*u[20]; 7630aa0ce06SHong Zhang wp[19]+= uik[35]*u[14]+ uik[36]*u[15]+ uik[37]*u[16]+ uik[38]*u[17]+ uik[39]*u[18]+ uik[40]*u[19]+ uik[41]*u[20]; 7640aa0ce06SHong Zhang wp[20]+= uik[42]*u[14]+ uik[43]*u[15]+ uik[44]*u[16]+ uik[45]*u[17]+ uik[46]*u[18]+ uik[47]*u[19]+ uik[48]*u[20]; 7650aa0ce06SHong Zhang 7660aa0ce06SHong Zhang wp[21]+= uik[0]*u[21] + uik[1]*u[22] + uik[2]*u[23] + uik[3]*u[24] + uik[4]*u[25] + uik[5]*u[26] + uik[6]*u[27]; 7670aa0ce06SHong Zhang wp[22]+= uik[7]*u[21] + uik[8]*u[22] + uik[9]*u[23]+ uik[10]*u[24]+ uik[11]*u[25]+ uik[12]*u[26]+ uik[13]*u[27]; 7680aa0ce06SHong Zhang wp[23]+= uik[14]*u[21]+ uik[15]*u[22]+ uik[16]*u[23]+ uik[17]*u[24]+ uik[18]*u[25]+ uik[19]*u[26]+ uik[20]*u[27]; 7690aa0ce06SHong Zhang wp[24]+= uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]; 7700aa0ce06SHong Zhang wp[25]+= uik[28]*u[21]+ uik[29]*u[22]+ uik[30]*u[23]+ uik[31]*u[24]+ uik[32]*u[25]+ uik[33]*u[26]+ uik[34]*u[27]; 7710aa0ce06SHong Zhang wp[26]+= uik[35]*u[21]+ uik[36]*u[22]+ uik[37]*u[23]+ uik[38]*u[24]+ uik[39]*u[25]+ uik[40]*u[26]+ uik[41]*u[27]; 7720aa0ce06SHong Zhang wp[27]+= uik[42]*u[21]+ uik[43]*u[22]+ uik[44]*u[23]+ uik[45]*u[24]+ uik[46]*u[25]+ uik[47]*u[26]+ uik[48]*u[27]; 7730aa0ce06SHong Zhang 7740aa0ce06SHong Zhang wp[28]+= uik[0]*u[28] + uik[1]*u[29] + uik[2]*u[30] + uik[3]*u[31] + uik[4]*u[32] + uik[5]*u[33] + uik[6]*u[34]; 7750aa0ce06SHong Zhang wp[29]+= uik[7]*u[28] + uik[8]*u[29] + uik[9]*u[30]+ uik[10]*u[31]+ uik[11]*u[32]+ uik[12]*u[33]+ uik[13]*u[34]; 7760aa0ce06SHong Zhang wp[30]+= uik[14]*u[28]+ uik[15]*u[29]+ uik[16]*u[30]+ uik[17]*u[31]+ uik[18]*u[32]+ uik[19]*u[33]+ uik[20]*u[34]; 7770aa0ce06SHong Zhang wp[31]+= uik[21]*u[28]+ uik[22]*u[29]+ uik[23]*u[30]+ uik[24]*u[31]+ uik[25]*u[32]+ uik[26]*u[33]+ uik[27]*u[34]; 7780aa0ce06SHong Zhang wp[32]+= uik[28]*u[28]+ uik[29]*u[29]+ uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]; 7790aa0ce06SHong Zhang wp[33]+= uik[35]*u[28]+ uik[36]*u[29]+ uik[37]*u[30]+ uik[38]*u[31]+ uik[39]*u[32]+ uik[40]*u[33]+ uik[41]*u[34]; 7800aa0ce06SHong Zhang wp[34]+= uik[42]*u[28]+ uik[43]*u[29]+ uik[44]*u[30]+ uik[45]*u[31]+ uik[46]*u[32]+ uik[47]*u[33]+ uik[48]*u[34]; 7810aa0ce06SHong Zhang 7820aa0ce06SHong Zhang wp[35]+= uik[0]*u[35] + uik[1]*u[36] + uik[2]*u[37] + uik[3]*u[38] + uik[4]*u[39] + uik[5]*u[40] + uik[6]*u[41]; 7830aa0ce06SHong Zhang wp[36]+= uik[7]*u[35] + uik[8]*u[36] + uik[9]*u[37]+ uik[10]*u[38]+ uik[11]*u[39]+ uik[12]*u[40]+ uik[13]*u[41]; 7840aa0ce06SHong Zhang wp[37]+= uik[14]*u[35]+ uik[15]*u[36]+ uik[16]*u[37]+ uik[17]*u[38]+ uik[18]*u[39]+ uik[19]*u[40]+ uik[20]*u[41]; 7850aa0ce06SHong Zhang wp[38]+= uik[21]*u[35]+ uik[22]*u[36]+ uik[23]*u[37]+ uik[24]*u[38]+ uik[25]*u[39]+ uik[26]*u[40]+ uik[27]*u[41]; 7860aa0ce06SHong Zhang wp[39]+= uik[28]*u[35]+ uik[29]*u[36]+ uik[30]*u[37]+ uik[31]*u[38]+ uik[32]*u[39]+ uik[33]*u[40]+ uik[34]*u[41]; 7870aa0ce06SHong Zhang wp[40]+= uik[35]*u[35]+ uik[36]*u[36]+ uik[37]*u[37]+ uik[38]*u[38]+ uik[39]*u[39]+ uik[40]*u[40]+ uik[41]*u[41]; 7880aa0ce06SHong Zhang wp[41]+= uik[42]*u[35]+ uik[43]*u[36]+ uik[44]*u[37]+ uik[45]*u[38]+ uik[46]*u[39]+ uik[47]*u[40]+ uik[48]*u[41]; 7890aa0ce06SHong Zhang 7900aa0ce06SHong Zhang wp[42]+= uik[0]*u[42] + uik[1]*u[43] + uik[2]*u[44] + uik[3]*u[45] + uik[4]*u[46] + uik[5]*u[47] + uik[6]*u[48]; 7910aa0ce06SHong Zhang wp[43]+= uik[7]*u[42] + uik[8]*u[43] + uik[9]*u[44]+ uik[10]*u[45]+ uik[11]*u[46]+ uik[12]*u[47]+ uik[13]*u[48]; 7920aa0ce06SHong Zhang wp[44]+= uik[14]*u[42]+ uik[15]*u[43]+ uik[16]*u[44]+ uik[17]*u[45]+ uik[18]*u[46]+ uik[19]*u[47]+ uik[20]*u[48]; 7930aa0ce06SHong Zhang wp[45]+= uik[21]*u[42]+ uik[22]*u[43]+ uik[23]*u[44]+ uik[24]*u[45]+ uik[25]*u[46]+ uik[26]*u[47]+ uik[27]*u[48]; 7940aa0ce06SHong Zhang wp[46]+= uik[28]*u[42]+ uik[29]*u[43]+ uik[30]*u[44]+ uik[31]*u[45]+ uik[32]*u[46]+ uik[33]*u[47]+ uik[34]*u[48]; 7950aa0ce06SHong Zhang wp[47]+= uik[35]*u[42]+ uik[36]*u[43]+ uik[37]*u[44]+ uik[38]*u[45]+ uik[39]*u[46]+ uik[40]*u[47]+ uik[41]*u[48]; 7960aa0ce06SHong Zhang wp[48]+= uik[42]*u[42]+ uik[43]*u[43]+ uik[44]*u[44]+ uik[45]*u[45]+ uik[46]*u[46]+ uik[47]*u[47]+ uik[48]*u[48]; 7970aa0ce06SHong Zhang } 7980aa0ce06SHong Zhang 7990aa0ce06SHong Zhang /* ... add i to row list for next nonzero entry */ 8000aa0ce06SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 8010aa0ce06SHong Zhang j = bj[jmin]; 8020aa0ce06SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 8030aa0ce06SHong Zhang } 8040aa0ce06SHong Zhang i = nexti; 8050aa0ce06SHong Zhang } 8060aa0ce06SHong Zhang 8070aa0ce06SHong Zhang /* save nonzero entries in k-th row of U ... */ 8080aa0ce06SHong Zhang 80949b5e25fSSatish Balay /* invert diagonal block */ 8100aa0ce06SHong Zhang d = ba+k*49; 8110aa0ce06SHong Zhang ierr = PetscMemcpy(d,dk,49*sizeof(MatScalar));CHKERRQ(ierr); 8120aa0ce06SHong Zhang ierr = Kernel_A_gets_inverse_A_7(d);CHKERRQ(ierr); 8130aa0ce06SHong Zhang 8140aa0ce06SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 8150aa0ce06SHong Zhang if (jmin < jmax) { 8160aa0ce06SHong Zhang for (j=jmin; j<jmax; j++){ 8170aa0ce06SHong Zhang vj = bj[j]; /* block col. index of U */ 8180aa0ce06SHong Zhang u = ba + j*49; 8190aa0ce06SHong Zhang wp = w + vj*49; 8200aa0ce06SHong Zhang for (k1=0; k1<49; k1++){ 8210aa0ce06SHong Zhang *u++ = *wp; 8220aa0ce06SHong Zhang *wp++ = 0.0; 8230aa0ce06SHong Zhang } 82449b5e25fSSatish Balay } 82549b5e25fSSatish Balay 8260aa0ce06SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 8270aa0ce06SHong Zhang il[k] = jmin; 8280aa0ce06SHong Zhang i = bj[jmin]; 8290aa0ce06SHong Zhang jl[k] = jl[i]; jl[i] = k; 8300aa0ce06SHong Zhang } 8310aa0ce06SHong Zhang } 8320aa0ce06SHong Zhang 8330aa0ce06SHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 8340aa0ce06SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 8350aa0ce06SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 8360aa0ce06SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 8370aa0ce06SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 8380aa0ce06SHong Zhang if (a->permute){ 8390aa0ce06SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 8400aa0ce06SHong Zhang } 8410aa0ce06SHong Zhang 8420aa0ce06SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 8430aa0ce06SHong Zhang C->factor = FACTOR_CHOLESKY; 84449b5e25fSSatish Balay C->assembled = PETSC_TRUE; 8450aa0ce06SHong Zhang C->preallocated = PETSC_TRUE; 84649b5e25fSSatish Balay PLogFlops(1.3333*343*b->mbs); /* from inverting diagonal blocks */ 84749b5e25fSSatish Balay PetscFunctionReturn(0); 84849b5e25fSSatish Balay } 84949b5e25fSSatish Balay 85049b5e25fSSatish Balay /* 85149b5e25fSSatish Balay Version for when blocks are 7 by 7 Using natural ordering 85249b5e25fSSatish Balay */ 85349b5e25fSSatish Balay #undef __FUNC__ 8546f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering" 8556f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering(Mat A,Mat *B) 85649b5e25fSSatish Balay { 85749b5e25fSSatish Balay Mat C = *B; 858ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 859ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 86071149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 861ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 862ab27746eSHong Zhang MatScalar *u,*d,*w,*wp; 86349b5e25fSSatish Balay 86449b5e25fSSatish Balay PetscFunctionBegin; 865*671cb588SHong Zhang 866ab27746eSHong Zhang /* initialization */ 867ab27746eSHong Zhang w = (MatScalar*)PetscMalloc(49*mbs*sizeof(MatScalar));CHKPTRQ(w); 868ab27746eSHong Zhang ierr = PetscMemzero(w,49*mbs*sizeof(MatScalar));CHKERRQ(ierr); 869ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 870ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 871ab27746eSHong Zhang for (i=0; i<mbs; i++) { 872ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 87349b5e25fSSatish Balay } 874ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(dk); 875ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(uik); 876ab27746eSHong Zhang 877ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 878ab27746eSHong Zhang 879ab27746eSHong Zhang /* for each row k */ 880ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 881ab27746eSHong Zhang 882ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 883ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 884ab27746eSHong Zhang if (jmin < jmax) { 885ab27746eSHong Zhang ap = aa + jmin*49; 886ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 887ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 888ab27746eSHong Zhang wp = w + vj*49; 889ab27746eSHong Zhang for (i=0; i<49; i++) *wp++ = *ap++; 89049b5e25fSSatish Balay } 89149b5e25fSSatish Balay } 892ab27746eSHong Zhang 893ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 894ab27746eSHong Zhang ierr = PetscMemcpy(dk,w+k*49,49*sizeof(MatScalar));CHKERRQ(ierr); 895ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 896ab27746eSHong Zhang 897ab27746eSHong Zhang while (i < mbs){ 898ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 899ab27746eSHong Zhang 900ab27746eSHong Zhang /* compute multiplier */ 901ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 902ab27746eSHong Zhang 903ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 904ab27746eSHong Zhang d = ba + i*49; 905ab27746eSHong Zhang u = ba + ili*49; 906ab27746eSHong Zhang 907ab27746eSHong Zhang uik[0] = -(d[0]*u[0] + d[7]*u[1]+ d[14]*u[2]+ d[21]*u[3]+ d[28]*u[4]+ d[35]*u[5]+ d[42]*u[6]); 908ab27746eSHong Zhang uik[1] = -(d[1]*u[0] + d[8]*u[1]+ d[15]*u[2]+ d[22]*u[3]+ d[29]*u[4]+ d[36]*u[5]+ d[43]*u[6]); 909ab27746eSHong Zhang uik[2] = -(d[2]*u[0] + d[9]*u[1]+ d[16]*u[2]+ d[23]*u[3]+ d[30]*u[4]+ d[37]*u[5]+ d[44]*u[6]); 910ab27746eSHong Zhang uik[3] = -(d[3]*u[0]+ d[10]*u[1]+ d[17]*u[2]+ d[24]*u[3]+ d[31]*u[4]+ d[38]*u[5]+ d[45]*u[6]); 911ab27746eSHong Zhang uik[4] = -(d[4]*u[0]+ d[11]*u[1]+ d[18]*u[2]+ d[25]*u[3]+ d[32]*u[4]+ d[39]*u[5]+ d[46]*u[6]); 912ab27746eSHong Zhang uik[5] = -(d[5]*u[0]+ d[12]*u[1]+ d[19]*u[2]+ d[26]*u[3]+ d[33]*u[4]+ d[40]*u[5]+ d[47]*u[6]); 913ab27746eSHong Zhang uik[6] = -(d[6]*u[0]+ d[13]*u[1]+ d[20]*u[2]+ d[27]*u[3]+ d[34]*u[4]+ d[41]*u[5]+ d[48]*u[6]); 914ab27746eSHong Zhang 915ab27746eSHong Zhang uik[7] = -(d[0]*u[7] + d[7]*u[8]+ d[14]*u[9]+ d[21]*u[10]+ d[28]*u[11]+ d[35]*u[12]+ d[42]*u[13]); 916ab27746eSHong Zhang uik[8] = -(d[1]*u[7] + d[8]*u[8]+ d[15]*u[9]+ d[22]*u[10]+ d[29]*u[11]+ d[36]*u[12]+ d[43]*u[13]); 917ab27746eSHong Zhang uik[9] = -(d[2]*u[7] + d[9]*u[8]+ d[16]*u[9]+ d[23]*u[10]+ d[30]*u[11]+ d[37]*u[12]+ d[44]*u[13]); 918ab27746eSHong Zhang uik[10]= -(d[3]*u[7]+ d[10]*u[8]+ d[17]*u[9]+ d[24]*u[10]+ d[31]*u[11]+ d[38]*u[12]+ d[45]*u[13]); 919ab27746eSHong Zhang uik[11]= -(d[4]*u[7]+ d[11]*u[8]+ d[18]*u[9]+ d[25]*u[10]+ d[32]*u[11]+ d[39]*u[12]+ d[46]*u[13]); 920ab27746eSHong Zhang uik[12]= -(d[5]*u[7]+ d[12]*u[8]+ d[19]*u[9]+ d[26]*u[10]+ d[33]*u[11]+ d[40]*u[12]+ d[47]*u[13]); 921ab27746eSHong Zhang uik[13]= -(d[6]*u[7]+ d[13]*u[8]+ d[20]*u[9]+ d[27]*u[10]+ d[34]*u[11]+ d[41]*u[12]+ d[48]*u[13]); 922ab27746eSHong Zhang 923ab27746eSHong Zhang uik[14]= -(d[0]*u[14] + d[7]*u[15]+ d[14]*u[16]+ d[21]*u[17]+ d[28]*u[18]+ d[35]*u[19]+ d[42]*u[20]); 924ab27746eSHong Zhang uik[15]= -(d[1]*u[14] + d[8]*u[15]+ d[15]*u[16]+ d[22]*u[17]+ d[29]*u[18]+ d[36]*u[19]+ d[43]*u[20]); 925ab27746eSHong Zhang uik[16]= -(d[2]*u[14] + d[9]*u[15]+ d[16]*u[16]+ d[23]*u[17]+ d[30]*u[18]+ d[37]*u[19]+ d[44]*u[20]); 926ab27746eSHong Zhang uik[17]= -(d[3]*u[14]+ d[10]*u[15]+ d[17]*u[16]+ d[24]*u[17]+ d[31]*u[18]+ d[38]*u[19]+ d[45]*u[20]); 927ab27746eSHong Zhang uik[18]= -(d[4]*u[14]+ d[11]*u[15]+ d[18]*u[16]+ d[25]*u[17]+ d[32]*u[18]+ d[39]*u[19]+ d[46]*u[20]); 928ab27746eSHong Zhang uik[19]= -(d[5]*u[14]+ d[12]*u[15]+ d[19]*u[16]+ d[26]*u[17]+ d[33]*u[18]+ d[40]*u[19]+ d[47]*u[20]); 929ab27746eSHong Zhang uik[20]= -(d[6]*u[14]+ d[13]*u[15]+ d[20]*u[16]+ d[27]*u[17]+ d[34]*u[18]+ d[41]*u[19]+ d[48]*u[20]); 930ab27746eSHong Zhang 931ab27746eSHong Zhang uik[21]= -(d[0]*u[21] + d[7]*u[22]+ d[14]*u[23]+ d[21]*u[24]+ d[28]*u[25]+ d[35]*u[26]+ d[42]*u[27]); 932ab27746eSHong Zhang uik[22]= -(d[1]*u[21] + d[8]*u[22]+ d[15]*u[23]+ d[22]*u[24]+ d[29]*u[25]+ d[36]*u[26]+ d[43]*u[27]); 933ab27746eSHong Zhang uik[23]= -(d[2]*u[21] + d[9]*u[22]+ d[16]*u[23]+ d[23]*u[24]+ d[30]*u[25]+ d[37]*u[26]+ d[44]*u[27]); 934ab27746eSHong Zhang uik[24]= -(d[3]*u[21]+ d[10]*u[22]+ d[17]*u[23]+ d[24]*u[24]+ d[31]*u[25]+ d[38]*u[26]+ d[45]*u[27]); 935ab27746eSHong Zhang uik[25]= -(d[4]*u[21]+ d[11]*u[22]+ d[18]*u[23]+ d[25]*u[24]+ d[32]*u[25]+ d[39]*u[26]+ d[46]*u[27]); 936ab27746eSHong Zhang uik[26]= -(d[5]*u[21]+ d[12]*u[22]+ d[19]*u[23]+ d[26]*u[24]+ d[33]*u[25]+ d[40]*u[26]+ d[47]*u[27]); 937ab27746eSHong Zhang uik[27]= -(d[6]*u[21]+ d[13]*u[22]+ d[20]*u[23]+ d[27]*u[24]+ d[34]*u[25]+ d[41]*u[26]+ d[48]*u[27]); 938ab27746eSHong Zhang 939ab27746eSHong Zhang uik[28]= -(d[0]*u[28] + d[7]*u[29]+ d[14]*u[30]+ d[21]*u[31]+ d[28]*u[32]+ d[35]*u[33]+ d[42]*u[34]); 940ab27746eSHong Zhang uik[29]= -(d[1]*u[28] + d[8]*u[29]+ d[15]*u[30]+ d[22]*u[31]+ d[29]*u[32]+ d[36]*u[33]+ d[43]*u[34]); 941ab27746eSHong Zhang uik[30]= -(d[2]*u[28] + d[9]*u[29]+ d[16]*u[30]+ d[23]*u[31]+ d[30]*u[32]+ d[37]*u[33]+ d[44]*u[34]); 942ab27746eSHong Zhang uik[31]= -(d[3]*u[28]+ d[10]*u[29]+ d[17]*u[30]+ d[24]*u[31]+ d[31]*u[32]+ d[38]*u[33]+ d[45]*u[34]); 943ab27746eSHong Zhang uik[32]= -(d[4]*u[28]+ d[11]*u[29]+ d[18]*u[30]+ d[25]*u[31]+ d[32]*u[32]+ d[39]*u[33]+ d[46]*u[34]); 944ab27746eSHong Zhang uik[33]= -(d[5]*u[28]+ d[12]*u[29]+ d[19]*u[30]+ d[26]*u[31]+ d[33]*u[32]+ d[40]*u[33]+ d[47]*u[34]); 945ab27746eSHong Zhang uik[34]= -(d[6]*u[28]+ d[13]*u[29]+ d[20]*u[30]+ d[27]*u[31]+ d[34]*u[32]+ d[41]*u[33]+ d[48]*u[34]); 946ab27746eSHong Zhang 947ab27746eSHong Zhang uik[35]= -(d[0]*u[35] + d[7]*u[36]+ d[14]*u[37]+ d[21]*u[38]+ d[28]*u[39]+ d[35]*u[40]+ d[42]*u[41]); 948ab27746eSHong Zhang uik[36]= -(d[1]*u[35] + d[8]*u[36]+ d[15]*u[37]+ d[22]*u[38]+ d[29]*u[39]+ d[36]*u[40]+ d[43]*u[41]); 949ab27746eSHong Zhang uik[37]= -(d[2]*u[35] + d[9]*u[36]+ d[16]*u[37]+ d[23]*u[38]+ d[30]*u[39]+ d[37]*u[40]+ d[44]*u[41]); 950ab27746eSHong Zhang uik[38]= -(d[3]*u[35]+ d[10]*u[36]+ d[17]*u[37]+ d[24]*u[38]+ d[31]*u[39]+ d[38]*u[40]+ d[45]*u[41]); 951ab27746eSHong Zhang uik[39]= -(d[4]*u[35]+ d[11]*u[36]+ d[18]*u[37]+ d[25]*u[38]+ d[32]*u[39]+ d[39]*u[40]+ d[46]*u[41]); 952ab27746eSHong Zhang uik[40]= -(d[5]*u[35]+ d[12]*u[36]+ d[19]*u[37]+ d[26]*u[38]+ d[33]*u[39]+ d[40]*u[40]+ d[47]*u[41]); 953ab27746eSHong Zhang uik[41]= -(d[6]*u[35]+ d[13]*u[36]+ d[20]*u[37]+ d[27]*u[38]+ d[34]*u[39]+ d[41]*u[40]+ d[48]*u[41]); 954ab27746eSHong Zhang 955ab27746eSHong Zhang uik[42]= -(d[0]*u[42] + d[7]*u[43]+ d[14]*u[44]+ d[21]*u[45]+ d[28]*u[46]+ d[35]*u[47]+ d[42]*u[48]); 956ab27746eSHong Zhang uik[43]= -(d[1]*u[42] + d[8]*u[43]+ d[15]*u[44]+ d[22]*u[45]+ d[29]*u[46]+ d[36]*u[47]+ d[43]*u[48]); 957ab27746eSHong Zhang uik[44]= -(d[2]*u[42] + d[9]*u[43]+ d[16]*u[44]+ d[23]*u[45]+ d[30]*u[46]+ d[37]*u[47]+ d[44]*u[48]); 958ab27746eSHong Zhang uik[45]= -(d[3]*u[42]+ d[10]*u[43]+ d[17]*u[44]+ d[24]*u[45]+ d[31]*u[46]+ d[38]*u[47]+ d[45]*u[48]); 959ab27746eSHong Zhang uik[46]= -(d[4]*u[42]+ d[11]*u[43]+ d[18]*u[44]+ d[25]*u[45]+ d[32]*u[46]+ d[39]*u[47]+ d[46]*u[48]); 960ab27746eSHong Zhang uik[47]= -(d[5]*u[42]+ d[12]*u[43]+ d[19]*u[44]+ d[26]*u[45]+ d[33]*u[46]+ d[40]*u[47]+ d[47]*u[48]); 961ab27746eSHong Zhang uik[48]= -(d[6]*u[42]+ d[13]*u[43]+ d[20]*u[44]+ d[27]*u[45]+ d[34]*u[46]+ d[41]*u[47]+ d[48]*u[48]); 962ab27746eSHong Zhang 963ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 964ab27746eSHong Zhang dk[0]+= uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6]; 965ab27746eSHong Zhang dk[1]+= uik[7]*u[0] + uik[8]*u[1] + uik[9]*u[2]+ uik[10]*u[3]+ uik[11]*u[4]+ uik[12]*u[5]+ uik[13]*u[6]; 966ab27746eSHong Zhang dk[2]+= uik[14]*u[0]+ uik[15]*u[1]+ uik[16]*u[2]+ uik[17]*u[3]+ uik[18]*u[4]+ uik[19]*u[5]+ uik[20]*u[6]; 967ab27746eSHong Zhang dk[3]+= uik[21]*u[0]+ uik[22]*u[1]+ uik[23]*u[2]+ uik[24]*u[3]+ uik[25]*u[4]+ uik[26]*u[5]+ uik[27]*u[6]; 968ab27746eSHong Zhang dk[4]+= uik[28]*u[0]+ uik[29]*u[1]+ uik[30]*u[2]+ uik[31]*u[3]+ uik[32]*u[4]+ uik[33]*u[5]+ uik[34]*u[6]; 969ab27746eSHong Zhang dk[5]+= uik[35]*u[0]+ uik[36]*u[1]+ uik[37]*u[2]+ uik[38]*u[3]+ uik[39]*u[4]+ uik[40]*u[5]+ uik[41]*u[6]; 970ab27746eSHong Zhang dk[6]+= uik[42]*u[0]+ uik[43]*u[1]+ uik[44]*u[2]+ uik[45]*u[3]+ uik[46]*u[4]+ uik[47]*u[5]+ uik[48]*u[6]; 971ab27746eSHong Zhang 972ab27746eSHong Zhang dk[7]+= uik[0]*u[7] + uik[1]*u[8] + uik[2]*u[9] + uik[3]*u[10] + uik[4]*u[11] + uik[5]*u[12] + uik[6]*u[13]; 973ab27746eSHong Zhang dk[8]+= uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]; 974ab27746eSHong Zhang dk[9]+= uik[14]*u[7]+ uik[15]*u[8]+ uik[16]*u[9]+ uik[17]*u[10]+ uik[18]*u[11]+ uik[19]*u[12]+ uik[20]*u[13]; 975ab27746eSHong Zhang dk[10]+=uik[21]*u[7]+ uik[22]*u[8]+ uik[23]*u[9]+ uik[24]*u[10]+ uik[25]*u[11]+ uik[26]*u[12]+ uik[27]*u[13]; 976ab27746eSHong Zhang dk[11]+=uik[28]*u[7]+ uik[29]*u[8]+ uik[30]*u[9]+ uik[31]*u[10]+ uik[32]*u[11]+ uik[33]*u[12]+ uik[34]*u[13]; 977ab27746eSHong Zhang dk[12]+=uik[35]*u[7]+ uik[36]*u[8]+ uik[37]*u[9]+ uik[38]*u[10]+ uik[39]*u[11]+ uik[40]*u[12]+ uik[41]*u[13]; 978ab27746eSHong Zhang dk[13]+=uik[42]*u[7]+ uik[43]*u[8]+ uik[44]*u[9]+ uik[45]*u[10]+ uik[46]*u[11]+ uik[47]*u[12]+ uik[48]*u[13]; 979ab27746eSHong Zhang 980ab27746eSHong Zhang dk[14]+= uik[0]*u[14] + uik[1]*u[15] + uik[2]*u[16] + uik[3]*u[17] + uik[4]*u[18] + uik[5]*u[19] + uik[6]*u[20]; 981ab27746eSHong Zhang dk[15]+= uik[7]*u[14] + uik[8]*u[15] + uik[9]*u[16]+ uik[10]*u[17]+ uik[11]*u[18]+ uik[12]*u[19]+ uik[13]*u[20]; 982ab27746eSHong Zhang dk[16]+= uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]; 983ab27746eSHong Zhang dk[17]+= uik[21]*u[14]+ uik[22]*u[15]+ uik[23]*u[16]+ uik[24]*u[17]+ uik[25]*u[18]+ uik[26]*u[19]+ uik[27]*u[20]; 984ab27746eSHong Zhang dk[18]+= uik[28]*u[14]+ uik[29]*u[15]+ uik[30]*u[16]+ uik[31]*u[17]+ uik[32]*u[18]+ uik[33]*u[19]+ uik[34]*u[20]; 985ab27746eSHong Zhang dk[19]+= uik[35]*u[14]+ uik[36]*u[15]+ uik[37]*u[16]+ uik[38]*u[17]+ uik[39]*u[18]+ uik[40]*u[19]+ uik[41]*u[20]; 986ab27746eSHong Zhang dk[20]+= uik[42]*u[14]+ uik[43]*u[15]+ uik[44]*u[16]+ uik[45]*u[17]+ uik[46]*u[18]+ uik[47]*u[19]+ uik[48]*u[20]; 987ab27746eSHong Zhang 988ab27746eSHong Zhang dk[21]+= uik[0]*u[21] + uik[1]*u[22] + uik[2]*u[23] + uik[3]*u[24] + uik[4]*u[25] + uik[5]*u[26] + uik[6]*u[27]; 989ab27746eSHong Zhang dk[22]+= uik[7]*u[21] + uik[8]*u[22] + uik[9]*u[23]+ uik[10]*u[24]+ uik[11]*u[25]+ uik[12]*u[26]+ uik[13]*u[27]; 990ab27746eSHong Zhang dk[23]+= uik[14]*u[21]+ uik[15]*u[22]+ uik[16]*u[23]+ uik[17]*u[24]+ uik[18]*u[25]+ uik[19]*u[26]+ uik[20]*u[27]; 991ab27746eSHong Zhang dk[24]+= uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]; 992ab27746eSHong Zhang dk[25]+= uik[28]*u[21]+ uik[29]*u[22]+ uik[30]*u[23]+ uik[31]*u[24]+ uik[32]*u[25]+ uik[33]*u[26]+ uik[34]*u[27]; 993ab27746eSHong Zhang dk[26]+= uik[35]*u[21]+ uik[36]*u[22]+ uik[37]*u[23]+ uik[38]*u[24]+ uik[39]*u[25]+ uik[40]*u[26]+ uik[41]*u[27]; 994ab27746eSHong Zhang dk[27]+= uik[42]*u[21]+ uik[43]*u[22]+ uik[44]*u[23]+ uik[45]*u[24]+ uik[46]*u[25]+ uik[47]*u[26]+ uik[48]*u[27]; 995ab27746eSHong Zhang 996ab27746eSHong Zhang dk[28]+= uik[0]*u[28] + uik[1]*u[29] + uik[2]*u[30] + uik[3]*u[31] + uik[4]*u[32] + uik[5]*u[33] + uik[6]*u[34]; 997ab27746eSHong Zhang dk[29]+= uik[7]*u[28] + uik[8]*u[29] + uik[9]*u[30]+ uik[10]*u[31]+ uik[11]*u[32]+ uik[12]*u[33]+ uik[13]*u[34]; 998ab27746eSHong Zhang dk[30]+= uik[14]*u[28]+ uik[15]*u[29]+ uik[16]*u[30]+ uik[17]*u[31]+ uik[18]*u[32]+ uik[19]*u[33]+ uik[20]*u[34]; 999ab27746eSHong Zhang dk[31]+= uik[21]*u[28]+ uik[22]*u[29]+ uik[23]*u[30]+ uik[24]*u[31]+ uik[25]*u[32]+ uik[26]*u[33]+ uik[27]*u[34]; 1000ab27746eSHong Zhang dk[32]+= uik[28]*u[28]+ uik[29]*u[29]+ uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]; 1001ab27746eSHong Zhang dk[33]+= uik[35]*u[28]+ uik[36]*u[29]+ uik[37]*u[30]+ uik[38]*u[31]+ uik[39]*u[32]+ uik[40]*u[33]+ uik[41]*u[34]; 1002ab27746eSHong Zhang dk[34]+= uik[42]*u[28]+ uik[43]*u[29]+ uik[44]*u[30]+ uik[45]*u[31]+ uik[46]*u[32]+ uik[47]*u[33]+ uik[48]*u[34]; 1003ab27746eSHong Zhang 1004ab27746eSHong Zhang dk[35]+= uik[0]*u[35] + uik[1]*u[36] + uik[2]*u[37] + uik[3]*u[38] + uik[4]*u[39] + uik[5]*u[40] + uik[6]*u[41]; 1005ab27746eSHong Zhang dk[36]+= uik[7]*u[35] + uik[8]*u[36] + uik[9]*u[37]+ uik[10]*u[38]+ uik[11]*u[39]+ uik[12]*u[40]+ uik[13]*u[41]; 1006ab27746eSHong Zhang dk[37]+= uik[14]*u[35]+ uik[15]*u[36]+ uik[16]*u[37]+ uik[17]*u[38]+ uik[18]*u[39]+ uik[19]*u[40]+ uik[20]*u[41]; 1007ab27746eSHong Zhang dk[38]+= uik[21]*u[35]+ uik[22]*u[36]+ uik[23]*u[37]+ uik[24]*u[38]+ uik[25]*u[39]+ uik[26]*u[40]+ uik[27]*u[41]; 1008ab27746eSHong Zhang dk[39]+= uik[28]*u[35]+ uik[29]*u[36]+ uik[30]*u[37]+ uik[31]*u[38]+ uik[32]*u[39]+ uik[33]*u[40]+ uik[34]*u[41]; 1009ab27746eSHong Zhang dk[40]+= uik[35]*u[35]+ uik[36]*u[36]+ uik[37]*u[37]+ uik[38]*u[38]+ uik[39]*u[39]+ uik[40]*u[40]+ uik[41]*u[41]; 1010ab27746eSHong Zhang dk[41]+= uik[42]*u[35]+ uik[43]*u[36]+ uik[44]*u[37]+ uik[45]*u[38]+ uik[46]*u[39]+ uik[47]*u[40]+ uik[48]*u[41]; 1011ab27746eSHong Zhang 1012ab27746eSHong Zhang dk[42]+= uik[0]*u[42] + uik[1]*u[43] + uik[2]*u[44] + uik[3]*u[45] + uik[4]*u[46] + uik[5]*u[47] + uik[6]*u[48]; 1013ab27746eSHong Zhang dk[43]+= uik[7]*u[42] + uik[8]*u[43] + uik[9]*u[44]+ uik[10]*u[45]+ uik[11]*u[46]+ uik[12]*u[47]+ uik[13]*u[48]; 1014ab27746eSHong Zhang dk[44]+= uik[14]*u[42]+ uik[15]*u[43]+ uik[16]*u[44]+ uik[17]*u[45]+ uik[18]*u[46]+ uik[19]*u[47]+ uik[20]*u[48]; 1015ab27746eSHong Zhang dk[45]+= uik[21]*u[42]+ uik[22]*u[43]+ uik[23]*u[44]+ uik[24]*u[45]+ uik[25]*u[46]+ uik[26]*u[47]+ uik[27]*u[48]; 1016ab27746eSHong Zhang dk[46]+= uik[28]*u[42]+ uik[29]*u[43]+ uik[30]*u[44]+ uik[31]*u[45]+ uik[32]*u[46]+ uik[33]*u[47]+ uik[34]*u[48]; 1017ab27746eSHong Zhang dk[47]+= uik[35]*u[42]+ uik[36]*u[43]+ uik[37]*u[44]+ uik[38]*u[45]+ uik[39]*u[46]+ uik[40]*u[47]+ uik[41]*u[48]; 1018ab27746eSHong Zhang dk[48]+= uik[42]*u[42]+ uik[43]*u[43]+ uik[44]*u[44]+ uik[45]*u[45]+ uik[46]*u[46]+ uik[47]*u[47]+ uik[48]*u[48]; 1019ab27746eSHong Zhang 1020ab27746eSHong Zhang /* update -U(i,k) */ 1021ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*49,uik,49*sizeof(MatScalar));CHKERRQ(ierr); 1022ab27746eSHong Zhang 1023ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 1024ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1025ab27746eSHong Zhang if (jmin < jmax){ 1026ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 1027ab27746eSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 1028ab27746eSHong Zhang wp = w + bj[j]*49; 1029ab27746eSHong Zhang u = ba + j*49; 1030ab27746eSHong Zhang 1031ab27746eSHong Zhang wp[0]+= uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6]; 1032ab27746eSHong Zhang wp[1]+= uik[7]*u[0] + uik[8]*u[1] + uik[9]*u[2]+ uik[10]*u[3]+ uik[11]*u[4]+ uik[12]*u[5]+ uik[13]*u[6]; 1033ab27746eSHong Zhang wp[2]+= uik[14]*u[0]+ uik[15]*u[1]+ uik[16]*u[2]+ uik[17]*u[3]+ uik[18]*u[4]+ uik[19]*u[5]+ uik[20]*u[6]; 1034ab27746eSHong Zhang wp[3]+= uik[21]*u[0]+ uik[22]*u[1]+ uik[23]*u[2]+ uik[24]*u[3]+ uik[25]*u[4]+ uik[26]*u[5]+ uik[27]*u[6]; 1035ab27746eSHong Zhang wp[4]+= uik[28]*u[0]+ uik[29]*u[1]+ uik[30]*u[2]+ uik[31]*u[3]+ uik[32]*u[4]+ uik[33]*u[5]+ uik[34]*u[6]; 1036ab27746eSHong Zhang wp[5]+= uik[35]*u[0]+ uik[36]*u[1]+ uik[37]*u[2]+ uik[38]*u[3]+ uik[39]*u[4]+ uik[40]*u[5]+ uik[41]*u[6]; 1037ab27746eSHong Zhang wp[6]+= uik[42]*u[0]+ uik[43]*u[1]+ uik[44]*u[2]+ uik[45]*u[3]+ uik[46]*u[4]+ uik[47]*u[5]+ uik[48]*u[6]; 1038ab27746eSHong Zhang 1039ab27746eSHong Zhang wp[7]+= uik[0]*u[7] + uik[1]*u[8] + uik[2]*u[9] + uik[3]*u[10] + uik[4]*u[11] + uik[5]*u[12] + uik[6]*u[13]; 1040ab27746eSHong Zhang wp[8]+= uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]; 1041ab27746eSHong Zhang wp[9]+= uik[14]*u[7]+ uik[15]*u[8]+ uik[16]*u[9]+ uik[17]*u[10]+ uik[18]*u[11]+ uik[19]*u[12]+ uik[20]*u[13]; 1042ab27746eSHong Zhang wp[10]+=uik[21]*u[7]+ uik[22]*u[8]+ uik[23]*u[9]+ uik[24]*u[10]+ uik[25]*u[11]+ uik[26]*u[12]+ uik[27]*u[13]; 1043ab27746eSHong Zhang wp[11]+=uik[28]*u[7]+ uik[29]*u[8]+ uik[30]*u[9]+ uik[31]*u[10]+ uik[32]*u[11]+ uik[33]*u[12]+ uik[34]*u[13]; 1044ab27746eSHong Zhang wp[12]+=uik[35]*u[7]+ uik[36]*u[8]+ uik[37]*u[9]+ uik[38]*u[10]+ uik[39]*u[11]+ uik[40]*u[12]+ uik[41]*u[13]; 1045ab27746eSHong Zhang wp[13]+=uik[42]*u[7]+ uik[43]*u[8]+ uik[44]*u[9]+ uik[45]*u[10]+ uik[46]*u[11]+ uik[47]*u[12]+ uik[48]*u[13]; 1046ab27746eSHong Zhang 1047ab27746eSHong Zhang wp[14]+= uik[0]*u[14] + uik[1]*u[15] + uik[2]*u[16] + uik[3]*u[17] + uik[4]*u[18] + uik[5]*u[19] + uik[6]*u[20]; 1048ab27746eSHong Zhang wp[15]+= uik[7]*u[14] + uik[8]*u[15] + uik[9]*u[16]+ uik[10]*u[17]+ uik[11]*u[18]+ uik[12]*u[19]+ uik[13]*u[20]; 1049ab27746eSHong Zhang wp[16]+= uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]; 1050ab27746eSHong Zhang wp[17]+= uik[21]*u[14]+ uik[22]*u[15]+ uik[23]*u[16]+ uik[24]*u[17]+ uik[25]*u[18]+ uik[26]*u[19]+ uik[27]*u[20]; 1051ab27746eSHong Zhang wp[18]+= uik[28]*u[14]+ uik[29]*u[15]+ uik[30]*u[16]+ uik[31]*u[17]+ uik[32]*u[18]+ uik[33]*u[19]+ uik[34]*u[20]; 1052ab27746eSHong Zhang wp[19]+= uik[35]*u[14]+ uik[36]*u[15]+ uik[37]*u[16]+ uik[38]*u[17]+ uik[39]*u[18]+ uik[40]*u[19]+ uik[41]*u[20]; 1053ab27746eSHong Zhang wp[20]+= uik[42]*u[14]+ uik[43]*u[15]+ uik[44]*u[16]+ uik[45]*u[17]+ uik[46]*u[18]+ uik[47]*u[19]+ uik[48]*u[20]; 1054ab27746eSHong Zhang 1055ab27746eSHong Zhang wp[21]+= uik[0]*u[21] + uik[1]*u[22] + uik[2]*u[23] + uik[3]*u[24] + uik[4]*u[25] + uik[5]*u[26] + uik[6]*u[27]; 1056ab27746eSHong Zhang wp[22]+= uik[7]*u[21] + uik[8]*u[22] + uik[9]*u[23]+ uik[10]*u[24]+ uik[11]*u[25]+ uik[12]*u[26]+ uik[13]*u[27]; 1057ab27746eSHong Zhang wp[23]+= uik[14]*u[21]+ uik[15]*u[22]+ uik[16]*u[23]+ uik[17]*u[24]+ uik[18]*u[25]+ uik[19]*u[26]+ uik[20]*u[27]; 1058ab27746eSHong Zhang wp[24]+= uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]; 1059ab27746eSHong Zhang wp[25]+= uik[28]*u[21]+ uik[29]*u[22]+ uik[30]*u[23]+ uik[31]*u[24]+ uik[32]*u[25]+ uik[33]*u[26]+ uik[34]*u[27]; 1060ab27746eSHong Zhang wp[26]+= uik[35]*u[21]+ uik[36]*u[22]+ uik[37]*u[23]+ uik[38]*u[24]+ uik[39]*u[25]+ uik[40]*u[26]+ uik[41]*u[27]; 1061ab27746eSHong Zhang wp[27]+= uik[42]*u[21]+ uik[43]*u[22]+ uik[44]*u[23]+ uik[45]*u[24]+ uik[46]*u[25]+ uik[47]*u[26]+ uik[48]*u[27]; 1062ab27746eSHong Zhang 1063ab27746eSHong Zhang wp[28]+= uik[0]*u[28] + uik[1]*u[29] + uik[2]*u[30] + uik[3]*u[31] + uik[4]*u[32] + uik[5]*u[33] + uik[6]*u[34]; 1064ab27746eSHong Zhang wp[29]+= uik[7]*u[28] + uik[8]*u[29] + uik[9]*u[30]+ uik[10]*u[31]+ uik[11]*u[32]+ uik[12]*u[33]+ uik[13]*u[34]; 1065ab27746eSHong Zhang wp[30]+= uik[14]*u[28]+ uik[15]*u[29]+ uik[16]*u[30]+ uik[17]*u[31]+ uik[18]*u[32]+ uik[19]*u[33]+ uik[20]*u[34]; 1066ab27746eSHong Zhang wp[31]+= uik[21]*u[28]+ uik[22]*u[29]+ uik[23]*u[30]+ uik[24]*u[31]+ uik[25]*u[32]+ uik[26]*u[33]+ uik[27]*u[34]; 1067ab27746eSHong Zhang wp[32]+= uik[28]*u[28]+ uik[29]*u[29]+ uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]; 1068ab27746eSHong Zhang wp[33]+= uik[35]*u[28]+ uik[36]*u[29]+ uik[37]*u[30]+ uik[38]*u[31]+ uik[39]*u[32]+ uik[40]*u[33]+ uik[41]*u[34]; 1069ab27746eSHong Zhang wp[34]+= uik[42]*u[28]+ uik[43]*u[29]+ uik[44]*u[30]+ uik[45]*u[31]+ uik[46]*u[32]+ uik[47]*u[33]+ uik[48]*u[34]; 1070ab27746eSHong Zhang 1071ab27746eSHong Zhang wp[35]+= uik[0]*u[35] + uik[1]*u[36] + uik[2]*u[37] + uik[3]*u[38] + uik[4]*u[39] + uik[5]*u[40] + uik[6]*u[41]; 1072ab27746eSHong Zhang wp[36]+= uik[7]*u[35] + uik[8]*u[36] + uik[9]*u[37]+ uik[10]*u[38]+ uik[11]*u[39]+ uik[12]*u[40]+ uik[13]*u[41]; 1073ab27746eSHong Zhang wp[37]+= uik[14]*u[35]+ uik[15]*u[36]+ uik[16]*u[37]+ uik[17]*u[38]+ uik[18]*u[39]+ uik[19]*u[40]+ uik[20]*u[41]; 1074ab27746eSHong Zhang wp[38]+= uik[21]*u[35]+ uik[22]*u[36]+ uik[23]*u[37]+ uik[24]*u[38]+ uik[25]*u[39]+ uik[26]*u[40]+ uik[27]*u[41]; 1075ab27746eSHong Zhang wp[39]+= uik[28]*u[35]+ uik[29]*u[36]+ uik[30]*u[37]+ uik[31]*u[38]+ uik[32]*u[39]+ uik[33]*u[40]+ uik[34]*u[41]; 1076ab27746eSHong Zhang wp[40]+= uik[35]*u[35]+ uik[36]*u[36]+ uik[37]*u[37]+ uik[38]*u[38]+ uik[39]*u[39]+ uik[40]*u[40]+ uik[41]*u[41]; 1077ab27746eSHong Zhang wp[41]+= uik[42]*u[35]+ uik[43]*u[36]+ uik[44]*u[37]+ uik[45]*u[38]+ uik[46]*u[39]+ uik[47]*u[40]+ uik[48]*u[41]; 1078ab27746eSHong Zhang 1079ab27746eSHong Zhang wp[42]+= uik[0]*u[42] + uik[1]*u[43] + uik[2]*u[44] + uik[3]*u[45] + uik[4]*u[46] + uik[5]*u[47] + uik[6]*u[48]; 1080ab27746eSHong Zhang wp[43]+= uik[7]*u[42] + uik[8]*u[43] + uik[9]*u[44]+ uik[10]*u[45]+ uik[11]*u[46]+ uik[12]*u[47]+ uik[13]*u[48]; 1081ab27746eSHong Zhang wp[44]+= uik[14]*u[42]+ uik[15]*u[43]+ uik[16]*u[44]+ uik[17]*u[45]+ uik[18]*u[46]+ uik[19]*u[47]+ uik[20]*u[48]; 1082ab27746eSHong Zhang wp[45]+= uik[21]*u[42]+ uik[22]*u[43]+ uik[23]*u[44]+ uik[24]*u[45]+ uik[25]*u[46]+ uik[26]*u[47]+ uik[27]*u[48]; 1083ab27746eSHong Zhang wp[46]+= uik[28]*u[42]+ uik[29]*u[43]+ uik[30]*u[44]+ uik[31]*u[45]+ uik[32]*u[46]+ uik[33]*u[47]+ uik[34]*u[48]; 1084ab27746eSHong Zhang wp[47]+= uik[35]*u[42]+ uik[36]*u[43]+ uik[37]*u[44]+ uik[38]*u[45]+ uik[39]*u[46]+ uik[40]*u[47]+ uik[41]*u[48]; 1085ab27746eSHong Zhang wp[48]+= uik[42]*u[42]+ uik[43]*u[43]+ uik[44]*u[44]+ uik[45]*u[45]+ uik[46]*u[46]+ uik[47]*u[47]+ uik[48]*u[48]; 108649b5e25fSSatish Balay } 1087ab27746eSHong Zhang 1088ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 1089ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1090ab27746eSHong Zhang j = bj[jmin]; 1091ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 109249b5e25fSSatish Balay } 1093ab27746eSHong Zhang i = nexti; 109449b5e25fSSatish Balay } 1095ab27746eSHong Zhang 1096ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 1097ab27746eSHong Zhang 109849b5e25fSSatish Balay /* invert diagonal block */ 1099ab27746eSHong Zhang d = ba+k*49; 1100ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,49*sizeof(MatScalar));CHKERRQ(ierr); 1101ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_7(d);CHKERRQ(ierr); 1102ab27746eSHong Zhang 1103ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1104ab27746eSHong Zhang if (jmin < jmax) { 1105ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 1106ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 1107ab27746eSHong Zhang u = ba + j*49; 1108ab27746eSHong Zhang wp = w + vj*49; 1109ab27746eSHong Zhang for (k1=0; k1<49; k1++){ 1110ab27746eSHong Zhang *u++ = *wp; 1111ab27746eSHong Zhang *wp++ = 0.0; 1112ab27746eSHong Zhang } 111349b5e25fSSatish Balay } 111449b5e25fSSatish Balay 1115ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1116ab27746eSHong Zhang il[k] = jmin; 1117ab27746eSHong Zhang i = bj[jmin]; 1118ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 1119ab27746eSHong Zhang } 1120ab27746eSHong Zhang } 1121ab27746eSHong Zhang 1122ab27746eSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 1123ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1124ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 1125ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1126ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 1127ab27746eSHong Zhang 1128ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 112949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 1130ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 113149b5e25fSSatish Balay PLogFlops(1.3333*343*b->mbs); /* from inverting diagonal blocks */ 113249b5e25fSSatish Balay PetscFunctionReturn(0); 113349b5e25fSSatish Balay } 113449b5e25fSSatish Balay 11359b0e81eeSHong Zhang /* Version for when blocks are 6 by 6 */ 113649b5e25fSSatish Balay #undef __FUNC__ 11376f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6" 11386f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_6(Mat A,Mat *B) 113949b5e25fSSatish Balay { 114049b5e25fSSatish Balay Mat C = *B; 11419b0e81eeSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 11429b0e81eeSHong Zhang IS perm = b->row; 11439b0e81eeSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 11449b0e81eeSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 11459b0e81eeSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 11469b0e81eeSHong Zhang MatScalar *u,*d,*w,*wp; 114749b5e25fSSatish Balay 114849b5e25fSSatish Balay PetscFunctionBegin; 11499b0e81eeSHong Zhang /* initialization */ 11509b0e81eeSHong Zhang w = (MatScalar*)PetscMalloc(36*mbs*sizeof(MatScalar));CHKPTRQ(w); 11519b0e81eeSHong Zhang ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 11529b0e81eeSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 11539b0e81eeSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 11549b0e81eeSHong Zhang for (i=0; i<mbs; i++) { 11559b0e81eeSHong Zhang jl[i] = mbs; il[0] = 0; 115649b5e25fSSatish Balay } 11579b0e81eeSHong Zhang dk = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(dk); 11589b0e81eeSHong Zhang uik = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(uik); 11599b0e81eeSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 11609b0e81eeSHong Zhang 11619b0e81eeSHong Zhang /* check permutation */ 11629b0e81eeSHong Zhang if (!a->permute){ 11639b0e81eeSHong Zhang ai = a->i; aj = a->j; aa = a->a; 11649b0e81eeSHong Zhang } else { 11659b0e81eeSHong Zhang ai = a->inew; aj = a->jnew; 11669b0e81eeSHong Zhang aa = (MatScalar*)PetscMalloc(36*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 11679b0e81eeSHong Zhang ierr = PetscMemcpy(aa,a->a,36*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 11689b0e81eeSHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 11699b0e81eeSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 11709b0e81eeSHong Zhang 11719b0e81eeSHong Zhang for (i=0; i<mbs; i++){ 11729b0e81eeSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 11739b0e81eeSHong Zhang for (j=jmin; j<jmax; j++){ 11749b0e81eeSHong Zhang while (a2anew[j] != j){ 11759b0e81eeSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 11769b0e81eeSHong Zhang for (k1=0; k1<36; k1++){ 11779b0e81eeSHong Zhang dk[k1] = aa[k*36+k1]; 11789b0e81eeSHong Zhang aa[k*36+k1] = aa[j*36+k1]; 11799b0e81eeSHong Zhang aa[j*36+k1] = dk[k1]; 118049b5e25fSSatish Balay } 118149b5e25fSSatish Balay } 11829b0e81eeSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 11839b0e81eeSHong Zhang if (i > aj[j]){ 11849b0e81eeSHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 11859b0e81eeSHong Zhang ap = aa + j*36; /* ptr to the beginning of j-th block of aa */ 11869b0e81eeSHong Zhang for (k=0; k<36; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 11879b0e81eeSHong Zhang for (k=0; k<6; k++){ /* j-th block of aa <- dk^T */ 11889b0e81eeSHong Zhang for (k1=0; k1<6; k1++) *ap++ = dk[k + 6*k1]; 118949b5e25fSSatish Balay } 119049b5e25fSSatish Balay } 119149b5e25fSSatish Balay } 11929b0e81eeSHong Zhang } 11939b0e81eeSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 11949b0e81eeSHong Zhang } 11959b0e81eeSHong Zhang 11969b0e81eeSHong Zhang /* for each row k */ 11979b0e81eeSHong Zhang for (k = 0; k<mbs; k++){ 11989b0e81eeSHong Zhang 11999b0e81eeSHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 12009b0e81eeSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 12019b0e81eeSHong Zhang if (jmin < jmax) { 12029b0e81eeSHong Zhang ap = aa + jmin*36; 12039b0e81eeSHong Zhang for (j = jmin; j < jmax; j++){ 12049b0e81eeSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 12059b0e81eeSHong Zhang wp = w + vj*36; 12069b0e81eeSHong Zhang for (i=0; i<36; i++) *wp++ = *ap++; 12079b0e81eeSHong Zhang } 12089b0e81eeSHong Zhang } 12099b0e81eeSHong Zhang 12109b0e81eeSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 12119b0e81eeSHong Zhang ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 12129b0e81eeSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 12139b0e81eeSHong Zhang 12149b0e81eeSHong Zhang while (i < mbs){ 12159b0e81eeSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 12169b0e81eeSHong Zhang 12179b0e81eeSHong Zhang /* compute multiplier */ 12189b0e81eeSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 12199b0e81eeSHong Zhang 12209b0e81eeSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 12219b0e81eeSHong Zhang d = ba + i*36; 12229b0e81eeSHong Zhang u = ba + ili*36; 12239b0e81eeSHong Zhang 12249b0e81eeSHong Zhang uik[0] = -(d[0]*u[0] + d[6]*u[1] + d[12]*u[2] + d[18]*u[3] + d[24]*u[4] + d[30]*u[5]); 12259b0e81eeSHong Zhang uik[1] = -(d[1]*u[0] + d[7]*u[1] + d[13]*u[2] + d[19]*u[3] + d[25]*u[4] + d[31]*u[5]); 12269b0e81eeSHong Zhang uik[2] = -(d[2]*u[0] + d[8]*u[1] + d[14]*u[2] + d[20]*u[3] + d[26]*u[4] + d[32]*u[5]); 12279b0e81eeSHong Zhang uik[3] = -(d[3]*u[0] + d[9]*u[1] + d[15]*u[2] + d[21]*u[3] + d[27]*u[4] + d[33]*u[5]); 12289b0e81eeSHong Zhang uik[4] = -(d[4]*u[0]+ d[10]*u[1] + d[16]*u[2] + d[22]*u[3] + d[28]*u[4] + d[34]*u[5]); 12299b0e81eeSHong Zhang uik[5] = -(d[5]*u[0]+ d[11]*u[1] + d[17]*u[2] + d[23]*u[3] + d[29]*u[4] + d[35]*u[5]); 12309b0e81eeSHong Zhang 12319b0e81eeSHong Zhang uik[6] = -(d[0]*u[6] + d[6]*u[7] + d[12]*u[8] + d[18]*u[9] + d[24]*u[10] + d[30]*u[11]); 12329b0e81eeSHong Zhang uik[7] = -(d[1]*u[6] + d[7]*u[7] + d[13]*u[8] + d[19]*u[9] + d[25]*u[10] + d[31]*u[11]); 12339b0e81eeSHong Zhang uik[8] = -(d[2]*u[6] + d[8]*u[7] + d[14]*u[8] + d[20]*u[9] + d[26]*u[10] + d[32]*u[11]); 12349b0e81eeSHong Zhang uik[9] = -(d[3]*u[6] + d[9]*u[7] + d[15]*u[8] + d[21]*u[9] + d[27]*u[10] + d[33]*u[11]); 12359b0e81eeSHong Zhang uik[10]= -(d[4]*u[6]+ d[10]*u[7] + d[16]*u[8] + d[22]*u[9] + d[28]*u[10] + d[34]*u[11]); 12369b0e81eeSHong Zhang uik[11]= -(d[5]*u[6]+ d[11]*u[7] + d[17]*u[8] + d[23]*u[9] + d[29]*u[10] + d[35]*u[11]); 12379b0e81eeSHong Zhang 12389b0e81eeSHong Zhang uik[12] = -(d[0]*u[12] + d[6]*u[13] + d[12]*u[14] + d[18]*u[15] + d[24]*u[16] + d[30]*u[17]); 12399b0e81eeSHong Zhang uik[13] = -(d[1]*u[12] + d[7]*u[13] + d[13]*u[14] + d[19]*u[15] + d[25]*u[16] + d[31]*u[17]); 12409b0e81eeSHong Zhang uik[14] = -(d[2]*u[12] + d[8]*u[13] + d[14]*u[14] + d[20]*u[15] + d[26]*u[16] + d[32]*u[17]); 12419b0e81eeSHong Zhang uik[15] = -(d[3]*u[12] + d[9]*u[13] + d[15]*u[14] + d[21]*u[15] + d[27]*u[16] + d[33]*u[17]); 12429b0e81eeSHong Zhang uik[16] = -(d[4]*u[12]+ d[10]*u[13] + d[16]*u[14] + d[22]*u[15] + d[28]*u[16] + d[34]*u[17]); 12439b0e81eeSHong Zhang uik[17] = -(d[5]*u[12]+ d[11]*u[13] + d[17]*u[14] + d[23]*u[15] + d[29]*u[16] + d[35]*u[17]); 12449b0e81eeSHong Zhang 12459b0e81eeSHong Zhang uik[18] = -(d[0]*u[18] + d[6]*u[19] + d[12]*u[20] + d[18]*u[21] + d[24]*u[22] + d[30]*u[23]); 12469b0e81eeSHong Zhang uik[19] = -(d[1]*u[18] + d[7]*u[19] + d[13]*u[20] + d[19]*u[21] + d[25]*u[22] + d[31]*u[23]); 12479b0e81eeSHong Zhang uik[20] = -(d[2]*u[18] + d[8]*u[19] + d[14]*u[20] + d[20]*u[21] + d[26]*u[22] + d[32]*u[23]); 12489b0e81eeSHong Zhang uik[21] = -(d[3]*u[18] + d[9]*u[19] + d[15]*u[20] + d[21]*u[21] + d[27]*u[22] + d[33]*u[23]); 12499b0e81eeSHong Zhang uik[22] = -(d[4]*u[18]+ d[10]*u[19] + d[16]*u[20] + d[22]*u[21] + d[28]*u[22] + d[34]*u[23]); 12509b0e81eeSHong Zhang uik[23] = -(d[5]*u[18]+ d[11]*u[19] + d[17]*u[20] + d[23]*u[21] + d[29]*u[22] + d[35]*u[23]); 12519b0e81eeSHong Zhang 12529b0e81eeSHong Zhang uik[24] = -(d[0]*u[24] + d[6]*u[25] + d[12]*u[26] + d[18]*u[27] + d[24]*u[28] + d[30]*u[29]); 12539b0e81eeSHong Zhang uik[25] = -(d[1]*u[24] + d[7]*u[25] + d[13]*u[26] + d[19]*u[27] + d[25]*u[28] + d[31]*u[29]); 12549b0e81eeSHong Zhang uik[26] = -(d[2]*u[24] + d[8]*u[25] + d[14]*u[26] + d[20]*u[27] + d[26]*u[28] + d[32]*u[29]); 12559b0e81eeSHong Zhang uik[27] = -(d[3]*u[24] + d[9]*u[25] + d[15]*u[26] + d[21]*u[27] + d[27]*u[28] + d[33]*u[29]); 12569b0e81eeSHong Zhang uik[28] = -(d[4]*u[24]+ d[10]*u[25] + d[16]*u[26] + d[22]*u[27] + d[28]*u[28] + d[34]*u[29]); 12579b0e81eeSHong Zhang uik[29] = -(d[5]*u[24]+ d[11]*u[25] + d[17]*u[26] + d[23]*u[27] + d[29]*u[28] + d[35]*u[29]); 12589b0e81eeSHong Zhang 12599b0e81eeSHong Zhang uik[30] = -(d[0]*u[30] + d[6]*u[31] + d[12]*u[32] + d[18]*u[33] + d[24]*u[34] + d[30]*u[35]); 12609b0e81eeSHong Zhang uik[31] = -(d[1]*u[30] + d[7]*u[31] + d[13]*u[32] + d[19]*u[33] + d[25]*u[34] + d[31]*u[35]); 12619b0e81eeSHong Zhang uik[32] = -(d[2]*u[30] + d[8]*u[31] + d[14]*u[32] + d[20]*u[33] + d[26]*u[34] + d[32]*u[35]); 12629b0e81eeSHong Zhang uik[33] = -(d[3]*u[30] + d[9]*u[31] + d[15]*u[32] + d[21]*u[33] + d[27]*u[34] + d[33]*u[35]); 12639b0e81eeSHong Zhang uik[34] = -(d[4]*u[30]+ d[10]*u[31] + d[16]*u[32] + d[22]*u[33] + d[28]*u[34] + d[34]*u[35]); 12649b0e81eeSHong Zhang uik[35] = -(d[5]*u[30]+ d[11]*u[31] + d[17]*u[32] + d[23]*u[33] + d[29]*u[34] + d[35]*u[35]); 12659b0e81eeSHong Zhang 12669b0e81eeSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 12679b0e81eeSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 12689b0e81eeSHong Zhang dk[1] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2] + uik[9]*u[3]+ uik[10]*u[4]+ uik[11]*u[5]; 12699b0e81eeSHong Zhang dk[2] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]+ uik[16]*u[4]+ uik[17]*u[5]; 12709b0e81eeSHong Zhang dk[3] += uik[18]*u[0]+ uik[19]*u[1]+ uik[20]*u[2]+ uik[21]*u[3]+ uik[22]*u[4]+ uik[23]*u[5]; 12719b0e81eeSHong Zhang dk[4] += uik[24]*u[0]+ uik[25]*u[1]+ uik[26]*u[2]+ uik[27]*u[3]+ uik[28]*u[4]+ uik[29]*u[5]; 12729b0e81eeSHong Zhang dk[5] += uik[30]*u[0]+ uik[31]*u[1]+ uik[32]*u[2]+ uik[33]*u[3]+ uik[34]*u[4]+ uik[35]*u[5]; 12739b0e81eeSHong Zhang 12749b0e81eeSHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8] + uik[3]*u[9] + uik[4]*u[10] + uik[5]*u[11]; 12759b0e81eeSHong Zhang dk[7] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]; 12769b0e81eeSHong Zhang dk[8] += uik[12]*u[6]+ uik[13]*u[7]+ uik[14]*u[8]+ uik[15]*u[9]+ uik[16]*u[10]+ uik[17]*u[11]; 12779b0e81eeSHong Zhang dk[9] += uik[18]*u[6]+ uik[19]*u[7]+ uik[20]*u[8]+ uik[21]*u[9]+ uik[22]*u[10]+ uik[23]*u[11]; 12789b0e81eeSHong Zhang dk[10]+= uik[24]*u[6]+ uik[25]*u[7]+ uik[26]*u[8]+ uik[27]*u[9]+ uik[28]*u[10]+ uik[29]*u[11]; 12799b0e81eeSHong Zhang dk[11]+= uik[30]*u[6]+ uik[31]*u[7]+ uik[32]*u[8]+ uik[33]*u[9]+ uik[34]*u[10]+ uik[35]*u[11]; 12809b0e81eeSHong Zhang 12819b0e81eeSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15] + uik[4]*u[16] + uik[5]*u[17]; 12829b0e81eeSHong Zhang dk[13]+= uik[6]*u[12] + uik[7]*u[13] + uik[8]*u[14] + uik[9]*u[15]+ uik[10]*u[16]+ uik[11]*u[17]; 12839b0e81eeSHong Zhang dk[14]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]; 12849b0e81eeSHong Zhang dk[15]+= uik[18]*u[12]+ uik[19]*u[13]+ uik[20]*u[14]+ uik[21]*u[15]+ uik[22]*u[16]+ uik[23]*u[17]; 12859b0e81eeSHong Zhang dk[16]+= uik[24]*u[12]+ uik[25]*u[13]+ uik[26]*u[14]+ uik[27]*u[15]+ uik[28]*u[16]+ uik[29]*u[17]; 12869b0e81eeSHong Zhang dk[17]+= uik[30]*u[12]+ uik[31]*u[13]+ uik[32]*u[14]+ uik[33]*u[15]+ uik[34]*u[16]+ uik[35]*u[17]; 12879b0e81eeSHong Zhang 12889b0e81eeSHong Zhang dk[18]+= uik[0]*u[18] + uik[1]*u[19] + uik[2]*u[20] + uik[3]*u[21] + uik[4]*u[22] + uik[5]*u[23]; 12899b0e81eeSHong Zhang dk[19]+= uik[6]*u[18] + uik[7]*u[19] + uik[8]*u[20] + uik[9]*u[21]+ uik[10]*u[22]+ uik[11]*u[23]; 12909b0e81eeSHong Zhang dk[20]+= uik[12]*u[18]+ uik[13]*u[19]+ uik[14]*u[20]+ uik[15]*u[21]+ uik[16]*u[22]+ uik[17]*u[23]; 12919b0e81eeSHong Zhang dk[21]+= uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]; 12929b0e81eeSHong Zhang dk[22]+= uik[24]*u[18]+ uik[25]*u[19]+ uik[26]*u[20]+ uik[27]*u[21]+ uik[28]*u[22]+ uik[29]*u[23]; 12939b0e81eeSHong Zhang dk[23]+= uik[30]*u[18]+ uik[31]*u[19]+ uik[32]*u[20]+ uik[33]*u[21]+ uik[34]*u[22]+ uik[35]*u[23]; 12949b0e81eeSHong Zhang 12959b0e81eeSHong Zhang dk[24]+= uik[0]*u[24] + uik[1]*u[25] + uik[2]*u[26] + uik[3]*u[27] + uik[4]*u[28] + uik[5]*u[29]; 12969b0e81eeSHong Zhang dk[25]+= uik[6]*u[24] + uik[7]*u[25] + uik[8]*u[26] + uik[9]*u[27]+ uik[10]*u[28]+ uik[11]*u[29]; 12979b0e81eeSHong Zhang dk[26]+= uik[12]*u[24]+ uik[13]*u[25]+ uik[14]*u[26]+ uik[15]*u[27]+ uik[16]*u[28]+ uik[17]*u[29]; 12989b0e81eeSHong Zhang dk[27]+= uik[18]*u[24]+ uik[19]*u[25]+ uik[20]*u[26]+ uik[21]*u[27]+ uik[22]*u[28]+ uik[23]*u[29]; 12999b0e81eeSHong Zhang dk[28]+= uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]+ uik[28]*u[28]+ uik[29]*u[29]; 13009b0e81eeSHong Zhang dk[29]+= uik[30]*u[24]+ uik[31]*u[25]+ uik[32]*u[26]+ uik[33]*u[27]+ uik[34]*u[28]+ uik[35]*u[29]; 13019b0e81eeSHong Zhang 13029b0e81eeSHong Zhang dk[30]+= uik[0]*u[30] + uik[1]*u[31] + uik[2]*u[32] + uik[3]*u[33] + uik[4]*u[34] + uik[5]*u[35]; 13039b0e81eeSHong Zhang dk[31]+= uik[6]*u[30] + uik[7]*u[31] + uik[8]*u[32] + uik[9]*u[33]+ uik[10]*u[34]+ uik[11]*u[35]; 13049b0e81eeSHong Zhang dk[32]+= uik[12]*u[30]+ uik[13]*u[31]+ uik[14]*u[32]+ uik[15]*u[33]+ uik[16]*u[34]+ uik[17]*u[35]; 13059b0e81eeSHong Zhang dk[33]+= uik[18]*u[30]+ uik[19]*u[31]+ uik[20]*u[32]+ uik[21]*u[33]+ uik[22]*u[34]+ uik[23]*u[35]; 13069b0e81eeSHong Zhang dk[34]+= uik[24]*u[30]+ uik[25]*u[31]+ uik[26]*u[32]+ uik[27]*u[33]+ uik[28]*u[34]+ uik[29]*u[35]; 13079b0e81eeSHong Zhang dk[35]+= uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]+ uik[35]*u[35]; 13089b0e81eeSHong Zhang 13099b0e81eeSHong Zhang /* update -U(i,k) */ 13109b0e81eeSHong Zhang ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 13119b0e81eeSHong Zhang 13129b0e81eeSHong Zhang /* add multiple of row i to k-th row ... */ 13139b0e81eeSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 13149b0e81eeSHong Zhang if (jmin < jmax){ 13159b0e81eeSHong Zhang for (j=jmin; j<jmax; j++) { 13169b0e81eeSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 13179b0e81eeSHong Zhang wp = w + bj[j]*36; 13189b0e81eeSHong Zhang u = ba + j*36; 13199b0e81eeSHong Zhang wp[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 13209b0e81eeSHong Zhang wp[1] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2] + uik[9]*u[3]+ uik[10]*u[4]+ uik[11]*u[5]; 13219b0e81eeSHong Zhang wp[2] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]+ uik[16]*u[4]+ uik[17]*u[5]; 13229b0e81eeSHong Zhang wp[3] += uik[18]*u[0]+ uik[19]*u[1]+ uik[20]*u[2]+ uik[21]*u[3]+ uik[22]*u[4]+ uik[23]*u[5]; 13239b0e81eeSHong Zhang wp[4] += uik[24]*u[0]+ uik[25]*u[1]+ uik[26]*u[2]+ uik[27]*u[3]+ uik[28]*u[4]+ uik[29]*u[5]; 13249b0e81eeSHong Zhang wp[5] += uik[30]*u[0]+ uik[31]*u[1]+ uik[32]*u[2]+ uik[33]*u[3]+ uik[34]*u[4]+ uik[35]*u[5]; 13259b0e81eeSHong Zhang 13269b0e81eeSHong Zhang wp[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8] + uik[3]*u[9] + uik[4]*u[10] + uik[5]*u[11]; 13279b0e81eeSHong Zhang wp[7] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]; 13289b0e81eeSHong Zhang wp[8] += uik[12]*u[6]+ uik[13]*u[7]+ uik[14]*u[8]+ uik[15]*u[9]+ uik[16]*u[10]+ uik[17]*u[11]; 13299b0e81eeSHong Zhang wp[9] += uik[18]*u[6]+ uik[19]*u[7]+ uik[20]*u[8]+ uik[21]*u[9]+ uik[22]*u[10]+ uik[23]*u[11]; 13309b0e81eeSHong Zhang wp[10]+= uik[24]*u[6]+ uik[25]*u[7]+ uik[26]*u[8]+ uik[27]*u[9]+ uik[28]*u[10]+ uik[29]*u[11]; 13319b0e81eeSHong Zhang wp[11]+= uik[30]*u[6]+ uik[31]*u[7]+ uik[32]*u[8]+ uik[33]*u[9]+ uik[34]*u[10]+ uik[35]*u[11]; 13329b0e81eeSHong Zhang 13339b0e81eeSHong Zhang wp[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15] + uik[4]*u[16] + uik[5]*u[17]; 13349b0e81eeSHong Zhang wp[13]+= uik[6]*u[12] + uik[7]*u[13] + uik[8]*u[14] + uik[9]*u[15]+ uik[10]*u[16]+ uik[11]*u[17]; 13359b0e81eeSHong Zhang wp[14]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]; 13369b0e81eeSHong Zhang wp[15]+= uik[18]*u[12]+ uik[19]*u[13]+ uik[20]*u[14]+ uik[21]*u[15]+ uik[22]*u[16]+ uik[23]*u[17]; 13379b0e81eeSHong Zhang wp[16]+= uik[24]*u[12]+ uik[25]*u[13]+ uik[26]*u[14]+ uik[27]*u[15]+ uik[28]*u[16]+ uik[29]*u[17]; 13389b0e81eeSHong Zhang wp[17]+= uik[30]*u[12]+ uik[31]*u[13]+ uik[32]*u[14]+ uik[33]*u[15]+ uik[34]*u[16]+ uik[35]*u[17]; 13399b0e81eeSHong Zhang 13409b0e81eeSHong Zhang wp[18]+= uik[0]*u[18] + uik[1]*u[19] + uik[2]*u[20] + uik[3]*u[21] + uik[4]*u[22] + uik[5]*u[23]; 13419b0e81eeSHong Zhang wp[19]+= uik[6]*u[18] + uik[7]*u[19] + uik[8]*u[20] + uik[9]*u[21]+ uik[10]*u[22]+ uik[11]*u[23]; 13429b0e81eeSHong Zhang wp[20]+= uik[12]*u[18]+ uik[13]*u[19]+ uik[14]*u[20]+ uik[15]*u[21]+ uik[16]*u[22]+ uik[17]*u[23]; 13439b0e81eeSHong Zhang wp[21]+= uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]; 13449b0e81eeSHong Zhang wp[22]+= uik[24]*u[18]+ uik[25]*u[19]+ uik[26]*u[20]+ uik[27]*u[21]+ uik[28]*u[22]+ uik[29]*u[23]; 13459b0e81eeSHong Zhang wp[23]+= uik[30]*u[18]+ uik[31]*u[19]+ uik[32]*u[20]+ uik[33]*u[21]+ uik[34]*u[22]+ uik[35]*u[23]; 13469b0e81eeSHong Zhang 13479b0e81eeSHong Zhang wp[24]+= uik[0]*u[24] + uik[1]*u[25] + uik[2]*u[26] + uik[3]*u[27] + uik[4]*u[28] + uik[5]*u[29]; 13489b0e81eeSHong Zhang wp[25]+= uik[6]*u[24] + uik[7]*u[25] + uik[8]*u[26] + uik[9]*u[27]+ uik[10]*u[28]+ uik[11]*u[29]; 13499b0e81eeSHong Zhang wp[26]+= uik[12]*u[24]+ uik[13]*u[25]+ uik[14]*u[26]+ uik[15]*u[27]+ uik[16]*u[28]+ uik[17]*u[29]; 13509b0e81eeSHong Zhang wp[27]+= uik[18]*u[24]+ uik[19]*u[25]+ uik[20]*u[26]+ uik[21]*u[27]+ uik[22]*u[28]+ uik[23]*u[29]; 13519b0e81eeSHong Zhang wp[28]+= uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]+ uik[28]*u[28]+ uik[29]*u[29]; 13529b0e81eeSHong Zhang wp[29]+= uik[30]*u[24]+ uik[31]*u[25]+ uik[32]*u[26]+ uik[33]*u[27]+ uik[34]*u[28]+ uik[35]*u[29]; 13539b0e81eeSHong Zhang 13549b0e81eeSHong Zhang wp[30]+= uik[0]*u[30] + uik[1]*u[31] + uik[2]*u[32] + uik[3]*u[33] + uik[4]*u[34] + uik[5]*u[35]; 13559b0e81eeSHong Zhang wp[31]+= uik[6]*u[30] + uik[7]*u[31] + uik[8]*u[32] + uik[9]*u[33]+ uik[10]*u[34]+ uik[11]*u[35]; 13569b0e81eeSHong Zhang wp[32]+= uik[12]*u[30]+ uik[13]*u[31]+ uik[14]*u[32]+ uik[15]*u[33]+ uik[16]*u[34]+ uik[17]*u[35]; 13579b0e81eeSHong Zhang wp[33]+= uik[18]*u[30]+ uik[19]*u[31]+ uik[20]*u[32]+ uik[21]*u[33]+ uik[22]*u[34]+ uik[23]*u[35]; 13589b0e81eeSHong Zhang wp[34]+= uik[24]*u[30]+ uik[25]*u[31]+ uik[26]*u[32]+ uik[27]*u[33]+ uik[28]*u[34]+ uik[29]*u[35]; 13599b0e81eeSHong Zhang wp[35]+= uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]+ uik[35]*u[35]; 13609b0e81eeSHong Zhang } 13619b0e81eeSHong Zhang 13629b0e81eeSHong Zhang /* ... add i to row list for next nonzero entry */ 13639b0e81eeSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 13649b0e81eeSHong Zhang j = bj[jmin]; 13659b0e81eeSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 13669b0e81eeSHong Zhang } 13679b0e81eeSHong Zhang i = nexti; 13689b0e81eeSHong Zhang } 13699b0e81eeSHong Zhang 13709b0e81eeSHong Zhang /* save nonzero entries in k-th row of U ... */ 13719b0e81eeSHong Zhang 137249b5e25fSSatish Balay /* invert diagonal block */ 13739b0e81eeSHong Zhang d = ba+k*36; 13749b0e81eeSHong Zhang ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 13759b0e81eeSHong Zhang ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 13769b0e81eeSHong Zhang 13779b0e81eeSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 13789b0e81eeSHong Zhang if (jmin < jmax) { 13799b0e81eeSHong Zhang for (j=jmin; j<jmax; j++){ 13809b0e81eeSHong Zhang vj = bj[j]; /* block col. index of U */ 13819b0e81eeSHong Zhang u = ba + j*36; 13829b0e81eeSHong Zhang wp = w + vj*36; 13839b0e81eeSHong Zhang for (k1=0; k1<36; k1++){ 13849b0e81eeSHong Zhang *u++ = *wp; 13859b0e81eeSHong Zhang *wp++ = 0.0; 13869b0e81eeSHong Zhang } 138749b5e25fSSatish Balay } 138849b5e25fSSatish Balay 13899b0e81eeSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 13909b0e81eeSHong Zhang il[k] = jmin; 13919b0e81eeSHong Zhang i = bj[jmin]; 13929b0e81eeSHong Zhang jl[k] = jl[i]; jl[i] = k; 13939b0e81eeSHong Zhang } 13949b0e81eeSHong Zhang } 13959b0e81eeSHong Zhang 13969b0e81eeSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 13979b0e81eeSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 13989b0e81eeSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 13999b0e81eeSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 14009b0e81eeSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 14019b0e81eeSHong Zhang if (a->permute){ 14029b0e81eeSHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 14039b0e81eeSHong Zhang } 14049b0e81eeSHong Zhang 14059b0e81eeSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 14069b0e81eeSHong Zhang C->factor = FACTOR_CHOLESKY; 140749b5e25fSSatish Balay C->assembled = PETSC_TRUE; 14089b0e81eeSHong Zhang C->preallocated = PETSC_TRUE; 140949b5e25fSSatish Balay PLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 141049b5e25fSSatish Balay PetscFunctionReturn(0); 141149b5e25fSSatish Balay } 14129b0e81eeSHong Zhang 141349b5e25fSSatish Balay /* 141449b5e25fSSatish Balay Version for when blocks are 6 by 6 Using natural ordering 141549b5e25fSSatish Balay */ 141649b5e25fSSatish Balay #undef __FUNC__ 14176f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering" 14186f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering(Mat A,Mat *B) 141949b5e25fSSatish Balay { 142049b5e25fSSatish Balay Mat C = *B; 1421ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1422ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 142371149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1424ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1425ab27746eSHong Zhang MatScalar *u,*d,*w,*wp; 142649b5e25fSSatish Balay 142749b5e25fSSatish Balay PetscFunctionBegin; 1428*671cb588SHong Zhang 1429ab27746eSHong Zhang /* initialization */ 1430ab27746eSHong Zhang w = (MatScalar*)PetscMalloc(36*mbs*sizeof(MatScalar));CHKPTRQ(w); 1431ab27746eSHong Zhang ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1432ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 1433ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 1434ab27746eSHong Zhang for (i=0; i<mbs; i++) { 1435ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 143649b5e25fSSatish Balay } 1437ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(dk); 1438ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(uik); 1439ab27746eSHong Zhang 1440ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 1441ab27746eSHong Zhang 1442ab27746eSHong Zhang /* for each row k */ 1443ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 1444ab27746eSHong Zhang 1445ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 1446ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 1447ab27746eSHong Zhang if (jmin < jmax) { 1448ab27746eSHong Zhang ap = aa + jmin*36; 1449ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 1450ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 1451ab27746eSHong Zhang wp = w + vj*36; 1452ab27746eSHong Zhang for (i=0; i<36; i++) *wp++ = *ap++; 145349b5e25fSSatish Balay } 145449b5e25fSSatish Balay } 1455ab27746eSHong Zhang 1456ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1457ab27746eSHong Zhang ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 1458ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1459ab27746eSHong Zhang 1460ab27746eSHong Zhang while (i < mbs){ 1461ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1462ab27746eSHong Zhang 1463ab27746eSHong Zhang /* compute multiplier */ 1464ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1465ab27746eSHong Zhang 1466ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1467ab27746eSHong Zhang d = ba + i*36; 1468ab27746eSHong Zhang u = ba + ili*36; 1469ab27746eSHong Zhang 1470ab27746eSHong Zhang uik[0] = -(d[0]*u[0] + d[6]*u[1] + d[12]*u[2] + d[18]*u[3] + d[24]*u[4] + d[30]*u[5]); 1471ab27746eSHong Zhang uik[1] = -(d[1]*u[0] + d[7]*u[1] + d[13]*u[2] + d[19]*u[3] + d[25]*u[4] + d[31]*u[5]); 1472ab27746eSHong Zhang uik[2] = -(d[2]*u[0] + d[8]*u[1] + d[14]*u[2] + d[20]*u[3] + d[26]*u[4] + d[32]*u[5]); 1473ab27746eSHong Zhang uik[3] = -(d[3]*u[0] + d[9]*u[1] + d[15]*u[2] + d[21]*u[3] + d[27]*u[4] + d[33]*u[5]); 1474ab27746eSHong Zhang uik[4] = -(d[4]*u[0]+ d[10]*u[1] + d[16]*u[2] + d[22]*u[3] + d[28]*u[4] + d[34]*u[5]); 1475ab27746eSHong Zhang uik[5] = -(d[5]*u[0]+ d[11]*u[1] + d[17]*u[2] + d[23]*u[3] + d[29]*u[4] + d[35]*u[5]); 1476ab27746eSHong Zhang 1477ab27746eSHong Zhang uik[6] = -(d[0]*u[6] + d[6]*u[7] + d[12]*u[8] + d[18]*u[9] + d[24]*u[10] + d[30]*u[11]); 1478ab27746eSHong Zhang uik[7] = -(d[1]*u[6] + d[7]*u[7] + d[13]*u[8] + d[19]*u[9] + d[25]*u[10] + d[31]*u[11]); 1479ab27746eSHong Zhang uik[8] = -(d[2]*u[6] + d[8]*u[7] + d[14]*u[8] + d[20]*u[9] + d[26]*u[10] + d[32]*u[11]); 1480ab27746eSHong Zhang uik[9] = -(d[3]*u[6] + d[9]*u[7] + d[15]*u[8] + d[21]*u[9] + d[27]*u[10] + d[33]*u[11]); 1481ab27746eSHong Zhang uik[10]= -(d[4]*u[6]+ d[10]*u[7] + d[16]*u[8] + d[22]*u[9] + d[28]*u[10] + d[34]*u[11]); 1482ab27746eSHong Zhang uik[11]= -(d[5]*u[6]+ d[11]*u[7] + d[17]*u[8] + d[23]*u[9] + d[29]*u[10] + d[35]*u[11]); 1483ab27746eSHong Zhang 1484ab27746eSHong Zhang uik[12] = -(d[0]*u[12] + d[6]*u[13] + d[12]*u[14] + d[18]*u[15] + d[24]*u[16] + d[30]*u[17]); 1485ab27746eSHong Zhang uik[13] = -(d[1]*u[12] + d[7]*u[13] + d[13]*u[14] + d[19]*u[15] + d[25]*u[16] + d[31]*u[17]); 1486ab27746eSHong Zhang uik[14] = -(d[2]*u[12] + d[8]*u[13] + d[14]*u[14] + d[20]*u[15] + d[26]*u[16] + d[32]*u[17]); 1487ab27746eSHong Zhang uik[15] = -(d[3]*u[12] + d[9]*u[13] + d[15]*u[14] + d[21]*u[15] + d[27]*u[16] + d[33]*u[17]); 1488ab27746eSHong Zhang uik[16] = -(d[4]*u[12]+ d[10]*u[13] + d[16]*u[14] + d[22]*u[15] + d[28]*u[16] + d[34]*u[17]); 1489ab27746eSHong Zhang uik[17] = -(d[5]*u[12]+ d[11]*u[13] + d[17]*u[14] + d[23]*u[15] + d[29]*u[16] + d[35]*u[17]); 1490ab27746eSHong Zhang 1491ab27746eSHong Zhang uik[18] = -(d[0]*u[18] + d[6]*u[19] + d[12]*u[20] + d[18]*u[21] + d[24]*u[22] + d[30]*u[23]); 1492ab27746eSHong Zhang uik[19] = -(d[1]*u[18] + d[7]*u[19] + d[13]*u[20] + d[19]*u[21] + d[25]*u[22] + d[31]*u[23]); 1493ab27746eSHong Zhang uik[20] = -(d[2]*u[18] + d[8]*u[19] + d[14]*u[20] + d[20]*u[21] + d[26]*u[22] + d[32]*u[23]); 1494ab27746eSHong Zhang uik[21] = -(d[3]*u[18] + d[9]*u[19] + d[15]*u[20] + d[21]*u[21] + d[27]*u[22] + d[33]*u[23]); 1495ab27746eSHong Zhang uik[22] = -(d[4]*u[18]+ d[10]*u[19] + d[16]*u[20] + d[22]*u[21] + d[28]*u[22] + d[34]*u[23]); 1496ab27746eSHong Zhang uik[23] = -(d[5]*u[18]+ d[11]*u[19] + d[17]*u[20] + d[23]*u[21] + d[29]*u[22] + d[35]*u[23]); 1497ab27746eSHong Zhang 1498ab27746eSHong Zhang uik[24] = -(d[0]*u[24] + d[6]*u[25] + d[12]*u[26] + d[18]*u[27] + d[24]*u[28] + d[30]*u[29]); 1499ab27746eSHong Zhang uik[25] = -(d[1]*u[24] + d[7]*u[25] + d[13]*u[26] + d[19]*u[27] + d[25]*u[28] + d[31]*u[29]); 1500ab27746eSHong Zhang uik[26] = -(d[2]*u[24] + d[8]*u[25] + d[14]*u[26] + d[20]*u[27] + d[26]*u[28] + d[32]*u[29]); 1501ab27746eSHong Zhang uik[27] = -(d[3]*u[24] + d[9]*u[25] + d[15]*u[26] + d[21]*u[27] + d[27]*u[28] + d[33]*u[29]); 1502ab27746eSHong Zhang uik[28] = -(d[4]*u[24]+ d[10]*u[25] + d[16]*u[26] + d[22]*u[27] + d[28]*u[28] + d[34]*u[29]); 1503ab27746eSHong Zhang uik[29] = -(d[5]*u[24]+ d[11]*u[25] + d[17]*u[26] + d[23]*u[27] + d[29]*u[28] + d[35]*u[29]); 1504ab27746eSHong Zhang 1505ab27746eSHong Zhang uik[30] = -(d[0]*u[30] + d[6]*u[31] + d[12]*u[32] + d[18]*u[33] + d[24]*u[34] + d[30]*u[35]); 1506ab27746eSHong Zhang uik[31] = -(d[1]*u[30] + d[7]*u[31] + d[13]*u[32] + d[19]*u[33] + d[25]*u[34] + d[31]*u[35]); 1507ab27746eSHong Zhang uik[32] = -(d[2]*u[30] + d[8]*u[31] + d[14]*u[32] + d[20]*u[33] + d[26]*u[34] + d[32]*u[35]); 1508ab27746eSHong Zhang uik[33] = -(d[3]*u[30] + d[9]*u[31] + d[15]*u[32] + d[21]*u[33] + d[27]*u[34] + d[33]*u[35]); 1509ab27746eSHong Zhang uik[34] = -(d[4]*u[30]+ d[10]*u[31] + d[16]*u[32] + d[22]*u[33] + d[28]*u[34] + d[34]*u[35]); 1510ab27746eSHong Zhang uik[35] = -(d[5]*u[30]+ d[11]*u[31] + d[17]*u[32] + d[23]*u[33] + d[29]*u[34] + d[35]*u[35]); 1511ab27746eSHong Zhang 1512ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1513ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 1514ab27746eSHong Zhang dk[1] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2] + uik[9]*u[3]+ uik[10]*u[4]+ uik[11]*u[5]; 1515ab27746eSHong Zhang dk[2] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]+ uik[16]*u[4]+ uik[17]*u[5]; 1516ab27746eSHong Zhang dk[3] += uik[18]*u[0]+ uik[19]*u[1]+ uik[20]*u[2]+ uik[21]*u[3]+ uik[22]*u[4]+ uik[23]*u[5]; 1517ab27746eSHong Zhang dk[4] += uik[24]*u[0]+ uik[25]*u[1]+ uik[26]*u[2]+ uik[27]*u[3]+ uik[28]*u[4]+ uik[29]*u[5]; 1518ab27746eSHong Zhang dk[5] += uik[30]*u[0]+ uik[31]*u[1]+ uik[32]*u[2]+ uik[33]*u[3]+ uik[34]*u[4]+ uik[35]*u[5]; 1519ab27746eSHong Zhang 1520ab27746eSHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8] + uik[3]*u[9] + uik[4]*u[10] + uik[5]*u[11]; 1521ab27746eSHong Zhang dk[7] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]; 1522ab27746eSHong Zhang dk[8] += uik[12]*u[6]+ uik[13]*u[7]+ uik[14]*u[8]+ uik[15]*u[9]+ uik[16]*u[10]+ uik[17]*u[11]; 1523ab27746eSHong Zhang dk[9] += uik[18]*u[6]+ uik[19]*u[7]+ uik[20]*u[8]+ uik[21]*u[9]+ uik[22]*u[10]+ uik[23]*u[11]; 1524ab27746eSHong Zhang dk[10]+= uik[24]*u[6]+ uik[25]*u[7]+ uik[26]*u[8]+ uik[27]*u[9]+ uik[28]*u[10]+ uik[29]*u[11]; 1525ab27746eSHong Zhang dk[11]+= uik[30]*u[6]+ uik[31]*u[7]+ uik[32]*u[8]+ uik[33]*u[9]+ uik[34]*u[10]+ uik[35]*u[11]; 1526ab27746eSHong Zhang 1527ab27746eSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15] + uik[4]*u[16] + uik[5]*u[17]; 1528ab27746eSHong Zhang dk[13]+= uik[6]*u[12] + uik[7]*u[13] + uik[8]*u[14] + uik[9]*u[15]+ uik[10]*u[16]+ uik[11]*u[17]; 1529ab27746eSHong Zhang dk[14]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]; 1530ab27746eSHong Zhang dk[15]+= uik[18]*u[12]+ uik[19]*u[13]+ uik[20]*u[14]+ uik[21]*u[15]+ uik[22]*u[16]+ uik[23]*u[17]; 1531ab27746eSHong Zhang dk[16]+= uik[24]*u[12]+ uik[25]*u[13]+ uik[26]*u[14]+ uik[27]*u[15]+ uik[28]*u[16]+ uik[29]*u[17]; 1532ab27746eSHong Zhang dk[17]+= uik[30]*u[12]+ uik[31]*u[13]+ uik[32]*u[14]+ uik[33]*u[15]+ uik[34]*u[16]+ uik[35]*u[17]; 1533ab27746eSHong Zhang 1534ab27746eSHong Zhang dk[18]+= uik[0]*u[18] + uik[1]*u[19] + uik[2]*u[20] + uik[3]*u[21] + uik[4]*u[22] + uik[5]*u[23]; 1535ab27746eSHong Zhang dk[19]+= uik[6]*u[18] + uik[7]*u[19] + uik[8]*u[20] + uik[9]*u[21]+ uik[10]*u[22]+ uik[11]*u[23]; 1536ab27746eSHong Zhang dk[20]+= uik[12]*u[18]+ uik[13]*u[19]+ uik[14]*u[20]+ uik[15]*u[21]+ uik[16]*u[22]+ uik[17]*u[23]; 1537ab27746eSHong Zhang dk[21]+= uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]; 1538ab27746eSHong Zhang dk[22]+= uik[24]*u[18]+ uik[25]*u[19]+ uik[26]*u[20]+ uik[27]*u[21]+ uik[28]*u[22]+ uik[29]*u[23]; 1539ab27746eSHong Zhang dk[23]+= uik[30]*u[18]+ uik[31]*u[19]+ uik[32]*u[20]+ uik[33]*u[21]+ uik[34]*u[22]+ uik[35]*u[23]; 1540ab27746eSHong Zhang 1541ab27746eSHong Zhang dk[24]+= uik[0]*u[24] + uik[1]*u[25] + uik[2]*u[26] + uik[3]*u[27] + uik[4]*u[28] + uik[5]*u[29]; 1542ab27746eSHong Zhang dk[25]+= uik[6]*u[24] + uik[7]*u[25] + uik[8]*u[26] + uik[9]*u[27]+ uik[10]*u[28]+ uik[11]*u[29]; 1543ab27746eSHong Zhang dk[26]+= uik[12]*u[24]+ uik[13]*u[25]+ uik[14]*u[26]+ uik[15]*u[27]+ uik[16]*u[28]+ uik[17]*u[29]; 1544ab27746eSHong Zhang dk[27]+= uik[18]*u[24]+ uik[19]*u[25]+ uik[20]*u[26]+ uik[21]*u[27]+ uik[22]*u[28]+ uik[23]*u[29]; 1545ab27746eSHong Zhang dk[28]+= uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]+ uik[28]*u[28]+ uik[29]*u[29]; 1546ab27746eSHong Zhang dk[29]+= uik[30]*u[24]+ uik[31]*u[25]+ uik[32]*u[26]+ uik[33]*u[27]+ uik[34]*u[28]+ uik[35]*u[29]; 1547ab27746eSHong Zhang 1548ab27746eSHong Zhang dk[30]+= uik[0]*u[30] + uik[1]*u[31] + uik[2]*u[32] + uik[3]*u[33] + uik[4]*u[34] + uik[5]*u[35]; 1549ab27746eSHong Zhang dk[31]+= uik[6]*u[30] + uik[7]*u[31] + uik[8]*u[32] + uik[9]*u[33]+ uik[10]*u[34]+ uik[11]*u[35]; 1550ab27746eSHong Zhang dk[32]+= uik[12]*u[30]+ uik[13]*u[31]+ uik[14]*u[32]+ uik[15]*u[33]+ uik[16]*u[34]+ uik[17]*u[35]; 1551ab27746eSHong Zhang dk[33]+= uik[18]*u[30]+ uik[19]*u[31]+ uik[20]*u[32]+ uik[21]*u[33]+ uik[22]*u[34]+ uik[23]*u[35]; 1552ab27746eSHong Zhang dk[34]+= uik[24]*u[30]+ uik[25]*u[31]+ uik[26]*u[32]+ uik[27]*u[33]+ uik[28]*u[34]+ uik[29]*u[35]; 1553ab27746eSHong Zhang dk[35]+= uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]+ uik[35]*u[35]; 1554ab27746eSHong Zhang 1555ab27746eSHong Zhang /* update -U(i,k) */ 1556ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 1557ab27746eSHong Zhang 1558ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 1559ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1560ab27746eSHong Zhang if (jmin < jmax){ 1561ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 1562ab27746eSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 1563ab27746eSHong Zhang wp = w + bj[j]*36; 1564ab27746eSHong Zhang u = ba + j*36; 1565ab27746eSHong Zhang wp[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 1566ab27746eSHong Zhang wp[1] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2] + uik[9]*u[3]+ uik[10]*u[4]+ uik[11]*u[5]; 1567ab27746eSHong Zhang wp[2] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]+ uik[16]*u[4]+ uik[17]*u[5]; 1568ab27746eSHong Zhang wp[3] += uik[18]*u[0]+ uik[19]*u[1]+ uik[20]*u[2]+ uik[21]*u[3]+ uik[22]*u[4]+ uik[23]*u[5]; 1569ab27746eSHong Zhang wp[4] += uik[24]*u[0]+ uik[25]*u[1]+ uik[26]*u[2]+ uik[27]*u[3]+ uik[28]*u[4]+ uik[29]*u[5]; 1570ab27746eSHong Zhang wp[5] += uik[30]*u[0]+ uik[31]*u[1]+ uik[32]*u[2]+ uik[33]*u[3]+ uik[34]*u[4]+ uik[35]*u[5]; 1571ab27746eSHong Zhang 1572ab27746eSHong Zhang wp[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8] + uik[3]*u[9] + uik[4]*u[10] + uik[5]*u[11]; 1573ab27746eSHong Zhang wp[7] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]; 1574ab27746eSHong Zhang wp[8] += uik[12]*u[6]+ uik[13]*u[7]+ uik[14]*u[8]+ uik[15]*u[9]+ uik[16]*u[10]+ uik[17]*u[11]; 1575ab27746eSHong Zhang wp[9] += uik[18]*u[6]+ uik[19]*u[7]+ uik[20]*u[8]+ uik[21]*u[9]+ uik[22]*u[10]+ uik[23]*u[11]; 1576ab27746eSHong Zhang wp[10]+= uik[24]*u[6]+ uik[25]*u[7]+ uik[26]*u[8]+ uik[27]*u[9]+ uik[28]*u[10]+ uik[29]*u[11]; 1577ab27746eSHong Zhang wp[11]+= uik[30]*u[6]+ uik[31]*u[7]+ uik[32]*u[8]+ uik[33]*u[9]+ uik[34]*u[10]+ uik[35]*u[11]; 1578ab27746eSHong Zhang 1579ab27746eSHong Zhang wp[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15] + uik[4]*u[16] + uik[5]*u[17]; 1580ab27746eSHong Zhang wp[13]+= uik[6]*u[12] + uik[7]*u[13] + uik[8]*u[14] + uik[9]*u[15]+ uik[10]*u[16]+ uik[11]*u[17]; 1581ab27746eSHong Zhang wp[14]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]; 1582ab27746eSHong Zhang wp[15]+= uik[18]*u[12]+ uik[19]*u[13]+ uik[20]*u[14]+ uik[21]*u[15]+ uik[22]*u[16]+ uik[23]*u[17]; 1583ab27746eSHong Zhang wp[16]+= uik[24]*u[12]+ uik[25]*u[13]+ uik[26]*u[14]+ uik[27]*u[15]+ uik[28]*u[16]+ uik[29]*u[17]; 1584ab27746eSHong Zhang wp[17]+= uik[30]*u[12]+ uik[31]*u[13]+ uik[32]*u[14]+ uik[33]*u[15]+ uik[34]*u[16]+ uik[35]*u[17]; 1585ab27746eSHong Zhang 1586ab27746eSHong Zhang wp[18]+= uik[0]*u[18] + uik[1]*u[19] + uik[2]*u[20] + uik[3]*u[21] + uik[4]*u[22] + uik[5]*u[23]; 1587ab27746eSHong Zhang wp[19]+= uik[6]*u[18] + uik[7]*u[19] + uik[8]*u[20] + uik[9]*u[21]+ uik[10]*u[22]+ uik[11]*u[23]; 1588ab27746eSHong Zhang wp[20]+= uik[12]*u[18]+ uik[13]*u[19]+ uik[14]*u[20]+ uik[15]*u[21]+ uik[16]*u[22]+ uik[17]*u[23]; 1589ab27746eSHong Zhang wp[21]+= uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]; 1590ab27746eSHong Zhang wp[22]+= uik[24]*u[18]+ uik[25]*u[19]+ uik[26]*u[20]+ uik[27]*u[21]+ uik[28]*u[22]+ uik[29]*u[23]; 1591ab27746eSHong Zhang wp[23]+= uik[30]*u[18]+ uik[31]*u[19]+ uik[32]*u[20]+ uik[33]*u[21]+ uik[34]*u[22]+ uik[35]*u[23]; 1592ab27746eSHong Zhang 1593ab27746eSHong Zhang wp[24]+= uik[0]*u[24] + uik[1]*u[25] + uik[2]*u[26] + uik[3]*u[27] + uik[4]*u[28] + uik[5]*u[29]; 1594ab27746eSHong Zhang wp[25]+= uik[6]*u[24] + uik[7]*u[25] + uik[8]*u[26] + uik[9]*u[27]+ uik[10]*u[28]+ uik[11]*u[29]; 1595ab27746eSHong Zhang wp[26]+= uik[12]*u[24]+ uik[13]*u[25]+ uik[14]*u[26]+ uik[15]*u[27]+ uik[16]*u[28]+ uik[17]*u[29]; 1596ab27746eSHong Zhang wp[27]+= uik[18]*u[24]+ uik[19]*u[25]+ uik[20]*u[26]+ uik[21]*u[27]+ uik[22]*u[28]+ uik[23]*u[29]; 1597ab27746eSHong Zhang wp[28]+= uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]+ uik[28]*u[28]+ uik[29]*u[29]; 1598ab27746eSHong Zhang wp[29]+= uik[30]*u[24]+ uik[31]*u[25]+ uik[32]*u[26]+ uik[33]*u[27]+ uik[34]*u[28]+ uik[35]*u[29]; 1599ab27746eSHong Zhang 1600ab27746eSHong Zhang wp[30]+= uik[0]*u[30] + uik[1]*u[31] + uik[2]*u[32] + uik[3]*u[33] + uik[4]*u[34] + uik[5]*u[35]; 1601ab27746eSHong Zhang wp[31]+= uik[6]*u[30] + uik[7]*u[31] + uik[8]*u[32] + uik[9]*u[33]+ uik[10]*u[34]+ uik[11]*u[35]; 1602ab27746eSHong Zhang wp[32]+= uik[12]*u[30]+ uik[13]*u[31]+ uik[14]*u[32]+ uik[15]*u[33]+ uik[16]*u[34]+ uik[17]*u[35]; 1603ab27746eSHong Zhang wp[33]+= uik[18]*u[30]+ uik[19]*u[31]+ uik[20]*u[32]+ uik[21]*u[33]+ uik[22]*u[34]+ uik[23]*u[35]; 1604ab27746eSHong Zhang wp[34]+= uik[24]*u[30]+ uik[25]*u[31]+ uik[26]*u[32]+ uik[27]*u[33]+ uik[28]*u[34]+ uik[29]*u[35]; 1605ab27746eSHong Zhang wp[35]+= uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]+ uik[35]*u[35]; 160649b5e25fSSatish Balay } 1607ab27746eSHong Zhang 1608ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 1609ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1610ab27746eSHong Zhang j = bj[jmin]; 1611ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 161249b5e25fSSatish Balay } 1613ab27746eSHong Zhang i = nexti; 161449b5e25fSSatish Balay } 1615ab27746eSHong Zhang 1616ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 1617ab27746eSHong Zhang 161849b5e25fSSatish Balay /* invert diagonal block */ 1619ab27746eSHong Zhang d = ba+k*36; 1620ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 1621ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 1622ab27746eSHong Zhang 1623ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1624ab27746eSHong Zhang if (jmin < jmax) { 1625ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 1626ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 1627ab27746eSHong Zhang u = ba + j*36; 1628ab27746eSHong Zhang wp = w + vj*36; 1629ab27746eSHong Zhang for (k1=0; k1<36; k1++){ 1630ab27746eSHong Zhang *u++ = *wp; 1631ab27746eSHong Zhang *wp++ = 0.0; 1632ab27746eSHong Zhang } 163349b5e25fSSatish Balay } 163449b5e25fSSatish Balay 1635ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1636ab27746eSHong Zhang il[k] = jmin; 1637ab27746eSHong Zhang i = bj[jmin]; 1638ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 1639ab27746eSHong Zhang } 1640ab27746eSHong Zhang } 1641ab27746eSHong Zhang 1642ab27746eSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 1643ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1644ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 1645ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1646ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 1647ab27746eSHong Zhang 1648ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 164949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 1650ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 165149b5e25fSSatish Balay PLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 165249b5e25fSSatish Balay PetscFunctionReturn(0); 165349b5e25fSSatish Balay } 165449b5e25fSSatish Balay 1655b59abe41SHong Zhang /* Version for when blocks are 5 by 5 */ 1656b59abe41SHong Zhang #undef __FUNC__ 1657b59abe41SHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_5" 1658b59abe41SHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_5(Mat A,Mat *B) 1659b59abe41SHong Zhang { 1660b59abe41SHong Zhang Mat C = *B; 1661b59abe41SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1662b59abe41SHong Zhang IS perm = b->row; 1663b59abe41SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1664b59abe41SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1665b59abe41SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1666b59abe41SHong Zhang MatScalar *u,*d,*rtmp,*rtmp_ptr; 1667b59abe41SHong Zhang 1668b59abe41SHong Zhang PetscFunctionBegin; 1669b59abe41SHong Zhang /* initialization */ 1670b59abe41SHong Zhang rtmp = (MatScalar*)PetscMalloc(25*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 1671b59abe41SHong Zhang ierr = PetscMemzero(rtmp,25*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1672b59abe41SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 1673b59abe41SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 1674b59abe41SHong Zhang for (i=0; i<mbs; i++) { 1675b59abe41SHong Zhang jl[i] = mbs; il[0] = 0; 1676b59abe41SHong Zhang } 1677b59abe41SHong Zhang dk = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(dk); 1678b59abe41SHong Zhang uik = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(uik); 1679b59abe41SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 1680b59abe41SHong Zhang 1681b59abe41SHong Zhang /* check permutation */ 1682b59abe41SHong Zhang if (!a->permute){ 1683b59abe41SHong Zhang ai = a->i; aj = a->j; aa = a->a; 1684b59abe41SHong Zhang } else { 1685b59abe41SHong Zhang ai = a->inew; aj = a->jnew; 1686b59abe41SHong Zhang aa = (MatScalar*)PetscMalloc(25*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 1687b59abe41SHong Zhang ierr = PetscMemcpy(aa,a->a,25*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 1688b59abe41SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 1689b59abe41SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 1690b59abe41SHong Zhang 1691b59abe41SHong Zhang for (i=0; i<mbs; i++){ 1692b59abe41SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 1693b59abe41SHong Zhang for (j=jmin; j<jmax; j++){ 1694b59abe41SHong Zhang while (a2anew[j] != j){ 1695b59abe41SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 1696b59abe41SHong Zhang for (k1=0; k1<25; k1++){ 1697b59abe41SHong Zhang dk[k1] = aa[k*25+k1]; 1698b59abe41SHong Zhang aa[k*25+k1] = aa[j*25+k1]; 1699b59abe41SHong Zhang aa[j*25+k1] = dk[k1]; 1700b59abe41SHong Zhang } 1701b59abe41SHong Zhang } 1702b59abe41SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 1703b59abe41SHong Zhang if (i > aj[j]){ 1704b59abe41SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 1705b59abe41SHong Zhang ap = aa + j*25; /* ptr to the beginning of j-th block of aa */ 1706b59abe41SHong Zhang for (k=0; k<25; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 1707b59abe41SHong Zhang for (k=0; k<5; k++){ /* j-th block of aa <- dk^T */ 1708b59abe41SHong Zhang for (k1=0; k1<5; k1++) *ap++ = dk[k + 5*k1]; 1709b59abe41SHong Zhang } 1710b59abe41SHong Zhang } 1711b59abe41SHong Zhang } 1712b59abe41SHong Zhang } 1713b59abe41SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 1714b59abe41SHong Zhang } 1715b59abe41SHong Zhang 1716b59abe41SHong Zhang /* for each row k */ 1717b59abe41SHong Zhang for (k = 0; k<mbs; k++){ 1718b59abe41SHong Zhang 1719b59abe41SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 1720b59abe41SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 1721b59abe41SHong Zhang if (jmin < jmax) { 1722b59abe41SHong Zhang ap = aa + jmin*25; 1723b59abe41SHong Zhang for (j = jmin; j < jmax; j++){ 1724b59abe41SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 1725b59abe41SHong Zhang rtmp_ptr = rtmp + vj*25; 1726b59abe41SHong Zhang for (i=0; i<25; i++) *rtmp_ptr++ = *ap++; 1727b59abe41SHong Zhang } 1728b59abe41SHong Zhang } 1729b59abe41SHong Zhang 1730b59abe41SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1731b59abe41SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*25,25*sizeof(MatScalar));CHKERRQ(ierr); 1732b59abe41SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1733b59abe41SHong Zhang 1734b59abe41SHong Zhang while (i < mbs){ 1735b59abe41SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1736b59abe41SHong Zhang 1737b59abe41SHong Zhang /* compute multiplier */ 1738b59abe41SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1739b59abe41SHong Zhang 1740b59abe41SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1741b59abe41SHong Zhang d = ba + i*25; 1742b59abe41SHong Zhang u = ba + ili*25; 1743b59abe41SHong Zhang 1744b59abe41SHong Zhang uik[0] = -(d[0]*u[0] + d[5]*u[1] + d[10]*u[2] + d[15]*u[3] + d[20]*u[4]); 1745b59abe41SHong Zhang uik[1] = -(d[1]*u[0] + d[6]*u[1] + d[11]*u[2] + d[16]*u[3] + d[21]*u[4]); 1746b59abe41SHong Zhang uik[2] = -(d[2]*u[0] + d[7]*u[1] + d[12]*u[2] + d[17]*u[3] + d[22]*u[4]); 1747b59abe41SHong Zhang uik[3] = -(d[3]*u[0] + d[8]*u[1] + d[13]*u[2] + d[18]*u[3] + d[23]*u[4]); 1748b59abe41SHong Zhang uik[4] = -(d[4]*u[0] + d[9]*u[1] + d[14]*u[2] + d[19]*u[3] + d[24]*u[4]); 1749b59abe41SHong Zhang 1750b59abe41SHong Zhang uik[5] = -(d[0]*u[5] + d[5]*u[6] + d[10]*u[7] + d[15]*u[8] + d[20]*u[9]); 1751b59abe41SHong Zhang uik[6] = -(d[1]*u[5] + d[6]*u[6] + d[11]*u[7] + d[16]*u[8] + d[21]*u[9]); 1752b59abe41SHong Zhang uik[7] = -(d[2]*u[5] + d[7]*u[6] + d[12]*u[7] + d[17]*u[8] + d[22]*u[9]); 1753b59abe41SHong Zhang uik[8] = -(d[3]*u[5] + d[8]*u[6] + d[13]*u[7] + d[18]*u[8] + d[23]*u[9]); 1754b59abe41SHong Zhang uik[9] = -(d[4]*u[5] + d[9]*u[6] + d[14]*u[7] + d[19]*u[8] + d[24]*u[9]); 1755b59abe41SHong Zhang 1756b59abe41SHong Zhang uik[10]= -(d[0]*u[10] + d[5]*u[11] + d[10]*u[12] + d[15]*u[13] + d[20]*u[14]); 1757b59abe41SHong Zhang uik[11]= -(d[1]*u[10] + d[6]*u[11] + d[11]*u[12] + d[16]*u[13] + d[21]*u[14]); 1758b59abe41SHong Zhang uik[12]= -(d[2]*u[10] + d[7]*u[11] + d[12]*u[12] + d[17]*u[13] + d[22]*u[14]); 1759b59abe41SHong Zhang uik[13]= -(d[3]*u[10] + d[8]*u[11] + d[13]*u[12] + d[18]*u[13] + d[23]*u[14]); 1760b59abe41SHong Zhang uik[14]= -(d[4]*u[10] + d[9]*u[11] + d[14]*u[12] + d[19]*u[13] + d[24]*u[14]); 1761b59abe41SHong Zhang 1762b59abe41SHong Zhang uik[15]= -(d[0]*u[15] + d[5]*u[16] + d[10]*u[17] + d[15]*u[18] + d[20]*u[19]); 1763b59abe41SHong Zhang uik[16]= -(d[1]*u[15] + d[6]*u[16] + d[11]*u[17] + d[16]*u[18] + d[21]*u[19]); 1764b59abe41SHong Zhang uik[17]= -(d[2]*u[15] + d[7]*u[16] + d[12]*u[17] + d[17]*u[18] + d[22]*u[19]); 1765b59abe41SHong Zhang uik[18]= -(d[3]*u[15] + d[8]*u[16] + d[13]*u[17] + d[18]*u[18] + d[23]*u[19]); 1766b59abe41SHong Zhang uik[19]= -(d[4]*u[15] + d[9]*u[16] + d[14]*u[17] + d[19]*u[18] + d[24]*u[19]); 1767b59abe41SHong Zhang 1768b59abe41SHong Zhang uik[20]= -(d[0]*u[20] + d[5]*u[21] + d[10]*u[22] + d[15]*u[23] + d[20]*u[24]); 1769b59abe41SHong Zhang uik[21]= -(d[1]*u[20] + d[6]*u[21] + d[11]*u[22] + d[16]*u[23] + d[21]*u[24]); 1770b59abe41SHong Zhang uik[22]= -(d[2]*u[20] + d[7]*u[21] + d[12]*u[22] + d[17]*u[23] + d[22]*u[24]); 1771b59abe41SHong Zhang uik[23]= -(d[3]*u[20] + d[8]*u[21] + d[13]*u[22] + d[18]*u[23] + d[23]*u[24]); 1772b59abe41SHong Zhang uik[24]= -(d[4]*u[20] + d[9]*u[21] + d[14]*u[22] + d[19]*u[23] + d[24]*u[24]); 1773b59abe41SHong Zhang 1774b59abe41SHong Zhang 1775936bb232SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1776936bb232SHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1777936bb232SHong Zhang dk[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1778936bb232SHong Zhang dk[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1779936bb232SHong Zhang dk[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1780936bb232SHong Zhang dk[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1781b59abe41SHong Zhang 1782936bb232SHong Zhang dk[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1783936bb232SHong Zhang dk[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1784936bb232SHong Zhang dk[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1785936bb232SHong Zhang dk[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1786936bb232SHong Zhang dk[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1787b59abe41SHong Zhang 1788936bb232SHong Zhang dk[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1789936bb232SHong Zhang dk[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1790936bb232SHong Zhang dk[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1791936bb232SHong Zhang dk[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1792936bb232SHong Zhang dk[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1793b59abe41SHong Zhang 1794936bb232SHong Zhang dk[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1795936bb232SHong Zhang dk[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1796936bb232SHong Zhang dk[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1797936bb232SHong Zhang dk[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1798936bb232SHong Zhang dk[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1799936bb232SHong Zhang 1800936bb232SHong Zhang dk[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1801936bb232SHong Zhang dk[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1802936bb232SHong Zhang dk[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1803936bb232SHong Zhang dk[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 18044f078a5fSHong Zhang dk[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1805b59abe41SHong Zhang 1806b59abe41SHong Zhang /* update -U(i,k) */ 1807b59abe41SHong Zhang ierr = PetscMemcpy(ba+ili*25,uik,25*sizeof(MatScalar));CHKERRQ(ierr); 1808b59abe41SHong Zhang 1809b59abe41SHong Zhang /* add multiple of row i to k-th row ... */ 1810b59abe41SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1811b59abe41SHong Zhang if (jmin < jmax){ 1812b59abe41SHong Zhang for (j=jmin; j<jmax; j++) { 1813b59abe41SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 1814b59abe41SHong Zhang rtmp_ptr = rtmp + bj[j]*25; 1815b59abe41SHong Zhang u = ba + j*25; 1816936bb232SHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1817936bb232SHong Zhang rtmp_ptr[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1818936bb232SHong Zhang rtmp_ptr[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1819936bb232SHong Zhang rtmp_ptr[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1820936bb232SHong Zhang rtmp_ptr[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1821b59abe41SHong Zhang 1822936bb232SHong Zhang rtmp_ptr[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1823936bb232SHong Zhang rtmp_ptr[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1824936bb232SHong Zhang rtmp_ptr[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1825936bb232SHong Zhang rtmp_ptr[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1826936bb232SHong Zhang rtmp_ptr[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1827b59abe41SHong Zhang 1828936bb232SHong Zhang rtmp_ptr[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1829936bb232SHong Zhang rtmp_ptr[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1830936bb232SHong Zhang rtmp_ptr[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1831936bb232SHong Zhang rtmp_ptr[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1832936bb232SHong Zhang rtmp_ptr[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1833b59abe41SHong Zhang 1834936bb232SHong Zhang rtmp_ptr[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1835936bb232SHong Zhang rtmp_ptr[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1836936bb232SHong Zhang rtmp_ptr[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1837936bb232SHong Zhang rtmp_ptr[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1838936bb232SHong Zhang rtmp_ptr[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1839936bb232SHong Zhang 1840936bb232SHong Zhang rtmp_ptr[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1841936bb232SHong Zhang rtmp_ptr[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1842936bb232SHong Zhang rtmp_ptr[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1843936bb232SHong Zhang rtmp_ptr[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 18444f078a5fSHong Zhang rtmp_ptr[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1845b59abe41SHong Zhang } 1846b59abe41SHong Zhang 1847b59abe41SHong Zhang /* ... add i to row list for next nonzero entry */ 1848b59abe41SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1849b59abe41SHong Zhang j = bj[jmin]; 1850b59abe41SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 1851b59abe41SHong Zhang } 