149b5e25fSSatish Balay /* Using Modified Sparse Row (MSR) storage. 249b5e25fSSatish Balay See page 85, "Iterative Methods ..." by Saad. */ 349b5e25fSSatish Balay 4*b0a32e0cSBarry Smith /*$Id: sbaijfact.c,v 1.51 2000/11/28 17:29:37 bsmith Exp bsmith $*/ 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; 23671cb588SHong Zhang PetscTruth perm_identity; 2449b5e25fSSatish Balay 2549b5e25fSSatish Balay PetscFunctionBegin; 26671cb588SHong Zhang if (!perm) { 27671cb588SHong Zhang ierr = ISCreateStride(PETSC_COMM_SELF,mbs,0,1,&perm);CHKERRQ(ierr); 28671cb588SHong Zhang } 2977f73120SHong Zhang PetscValidHeaderSpecific(perm,IS_COOKIE); 302d007305SHong Zhang 31cb718733SHong Zhang /* check whether perm is the identity mapping */ 32671cb588SHong Zhang ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 33671cb588SHong Zhang if (!perm_identity) a->permute = PETSC_TRUE; 342d007305SHong Zhang 35671cb588SHong Zhang ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 36671cb588SHong Zhang 37671cb588SHong 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 */ 47*b0a32e0cSBarry Smith ierr = PetscMalloc((mbs+1)*sizeof(int),& iu );CHKERRQ(ierr); 4849b5e25fSSatish Balay umax = (int)(f*ai[mbs] + 1); umax += mbs + 1; 49*b0a32e0cSBarry Smith ierr = PetscMalloc(umax*sizeof(int),&( ju ));CHKERRQ(ierr); 5049b5e25fSSatish Balay iu[0] = mbs+1; 5149b5e25fSSatish Balay juptr = mbs; 52*b0a32e0cSBarry Smith ierr = PetscMalloc(mbs*sizeof(int),&( jl ));CHKERRQ(ierr); 53*b0a32e0cSBarry Smith ierr = PetscMalloc(mbs*sizeof(int),&( q ));CHKERRQ(ierr); 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 */ 120*b0a32e0cSBarry Smith ierr = PetscMalloc(umax*sizeof(int),&( jutmp ));CHKERRQ(ierr); 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]); 138*b0a32e0cSBarry Smith PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Reallocs %d Fill ratio:given %g needed %g\n",realloc,f,af); 139*b0a32e0cSBarry Smith PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Run with -pc_lu_fill %g or use \n",af); 140*b0a32e0cSBarry Smith PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:PCLUSetFill(pc,%g);\n",af); 141*b0a32e0cSBarry Smith PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:for best performance.\n"); 14249b5e25fSSatish Balay } else { 143*b0a32e0cSBarry Smith PetscLogInfo(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); 152*b0a32e0cSBarry Smith /* PetscLogObjectParent(*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); 159*b0a32e0cSBarry Smith b->a = (MatScalar*)PetscMalloc((iu[mbs]+1)*sizeof(MatScalar)*bs2);CHKERRQ(ierr); 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); 169*b0a32e0cSBarry Smith ierr = PetscMalloc((bs*mbs+bs)*sizeof(Scalar),& b->solve_work );CHKERRQ(ierr); 17049b5e25fSSatish Balay /* In b structure: Free imax, ilen, old a, old j. 17149b5e25fSSatish Balay Allocate idnew, solve_work, new a, new j */ 172*b0a32e0cSBarry Smith PetscLogObjectMemory(*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 184671cb588SHong Zhang if (perm_identity){ 185671cb588SHong Zhang switch (bs) { 186671cb588SHong Zhang case 1: 187671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering; 188671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering; 189*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=1\n"); 190671cb588SHong Zhang break; 191671cb588SHong Zhang case 2: 192671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering; 193671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_2_NaturalOrdering; 194*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=2\n"); 195671cb588SHong Zhang break; 196671cb588SHong Zhang case 3: 197671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering; 198671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_3_NaturalOrdering; 199*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:sing special in-place natural ordering factor and solve BS=3\n"); 200671cb588SHong Zhang break; 201671cb588SHong Zhang case 4: 202671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering; 203671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_4_NaturalOrdering; 204*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=4\n"); 205671cb588SHong Zhang break; 206671cb588SHong Zhang case 5: 207671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering; 208671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_5_NaturalOrdering; 209*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=5\n"); 210671cb588SHong Zhang break; 211671cb588SHong Zhang case 6: 212671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering; 213671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_6_NaturalOrdering; 214*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=6\n"); 215671cb588SHong Zhang break; 216671cb588SHong Zhang case 7: 217671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering; 218671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_7_NaturalOrdering; 219*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=7\n"); 220671cb588SHong Zhang break; 221671cb588SHong Zhang default: 222671cb588SHong Zhang (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering; 223671cb588SHong Zhang (*B)->ops->solve = MatSolve_SeqSBAIJ_N_NaturalOrdering; 224*b0a32e0cSBarry Smith PetscLogInfo(A,"MatIncompleteCholeskyFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS>7\n"); 225671cb588SHong Zhang break; 226671cb588SHong Zhang } 227671cb588SHong Zhang } 228671cb588SHong Zhang 22949b5e25fSSatish Balay PetscFunctionReturn(0); 23049b5e25fSSatish Balay } 23149b5e25fSSatish Balay 23235d8aa7fSBarry Smith 23349b5e25fSSatish Balay #undef __FUNC__ 2346f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_N" 2356f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_N(Mat A,Mat *B) 23649b5e25fSSatish Balay { 23749b5e25fSSatish Balay Mat C = *B; 2384c16a6a6SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2394c16a6a6SHong Zhang IS perm = b->row; 2404c16a6a6SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2414c16a6a6SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2424c16a6a6SHong Zhang int bs=a->bs,bs2 = a->bs2; 2434c16a6a6SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2444c16a6a6SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 24528de702eSHong Zhang MatScalar *work; 2464c16a6a6SHong Zhang int *pivots; 2474c16a6a6SHong Zhang 2484c16a6a6SHong Zhang PetscFunctionBegin; 2499706f043SHong Zhang 2504c16a6a6SHong Zhang /* initialization */ 251*b0a32e0cSBarry Smith ierr = PetscMalloc(bs2*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 2524c16a6a6SHong Zhang ierr = PetscMemzero(rtmp,bs2*mbs*sizeof(MatScalar));CHKERRQ(ierr); 253*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 25428de702eSHong Zhang jl = il + mbs; 2554c16a6a6SHong Zhang for (i=0; i<mbs; i++) { 2564c16a6a6SHong Zhang jl[i] = mbs; il[0] = 0; 2574c16a6a6SHong Zhang } 258*b0a32e0cSBarry Smith ierr = PetscMalloc((2*bs2+bs)*sizeof(MatScalar),& dk );CHKERRQ(ierr); 25928de702eSHong Zhang uik = dk + bs2; 26028de702eSHong Zhang work = uik + bs2; 261*b0a32e0cSBarry Smith ierr = PetscMalloc(bs*sizeof(int),&( pivots));CHKERRQ(ierr); 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; 270*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(bs2*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2714c16a6a6SHong Zhang ierr = PetscMemcpy(aa,a->a,bs2*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 272*b0a32e0cSBarry Smith a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 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(dk);CHKERRQ(ierr); 3834c16a6a6SHong Zhang ierr = PetscFree(pivots);CHKERRQ(ierr); 3844c16a6a6SHong Zhang if (a->permute){ 3854c16a6a6SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 3864c16a6a6SHong Zhang } 3874c16a6a6SHong Zhang 3884c16a6a6SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 3894c16a6a6SHong Zhang C->factor = FACTOR_CHOLESKY; 3904c16a6a6SHong Zhang C->assembled = PETSC_TRUE; 3914c16a6a6SHong Zhang C->preallocated = PETSC_TRUE; 392*b0a32e0cSBarry Smith PetscLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 3934c16a6a6SHong Zhang PetscFunctionReturn(0); 3944c16a6a6SHong Zhang } 395d4edadd4SHong Zhang 396671cb588SHong Zhang #undef __FUNC__ 397671cb588SHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering" 398671cb588SHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering(Mat A,Mat *B) 399671cb588SHong Zhang { 400671cb588SHong Zhang Mat C = *B; 