1852b59abe41SHong Zhang i = nexti; 1853b59abe41SHong Zhang } 1854b59abe41SHong Zhang 1855b59abe41SHong Zhang /* save nonzero entries in k-th row of U ... */ 1856b59abe41SHong Zhang 1857b59abe41SHong Zhang /* invert diagonal block */ 1858b59abe41SHong Zhang d = ba+k*25; 1859b59abe41SHong Zhang ierr = PetscMemcpy(d,dk,25*sizeof(MatScalar));CHKERRQ(ierr); 1860b59abe41SHong Zhang ierr = Kernel_A_gets_inverse_A_5(d);CHKERRQ(ierr); 1861b59abe41SHong Zhang 1862b59abe41SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1863b59abe41SHong Zhang if (jmin < jmax) { 1864b59abe41SHong Zhang for (j=jmin; j<jmax; j++){ 1865b59abe41SHong Zhang vj = bj[j]; /* block col. index of U */ 1866b59abe41SHong Zhang u = ba + j*25; 1867b59abe41SHong Zhang rtmp_ptr = rtmp + vj*25; 1868b59abe41SHong Zhang for (k1=0; k1<25; k1++){ 1869b59abe41SHong Zhang *u++ = *rtmp_ptr; 1870b59abe41SHong Zhang *rtmp_ptr++ = 0.0; 1871b59abe41SHong Zhang } 1872b59abe41SHong Zhang } 1873b59abe41SHong Zhang 1874b59abe41SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1875b59abe41SHong Zhang il[k] = jmin; 1876b59abe41SHong Zhang i = bj[jmin]; 1877b59abe41SHong Zhang jl[k] = jl[i]; jl[i] = k; 1878b59abe41SHong Zhang } 1879b59abe41SHong Zhang } 1880b59abe41SHong Zhang 1881b59abe41SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 1882b59abe41SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1883b59abe41SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 1884b59abe41SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1885b59abe41SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 1886b59abe41SHong Zhang if (a->permute){ 1887b59abe41SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 1888b59abe41SHong Zhang } 1889b59abe41SHong Zhang 1890b59abe41SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 1891b59abe41SHong Zhang C->factor = FACTOR_CHOLESKY; 1892b59abe41SHong Zhang C->assembled = PETSC_TRUE; 1893b59abe41SHong Zhang C->preallocated = PETSC_TRUE; 1894b59abe41SHong Zhang PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 1895b59abe41SHong Zhang PetscFunctionReturn(0); 1896b59abe41SHong Zhang } 189749b5e25fSSatish Balay 189849b5e25fSSatish Balay /* 189949b5e25fSSatish Balay Version for when blocks are 5 by 5 Using natural ordering 190049b5e25fSSatish Balay */ 190149b5e25fSSatish Balay #undef __FUNC__ 19026f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering" 19036f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering(Mat A,Mat *B) 190449b5e25fSSatish Balay { 190549b5e25fSSatish Balay Mat C = *B; 1906ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1907ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 190871149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1909ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1910ab27746eSHong Zhang MatScalar *u,*d,*rtmp,*rtmp_ptr; 191149b5e25fSSatish Balay 191249b5e25fSSatish Balay PetscFunctionBegin; 1913*671cb588SHong Zhang 1914ab27746eSHong Zhang /* initialization */ 1915ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(25*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 1916ab27746eSHong Zhang ierr = PetscMemzero(rtmp,25*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1917ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 1918ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 1919ab27746eSHong Zhang for (i=0; i<mbs; i++) { 1920ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 192149b5e25fSSatish Balay } 1922ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(dk); 1923ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(uik); 1924ab27746eSHong Zhang 1925ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 1926ab27746eSHong Zhang 1927ab27746eSHong Zhang /* for each row k */ 1928ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 1929ab27746eSHong Zhang 1930ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 1931ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 1932ab27746eSHong Zhang if (jmin < jmax) { 1933ab27746eSHong Zhang ap = aa + jmin*25; 1934ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 1935ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 1936ab27746eSHong Zhang rtmp_ptr = rtmp + vj*25; 1937ab27746eSHong Zhang for (i=0; i<25; i++) *rtmp_ptr++ = *ap++; 193849b5e25fSSatish Balay } 193949b5e25fSSatish Balay } 1940ab27746eSHong Zhang 1941ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1942ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*25,25*sizeof(MatScalar));CHKERRQ(ierr); 1943ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1944ab27746eSHong Zhang 1945ab27746eSHong Zhang while (i < mbs){ 1946ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1947ab27746eSHong Zhang 1948ab27746eSHong Zhang /* compute multiplier */ 1949ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1950ab27746eSHong Zhang 1951ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1952ab27746eSHong Zhang d = ba + i*25; 1953ab27746eSHong Zhang u = ba + ili*25; 1954ab27746eSHong Zhang 1955ab27746eSHong Zhang uik[0] = -(d[0]*u[0] + d[5]*u[1] + d[10]*u[2] + d[15]*u[3] + d[20]*u[4]); 1956ab27746eSHong Zhang uik[1] = -(d[1]*u[0] + d[6]*u[1] + d[11]*u[2] + d[16]*u[3] + d[21]*u[4]); 1957ab27746eSHong Zhang uik[2] = -(d[2]*u[0] + d[7]*u[1] + d[12]*u[2] + d[17]*u[3] + d[22]*u[4]); 1958ab27746eSHong Zhang uik[3] = -(d[3]*u[0] + d[8]*u[1] + d[13]*u[2] + d[18]*u[3] + d[23]*u[4]); 1959ab27746eSHong Zhang uik[4] = -(d[4]*u[0] + d[9]*u[1] + d[14]*u[2] + d[19]*u[3] + d[24]*u[4]); 1960ab27746eSHong Zhang 1961ab27746eSHong Zhang uik[5] = -(d[0]*u[5] + d[5]*u[6] + d[10]*u[7] + d[15]*u[8] + d[20]*u[9]); 1962ab27746eSHong Zhang uik[6] = -(d[1]*u[5] + d[6]*u[6] + d[11]*u[7] + d[16]*u[8] + d[21]*u[9]); 1963ab27746eSHong Zhang uik[7] = -(d[2]*u[5] + d[7]*u[6] + d[12]*u[7] + d[17]*u[8] + d[22]*u[9]); 1964ab27746eSHong Zhang uik[8] = -(d[3]*u[5] + d[8]*u[6] + d[13]*u[7] + d[18]*u[8] + d[23]*u[9]); 1965ab27746eSHong Zhang uik[9] = -(d[4]*u[5] + d[9]*u[6] + d[14]*u[7] + d[19]*u[8] + d[24]*u[9]); 1966ab27746eSHong Zhang 1967ab27746eSHong Zhang uik[10]= -(d[0]*u[10] + d[5]*u[11] + d[10]*u[12] + d[15]*u[13] + d[20]*u[14]); 1968ab27746eSHong Zhang uik[11]= -(d[1]*u[10] + d[6]*u[11] + d[11]*u[12] + d[16]*u[13] + d[21]*u[14]); 1969ab27746eSHong Zhang uik[12]= -(d[2]*u[10] + d[7]*u[11] + d[12]*u[12] + d[17]*u[13] + d[22]*u[14]); 1970ab27746eSHong Zhang uik[13]= -(d[3]*u[10] + d[8]*u[11] + d[13]*u[12] + d[18]*u[13] + d[23]*u[14]); 1971ab27746eSHong Zhang uik[14]= -(d[4]*u[10] + d[9]*u[11] + d[14]*u[12] + d[19]*u[13] + d[24]*u[14]); 1972ab27746eSHong Zhang 1973ab27746eSHong Zhang uik[15]= -(d[0]*u[15] + d[5]*u[16] + d[10]*u[17] + d[15]*u[18] + d[20]*u[19]); 1974ab27746eSHong Zhang uik[16]= -(d[1]*u[15] + d[6]*u[16] + d[11]*u[17] + d[16]*u[18] + d[21]*u[19]); 1975ab27746eSHong Zhang uik[17]= -(d[2]*u[15] + d[7]*u[16] + d[12]*u[17] + d[17]*u[18] + d[22]*u[19]); 1976ab27746eSHong Zhang uik[18]= -(d[3]*u[15] + d[8]*u[16] + d[13]*u[17] + d[18]*u[18] + d[23]*u[19]); 1977ab27746eSHong Zhang uik[19]= -(d[4]*u[15] + d[9]*u[16] + d[14]*u[17] + d[19]*u[18] + d[24]*u[19]); 1978ab27746eSHong Zhang 1979ab27746eSHong Zhang uik[20]= -(d[0]*u[20] + d[5]*u[21] + d[10]*u[22] + d[15]*u[23] + d[20]*u[24]); 1980ab27746eSHong Zhang uik[21]= -(d[1]*u[20] + d[6]*u[21] + d[11]*u[22] + d[16]*u[23] + d[21]*u[24]); 1981ab27746eSHong Zhang uik[22]= -(d[2]*u[20] + d[7]*u[21] + d[12]*u[22] + d[17]*u[23] + d[22]*u[24]); 1982ab27746eSHong Zhang uik[23]= -(d[3]*u[20] + d[8]*u[21] + d[13]*u[22] + d[18]*u[23] + d[23]*u[24]); 1983ab27746eSHong Zhang uik[24]= -(d[4]*u[20] + d[9]*u[21] + d[14]*u[22] + d[19]*u[23] + d[24]*u[24]); 1984ab27746eSHong Zhang 1985ab27746eSHong Zhang 1986ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1987ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1988ab27746eSHong Zhang dk[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1989ab27746eSHong Zhang dk[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1990ab27746eSHong Zhang dk[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1991ab27746eSHong Zhang dk[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1992ab27746eSHong Zhang 1993ab27746eSHong Zhang dk[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1994ab27746eSHong Zhang dk[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1995ab27746eSHong Zhang dk[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1996ab27746eSHong Zhang dk[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1997ab27746eSHong Zhang dk[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1998ab27746eSHong Zhang 1999ab27746eSHong Zhang dk[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 2000ab27746eSHong Zhang dk[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 2001ab27746eSHong Zhang dk[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 2002ab27746eSHong Zhang dk[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 2003ab27746eSHong Zhang dk[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 2004ab27746eSHong Zhang 2005ab27746eSHong Zhang dk[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 2006ab27746eSHong Zhang dk[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 2007ab27746eSHong Zhang dk[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 2008ab27746eSHong Zhang dk[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 2009ab27746eSHong Zhang dk[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 2010ab27746eSHong Zhang 2011ab27746eSHong Zhang dk[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 2012ab27746eSHong Zhang dk[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 2013ab27746eSHong Zhang dk[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 2014ab27746eSHong Zhang dk[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 2015ab27746eSHong Zhang dk[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 2016ab27746eSHong Zhang 2017ab27746eSHong Zhang /* update -U(i,k) */ 2018ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*25,uik,25*sizeof(MatScalar));CHKERRQ(ierr); 2019ab27746eSHong Zhang 2020ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2021ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2022ab27746eSHong Zhang if (jmin < jmax){ 2023ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2024ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2025ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*25; 2026ab27746eSHong Zhang u = ba + j*25; 2027ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 2028ab27746eSHong Zhang rtmp_ptr[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 2029ab27746eSHong Zhang rtmp_ptr[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 2030ab27746eSHong Zhang rtmp_ptr[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 2031ab27746eSHong Zhang rtmp_ptr[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 2032ab27746eSHong Zhang 2033ab27746eSHong Zhang rtmp_ptr[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 2034ab27746eSHong Zhang rtmp_ptr[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 2035ab27746eSHong Zhang rtmp_ptr[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 2036ab27746eSHong Zhang rtmp_ptr[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 2037ab27746eSHong Zhang rtmp_ptr[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 2038ab27746eSHong Zhang 2039ab27746eSHong Zhang rtmp_ptr[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 2040ab27746eSHong Zhang rtmp_ptr[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 2041ab27746eSHong Zhang rtmp_ptr[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 2042ab27746eSHong Zhang rtmp_ptr[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 2043ab27746eSHong Zhang rtmp_ptr[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 2044ab27746eSHong Zhang 2045ab27746eSHong Zhang rtmp_ptr[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 2046ab27746eSHong Zhang rtmp_ptr[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 2047ab27746eSHong Zhang rtmp_ptr[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 2048ab27746eSHong Zhang rtmp_ptr[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 2049ab27746eSHong Zhang rtmp_ptr[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 2050ab27746eSHong Zhang 2051ab27746eSHong Zhang rtmp_ptr[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 2052ab27746eSHong Zhang rtmp_ptr[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 2053ab27746eSHong Zhang rtmp_ptr[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 2054ab27746eSHong Zhang rtmp_ptr[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 2055ab27746eSHong Zhang rtmp_ptr[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 205649b5e25fSSatish Balay } 2057ab27746eSHong Zhang 2058ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2059ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2060ab27746eSHong Zhang j = bj[jmin]; 2061ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 206249b5e25fSSatish Balay } 2063ab27746eSHong Zhang i = nexti; 206449b5e25fSSatish Balay } 2065ab27746eSHong Zhang 2066ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2067ab27746eSHong Zhang 206849b5e25fSSatish Balay /* invert diagonal block */ 2069ab27746eSHong Zhang d = ba+k*25; 2070ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,25*sizeof(MatScalar));CHKERRQ(ierr); 2071ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_5(d);CHKERRQ(ierr); 2072ab27746eSHong Zhang 2073ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2074ab27746eSHong Zhang if (jmin < jmax) { 2075ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2076ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2077ab27746eSHong Zhang u = ba + j*25; 2078ab27746eSHong Zhang rtmp_ptr = rtmp + vj*25; 2079ab27746eSHong Zhang for (k1=0; k1<25; k1++){ 2080ab27746eSHong Zhang *u++ = *rtmp_ptr; 2081ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2082ab27746eSHong Zhang } 2083ab27746eSHong Zhang } 2084ab27746eSHong Zhang 2085ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2086ab27746eSHong Zhang il[k] = jmin; 2087ab27746eSHong Zhang i = bj[jmin]; 2088ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2089ab27746eSHong Zhang } 209049b5e25fSSatish Balay } 209149b5e25fSSatish Balay 209249b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2093ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2094ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2095ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2096ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2097ab27746eSHong Zhang 2098ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 209949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 2100ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 210149b5e25fSSatish Balay PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 210249b5e25fSSatish Balay PetscFunctionReturn(0); 210349b5e25fSSatish Balay } 210449b5e25fSSatish Balay 210549b5e25fSSatish Balay /* 210649b5e25fSSatish Balay Version for when blocks are 4 by 4 Using natural ordering 210749b5e25fSSatish Balay */ 210849b5e25fSSatish Balay #undef __FUNC__ 21096f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering" 21106f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering(Mat A,Mat *B) 211149b5e25fSSatish Balay { 211249b5e25fSSatish Balay Mat C = *B; 2113ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2114ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 211571149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2116ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2117ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 211849b5e25fSSatish Balay 211949b5e25fSSatish Balay PetscFunctionBegin; 2120*671cb588SHong Zhang 2121ab27746eSHong Zhang /* initialization */ 2122ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(16*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2123ab27746eSHong Zhang ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2124ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2125ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2126ab27746eSHong Zhang for (i=0; i<mbs; i++) { 2127ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 212849b5e25fSSatish Balay } 2129ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(dk); 2130ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(uik); 2131ab27746eSHong Zhang 2132ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2133ab27746eSHong Zhang 2134ab27746eSHong Zhang /* for each row k */ 2135ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 2136ab27746eSHong Zhang 2137ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 2138ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 2139ab27746eSHong Zhang if (jmin < jmax) { 2140ab27746eSHong Zhang ap = aa + jmin*16; 2141ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 2142ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 2143ab27746eSHong Zhang rtmp_ptr = rtmp + vj*16; 2144ab27746eSHong Zhang for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 214549b5e25fSSatish Balay } 214649b5e25fSSatish Balay } 2147ab27746eSHong Zhang 2148ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2149ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 2150ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2151ab27746eSHong Zhang 2152ab27746eSHong Zhang while (i < mbs){ 2153ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2154ab27746eSHong Zhang 2155ab27746eSHong Zhang /* compute multiplier */ 2156ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2157ab27746eSHong Zhang 2158ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2159ab27746eSHong Zhang diag = ba + i*16; 2160ab27746eSHong Zhang u = ba + ili*16; 2161ab27746eSHong Zhang 2162ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 2163ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 2164ab27746eSHong Zhang uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 2165ab27746eSHong Zhang uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 2166ab27746eSHong Zhang 2167ab27746eSHong Zhang uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 2168ab27746eSHong Zhang uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 2169ab27746eSHong Zhang uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 2170ab27746eSHong Zhang uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 2171ab27746eSHong Zhang 2172ab27746eSHong Zhang uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 2173ab27746eSHong Zhang uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 2174ab27746eSHong Zhang uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 2175ab27746eSHong Zhang uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 2176ab27746eSHong Zhang 2177ab27746eSHong Zhang uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 2178ab27746eSHong Zhang uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 2179ab27746eSHong Zhang uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 2180ab27746eSHong Zhang uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 2181ab27746eSHong Zhang 2182ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2183ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2184ab27746eSHong Zhang dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2185ab27746eSHong Zhang dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2186ab27746eSHong Zhang dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2187ab27746eSHong Zhang 2188ab27746eSHong Zhang dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2189ab27746eSHong Zhang dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2190ab27746eSHong Zhang dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2191ab27746eSHong Zhang dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2192ab27746eSHong Zhang 2193ab27746eSHong Zhang dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2194ab27746eSHong Zhang dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2195ab27746eSHong Zhang dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2196ab27746eSHong Zhang dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2197ab27746eSHong Zhang 2198ab27746eSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2199ab27746eSHong Zhang dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2200ab27746eSHong Zhang dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2201ab27746eSHong Zhang dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2202ab27746eSHong Zhang 2203ab27746eSHong Zhang /* update -U(i,k) */ 2204ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 2205ab27746eSHong Zhang 2206ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2207ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2208ab27746eSHong Zhang if (jmin < jmax){ 2209ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2210ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2211ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*16; 2212ab27746eSHong Zhang u = ba + j*16; 2213ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2214ab27746eSHong