401671cb588SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 402671cb588SHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 4031ad1de41SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 404671cb588SHong Zhang int bs=a->bs,bs2 = a->bs2; 405671cb588SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 406671cb588SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 40728de702eSHong Zhang MatScalar *work; 408671cb588SHong Zhang int *pivots; 409671cb588SHong Zhang 410671cb588SHong Zhang PetscFunctionBegin; 411671cb588SHong Zhang 412671cb588SHong Zhang /* initialization */ 413671cb588SHong Zhang 414*b0a32e0cSBarry Smith ierr = PetscMalloc(bs2*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 415671cb588SHong Zhang ierr = PetscMemzero(rtmp,bs2*mbs*sizeof(MatScalar));CHKERRQ(ierr); 416*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 41728de702eSHong Zhang jl = il + mbs; 418671cb588SHong Zhang for (i=0; i<mbs; i++) { 419671cb588SHong Zhang jl[i] = mbs; il[0] = 0; 420671cb588SHong Zhang } 421*b0a32e0cSBarry Smith ierr = PetscMalloc((2*bs2+bs)*sizeof(MatScalar),& dk );CHKERRQ(ierr); 42228de702eSHong Zhang uik = dk + bs2; 42328de702eSHong Zhang work = uik + bs2; 424*b0a32e0cSBarry Smith ierr = PetscMalloc(bs*sizeof(int),&( pivots));CHKERRQ(ierr); 425671cb588SHong Zhang 426671cb588SHong Zhang ai = a->i; aj = a->j; aa = a->a; 427671cb588SHong Zhang 428671cb588SHong Zhang /* for each row k */ 429671cb588SHong Zhang for (k = 0; k<mbs; k++){ 430671cb588SHong Zhang 431671cb588SHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 432671cb588SHong Zhang jmin = ai[k]; jmax = ai[k+1]; 433671cb588SHong Zhang if (jmin < jmax) { 434671cb588SHong Zhang ap = aa + jmin*bs2; 435671cb588SHong Zhang for (j = jmin; j < jmax; j++){ 436671cb588SHong Zhang vj = aj[j]; /* block col. index */ 437671cb588SHong Zhang rtmp_ptr = rtmp + vj*bs2; 438671cb588SHong Zhang for (i=0; i<bs2; i++) *rtmp_ptr++ = *ap++; 439671cb588SHong Zhang } 440671cb588SHong Zhang } 441671cb588SHong Zhang 442671cb588SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 443671cb588SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 444671cb588SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 445671cb588SHong Zhang 446671cb588SHong Zhang while (i < mbs){ 447671cb588SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 448671cb588SHong Zhang 449671cb588SHong Zhang /* compute multiplier */ 450671cb588SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 451671cb588SHong Zhang 452671cb588SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 453671cb588SHong Zhang diag = ba + i*bs2; 454671cb588SHong Zhang u = ba + ili*bs2; 455671cb588SHong Zhang ierr = PetscMemzero(uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 456671cb588SHong Zhang Kernel_A_gets_A_minus_B_times_C(bs,uik,diag,u); 457671cb588SHong Zhang 458671cb588SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 459671cb588SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,dk,uik,u); 460671cb588SHong Zhang 461671cb588SHong Zhang /* update -U(i,k) */ 462671cb588SHong Zhang ierr = PetscMemcpy(ba+ili*bs2,uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 463671cb588SHong Zhang 464671cb588SHong Zhang /* add multiple of row i to k-th row ... */ 465671cb588SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 466671cb588SHong Zhang if (jmin < jmax){ 467671cb588SHong Zhang for (j=jmin; j<jmax; j++) { 468671cb588SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 469671cb588SHong Zhang rtmp_ptr = rtmp + bj[j]*bs2; 470671cb588SHong Zhang u = ba + j*bs2; 471671cb588SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,rtmp_ptr,uik,u); 472671cb588SHong Zhang } 473671cb588SHong Zhang 474671cb588SHong Zhang /* ... add i to row list for next nonzero entry */ 475671cb588SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 476671cb588SHong Zhang j = bj[jmin]; 477671cb588SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 478671cb588SHong Zhang } 479671cb588SHong Zhang i = nexti; 480671cb588SHong Zhang } 481671cb588SHong Zhang 482671cb588SHong Zhang /* save nonzero entries in k-th row of U ... */ 483671cb588SHong Zhang 484671cb588SHong Zhang /* invert diagonal block */ 485671cb588SHong Zhang diag = ba+k*bs2; 486671cb588SHong Zhang ierr = PetscMemcpy(diag,dk,bs2*sizeof(MatScalar));CHKERRQ(ierr); 487671cb588SHong Zhang Kernel_A_gets_inverse_A(bs,diag,pivots,work); 488671cb588SHong Zhang 489671cb588SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 490671cb588SHong Zhang if (jmin < jmax) { 491671cb588SHong Zhang for (j=jmin; j<jmax; j++){ 492671cb588SHong Zhang vj = bj[j]; /* block col. index of U */ 493671cb588SHong Zhang u = ba + j*bs2; 494671cb588SHong Zhang rtmp_ptr = rtmp + vj*bs2; 495671cb588SHong Zhang for (k1=0; k1<bs2; k1++){ 496671cb588SHong Zhang *u++ = *rtmp_ptr; 497671cb588SHong Zhang *rtmp_ptr++ = 0.0; 498671cb588SHong Zhang } 499671cb588SHong Zhang } 500671cb588SHong Zhang 501671cb588SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 502671cb588SHong Zhang il[k] = jmin; 503671cb588SHong Zhang i = bj[jmin]; 504671cb588SHong Zhang jl[k] = jl[i]; jl[i] = k; 505671cb588SHong Zhang } 506671cb588SHong Zhang } 507671cb588SHong Zhang 508671cb588SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 509671cb588SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 510671cb588SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 511671cb588SHong Zhang ierr = PetscFree(pivots);CHKERRQ(ierr); 512671cb588SHong Zhang 513671cb588SHong Zhang C->factor = FACTOR_CHOLESKY; 514671cb588SHong Zhang C->assembled = PETSC_TRUE; 515671cb588SHong Zhang C->preallocated = PETSC_TRUE; 516*b0a32e0cSBarry Smith PetscLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 517671cb588SHong Zhang PetscFunctionReturn(0); 518671cb588SHong Zhang } 519671cb588SHong Zhang 5200aa0ce06SHong Zhang /* Version for when blocks are 7 by 7 */ 52149b5e25fSSatish Balay #undef __FUNC__ 5226f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_7" 5236f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_7(Mat A,Mat *B) 52449b5e25fSSatish Balay { 52549b5e25fSSatish Balay Mat C = *B; 5260aa0ce06SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 5270aa0ce06SHong Zhang IS perm = b->row; 5280aa0ce06SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 5290aa0ce06SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 5300aa0ce06SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 5310aa0ce06SHong Zhang MatScalar *u,*d,*w,*wp; 53249b5e25fSSatish Balay 53349b5e25fSSatish Balay PetscFunctionBegin; 5340aa0ce06SHong Zhang /* initialization */ 535*b0a32e0cSBarry Smith ierr = PetscMalloc(49*mbs*sizeof(MatScalar),&( w ));CHKERRQ(ierr); 5360aa0ce06SHong Zhang ierr = PetscMemzero(w,49*mbs*sizeof(MatScalar));CHKERRQ(ierr); 537*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 53828de702eSHong Zhang jl = il + mbs; 5390aa0ce06SHong Zhang for (i=0; i<mbs; i++) { 5400aa0ce06SHong Zhang jl[i] = mbs; il[0] = 0; 54149b5e25fSSatish Balay } 542*b0a32e0cSBarry Smith ierr = PetscMalloc(98*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 54328de702eSHong Zhang uik = dk + 49; 5440aa0ce06SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 5450aa0ce06SHong Zhang 5460aa0ce06SHong Zhang /* check permutation */ 5470aa0ce06SHong Zhang if (!a->permute){ 5480aa0ce06SHong Zhang ai = a->i; aj = a->j; aa = a->a; 5490aa0ce06SHong Zhang } else { 5500aa0ce06SHong Zhang ai = a->inew; aj = a->jnew; 551*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(49*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 5520aa0ce06SHong Zhang ierr = PetscMemcpy(aa,a->a,49*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 553*b0a32e0cSBarry Smith a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 5540aa0ce06SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 5550aa0ce06SHong Zhang 5560aa0ce06SHong Zhang for (i=0; i<mbs; i++){ 5570aa0ce06SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 5580aa0ce06SHong Zhang for (j=jmin; j<jmax; j++){ 5590aa0ce06SHong Zhang while (a2anew[j] != j){ 5600aa0ce06SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 5610aa0ce06SHong Zhang for (k1=0; k1<49; k1++){ 5620aa0ce06SHong Zhang dk[k1] = aa[k*49+k1]; 5630aa0ce06SHong Zhang aa[k*49+k1] = aa[j*49+k1]; 5640aa0ce06SHong Zhang aa[j*49+k1] = dk[k1]; 56549b5e25fSSatish Balay } 56649b5e25fSSatish Balay } 5670aa0ce06SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 5680aa0ce06SHong Zhang if (i > aj[j]){ 5690aa0ce06SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 5700aa0ce06SHong Zhang ap = aa + j*49; /* ptr to the beginning of j-th block of aa */ 5710aa0ce06SHong Zhang for (k=0; k<49; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 5720aa0ce06SHong Zhang for (k=0; k<7; k++){ /* j-th block of aa <- dk^T */ 5730aa0ce06SHong Zhang for (k1=0; k1<7; k1++) *ap++ = dk[k + 7*k1]; 57449b5e25fSSatish Balay } 57549b5e25fSSatish Balay } 57649b5e25fSSatish Balay } 5770aa0ce06SHong Zhang } 5780aa0ce06SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 5790aa0ce06SHong Zhang } 5800aa0ce06SHong Zhang 5810aa0ce06SHong Zhang /* for each row k */ 5820aa0ce06SHong Zhang for (k = 0; k<mbs; k++){ 5830aa0ce06SHong Zhang 5840aa0ce06SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 5850aa0ce06SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 5860aa0ce06SHong Zhang if (jmin < jmax) { 5870aa0ce06SHong Zhang ap = aa + jmin*49; 5880aa0ce06SHong Zhang for (j = jmin; j < jmax; j++){ 5890aa0ce06SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 5900aa0ce06SHong Zhang wp = w + vj*49; 5910aa0ce06SHong Zhang for (i=0; i<49; i++) *wp++ = *ap++; 5920aa0ce06SHong Zhang } 5930aa0ce06SHong Zhang } 5940aa0ce06SHong Zhang 5950aa0ce06SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 5960aa0ce06SHong Zhang ierr = PetscMemcpy(dk,w+k*49,49*sizeof(MatScalar));CHKERRQ(ierr); 5970aa0ce06SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 5980aa0ce06SHong Zhang 5990aa0ce06SHong Zhang while (i < mbs){ 6000aa0ce06SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 6010aa0ce06SHong Zhang 6020aa0ce06SHong Zhang /* compute multiplier */ 6030aa0ce06SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 6040aa0ce06SHong Zhang 6050aa0ce06SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 6060aa0ce06SHong Zhang d = ba + i*49; 6070aa0ce06SHong Zhang u = ba + ili*49; 6080aa0ce06SHong Zhang 6090aa0ce06SHong 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]); 6100aa0ce06SHong 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]); 6110aa0ce06SHong 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]); 6120aa0ce06SHong 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]); 6130aa0ce06SHong 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]); 6140aa0ce06SHong 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]); 6150aa0ce06SHong 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]); 6160aa0ce06SHong Zhang 6170aa0ce06SHong 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]); 6180aa0ce06SHong 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]); 6190aa0ce06SHong 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]); 6200aa0ce06SHong 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]); 6210aa0ce06SHong 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]); 6220aa0ce06SHong 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]); 6230aa0ce06SHong 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]); 6240aa0ce06SHong Zhang 6250aa0ce06SHong 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]); 6260aa0ce06SHong 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]); 6270aa0ce06SHong 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]); 6280aa0ce06SHong 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]); 6290aa0ce06SHong 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]); 6300aa0ce06SHong 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]); 6310aa0ce06SHong 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]); 6320aa0ce06SHong Zhang 6330aa0ce06SHong 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]); 6340aa0ce06SHong 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]); 6350aa0ce06SHong 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]); 6360aa0ce06SHong 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]); 6370aa0ce06SHong 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]); 6380aa0ce06SHong 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]); 6390aa0ce06SHong 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]); 6400aa0ce06SHong Zhang 6410aa0ce06SHong 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]); 6420aa0ce06SHong 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]); 6430aa0ce06SHong 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]); 6440aa0ce06SHong 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]); 6450aa0ce06SHong 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]); 6460aa0ce06SHong 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]); 6470aa0ce06SHong 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]); 6480aa0ce06SHong Zhang 6490aa0ce06SHong 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]); 6500aa0ce06SHong 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]); 6510aa0ce06SHong 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]); 6520aa0ce06SHong 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]); 6530aa0ce06SHong 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]); 6540aa0ce06SHong 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]); 6550aa0ce06SHong 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]); 6560aa0ce06SHong Zhang 6570aa0ce06SHong 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]); 6580aa0ce06SHong 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]); 6590aa0ce06SHong 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]); 6600aa0ce06SHong 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]); 6610aa0ce06SHong 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]); 6620aa0ce06SHong 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]); 6630aa0ce06SHong 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]); 6640aa0ce06SHong Zhang 6650aa0ce06SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 6660aa0ce06SHong 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]; 6670aa0ce06SHong 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]; 6680aa0ce06SHong 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]; 6690aa0ce06SHong 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]; 6700aa0ce06SHong 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]; 6710aa0ce06SHong 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]; 6720aa0ce06SHong 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]; 6730aa0ce06SHong Zhang 6740aa0ce06SHong 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]; 6750aa0ce06SHong 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]; 6760aa0ce06SHong 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]; 6770aa0ce06SHong 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]; 6780aa0ce06SHong 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]; 6790aa0ce06SHong 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]; 6800aa0ce06SHong 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]; 6810aa0ce06SHong Zhang 6820aa0ce06SHong 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]; 6830aa0ce06SHong 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]; 6840aa0ce06SHong 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]; 6850aa0ce06SHong 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]; 6860aa0ce06SHong 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]; 6870aa0ce06SHong 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]; 6880aa0ce06SHong 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]; 6890aa0ce06SHong Zhang 6900aa0ce06SHong 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]; 6910aa0ce06SHong 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]; 6920aa0ce06SHong 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]; 6930aa0ce06SHong 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]; 6940aa0ce06SHong 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]; 6950aa0ce06SHong 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]; 6960aa0ce06SHong 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]; 6970aa0ce06SHong Zhang 6980aa0ce06SHong 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]; 6990aa0ce06SHong 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]; 7000aa0ce06SHong 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]; 7010aa0ce06SHong 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]; 7020aa0ce06SHong 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]; 7030aa0ce06SHong 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]; 