Zhang rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2215ab27746eSHong Zhang rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2216ab27746eSHong Zhang rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2217ab27746eSHong Zhang 2218ab27746eSHong Zhang rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2219ab27746eSHong Zhang rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2220ab27746eSHong Zhang rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2221ab27746eSHong Zhang rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2222ab27746eSHong Zhang 2223ab27746eSHong Zhang rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2224ab27746eSHong Zhang rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2225ab27746eSHong Zhang rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2226ab27746eSHong Zhang rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2227ab27746eSHong Zhang 2228ab27746eSHong Zhang rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2229ab27746eSHong Zhang rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2230ab27746eSHong Zhang rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2231ab27746eSHong Zhang rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 223249b5e25fSSatish Balay } 2233ab27746eSHong Zhang 2234ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2235ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2236ab27746eSHong Zhang j = bj[jmin]; 2237ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 223849b5e25fSSatish Balay } 2239ab27746eSHong Zhang i = nexti; 224049b5e25fSSatish Balay } 2241ab27746eSHong Zhang 2242ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2243ab27746eSHong Zhang 224449b5e25fSSatish Balay /* invert diagonal block */ 2245ab27746eSHong Zhang diag = ba+k*16; 2246ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 2247ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 2248ab27746eSHong Zhang 2249ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2250ab27746eSHong Zhang if (jmin < jmax) { 2251ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2252ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2253ab27746eSHong Zhang u = ba + j*16; 2254ab27746eSHong Zhang rtmp_ptr = rtmp + vj*16; 2255ab27746eSHong Zhang for (k1=0; k1<16; k1++){ 2256ab27746eSHong Zhang *u++ = *rtmp_ptr; 2257ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2258ab27746eSHong Zhang } 2259ab27746eSHong Zhang } 2260ab27746eSHong Zhang 2261ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2262ab27746eSHong Zhang il[k] = jmin; 2263ab27746eSHong Zhang i = bj[jmin]; 2264ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2265ab27746eSHong Zhang } 226649b5e25fSSatish Balay } 226749b5e25fSSatish Balay 226849b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2269ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2270ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2271ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2272ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2273ab27746eSHong Zhang 2274ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 227549b5e25fSSatish Balay C->assembled = PETSC_TRUE; 22767a05c331SHong Zhang C->preallocated = PETSC_TRUE; 227749b5e25fSSatish Balay PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 227849b5e25fSSatish Balay PetscFunctionReturn(0); 227949b5e25fSSatish Balay } 228049b5e25fSSatish Balay 2281effa298cSHong Zhang /* Version for when blocks are 4 by 4 */ 2282effa298cSHong Zhang #undef __FUNC__ 2283effa298cSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4" 2284effa298cSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_4(Mat A,Mat *B) 2285effa298cSHong Zhang { 2286effa298cSHong Zhang Mat C = *B; 2287effa298cSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2288effa298cSHong Zhang IS perm = b->row; 2289effa298cSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2290effa298cSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2291effa298cSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2292effa298cSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 2293effa298cSHong Zhang 2294effa298cSHong Zhang PetscFunctionBegin; 2295effa298cSHong Zhang /* initialization */ 2296effa298cSHong Zhang rtmp = (MatScalar*)PetscMalloc(16*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2297effa298cSHong Zhang ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2298effa298cSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2299effa298cSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2300effa298cSHong Zhang for (i=0; i<mbs; i++) { 2301effa298cSHong Zhang jl[i] = mbs; il[0] = 0; 2302effa298cSHong Zhang } 2303effa298cSHong Zhang dk = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(dk); 2304effa298cSHong Zhang uik = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(uik); 2305effa298cSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2306effa298cSHong Zhang 2307effa298cSHong Zhang /* check permutation */ 2308effa298cSHong Zhang if (!a->permute){ 2309effa298cSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2310effa298cSHong Zhang } else { 2311effa298cSHong Zhang ai = a->inew; aj = a->jnew; 2312effa298cSHong Zhang aa = (MatScalar*)PetscMalloc(16*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2313effa298cSHong Zhang ierr = PetscMemcpy(aa,a->a,16*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2314effa298cSHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2315effa298cSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2316effa298cSHong Zhang 2317effa298cSHong Zhang for (i=0; i<mbs; i++){ 2318effa298cSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2319effa298cSHong Zhang for (j=jmin; j<jmax; j++){ 2320effa298cSHong Zhang while (a2anew[j] != j){ 2321effa298cSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2322effa298cSHong Zhang for (k1=0; k1<16; k1++){ 2323effa298cSHong Zhang dk[k1] = aa[k*16+k1]; 2324effa298cSHong Zhang aa[k*16+k1] = aa[j*16+k1]; 2325effa298cSHong Zhang aa[j*16+k1] = dk[k1]; 2326effa298cSHong Zhang } 2327effa298cSHong Zhang } 2328effa298cSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2329effa298cSHong Zhang if (i > aj[j]){ 2330effa298cSHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2331effa298cSHong Zhang ap = aa + j*16; /* ptr to the beginning of j-th block of aa */ 2332effa298cSHong Zhang for (k=0; k<16; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2333effa298cSHong Zhang for (k=0; k<4; k++){ /* j-th block of aa <- dk^T */ 2334effa298cSHong Zhang for (k1=0; k1<4; k1++) *ap++ = dk[k + 4*k1]; 2335effa298cSHong Zhang } 2336effa298cSHong Zhang } 2337effa298cSHong Zhang } 2338effa298cSHong Zhang } 2339effa298cSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2340effa298cSHong Zhang } 2341effa298cSHong Zhang 2342effa298cSHong Zhang /* for each row k */ 2343effa298cSHong Zhang for (k = 0; k<mbs; k++){ 2344effa298cSHong Zhang 2345effa298cSHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2346effa298cSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 2347effa298cSHong Zhang if (jmin < jmax) { 2348effa298cSHong Zhang ap = aa + jmin*16; 2349effa298cSHong Zhang for (j = jmin; j < jmax; j++){ 2350effa298cSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2351effa298cSHong Zhang rtmp_ptr = rtmp + vj*16; 2352effa298cSHong Zhang for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 2353effa298cSHong Zhang } 2354effa298cSHong Zhang } 2355effa298cSHong Zhang 2356effa298cSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2357effa298cSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 2358effa298cSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2359effa298cSHong Zhang 2360effa298cSHong Zhang while (i < mbs){ 2361effa298cSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2362effa298cSHong Zhang 2363effa298cSHong Zhang /* compute multiplier */ 2364effa298cSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2365effa298cSHong Zhang 2366effa298cSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2367effa298cSHong Zhang diag = ba + i*16; 2368effa298cSHong Zhang u = ba + ili*16; 2369effa298cSHong Zhang 2370effa298cSHong Zhang uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 2371effa298cSHong Zhang uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 2372effa298cSHong Zhang uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 2373effa298cSHong Zhang uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 2374effa298cSHong Zhang 2375effa298cSHong Zhang uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 2376effa298cSHong Zhang uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 2377effa298cSHong Zhang uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 2378effa298cSHong Zhang uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 2379effa298cSHong Zhang 2380effa298cSHong Zhang uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 2381effa298cSHong Zhang uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 2382effa298cSHong Zhang uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 2383effa298cSHong Zhang uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 2384effa298cSHong Zhang 2385effa298cSHong Zhang uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 2386effa298cSHong Zhang uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 2387effa298cSHong Zhang uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 2388effa298cSHong Zhang uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 2389effa298cSHong Zhang 2390effa298cSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2391effa298cSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2392effa298cSHong Zhang dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2393effa298cSHong Zhang dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2394effa298cSHong Zhang dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2395effa298cSHong Zhang 2396effa298cSHong Zhang dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2397effa298cSHong Zhang dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2398effa298cSHong Zhang dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2399effa298cSHong Zhang dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2400effa298cSHong Zhang 2401effa298cSHong Zhang dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2402effa298cSHong Zhang dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2403effa298cSHong Zhang dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2404effa298cSHong Zhang dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2405effa298cSHong Zhang 2406effa298cSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2407effa298cSHong Zhang dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2408effa298cSHong Zhang dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2409effa298cSHong Zhang dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2410effa298cSHong Zhang 2411effa298cSHong Zhang /* update -U(i,k) */ 2412effa298cSHong Zhang ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 2413effa298cSHong Zhang 2414effa298cSHong Zhang /* add multiple of row i to k-th row ... */ 2415effa298cSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2416effa298cSHong Zhang if (jmin < jmax){ 2417effa298cSHong Zhang for (j=jmin; j<jmax; j++) { 2418effa298cSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2419effa298cSHong Zhang rtmp_ptr = rtmp + bj[j]*16; 2420effa298cSHong Zhang u = ba + j*16; 2421effa298cSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2422effa298cSHong Zhang rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2423effa298cSHong Zhang rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2424effa298cSHong Zhang rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2425effa298cSHong Zhang 2426effa298cSHong Zhang rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2427effa298cSHong Zhang rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2428effa298cSHong Zhang rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2429effa298cSHong Zhang rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2430effa298cSHong Zhang 2431effa298cSHong Zhang rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2432effa298cSHong Zhang rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2433effa298cSHong Zhang rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2434effa298cSHong Zhang rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2435effa298cSHong Zhang 2436effa298cSHong Zhang rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2437effa298cSHong Zhang rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2438effa298cSHong Zhang rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2439effa298cSHong Zhang rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2440effa298cSHong Zhang } 2441effa298cSHong Zhang 2442effa298cSHong Zhang /* ... add i to row list for next nonzero entry */ 2443effa298cSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2444effa298cSHong Zhang j = bj[jmin]; 2445effa298cSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 2446effa298cSHong Zhang } 2447effa298cSHong Zhang i = nexti; 2448effa298cSHong Zhang } 2449effa298cSHong Zhang 2450effa298cSHong Zhang /* save nonzero entries in k-th row of U ... */ 2451effa298cSHong Zhang 2452effa298cSHong Zhang /* invert diagonal block */ 2453effa298cSHong Zhang diag = ba+k*16; 2454effa298cSHong Zhang ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 2455effa298cSHong Zhang ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 2456effa298cSHong Zhang 2457effa298cSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2458effa298cSHong Zhang if (jmin < jmax) { 2459effa298cSHong Zhang for (j=jmin; j<jmax; j++){ 2460effa298cSHong Zhang vj = bj[j]; /* block col. index of U */ 2461effa298cSHong Zhang u = ba + j*16; 2462effa298cSHong Zhang rtmp_ptr = rtmp + vj*16; 2463effa298cSHong Zhang for (k1=0; k1<16; k1++){ 2464effa298cSHong Zhang *u++ = *rtmp_ptr; 2465effa298cSHong Zhang *rtmp_ptr++ = 0.0; 2466effa298cSHong Zhang } 2467effa298cSHong Zhang } 2468effa298cSHong Zhang 2469effa298cSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2470effa298cSHong Zhang il[k] = jmin; 2471effa298cSHong Zhang i = bj[jmin]; 2472effa298cSHong Zhang jl[k] = jl[i]; jl[i] = k; 2473effa298cSHong Zhang } 2474effa298cSHong Zhang } 2475effa298cSHong Zhang 2476effa298cSHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 2477effa298cSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2478effa298cSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2479effa298cSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2480effa298cSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2481effa298cSHong Zhang if (a->permute){ 2482effa298cSHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 2483effa298cSHong Zhang } 2484effa298cSHong Zhang 2485effa298cSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 2486effa298cSHong Zhang C->factor = FACTOR_CHOLESKY; 2487effa298cSHong Zhang C->assembled = PETSC_TRUE; 2488effa298cSHong Zhang C->preallocated = PETSC_TRUE; 2489effa298cSHong Zhang PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 2490effa298cSHong Zhang PetscFunctionReturn(0); 2491effa298cSHong Zhang } 2492effa298cSHong Zhang 2493a1723e09SHong Zhang /* Version for when blocks are 3 by 3 */ 249449b5e25fSSatish Balay #undef __FUNC__ 24956f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_3" 24966f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_3(Mat A,Mat *B) 249749b5e25fSSatish Balay { 249849b5e25fSSatish Balay Mat C = *B; 2499a1723e09SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2500a1723e09SHong Zhang IS perm = b->row; 2501a1723e09SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2502a1723e09SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2503a1723e09SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2504a1723e09SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 250549b5e25fSSatish Balay 250649b5e25fSSatish Balay PetscFunctionBegin; 2507*671cb588SHong Zhang 2508a1723e09SHong Zhang /* initialization */ 2509a1723e09SHong Zhang rtmp = (MatScalar*)PetscMalloc(9*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2510a1723e09SHong Zhang ierr = PetscMemzero(rtmp,9*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2511a1723e09SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2512a1723e09SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2513a1723e09SHong Zhang for (i=0; i<mbs; i++) { 2514a1723e09SHong Zhang jl[i] = mbs; il[0] = 0; 2515a1723e09SHong Zhang } 2516a1723e09SHong Zhang dk = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(dk); 2517a1723e09SHong Zhang uik = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(uik); 2518a1723e09SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 251949b5e25fSSatish Balay 2520a1723e09SHong Zhang /* check permutation */ 2521a1723e09SHong Zhang if (!a->permute){ 2522a1723e09SHong Zhang ai = a->i; aj = a->j; aa = a->a; 2523a1723e09SHong Zhang } else { 2524a1723e09SHong Zhang ai = a->inew; aj = a->jnew; 2525a1723e09SHong Zhang aa = (MatScalar*)PetscMalloc(9*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2526a1723e09SHong Zhang ierr = PetscMemcpy(aa,a->a,9*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2527a1723e09SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2528a1723e09SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 252949b5e25fSSatish Balay 2530a1723e09SHong Zhang for (i=0; i<mbs; i++){ 2531a1723e09SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2532a1723e09SHong Zhang for (j=jmin; j<jmax; j++){ 2533a1723e09SHong Zhang while (a2anew[j] != j){ 2534a1723e09SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2535a1723e09SHong Zhang for (k1=0; k1<9; k1++){ 2536a1723e09SHong Zhang dk[k1] = aa[k*9+k1]; 2537a1723e09SHong Zhang aa[k*9+k1] = aa[j*9+k1]; 2538a1723e09SHong Zhang aa[j*9+k1] = dk[k1]; 2539a1723e09SHong Zhang } 2540a1723e09SHong Zhang } 2541a1723e09SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2542a1723e09SHong Zhang if (i > aj[j]){ 254347c40530SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2544a1723e09SHong Zhang ap = aa + j*9; /* ptr to the beginning of j-th block of aa */ 2545a1723e09SHong Zhang for (k=0; k<9; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2546a1723e09SHong Zhang for (k=0; k<3; k++){ /* j-th block of aa <- dk^T */ 254747c40530SHong Zhang for (k1=0; k1<3; k1++) *ap++ = dk[k + 3*k1]; 2548a1723e09SHong Zhang } 2549a1723e09SHong Zhang } 2550a1723e09SHong Zhang } 2551a1723e09SHong Zhang } 2552a1723e09SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2553a1723e09SHong Zhang } 255449b5e25fSSatish Balay 2555a1723e09SHong Zhang /* for each row k */ 2556a1723e09SHong Zhang for (k = 0; k<mbs; k++){ 255749b5e25fSSatish Balay 2558a1723e09SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2559a1723e09SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 2560a1723e09SHong Zhang if (jmin < jmax) { 2561a1723e09SHong Zhang ap = aa + jmin*9; 2562a1723e09SHong Zhang for (j = jmin; j < jmax; j++){ 2563a1723e09SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2564a1723e09SHong Zhang rtmp_ptr = rtmp + vj*9; 2565a1723e09SHong Zhang for (i=0; i<9; i++) *rtmp_ptr++ = *ap++; 2566a1723e09SHong Zhang } 2567a1723e09SHong Zhang } 256849b5e25fSSatish Balay 2569a1723e09SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2570a1723e09SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*9,9*sizeof(MatScalar));CHKERRQ(ierr); 2571a1723e09SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2572a1723e09SHong Zhang 2573a1723e09SHong Zhang while (i < mbs){ 2574a1723e09SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2575a1723e09SHong Zhang 2576a1723e09SHong Zhang /* compute multiplier */ 2577a1723e09SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2578a1723e09SHong Zhang 2579a1723e09SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2580a1723e09SHong Zhang diag = ba + i*9; 2581a1723e09SHong Zhang u = ba + ili*9; 2582a1723e09SHong Zhang 2583a1723e09SHong Zhang uik[0] = -(diag[0]*u[0] + diag[3]*u[1] + diag[6]*u[2]); 2584a1723e09SHong Zhang uik[1] = -(diag[1]*u[0] + diag[4]*u[1] + diag[7]*u[2]); 2585a1723e09SHong Zhang uik[2] = -(diag[2]*u[0] + diag[5]*u[1] + diag[8]*u[2]); 2586a1723e09SHong Zhang 2587a1723e09SHong Zhang uik[3] = -(diag[0]*u[3] + diag[3]*u[4] + diag[6]*u[5]); 2588a1723e09SHong Zhang uik[4] = -(diag[1]*u[3] + diag[4]*u[4] + diag[7]*u[5]); 2589a1723e09SHong Zhang uik[5] = -(diag[2]*u[3] + diag[5]*u[4] + diag[8]*u[5]); 2590a1723e09SHong Zhang 2591a1723e09SHong Zhang uik[6] = -(diag[0]*u[6] + diag[3]*u[7] + diag[6]*u[8]); 2592a1723e09SHong Zhang uik[7] = -(diag[1]*u[6] + diag[4]*u[7] + diag[7]*u[8]); 2593a1723e09SHong Zhang uik[8] = -(diag[2]*u[6] + diag[5]*u[7] + diag[8]*u[8]); 2594a1723e09SHong Zhang 2595a1723e09SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2596a1723e09SHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2597a1723e09SHong