7040aa0ce06SHong 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]; 7050aa0ce06SHong Zhang 7060aa0ce06SHong 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]; 7070aa0ce06SHong 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]; 7080aa0ce06SHong 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]; 7090aa0ce06SHong 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]; 7100aa0ce06SHong 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]; 7110aa0ce06SHong 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]; 7120aa0ce06SHong 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]; 7130aa0ce06SHong Zhang 7140aa0ce06SHong 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]; 7150aa0ce06SHong 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]; 7160aa0ce06SHong 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]; 7170aa0ce06SHong 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]; 7180aa0ce06SHong 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]; 7190aa0ce06SHong 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]; 7200aa0ce06SHong 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]; 7210aa0ce06SHong Zhang 7220aa0ce06SHong Zhang /* update -U(i,k) */ 7230aa0ce06SHong Zhang ierr = PetscMemcpy(ba+ili*49,uik,49*sizeof(MatScalar));CHKERRQ(ierr); 7240aa0ce06SHong Zhang 7250aa0ce06SHong Zhang /* add multiple of row i to k-th row ... */ 7260aa0ce06SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 7270aa0ce06SHong Zhang if (jmin < jmax){ 7280aa0ce06SHong Zhang for (j=jmin; j<jmax; j++) { 7290aa0ce06SHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 7300aa0ce06SHong Zhang wp = w + bj[j]*49; 7310aa0ce06SHong Zhang u = ba + j*49; 7320aa0ce06SHong Zhang 7330aa0ce06SHong 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]; 7340aa0ce06SHong 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]; 7350aa0ce06SHong 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]; 7360aa0ce06SHong 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]; 7370aa0ce06SHong 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]; 7380aa0ce06SHong 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]; 7390aa0ce06SHong 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]; 7400aa0ce06SHong Zhang 7410aa0ce06SHong 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]; 7420aa0ce06SHong 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]; 7430aa0ce06SHong 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]; 7440aa0ce06SHong 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]; 7450aa0ce06SHong 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]; 7460aa0ce06SHong 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]; 7470aa0ce06SHong 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]; 7480aa0ce06SHong Zhang 7490aa0ce06SHong 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]; 7500aa0ce06SHong 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]; 7510aa0ce06SHong 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]; 7520aa0ce06SHong 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]; 7530aa0ce06SHong 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]; 7540aa0ce06SHong 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]; 7550aa0ce06SHong 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]; 7560aa0ce06SHong Zhang 7570aa0ce06SHong 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]; 7580aa0ce06SHong 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]; 7590aa0ce06SHong 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]; 7600aa0ce06SHong 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]; 7610aa0ce06SHong 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]; 7620aa0ce06SHong 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]; 7630aa0ce06SHong 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]; 7640aa0ce06SHong Zhang 7650aa0ce06SHong 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]; 7660aa0ce06SHong 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]; 7670aa0ce06SHong 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]; 7680aa0ce06SHong 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]; 7690aa0ce06SHong 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]; 7700aa0ce06SHong 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]; 7710aa0ce06SHong 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]; 7720aa0ce06SHong Zhang 7730aa0ce06SHong 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]; 7740aa0ce06SHong 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]; 7750aa0ce06SHong 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]; 7760aa0ce06SHong 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]; 7770aa0ce06SHong 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]; 7780aa0ce06SHong 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]; 7790aa0ce06SHong 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]; 7800aa0ce06SHong Zhang 7810aa0ce06SHong 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]; 7820aa0ce06SHong 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]; 7830aa0ce06SHong 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]; 7840aa0ce06SHong 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]; 7850aa0ce06SHong 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]; 7860aa0ce06SHong 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]; 7870aa0ce06SHong 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]; 7880aa0ce06SHong Zhang } 7890aa0ce06SHong Zhang 7900aa0ce06SHong Zhang /* ... add i to row list for next nonzero entry */ 7910aa0ce06SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 7920aa0ce06SHong Zhang j = bj[jmin]; 7930aa0ce06SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 7940aa0ce06SHong Zhang } 7950aa0ce06SHong Zhang i = nexti; 7960aa0ce06SHong Zhang } 7970aa0ce06SHong Zhang 7980aa0ce06SHong Zhang /* save nonzero entries in k-th row of U ... */ 7990aa0ce06SHong Zhang 80049b5e25fSSatish Balay /* invert diagonal block */ 8010aa0ce06SHong Zhang d = ba+k*49; 8020aa0ce06SHong Zhang ierr = PetscMemcpy(d,dk,49*sizeof(MatScalar));CHKERRQ(ierr); 8030aa0ce06SHong Zhang ierr = Kernel_A_gets_inverse_A_7(d);CHKERRQ(ierr); 8040aa0ce06SHong Zhang 8050aa0ce06SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 8060aa0ce06SHong Zhang if (jmin < jmax) { 8070aa0ce06SHong Zhang for (j=jmin; j<jmax; j++){ 8080aa0ce06SHong Zhang vj = bj[j]; /* block col. index of U */ 8090aa0ce06SHong Zhang u = ba + j*49; 8100aa0ce06SHong Zhang wp = w + vj*49; 8110aa0ce06SHong Zhang for (k1=0; k1<49; k1++){ 8120aa0ce06SHong Zhang *u++ = *wp; 8130aa0ce06SHong Zhang *wp++ = 0.0; 8140aa0ce06SHong Zhang } 81549b5e25fSSatish Balay } 81649b5e25fSSatish Balay 8170aa0ce06SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 8180aa0ce06SHong Zhang il[k] = jmin; 8190aa0ce06SHong Zhang i = bj[jmin]; 8200aa0ce06SHong Zhang jl[k] = jl[i]; jl[i] = k; 8210aa0ce06SHong Zhang } 8220aa0ce06SHong Zhang } 8230aa0ce06SHong Zhang 8240aa0ce06SHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 8250aa0ce06SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 8260aa0ce06SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 8270aa0ce06SHong Zhang if (a->permute){ 8280aa0ce06SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 8290aa0ce06SHong Zhang } 8300aa0ce06SHong Zhang 8310aa0ce06SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 8320aa0ce06SHong Zhang C->factor = FACTOR_CHOLESKY; 83349b5e25fSSatish Balay C->assembled = PETSC_TRUE; 8340aa0ce06SHong Zhang C->preallocated = PETSC_TRUE; 835*b0a32e0cSBarry Smith PetscLogFlops(1.3333*343*b->mbs); /* from inverting diagonal blocks */ 83649b5e25fSSatish Balay PetscFunctionReturn(0); 83749b5e25fSSatish Balay } 83849b5e25fSSatish Balay 83949b5e25fSSatish Balay /* 84049b5e25fSSatish Balay Version for when blocks are 7 by 7 Using natural ordering 84149b5e25fSSatish Balay */ 84249b5e25fSSatish Balay #undef __FUNC__ 8436f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering" 8446f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering(Mat A,Mat *B) 84549b5e25fSSatish Balay { 84649b5e25fSSatish Balay Mat C = *B; 847ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 848ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 84971149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 850ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 851ab27746eSHong Zhang MatScalar *u,*d,*w,*wp; 85249b5e25fSSatish Balay 85349b5e25fSSatish Balay PetscFunctionBegin; 854671cb588SHong Zhang 855ab27746eSHong Zhang /* initialization */ 856*b0a32e0cSBarry Smith ierr = PetscMalloc(49*mbs*sizeof(MatScalar),&( w ));CHKERRQ(ierr); 857ab27746eSHong Zhang ierr = PetscMemzero(w,49*mbs*sizeof(MatScalar));CHKERRQ(ierr); 858*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 85928de702eSHong Zhang jl = il + mbs; 860ab27746eSHong Zhang for (i=0; i<mbs; i++) { 861ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 86249b5e25fSSatish Balay } 863*b0a32e0cSBarry Smith ierr = PetscMalloc(98*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 86428de702eSHong Zhang uik = dk + 49; 865ab27746eSHong Zhang 866ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 867ab27746eSHong Zhang 868ab27746eSHong Zhang /* for each row k */ 869ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 870ab27746eSHong Zhang 871ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 872ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 873ab27746eSHong Zhang if (jmin < jmax) { 874ab27746eSHong Zhang ap = aa + jmin*49; 875ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 876ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 877ab27746eSHong Zhang wp = w + vj*49; 878ab27746eSHong Zhang for (i=0; i<49; i++) *wp++ = *ap++; 87949b5e25fSSatish Balay } 88049b5e25fSSatish Balay } 881ab27746eSHong Zhang 882ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 883ab27746eSHong Zhang ierr = PetscMemcpy(dk,w+k*49,49*sizeof(MatScalar));CHKERRQ(ierr); 884ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 885ab27746eSHong Zhang 886ab27746eSHong Zhang while (i < mbs){ 887ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 888ab27746eSHong Zhang 889ab27746eSHong Zhang /* compute multiplier */ 890ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 891ab27746eSHong Zhang 892ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 893ab27746eSHong Zhang d = ba + i*49; 894ab27746eSHong Zhang u = ba + ili*49; 895ab27746eSHong Zhang 896ab27746eSHong 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]); 897ab27746eSHong 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]); 898ab27746eSHong 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]); 899ab27746eSHong 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]); 900ab27746eSHong 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]); 901ab27746eSHong 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]); 902ab27746eSHong 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]); 903ab27746eSHong Zhang 904ab27746eSHong 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]); 905ab27746eSHong 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]); 906ab27746eSHong 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]); 907ab27746eSHong 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]); 908ab27746eSHong 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]); 909ab27746eSHong 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]); 910ab27746eSHong 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]); 911ab27746eSHong Zhang 912ab27746eSHong 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]); 913ab27746eSHong 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]); 914ab27746eSHong 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]); 915ab27746eSHong 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]); 916ab27746eSHong 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]); 917ab27746eSHong 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]); 918ab27746eSHong 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]); 919ab27746eSHong Zhang 920ab27746eSHong 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]); 921ab27746eSHong 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]); 922ab27746eSHong 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]); 923ab27746eSHong 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]); 924ab27746eSHong 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]); 925ab27746eSHong 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]); 926ab27746eSHong 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]); 927ab27746eSHong Zhang 928ab27746eSHong 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]); 929ab27746eSHong 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]); 930ab27746eSHong 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]); 931ab27746eSHong 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]); 932ab27746eSHong 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]); 933ab27746eSHong 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]); 934ab27746eSHong 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]); 935ab27746eSHong Zhang 936ab27746eSHong 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]); 937ab27746eSHong 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]); 938ab27746eSHong 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]); 939ab27746eSHong 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]); 940ab27746eSHong 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]); 941ab27746eSHong 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]); 942ab27746eSHong 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]); 943ab27746eSHong Zhang 944ab27746eSHong 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]); 945ab27746eSHong 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]); 946ab27746eSHong 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]); 947ab27746eSHong 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]); 948ab27746eSHong 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]); 949ab27746eSHong 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]); 950ab27746eSHong 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]); 951ab27746eSHong Zhang 952ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 953ab27746eSHong 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]; 954ab27746eSHong 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]; 955ab27746eSHong 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]; 956ab27746eSHong 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]; 957ab27746eSHong 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]; 958ab27746eSHong 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]; 959ab27746eSHong 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]; 960ab27746eSHong Zhang 961ab27746eSHong 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]; 962ab27746eSHong 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]; 963ab27746eSHong 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]; 964ab27746eSHong 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]; 965ab27746eSHong 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]; 966ab27746eSHong 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]; 967ab27746eSHong 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]; 968ab27746eSHong Zhang 969ab27746eSHong 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]; 970ab27746eSHong 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]; 971ab27746eSHong 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]; 972ab27746eSHong 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]; 973ab27746eSHong 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]; 974ab27746eSHong 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]; 975ab27746eSHong 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]; 976ab27746eSHong Zhang 977ab27746eSHong 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]; 978ab27746eSHong 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]; 979ab27746eSHong 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]; 980ab27746eSHong 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]; 981ab27746eSHong 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]; 982ab27746eSHong 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]; 983ab27746eSHong 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]; 984ab27746eSHong Zhang 985ab27746eSHong 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]; 986ab27746eSHong 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]; 987ab27746eSHong 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]; 988ab27746eSHong 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]; 989ab27746eSHong 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]; 990ab27746eSHong 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]; 991ab27746eSHong 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]; 992ab27746eSHong Zhang 993ab27746eSHong 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]; 994ab27746eSHong 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]; 995ab27746eSHong 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]; 996ab27746eSHong 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]; 997ab27746eSHong 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]; 998ab27746eSHong 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]; 999ab27746eSHong 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]; 1000ab27746eSHong Zhang 1001ab27746eSHong 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]; 1002ab27746eSHong 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]; 1003ab27746eSHong 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]; 1004ab27746eSHong 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]; 1005ab27746eSHong 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]; 1006ab27746eSHong 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]; 1007ab27746eSHong 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]; 1008ab27746eSHong Zhang 1009ab27746eSHong Zhang /* update -U(i,k) */ 1010ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*49,uik,49*sizeof(MatScalar));CHKERRQ(ierr); 1011ab27746eSHong Zhang 1012ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 1013ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1014ab27746eSHong Zhang if (jmin < jmax){ 1015ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 1016ab27746eSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 1017ab27746eSHong Zhang wp = w + bj[j]*49; 1018ab27746eSHong Zhang u = ba + j*49; 1019ab27746eSHong Zhang 1020ab27746eSHong 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]; 1021ab27746eSHong 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]; 1022ab27746eSHong 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]; 1023ab27746eSHong 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]; 1024ab27746eSHong 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]; 1025ab27746eSHong 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]; 1026ab27746eSHong 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]; 1027ab27746eSHong Zhang 1028ab27746eSHong 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]; 1029ab27746eSHong 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]; 1030ab27746eSHong 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]; 1031ab27746eSHong 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]; 1032ab27746eSHong 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]; 1033ab27746eSHong 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]; 1034ab27746eSHong 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]; 1035ab27746eSHong Zhang 1036ab27746eSHong 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]; 1037ab27746eSHong 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]; 1038ab27746eSHong 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]; 1039ab27746eSHong 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]; 1040ab27746eSHong 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]; 1041ab27746eSHong 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]; 1042ab27746eSHong 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]; 1043ab27746eSHong Zhang 1044ab27746eSHong 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]; 1045ab27746eSHong 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]; 1046ab27746eSHong 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]; 1047ab27746eSHong 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]; 1048ab27746eSHong 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]; 1049ab27746eSHong 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]; 1050ab27746eSHong 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]; 1051ab27746eSHong Zhang 1052ab27746eSHong 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]; 1053ab27746eSHong 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]; 1054ab27746eSHong 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]; 1055ab27746eSHong 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]; 1056ab27746eSHong 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]; 1057ab27746eSHong 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]; 1058ab27746eSHong 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]; 1059ab27746eSHong Zhang 1060ab27746eSHong 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]; 1061ab27746eSHong 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]; 1062ab27746eSHong 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]; 1063ab27746eSHong 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]; 1064ab27746eSHong 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]; 1065ab27746eSHong 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]; 1066ab27746eSHong 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]; 1067ab27746eSHong Zhang 1068ab27746eSHong 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]; 1069ab27746eSHong 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]; 1070ab27746eSHong 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]; 1071ab27746eSHong 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]; 1072ab27746eSHong 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]; 1073ab27746eSHong 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]; 1074ab27746eSHong 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]; 107549b5e25fSSatish Balay } 1076ab27746eSHong Zhang 1077ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 1078ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1079ab27746eSHong Zhang j = bj[jmin]; 1080ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 108149b5e25fSSatish Balay } 1082ab27746eSHong Zhang i = nexti; 108349b5e25fSSatish Balay } 1084ab27746eSHong Zhang 1085ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 1086ab27746eSHong Zhang 108749b5e25fSSatish Balay /* invert diagonal block */ 1088ab27746eSHong Zhang d = ba+k*49; 1089ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,49*sizeof(MatScalar));CHKERRQ(ierr); 1090ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_7(d);CHKERRQ(ierr); 1091ab27746eSHong Zhang 1092ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1093ab27746eSHong Zhang if (jmin < jmax) { 1094ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 1095ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 1096ab27746eSHong Zhang u = ba + j*49; 1097ab27746eSHong Zhang wp = w + vj*49; 1098ab27746eSHong Zhang for (k1=0; k1<49; k1++){ 1099ab27746eSHong Zhang *u++ = *wp; 1100ab27746eSHong Zhang *wp++ = 0.0; 1101ab27746eSHong Zhang } 110249b5e25fSSatish Balay } 110349b5e25fSSatish Balay 1104ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1105ab27746eSHong Zhang il[k] = jmin; 1106ab27746eSHong Zhang i = bj[jmin]; 1107ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 1108ab27746eSHong Zhang } 1109ab27746eSHong Zhang } 1110ab27746eSHong Zhang 1111ab27746eSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 1112ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1113ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1114ab27746eSHong Zhang 1115ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 111649b5e25fSSatish Balay C->assembled = PETSC_TRUE; 1117ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 1118*b0a32e0cSBarry Smith PetscLogFlops(1.3333*343*b->mbs); /* from inverting diagonal blocks */ 111949b5e25fSSatish Balay PetscFunctionReturn(0); 112049b5e25fSSatish Balay } 112149b5e25fSSatish Balay 11229b0e81eeSHong Zhang /* Version for when blocks are 6 by 6 */ 112349b5e25fSSatish Balay #undef __FUNC__ 11246f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6" 11256f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_6(Mat A,Mat *B) 112649b5e25fSSatish Balay { 112749b5e25fSSatish Balay Mat C = *B; 11289b0e81eeSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 11299b0e81eeSHong Zhang IS perm = b->row; 11309b0e81eeSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 11319b0e81eeSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 11329b0e81eeSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 11339b0e81eeSHong Zhang MatScalar *u,*d,*w,*wp; 113449b5e25fSSatish Balay 113549b5e25fSSatish