Zhang dk[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2598a1723e09SHong Zhang dk[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2599a1723e09SHong Zhang 2600a1723e09SHong Zhang dk[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2601a1723e09SHong Zhang dk[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2602a1723e09SHong Zhang dk[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2603a1723e09SHong Zhang 2604a1723e09SHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2605a1723e09SHong Zhang dk[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2606a1723e09SHong Zhang dk[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 2607a1723e09SHong Zhang 2608a1723e09SHong Zhang /* update -U(i,k) */ 2609a1723e09SHong Zhang ierr = PetscMemcpy(ba+ili*9,uik,9*sizeof(MatScalar));CHKERRQ(ierr); 2610a1723e09SHong Zhang 2611a1723e09SHong Zhang /* add multiple of row i to k-th row ... */ 2612a1723e09SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2613a1723e09SHong Zhang if (jmin < jmax){ 2614a1723e09SHong Zhang for (j=jmin; j<jmax; j++) { 2615a1723e09SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2616a1723e09SHong Zhang rtmp_ptr = rtmp + bj[j]*9; 2617a1723e09SHong Zhang u = ba + j*9; 2618a1723e09SHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2619a1723e09SHong Zhang rtmp_ptr[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2620a1723e09SHong Zhang rtmp_ptr[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2621a1723e09SHong Zhang 2622a1723e09SHong Zhang rtmp_ptr[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2623a1723e09SHong Zhang rtmp_ptr[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2624a1723e09SHong Zhang rtmp_ptr[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2625a1723e09SHong Zhang 2626a1723e09SHong Zhang rtmp_ptr[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2627a1723e09SHong Zhang rtmp_ptr[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2628a1723e09SHong Zhang rtmp_ptr[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 262949b5e25fSSatish Balay } 2630a1723e09SHong Zhang 2631a1723e09SHong Zhang /* ... add i to row list for next nonzero entry */ 2632a1723e09SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2633a1723e09SHong Zhang j = bj[jmin]; 2634a1723e09SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 263549b5e25fSSatish Balay } 2636a1723e09SHong Zhang i = nexti; 263749b5e25fSSatish Balay } 2638a1723e09SHong Zhang 2639a1723e09SHong Zhang /* save nonzero entries in k-th row of U ... */ 2640a1723e09SHong Zhang 264149b5e25fSSatish Balay /* invert diagonal block */ 2642a1723e09SHong Zhang diag = ba+k*9; 2643a1723e09SHong Zhang ierr = PetscMemcpy(diag,dk,9*sizeof(MatScalar));CHKERRQ(ierr); 2644a1723e09SHong Zhang ierr = Kernel_A_gets_inverse_A_3(diag);CHKERRQ(ierr); 2645a1723e09SHong Zhang 2646a1723e09SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2647a1723e09SHong Zhang if (jmin < jmax) { 2648a1723e09SHong Zhang for (j=jmin; j<jmax; j++){ 2649a1723e09SHong Zhang vj = bj[j]; /* block col. index of U */ 2650a1723e09SHong Zhang u = ba + j*9; 2651a1723e09SHong Zhang rtmp_ptr = rtmp + vj*9; 2652a1723e09SHong Zhang for (k1=0; k1<9; k1++){ 2653a1723e09SHong Zhang *u++ = *rtmp_ptr; 2654a1723e09SHong Zhang *rtmp_ptr++ = 0.0; 2655a1723e09SHong Zhang } 2656a1723e09SHong Zhang } 2657a1723e09SHong Zhang 2658a1723e09SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2659a1723e09SHong Zhang il[k] = jmin; 2660a1723e09SHong Zhang i = bj[jmin]; 2661a1723e09SHong Zhang jl[k] = jl[i]; jl[i] = k; 2662a1723e09SHong Zhang } 266349b5e25fSSatish Balay } 266449b5e25fSSatish Balay 266549b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2666a1723e09SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2667a1723e09SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2668a1723e09SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2669a1723e09SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2670a1723e09SHong Zhang if (a->permute){ 2671a1723e09SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 2672a1723e09SHong Zhang } 2673a1723e09SHong Zhang 2674a1723e09SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 2675a1723e09SHong Zhang C->factor = FACTOR_CHOLESKY; 267649b5e25fSSatish Balay C->assembled = PETSC_TRUE; 26777a05c331SHong Zhang C->preallocated = PETSC_TRUE; 2678effa298cSHong Zhang PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 267949b5e25fSSatish Balay PetscFunctionReturn(0); 268049b5e25fSSatish Balay } 2681a1723e09SHong Zhang 268249b5e25fSSatish Balay /* 268349b5e25fSSatish Balay Version for when blocks are 3 by 3 Using natural ordering 268449b5e25fSSatish Balay */ 268549b5e25fSSatish Balay #undef __FUNC__ 26866f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering" 26876f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering(Mat A,Mat *B) 268849b5e25fSSatish Balay { 268949b5e25fSSatish Balay Mat C = *B; 2690ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 269171149678SHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 269271149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2693ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2694ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 269549b5e25fSSatish Balay 269649b5e25fSSatish Balay PetscFunctionBegin; 2697*671cb588SHong Zhang 2698ab27746eSHong Zhang /* initialization */ 2699ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(9*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2700ab27746eSHong Zhang ierr = PetscMemzero(rtmp,9*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2701ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2702ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2703ab27746eSHong Zhang for (i=0; i<mbs; i++) { 2704ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 270549b5e25fSSatish Balay } 2706ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(dk); 2707ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(uik); 270849b5e25fSSatish Balay 2709ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 271049b5e25fSSatish Balay 2711ab27746eSHong Zhang /* for each row k */ 2712ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 271349b5e25fSSatish Balay 2714ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 2715ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 2716ab27746eSHong Zhang if (jmin < jmax) { 2717ab27746eSHong Zhang ap = aa + jmin*9; 2718ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 2719ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 2720ab27746eSHong Zhang rtmp_ptr = rtmp + vj*9; 2721ab27746eSHong Zhang for (i=0; i<9; i++) *rtmp_ptr++ = *ap++; 2722ab27746eSHong Zhang } 2723ab27746eSHong Zhang } 272449b5e25fSSatish Balay 2725ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2726ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*9,9*sizeof(MatScalar));CHKERRQ(ierr); 2727ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 272849b5e25fSSatish Balay 2729ab27746eSHong Zhang while (i < mbs){ 2730ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2731ab27746eSHong Zhang 2732ab27746eSHong Zhang /* compute multiplier */ 2733ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2734ab27746eSHong Zhang 2735ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2736ab27746eSHong Zhang diag = ba + i*9; 2737ab27746eSHong Zhang u = ba + ili*9; 2738ab27746eSHong Zhang 2739ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[3]*u[1] + diag[6]*u[2]); 2740ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[4]*u[1] + diag[7]*u[2]); 2741ab27746eSHong Zhang uik[2] = -(diag[2]*u[0] + diag[5]*u[1] + diag[8]*u[2]); 2742ab27746eSHong Zhang 2743ab27746eSHong Zhang uik[3] = -(diag[0]*u[3] + diag[3]*u[4] + diag[6]*u[5]); 2744ab27746eSHong Zhang uik[4] = -(diag[1]*u[3] + diag[4]*u[4] + diag[7]*u[5]); 2745ab27746eSHong Zhang uik[5] = -(diag[2]*u[3] + diag[5]*u[4] + diag[8]*u[5]); 2746ab27746eSHong Zhang 2747ab27746eSHong Zhang uik[6] = -(diag[0]*u[6] + diag[3]*u[7] + diag[6]*u[8]); 2748ab27746eSHong Zhang uik[7] = -(diag[1]*u[6] + diag[4]*u[7] + diag[7]*u[8]); 2749ab27746eSHong Zhang uik[8] = -(diag[2]*u[6] + diag[5]*u[7] + diag[8]*u[8]); 2750ab27746eSHong Zhang 2751ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2752ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2753ab27746eSHong Zhang dk[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2754ab27746eSHong Zhang dk[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2755ab27746eSHong Zhang 2756ab27746eSHong Zhang dk[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2757ab27746eSHong Zhang dk[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2758ab27746eSHong Zhang dk[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2759ab27746eSHong Zhang 2760ab27746eSHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2761ab27746eSHong Zhang dk[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2762ab27746eSHong Zhang dk[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 2763ab27746eSHong Zhang 2764ab27746eSHong Zhang /* update -U(i,k) */ 2765ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*9,uik,9*sizeof(MatScalar));CHKERRQ(ierr); 2766ab27746eSHong Zhang 2767ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2768ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2769ab27746eSHong Zhang if (jmin < jmax){ 2770ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2771ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2772ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*9; 2773ab27746eSHong Zhang u = ba + j*9; 2774ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2775ab27746eSHong Zhang rtmp_ptr[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2776ab27746eSHong Zhang rtmp_ptr[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2777ab27746eSHong Zhang 2778ab27746eSHong Zhang rtmp_ptr[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2779ab27746eSHong Zhang rtmp_ptr[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2780ab27746eSHong Zhang rtmp_ptr[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2781ab27746eSHong Zhang 2782ab27746eSHong Zhang rtmp_ptr[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2783ab27746eSHong Zhang rtmp_ptr[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2784ab27746eSHong Zhang rtmp_ptr[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 278549b5e25fSSatish Balay } 2786ab27746eSHong Zhang 2787ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2788ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2789ab27746eSHong Zhang j = bj[jmin]; 2790ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 279149b5e25fSSatish Balay } 2792ab27746eSHong Zhang i = nexti; 279349b5e25fSSatish Balay } 2794ab27746eSHong Zhang 2795ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2796ab27746eSHong Zhang 279749b5e25fSSatish Balay /* invert diagonal block */ 2798ab27746eSHong Zhang diag = ba+k*9; 2799ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,9*sizeof(MatScalar));CHKERRQ(ierr); 2800ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_3(diag);CHKERRQ(ierr); 2801ab27746eSHong Zhang 2802ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2803ab27746eSHong Zhang if (jmin < jmax) { 2804ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2805ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2806ab27746eSHong Zhang u = ba + j*9; 2807ab27746eSHong Zhang rtmp_ptr = rtmp + vj*9; 2808ab27746eSHong Zhang for (k1=0; k1<9; k1++){ 2809ab27746eSHong Zhang *u++ = *rtmp_ptr; 2810ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2811ab27746eSHong Zhang } 2812ab27746eSHong Zhang } 2813ab27746eSHong Zhang 2814ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2815ab27746eSHong Zhang il[k] = jmin; 2816ab27746eSHong Zhang i = bj[jmin]; 2817ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2818ab27746eSHong Zhang } 281949b5e25fSSatish Balay } 282049b5e25fSSatish Balay 282149b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2822ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2823ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2824ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2825ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2826ab27746eSHong Zhang 2827effa298cSHong Zhang C->factor = FACTOR_CHOLESKY; 282849b5e25fSSatish Balay C->assembled = PETSC_TRUE; 28295c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 283049b5e25fSSatish Balay PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 283149b5e25fSSatish Balay PetscFunctionReturn(0); 283249b5e25fSSatish Balay } 283349b5e25fSSatish Balay 283449b5e25fSSatish Balay /* 2835fcf159c0SHong Zhang Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP. 2836cc0c071aSHong Zhang Version for blocks 2 by 2. 283749b5e25fSSatish Balay */ 283849b5e25fSSatish Balay #undef __FUNC__ 28396f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_2" 28406f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_2(Mat A,Mat *B) 284149b5e25fSSatish Balay { 284249b5e25fSSatish Balay Mat C = *B; 284391602c66SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2844cc0c071aSHong Zhang IS perm = b->row; 2845cc0c071aSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2846cc0c071aSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2847a1723e09SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2848cc0c071aSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 284949b5e25fSSatish Balay 285049b5e25fSSatish Balay PetscFunctionBegin; 2851*671cb588SHong Zhang 285291602c66SHong Zhang /* initialization */ 285391602c66SHong Zhang /* il and jl record the first nonzero element in each row of the accessing 285491602c66SHong Zhang window U(0:k, k:mbs-1). 285591602c66SHong Zhang jl: list of rows to be added to uneliminated rows 285691602c66SHong Zhang i>= k: jl(i) is the first row to be added to row i 285791602c66SHong Zhang i< k: jl(i) is the row following row i in some list of rows 285891602c66SHong Zhang jl(i) = mbs indicates the end of a list 285991602c66SHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 286091602c66SHong Zhang row i of U */ 2861cc0c071aSHong Zhang rtmp = (MatScalar*)PetscMalloc(4*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2862cc0c071aSHong Zhang ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr); 286391602c66SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 286491602c66SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 286591602c66SHong Zhang for (i=0; i<mbs; i++) { 28663845f261SHong Zhang jl[i] = mbs; il[0] = 0; 286791602c66SHong Zhang } 2868cc0c071aSHong Zhang dk = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(dk); 2869cc0c071aSHong Zhang uik = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(uik); 2870cc0c071aSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2871cc0c071aSHong Zhang 2872cc0c071aSHong Zhang /* check permutation */ 2873cc0c071aSHong Zhang if (!a->permute){ 2874cc0c071aSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2875cc0c071aSHong Zhang } else { 2876cc0c071aSHong Zhang ai = a->inew; aj = a->jnew; 2877cc0c071aSHong Zhang aa = (MatScalar*)PetscMalloc(4*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2878cc0c071aSHong Zhang ierr = PetscMemcpy(aa,a->a,4*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2879cc0c071aSHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2880cc0c071aSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2881cc0c071aSHong Zhang 2882cc0c071aSHong Zhang for (i=0; i<mbs; i++){ 2883cc0c071aSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2884cc0c071aSHong Zhang for (j=jmin; j<jmax; j++){ 2885cc0c071aSHong Zhang while (a2anew[j] != j){ 2886cc0c071aSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2887cc0c071aSHong Zhang for (k1=0; k1<4; k1++){ 2888cc0c071aSHong Zhang dk[k1] = aa[k*4+k1]; 2889cc0c071aSHong Zhang aa[k*4+k1] = aa[j*4+k1]; 2890cc0c071aSHong Zhang aa[j*4+k1] = dk[k1]; 2891cc0c071aSHong Zhang } 2892cc0c071aSHong Zhang } 2893cc0c071aSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2894cc0c071aSHong Zhang if (i > aj[j]){ 2895a1723e09SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2896cc0c071aSHong Zhang ap = aa + j*4; /* ptr to the beginning of the block */ 2897cc0c071aSHong Zhang dk[1] = ap[1]; /* swap ap[1] and ap[2] */ 2898cc0c071aSHong Zhang ap[1] = ap[2]; 2899cc0c071aSHong Zhang ap[2] = dk[1]; 2900cc0c071aSHong Zhang } 2901cc0c071aSHong Zhang } 2902cc0c071aSHong Zhang } 2903cc0c071aSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2904cc0c071aSHong Zhang } 29053845f261SHong Zhang 290691602c66SHong Zhang /* for each row k */ 290791602c66SHong Zhang for (k = 0; k<mbs; k++){ 290891602c66SHong Zhang 290991602c66SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2910cc0c071aSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 291191602c66SHong Zhang if (jmin < jmax) { 2912cc0c071aSHong Zhang ap = aa + jmin*4; 291391602c66SHong Zhang for (j = jmin; j < jmax; j++){ 2914cc0c071aSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2915cc0c071aSHong Zhang rtmp_ptr = rtmp + vj*4; 2916cc0c071aSHong Zhang for (i=0; i<4; i++) *rtmp_ptr++ = *ap++; 291791602c66SHong Zhang } 291891602c66SHong Zhang } 291991602c66SHong Zhang 292091602c66SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2921cc0c071aSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr); 292291602c66SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 292391602c66SHong Zhang 292491602c66SHong Zhang while (i < mbs){ 292591602c66SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 292691602c66SHong Zhang 29273845f261SHong Zhang /* compute multiplier */ 292891602c66SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 29293845f261SHong Zhang 29303845f261SHong Zhang /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */ 2931cc0c071aSHong Zhang diag = ba + i*4; 2932cc0c071aSHong Zhang u = ba + ili*4; 2933cc0c071aSHong Zhang uik[0] = -(diag[0]*u[0] + diag[2]*u[1]); 2934cc0c071aSHong Zhang uik[1] = -(diag[1]*u[0] + diag[3]*u[1]); 2935cc0c071aSHong Zhang uik[2] = -(diag[0]*u[2] + diag[2]*u[3]); 2936cc0c071aSHong Zhang uik[3] = -(diag[1]*u[2] + diag[3]*u[3]); 29373845f261SHong Zhang 29383845f261SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */ 2939cc0c071aSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1]; 2940cc0c071aSHong Zhang dk[1] += uik[2]*u[0] + uik[3]*u[1]; 2941cc0c071aSHong Zhang dk[2] += uik[0]*u[2] + uik[1]*u[3]; 2942cc0c071aSHong Zhang dk[3] += uik[2]*u[2] + uik[3]*u[3]; 29433845f261SHong Zhang 29443845f261SHong Zhang /* update -U(i,k): ba[ili] = uik */ 2945cc0c071aSHong Zhang ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr); 294691602c66SHong Zhang 294791602c66SHong Zhang /* add multiple of row i to k-th row ... */ 294891602c66SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 294991602c66SHong Zhang if (jmin < jmax){ 29503845f261SHong Zhang for (j=jmin; j<jmax; j++) { 29513845f261SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */ 2952cc0c071aSHong Zhang rtmp_ptr = rtmp + bj[j]*4; 2953cc0c071aSHong Zhang u = ba + j*4; 2954cc0c071aSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1]; 2955cc0c071aSHong Zhang rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1]; 2956cc0c071aSHong Zhang rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3]; 2957cc0c071aSHong Zhang rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3]; 29583845f261SHong Zhang } 29593845f261SHong Zhang 296091602c66SHong Zhang /* ... add i to row list for next nonzero entry */ 296191602c66SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 296291602c66SHong Zhang j = bj[jmin]; 296391602c66SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 296491602c66SHong Zhang } 2965a1723e09SHong Zhang i = nexti; 296691602c66SHong Zhang } 2967cc0c071aSHong Zhang 296891602c66SHong Zhang /* save nonzero entries in k-th row of U ... */ 29693845f261SHong Zhang 2970cc0c071aSHong Zhang /* invert diagonal block */ 2971cc0c071aSHong Zhang diag = ba+k*4; 2972cc0c071aSHong Zhang ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr); 29733845f261SHong Zhang ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr); 29743845f261SHong Zhang 297591602c66SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 297691602c66SHong Zhang if (jmin < jmax) { 297791602c66SHong Zhang for (j=jmin; j<jmax; j++){ 2978cc0c071aSHong Zhang vj = bj[j]; /* block col. index of U */ 2979cc0c071aSHong Zhang u = ba + j*4; 2980cc0c071aSHong Zhang rtmp_ptr = rtmp + vj*4; 2981cc0c071aSHong Zhang for (k1=0; k1<4; k1++){ 2982cc0c071aSHong Zhang *u++ = *rtmp_ptr; 2983cc0c071aSHong Zhang *rtmp_ptr++ = 0.0; 2984cc0c071aSHong Zhang } 2985cc0c071aSHong Zhang } 29863845f261SHong Zhang 298791602c66SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 298891602c66SHong Zhang il[k] = jmin; 298991602c66SHong Zhang i = bj[jmin]; 299091602c66SHong Zhang jl[k] = jl[i]; jl[i] = k; 299191602c66SHong Zhang } 299291602c66SHong Zhang } 29933845f261SHong Zhang 299449b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 299591602c66SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 299691602c66SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 29973845f261SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 29983845f261SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 299991602c66SHong Zhang if (a->permute) { 300091602c66SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 300191602c66SHong Zhang } 3002cc0c071aSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 300391602c66SHong Zhang C->factor = FACTOR_CHOLESKY; 300449b5e25fSSatish Balay C->assembled = PETSC_TRUE; 30055c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 3006effa298cSHong Zhang PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 300749b5e25fSSatish Balay PetscFunctionReturn(0); 300849b5e25fSSatish Balay } 300991602c66SHong Zhang 301049b5e25fSSatish Balay /* 301149b5e25fSSatish Balay Version for when blocks are 2 by 2 Using natural ordering 301249b5e25fSSatish Balay */ 301349b5e25fSSatish Balay #undef __FUNC__ 30146f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering" 30156f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering(Mat A,Mat *B) 301649b5e25fSSatish Balay { 301749b5e25fSSatish Balay Mat C = *B; 3018ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 3019ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 302071149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 3021ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 3022ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 302349b5e25fSSatish Balay 302449b5e25fSSatish Balay PetscFunctionBegin; 3025*671cb588SHong Zhang 3026ab27746eSHong Zhang /* initialization */ 3027ab27746eSHong Zhang /* il and jl record the first nonzero element in each row of the accessing 3028ab27746eSHong Zhang window U(0:k, k:mbs-1). 