Balay PetscFunctionBegin; 11369b0e81eeSHong Zhang /* initialization */ 1137*b0a32e0cSBarry Smith ierr = PetscMalloc(36*mbs*sizeof(MatScalar),&( w ));CHKERRQ(ierr); 11389b0e81eeSHong Zhang ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1139*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 114028de702eSHong Zhang jl = il + mbs; 11419b0e81eeSHong Zhang for (i=0; i<mbs; i++) { 11429b0e81eeSHong Zhang jl[i] = mbs; il[0] = 0; 114349b5e25fSSatish Balay } 1144*b0a32e0cSBarry Smith ierr = PetscMalloc(72*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 114528de702eSHong Zhang uik = dk + 36; 11469b0e81eeSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 11479b0e81eeSHong Zhang 11489b0e81eeSHong Zhang /* check permutation */ 11499b0e81eeSHong Zhang if (!a->permute){ 11509b0e81eeSHong Zhang ai = a->i; aj = a->j; aa = a->a; 11519b0e81eeSHong Zhang } else { 11529b0e81eeSHong Zhang ai = a->inew; aj = a->jnew; 1153*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(36*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 11549b0e81eeSHong Zhang ierr = PetscMemcpy(aa,a->a,36*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 1155*b0a32e0cSBarry Smith a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 11569b0e81eeSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 11579b0e81eeSHong Zhang 11589b0e81eeSHong Zhang for (i=0; i<mbs; i++){ 11599b0e81eeSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 11609b0e81eeSHong Zhang for (j=jmin; j<jmax; j++){ 11619b0e81eeSHong Zhang while (a2anew[j] != j){ 11629b0e81eeSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 11639b0e81eeSHong Zhang for (k1=0; k1<36; k1++){ 11649b0e81eeSHong Zhang dk[k1] = aa[k*36+k1]; 11659b0e81eeSHong Zhang aa[k*36+k1] = aa[j*36+k1]; 11669b0e81eeSHong Zhang aa[j*36+k1] = dk[k1]; 116749b5e25fSSatish Balay } 116849b5e25fSSatish Balay } 11699b0e81eeSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 11709b0e81eeSHong Zhang if (i > aj[j]){ 11719b0e81eeSHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 11729b0e81eeSHong Zhang ap = aa + j*36; /* ptr to the beginning of j-th block of aa */ 11739b0e81eeSHong Zhang for (k=0; k<36; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 11749b0e81eeSHong Zhang for (k=0; k<6; k++){ /* j-th block of aa <- dk^T */ 11759b0e81eeSHong Zhang for (k1=0; k1<6; k1++) *ap++ = dk[k + 6*k1]; 117649b5e25fSSatish Balay } 117749b5e25fSSatish Balay } 117849b5e25fSSatish Balay } 11799b0e81eeSHong Zhang } 11809b0e81eeSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 11819b0e81eeSHong Zhang } 11829b0e81eeSHong Zhang 11839b0e81eeSHong Zhang /* for each row k */ 11849b0e81eeSHong Zhang for (k = 0; k<mbs; k++){ 11859b0e81eeSHong Zhang 11869b0e81eeSHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 11879b0e81eeSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 11889b0e81eeSHong Zhang if (jmin < jmax) { 11899b0e81eeSHong Zhang ap = aa + jmin*36; 11909b0e81eeSHong Zhang for (j = jmin; j < jmax; j++){ 11919b0e81eeSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 11929b0e81eeSHong Zhang wp = w + vj*36; 11939b0e81eeSHong Zhang for (i=0; i<36; i++) *wp++ = *ap++; 11949b0e81eeSHong Zhang } 11959b0e81eeSHong Zhang } 11969b0e81eeSHong Zhang 11979b0e81eeSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 11989b0e81eeSHong Zhang ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 11999b0e81eeSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 12009b0e81eeSHong Zhang 12019b0e81eeSHong Zhang while (i < mbs){ 12029b0e81eeSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 12039b0e81eeSHong Zhang 12049b0e81eeSHong Zhang /* compute multiplier */ 12059b0e81eeSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 12069b0e81eeSHong Zhang 12079b0e81eeSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 12089b0e81eeSHong Zhang d = ba + i*36; 12099b0e81eeSHong Zhang u = ba + ili*36; 12109b0e81eeSHong Zhang 12119b0e81eeSHong 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]); 12129b0e81eeSHong 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]); 12139b0e81eeSHong 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]); 12149b0e81eeSHong 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]); 12159b0e81eeSHong 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]); 12169b0e81eeSHong 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]); 12179b0e81eeSHong Zhang 12189b0e81eeSHong 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]); 12199b0e81eeSHong 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]); 12209b0e81eeSHong 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]); 12219b0e81eeSHong 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]); 12229b0e81eeSHong 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]); 12239b0e81eeSHong 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]); 12249b0e81eeSHong Zhang 12259b0e81eeSHong 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]); 12269b0e81eeSHong 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]); 12279b0e81eeSHong 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]); 12289b0e81eeSHong 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]); 12299b0e81eeSHong 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]); 12309b0e81eeSHong 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]); 12319b0e81eeSHong Zhang 12329b0e81eeSHong 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]); 12339b0e81eeSHong 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]); 12349b0e81eeSHong 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]); 12359b0e81eeSHong 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]); 12369b0e81eeSHong 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]); 12379b0e81eeSHong 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]); 12389b0e81eeSHong Zhang 12399b0e81eeSHong 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]); 12409b0e81eeSHong 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]); 12419b0e81eeSHong 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]); 12429b0e81eeSHong 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]); 12439b0e81eeSHong 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]); 12449b0e81eeSHong 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]); 12459b0e81eeSHong Zhang 12469b0e81eeSHong 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]); 12479b0e81eeSHong 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]); 12489b0e81eeSHong 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]); 12499b0e81eeSHong 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]); 12509b0e81eeSHong 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]); 12519b0e81eeSHong 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]); 12529b0e81eeSHong Zhang 12539b0e81eeSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 12549b0e81eeSHong 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]; 12559b0e81eeSHong 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]; 12569b0e81eeSHong 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]; 12579b0e81eeSHong 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]; 12589b0e81eeSHong 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]; 12599b0e81eeSHong 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]; 12609b0e81eeSHong Zhang 12619b0e81eeSHong 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]; 12629b0e81eeSHong 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]; 12639b0e81eeSHong 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]; 12649b0e81eeSHong 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]; 12659b0e81eeSHong 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]; 12669b0e81eeSHong 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]; 12679b0e81eeSHong Zhang 12689b0e81eeSHong 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]; 12699b0e81eeSHong 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]; 12709b0e81eeSHong 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]; 12719b0e81eeSHong 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]; 12729b0e81eeSHong 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]; 12739b0e81eeSHong 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]; 12749b0e81eeSHong Zhang 12759b0e81eeSHong 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]; 12769b0e81eeSHong 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]; 12779b0e81eeSHong 