3029ab27746eSHong Zhang jl: list of rows to be added to uneliminated rows 3030ab27746eSHong Zhang i>= k: jl(i) is the first row to be added to row i 3031ab27746eSHong Zhang i< k: jl(i) is the row following row i in some list of rows 3032ab27746eSHong Zhang jl(i) = mbs indicates the end of a list 3033ab27746eSHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 3034ab27746eSHong Zhang row i of U */ 3035ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(4*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 3036ab27746eSHong Zhang ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr); 3037ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 3038ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 3039ab27746eSHong Zhang for (i=0; i<mbs; i++) { 3040ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 304149b5e25fSSatish Balay } 3042ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(dk); 3043ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(uik); 3044ab27746eSHong Zhang 3045ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 3046ab27746eSHong Zhang 3047ab27746eSHong Zhang /* for each row k */ 3048ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 3049ab27746eSHong Zhang 3050ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 3051ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 3052ab27746eSHong Zhang if (jmin < jmax) { 3053ab27746eSHong Zhang ap = aa + jmin*4; 3054ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 3055ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 3056ab27746eSHong Zhang rtmp_ptr = rtmp + vj*4; 3057ab27746eSHong Zhang for (i=0; i<4; i++) *rtmp_ptr++ = *ap++; 305849b5e25fSSatish Balay } 305949b5e25fSSatish Balay } 3060ab27746eSHong Zhang 3061ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 3062ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr); 3063ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 3064ab27746eSHong Zhang 3065ab27746eSHong Zhang while (i < mbs){ 3066ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 3067ab27746eSHong Zhang 3068ab27746eSHong Zhang /* compute multiplier */ 3069ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 3070ab27746eSHong Zhang 3071ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */ 3072ab27746eSHong Zhang diag = ba + i*4; 3073ab27746eSHong Zhang u = ba + ili*4; 3074ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[2]*u[1]); 3075ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[3]*u[1]); 3076ab27746eSHong Zhang uik[2] = -(diag[0]*u[2] + diag[2]*u[3]); 3077ab27746eSHong Zhang uik[3] = -(diag[1]*u[2] + diag[3]*u[3]); 3078ab27746eSHong Zhang 3079ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */ 3080ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1]; 3081ab27746eSHong Zhang dk[1] += uik[2]*u[0] + uik[3]*u[1]; 3082ab27746eSHong Zhang dk[2] += uik[0]*u[2] + uik[1]*u[3]; 3083ab27746eSHong Zhang dk[3] += uik[2]*u[2] + uik[3]*u[3]; 3084ab27746eSHong Zhang 3085ab27746eSHong Zhang /* update -U(i,k): ba[ili] = uik */ 3086ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr); 3087ab27746eSHong Zhang 3088ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 3089ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 3090ab27746eSHong Zhang if (jmin < jmax){ 3091ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 3092ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */ 3093ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*4; 3094ab27746eSHong Zhang u = ba + j*4; 3095ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1]; 3096ab27746eSHong Zhang rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1]; 3097ab27746eSHong Zhang rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3]; 3098ab27746eSHong Zhang rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3]; 309949b5e25fSSatish Balay } 3100ab27746eSHong Zhang 3101ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 3102ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 3103ab27746eSHong Zhang j = bj[jmin]; 3104ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 310549b5e25fSSatish Balay } 3106ab27746eSHong Zhang i = nexti; 310749b5e25fSSatish Balay } 3108ab27746eSHong Zhang 3109ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 3110ab27746eSHong Zhang 311149b5e25fSSatish Balay /* invert diagonal block */ 3112ab27746eSHong Zhang diag = ba+k*4; 3113ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr); 3114ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr); 3115ab27746eSHong Zhang 3116ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 3117ab27746eSHong Zhang if (jmin < jmax) { 3118ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 3119ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 3120ab27746eSHong Zhang u = ba + j*4; 3121ab27746eSHong Zhang rtmp_ptr = rtmp + vj*4; 3122ab27746eSHong Zhang for (k1=0; k1<4; k1++){ 3123ab27746eSHong Zhang *u++ = *rtmp_ptr; 3124ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 3125ab27746eSHong Zhang } 3126ab27746eSHong Zhang } 3127ab27746eSHong Zhang 3128ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 3129ab27746eSHong Zhang il[k] = jmin; 3130ab27746eSHong Zhang i = bj[jmin]; 3131ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 3132ab27746eSHong Zhang } 313349b5e25fSSatish Balay } 313449b5e25fSSatish Balay 313549b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 3136ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 3137ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 3138ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 3139ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 3140ab27746eSHong Zhang 3141ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 314249b5e25fSSatish Balay C->assembled = PETSC_TRUE; 31435c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 314449b5e25fSSatish Balay PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 314549b5e25fSSatish Balay PetscFunctionReturn(0); 314649b5e25fSSatish Balay } 314749b5e25fSSatish Balay 314849b5e25fSSatish Balay /* 31495c0bcdfcSHong Zhang Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP. 315091602c66SHong Zhang Version for blocks are 1 by 1. 315149b5e25fSSatish Balay */ 315249b5e25fSSatish Balay #undef __FUNC__ 31536f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_1" 31546f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_1(Mat A,Mat *B) 315549b5e25fSSatish Balay { 315649b5e25fSSatish Balay Mat C = *B; 315749b5e25fSSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 315849b5e25fSSatish Balay IS ip = b->row; 3159351d0355SHong Zhang int *rip,ierr,i,j,mbs = a->mbs,*bi = b->i,*bj = b->j; 3160cb718733SHong Zhang int *ai,*aj,*r; 3161ab27746eSHong Zhang int k,jmin,jmax,*jl,*il,vj,nexti,ili; 3162066653e3SSatish Balay MatScalar *rtmp; 31632d007305SHong Zhang MatScalar *ba = b->a,*aa,ak; 316449b5e25fSSatish Balay MatScalar dk,uikdi; 316549b5e25fSSatish Balay 316649b5e25fSSatish Balay PetscFunctionBegin; 316749b5e25fSSatish Balay ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 3168cb718733SHong Zhang if (!a->permute){ 31692d007305SHong Zhang ai = a->i; aj = a->j; aa = a->a; 31702d007305SHong Zhang } else { 31712d007305SHong Zhang ai = a->inew; aj = a->jnew; 31722d007305SHong Zhang aa = (MatScalar*)PetscMalloc(ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 3173cb718733SHong Zhang ierr = PetscMemcpy(aa,a->a,ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 31742d007305SHong Zhang r = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(r); 31752d007305SHong Zhang ierr= PetscMemcpy(r,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 31762d007305SHong Zhang 31772d007305SHong Zhang jmin = ai[0]; jmax = ai[mbs]; 31782d007305SHong Zhang for (j=jmin; j<jmax; j++){ 3179cb718733SHong Zhang while (r[j] != j){ 31802d007305SHong Zhang k = r[j]; r[j] = r[k]; r[k] = k; 31812d007305SHong Zhang ak = aa[k]; aa[k] = aa[j]; aa[j] = ak; 31822d007305SHong Zhang } 31832d007305SHong Zhang } 31842d007305SHong Zhang ierr = PetscFree(r);CHKERRA(ierr); 31852d007305SHong Zhang } 31862d007305SHong Zhang 318791602c66SHong Zhang /* initialization */ 318849b5e25fSSatish Balay /* il and jl record the first nonzero element in each row of the accessing 318949b5e25fSSatish Balay window U(0:k, k:mbs-1). 319049b5e25fSSatish Balay jl: list of rows to be added to uneliminated rows 319149b5e25fSSatish Balay i>= k: jl(i) is the first row to be added to row i 319249b5e25fSSatish Balay i< k: jl(i) is the row following row i in some list of rows 319349b5e25fSSatish Balay jl(i) = mbs indicates the end of a list 319449b5e25fSSatish Balay il(i): points to the first nonzero element in columns k,...,mbs-1 of 319549b5e25fSSatish Balay row i of U */ 319649b5e25fSSatish Balay rtmp = (MatScalar*)PetscMalloc(mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 319749b5e25fSSatish Balay il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 319849b5e25fSSatish Balay jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 319949b5e25fSSatish Balay for (i=0; i<mbs; i++) { 320049b5e25fSSatish Balay rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0; 320149b5e25fSSatish Balay } 320249b5e25fSSatish Balay 320391602c66SHong Zhang /* for each row k */ 320449b5e25fSSatish Balay for (k = 0; k<mbs; k++){ 320549b5e25fSSatish Balay 320691602c66SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 320749b5e25fSSatish Balay jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 320849b5e25fSSatish Balay if (jmin < jmax) { 320949b5e25fSSatish Balay for (j = jmin; j < jmax; j++){ 3210351d0355SHong Zhang vj = rip[aj[j]]; 3211ab27746eSHong Zhang rtmp[vj] = aa[j]; 321249b5e25fSSatish Balay } 321349b5e25fSSatish Balay } 321449b5e25fSSatish Balay 321591602c66SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 321649b5e25fSSatish Balay dk = rtmp[k]; 321749b5e25fSSatish Balay i = jl[k]; /* first row to be added to k_th row */ 321849b5e25fSSatish Balay 321949b5e25fSSatish Balay while (i < mbs){ 322049b5e25fSSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 322149b5e25fSSatish Balay 322291602c66SHong Zhang /* compute multiplier, update D(k) and U(i,k) */ 322349b5e25fSSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 322449b5e25fSSatish Balay uikdi = - ba[ili]*ba[i]; 322549b5e25fSSatish Balay dk += uikdi*ba[ili]; 3226658e7b3eSHong Zhang ba[ili] = uikdi; /* -U(i,k) */ 322749b5e25fSSatish Balay 322891602c66SHong Zhang /* add multiple of row i to k-th row ... */ 322949b5e25fSSatish Balay jmin = ili + 1; jmax = bi[i+1]; 323049b5e25fSSatish Balay if (jmin < jmax){ 323149b5e25fSSatish Balay for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 323291602c66SHong Zhang /* ... add i to row list for next nonzero entry */ 323349b5e25fSSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 323449b5e25fSSatish Balay j = bj[jmin]; 323549b5e25fSSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 323649b5e25fSSatish Balay } 3237ab27746eSHong Zhang i = nexti; 323849b5e25fSSatish Balay } 323949b5e25fSSatish Balay 324091602c66SHong Zhang /* check for zero pivot and save diagoanl element */ 324149b5e25fSSatish Balay if (dk == 0.0){ 324229bbc08cSBarry Smith SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot"); 3243*671cb588SHong Zhang /* 32448be8c0c7SHong Zhang }else if (PetscRealPart(dk) < 0){ 32458be8c0c7SHong Zhang ierr = PetscPrintf(PETSC_COMM_SELF,"Negative pivot: d[%d] = %g\n",k,dk); 3246*671cb588SHong Zhang */ 324749b5e25fSSatish Balay } 324849b5e25fSSatish Balay 324991602c66SHong Zhang /* save nonzero entries in k-th row of U ... */ 3250f3dd7b2dSKris Buschelman ba[k] = 1.0/dk; 325149b5e25fSSatish Balay jmin = bi[k]; jmax = bi[k+1]; 325249b5e25fSSatish Balay if (jmin < jmax) { 325349b5e25fSSatish Balay for (j=jmin; j<jmax; j++){ 3254cc0c071aSHong Zhang vj = bj[j]; ba[j] = rtmp[vj]; rtmp[vj] = 0.0; 325549b5e25fSSatish Balay } 325691602c66SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 325749b5e25fSSatish Balay il[k] = jmin; 325849b5e25fSSatish Balay i = bj[jmin]; 325949b5e25fSSatish Balay jl[k] = jl[i]; jl[i] = k; 326049b5e25fSSatish Balay } 326149b5e25fSSatish Balay } 326249b5e25fSSatish Balay 326349b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 326449b5e25fSSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 326549b5e25fSSatish Balay ierr = PetscFree(jl);CHKERRQ(ierr); 3266cb718733SHong Zhang if (a->permute){ 3267cb718733SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 3268cb718733SHong Zhang } 326949b5e25fSSatish Balay 327049b5e25fSSatish Balay ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 32718be8c0c7SHong Zhang C->factor = FACTOR_CHOLESKY; 327249b5e25fSSatish Balay C->assembled = PETSC_TRUE; 32735c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 327449b5e25fSSatish Balay PLogFlops(b->mbs); 327549b5e25fSSatish Balay PetscFunctionReturn(0); 327649b5e25fSSatish Balay } 327749b5e25fSSatish Balay 3278*671cb588SHong Zhang /* 3279*671cb588SHong Zhang Version for when blocks are 1 by 1 Using natural ordering 3280*671cb588SHong Zhang */ 3281*671cb588SHong Zhang #undef __FUNC__ 3282*671cb588SHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering" 3283*671cb588SHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering(Mat A,Mat *B) 3284*671cb588SHong Zhang { 3285*671cb588SHong Zhang Mat C = *B; 3286*671cb588SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 3287*671cb588SHong Zhang int ierr,i,j,mbs = a->mbs,*bi = b->i,*bj = b->j; 3288*671cb588SHong Zhang int *ai,*aj,*r; 3289*671cb588SHong Zhang int k,jmin,jmax,*jl,*il,vj,nexti,ili; 3290*671cb588SHong Zhang MatScalar *rtmp; 3291*671cb588SHong Zhang MatScalar *ba = b->a,*aa,ak; 3292*671cb588SHong Zhang MatScalar dk,uikdi; 3293*671cb588SHong Zhang 3294*671cb588SHong Zhang PetscFunctionBegin; 3295*671cb588SHong Zhang 3296*671cb588SHong Zhang /* initialization */ 3297*671cb588SHong Zhang /* il and jl record the first nonzero element in each row of the accessing 3298*671cb588SHong Zhang window U(0:k, k:mbs-1). 3299*671cb588SHong Zhang jl: list of rows to be added to uneliminated rows 3300*671cb588SHong Zhang i>= k: jl(i) is the first row to be added to row i 3301*671cb588SHong Zhang i< k: jl(i) is the row following row i in some list of rows 3302*671cb588SHong Zhang jl(i) = mbs indicates the end of a list 3303*671cb588SHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 3304*671cb588SHong Zhang row i of U */ 3305*671cb588SHong Zhang rtmp = (MatScalar*)PetscMalloc(mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 3306*671cb588SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 3307*671cb588SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 3308*671cb588SHong Zhang for (i=0; i<mbs; i++) { 3309*671cb588SHong Zhang rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0; 3310*671cb588SHong Zhang } 3311*671cb588SHong Zhang 3312*671cb588SHong Zhang ai = a->i; aj = a->j; aa = a->a; 3313*671cb588SHong Zhang 3314*671cb588SHong Zhang /* for each row k */ 3315*671cb588SHong Zhang for (k = 0; k<mbs; k++){ 3316*671cb588SHong Zhang 3317*671cb588SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 3318*671cb588SHong Zhang jmin = ai[k]; jmax = ai[k+1]; 3319*671cb588SHong Zhang if (jmin < jmax) { 3320*671cb588SHong Zhang for (j = jmin; j < jmax; j++){ 3321*671cb588SHong Zhang vj = aj[j]; 3322*671cb588SHong Zhang rtmp[vj] = aa[j]; 3323*671cb588SHong Zhang } 3324*671cb588SHong Zhang } 3325*671cb588SHong Zhang 3326*671cb588SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 3327*671cb588SHong Zhang dk = rtmp[k]; 3328*671cb588SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 3329*671cb588SHong Zhang 3330*671cb588SHong Zhang while (i < mbs){ 3331*671cb588SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 3332*671cb588SHong Zhang 3333*671cb588SHong Zhang /* compute multiplier, update D(k) and U(i,k) */ 3334*671cb588SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 3335*671cb588SHong Zhang uikdi = - ba[ili]*ba[i]; 3336*671cb588SHong Zhang dk += uikdi*ba[ili]; 3337*671cb588SHong Zhang ba[ili] = uikdi; /* -U(i,k) */ 3338*671cb588SHong Zhang 3339*671cb588SHong Zhang /* add multiple of row i to k-th row ... */ 3340*671cb588SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 3341*671cb588SHong Zhang if (jmin < jmax){ 3342*671cb588SHong Zhang for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 3343*671cb588SHong Zhang /* ... add i to row list for next nonzero entry */ 3344*671cb588SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 3345*671cb588SHong Zhang j = bj[jmin]; 3346*671cb588SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 3347*671cb588SHong Zhang } 3348*671cb588SHong Zhang i = nexti; 3349*671cb588SHong Zhang } 3350*671cb588SHong Zhang 3351*671cb588SHong Zhang /* check for zero pivot and save diagoanl element */ 3352*671cb588SHong Zhang if (dk == 0.0){ 3353*671cb588SHong Zhang SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot"); 3354*671cb588SHong Zhang /* 3355*671cb588SHong Zhang }else if (PetscRealPart(dk) < 0){ 3356*671cb588SHong Zhang ierr = PetscPrintf(PETSC_COMM_SELF,"Negative pivot: d[%d] = %g\n",k,dk); 3357*671cb588SHong Zhang */ 3358*671cb588SHong Zhang } 3359*671cb588SHong Zhang 3360*671cb588SHong Zhang /* save nonzero entries in k-th row of U ... */ 3361*671cb588SHong Zhang ba[k] = 1.0/dk; 3362*671cb588SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 3363*671cb588SHong Zhang if (jmin < jmax) { 3364*671cb588SHong Zhang for (j=jmin; j<jmax; j++){ 3365*671cb588SHong Zhang vj = bj[j]; ba[j] = rtmp[vj]; rtmp[vj] = 0.0; 3366*671cb588SHong Zhang } 3367*671cb588SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 3368*671cb588SHong Zhang il[k] = jmin; 3369*671cb588SHong Zhang i = bj[jmin]; 3370*671cb588SHong Zhang jl[k] = jl[i]; jl[i] = k; 3371*671cb588SHong Zhang } 3372*671cb588SHong Zhang } 3373*671cb588SHong Zhang 3374*671cb588SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 3375*671cb588SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 3376*671cb588SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 3377*671cb588SHong Zhang 3378*671cb588SHong Zhang C->factor = FACTOR_CHOLESKY; 3379*671cb588SHong Zhang C->assembled = PETSC_TRUE; 3380*671cb588SHong Zhang C->preallocated = PETSC_TRUE; 3381*671cb588SHong Zhang PLogFlops(b->mbs); 3382*671cb588SHong Zhang PetscFunctionReturn(0); 3383*671cb588SHong Zhang } 3384*671cb588SHong Zhang 338549b5e25fSSatish Balay #undef __FUNC__ 33866f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactor_SeqSBAIJ" 3387b8b23757SHong Zhang int MatCholeskyFactor_SeqSBAIJ(Mat A,IS perm,PetscReal f) 338849b5e25fSSatish Balay { 33894445ddedSHong Zhang int ierr; 339049b5e25fSSatish Balay Mat C; 339149b5e25fSSatish Balay 339249b5e25fSSatish Balay PetscFunctionBegin; 3393b8b23757SHong Zhang ierr = MatCholeskyFactorSymbolic(A,perm,f,&C);CHKERRQ(ierr); 3394a4ada70bSHong Zhang ierr = MatCholeskyFactorNumeric(A,&C);CHKERRQ(ierr); 33954445ddedSHong Zhang ierr = MatHeaderCopy(A,C);CHKERRQ(ierr); 339649b5e25fSSatish Balay PetscFunctionReturn(0); 339749b5e25fSSatish Balay } 339849b5e25fSSatish Balay 339949b5e25fSSatish Balay 3400