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]; 12789b0e81eeSHong 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]; 12799b0e81eeSHong 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]; 12809b0e81eeSHong 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]; 12819b0e81eeSHong Zhang 12829b0e81eeSHong 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]; 12839b0e81eeSHong 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]; 12849b0e81eeSHong 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]; 12859b0e81eeSHong 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]; 12869b0e81eeSHong 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]; 12879b0e81eeSHong 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]; 12889b0e81eeSHong Zhang 12899b0e81eeSHong 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]; 12909b0e81eeSHong 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]; 12919b0e81eeSHong 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]; 12929b0e81eeSHong 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]; 12939b0e81eeSHong 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]; 12949b0e81eeSHong 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]; 12959b0e81eeSHong Zhang 12969b0e81eeSHong Zhang /* update -U(i,k) */ 12979b0e81eeSHong Zhang ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 12989b0e81eeSHong Zhang 12999b0e81eeSHong Zhang /* add multiple of row i to k-th row ... */ 13009b0e81eeSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 13019b0e81eeSHong Zhang if (jmin < jmax){ 13029b0e81eeSHong Zhang for (j=jmin; j<jmax; j++) { 13039b0e81eeSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 13049b0e81eeSHong Zhang wp = w + bj[j]*36; 13059b0e81eeSHong Zhang u = ba + j*36; 13069b0e81eeSHong 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]; 13079b0e81eeSHong 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]; 13089b0e81eeSHong 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]; 13099b0e81eeSHong 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]; 13109b0e81eeSHong 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]; 13119b0e81eeSHong 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]; 13129b0e81eeSHong Zhang 13139b0e81eeSHong 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]; 13149b0e81eeSHong 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]; 13159b0e81eeSHong 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]; 13169b0e81eeSHong 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]; 13179b0e81eeSHong 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]; 13189b0e81eeSHong 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]; 13199b0e81eeSHong Zhang 13209b0e81eeSHong 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]; 13219b0e81eeSHong 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]; 13229b0e81eeSHong 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]; 13239b0e81eeSHong 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]; 13249b0e81eeSHong 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]; 13259b0e81eeSHong 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]; 13269b0e81eeSHong Zhang 13279b0e81eeSHong 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]; 13289b0e81eeSHong 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]; 13299b0e81eeSHong 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]; 13309b0e81eeSHong 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]; 13319b0e81eeSHong 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]; 13329b0e81eeSHong 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]; 13339b0e81eeSHong Zhang 13349b0e81eeSHong 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]; 13359b0e81eeSHong 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]; 13369b0e81eeSHong 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]; 13379b0e81eeSHong 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]; 13389b0e81eeSHong 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]; 13399b0e81eeSHong 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]; 13409b0e81eeSHong Zhang 13419b0e81eeSHong 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]; 13429b0e81eeSHong 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]; 13439b0e81eeSHong 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]; 13449b0e81eeSHong 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]; 13459b0e81eeSHong 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]; 13469b0e81eeSHong 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]; 13479b0e81eeSHong Zhang } 13489b0e81eeSHong Zhang 13499b0e81eeSHong Zhang /* ... add i to row list for next nonzero entry */ 13509b0e81eeSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 13519b0e81eeSHong Zhang j = bj[jmin]; 13529b0e81eeSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 13539b0e81eeSHong Zhang } 13549b0e81eeSHong Zhang i = nexti; 13559b0e81eeSHong Zhang } 13569b0e81eeSHong Zhang 13579b0e81eeSHong Zhang /* save nonzero entries in k-th row of U ... */ 13589b0e81eeSHong Zhang 135949b5e25fSSatish Balay /* invert diagonal block */ 13609b0e81eeSHong Zhang d = ba+k*36; 13619b0e81eeSHong Zhang ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 13629b0e81eeSHong Zhang ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 13639b0e81eeSHong Zhang 13649b0e81eeSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 13659b0e81eeSHong Zhang if (jmin < jmax) { 13669b0e81eeSHong Zhang for (j=jmin; j<jmax; j++){ 13679b0e81eeSHong Zhang vj = bj[j]; /* block col. index of U */ 13689b0e81eeSHong Zhang u = ba + j*36; 13699b0e81eeSHong Zhang wp = w + vj*36; 13709b0e81eeSHong Zhang for (k1=0; k1<36; k1++){ 13719b0e81eeSHong Zhang *u++ = *wp; 13729b0e81eeSHong Zhang *wp++ = 0.0; 13739b0e81eeSHong Zhang } 137449b5e25fSSatish Balay } 137549b5e25fSSatish Balay 13769b0e81eeSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 13779b0e81eeSHong Zhang il[k] = jmin; 13789b0e81eeSHong Zhang i = bj[jmin]; 13799b0e81eeSHong Zhang jl[k] = jl[i]; jl[i] = k; 13809b0e81eeSHong Zhang } 13819b0e81eeSHong Zhang } 13829b0e81eeSHong Zhang 13839b0e81eeSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 13849b0e81eeSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 13859b0e81eeSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 13869b0e81eeSHong Zhang if (a->permute){ 13879b0e81eeSHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 13889b0e81eeSHong Zhang } 13899b0e81eeSHong Zhang 13909b0e81eeSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 13919b0e81eeSHong Zhang C->factor = FACTOR_CHOLESKY; 139249b5e25fSSatish Balay C->assembled = PETSC_TRUE; 13939b0e81eeSHong Zhang C->preallocated = PETSC_TRUE; 1394*b0a32e0cSBarry Smith PetscLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 139549b5e25fSSatish Balay PetscFunctionReturn(0); 139649b5e25fSSatish Balay } 13979b0e81eeSHong Zhang 139849b5e25fSSatish Balay /* 139949b5e25fSSatish Balay Version for when blocks are 6 by 6 Using natural ordering 140049b5e25fSSatish Balay */ 140149b5e25fSSatish Balay #undef __FUNC__ 14026f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering" 14036f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering(Mat A,Mat *B) 140449b5e25fSSatish Balay { 140549b5e25fSSatish Balay Mat C = *B; 1406ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1407ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 140871149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1409ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1410ab27746eSHong Zhang MatScalar *u,*d,*w,*wp; 141149b5e25fSSatish Balay 141249b5e25fSSatish Balay PetscFunctionBegin; 1413671cb588SHong Zhang 1414ab27746eSHong Zhang /* initialization */ 1415*b0a32e0cSBarry Smith ierr = PetscMalloc(36*mbs*sizeof(MatScalar),&( w ));CHKERRQ(ierr); 1416ab27746eSHong Zhang ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1417*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 141828de702eSHong Zhang jl = il + mbs; 1419ab27746eSHong Zhang for (i=0; i<mbs; i++) { 1420ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 142149b5e25fSSatish Balay } 1422*b0a32e0cSBarry Smith ierr = PetscMalloc(72*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 142328de702eSHong Zhang uik = dk + 36; 1424ab27746eSHong Zhang 1425ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 1426ab27746eSHong Zhang 1427ab27746eSHong Zhang /* for each row k */ 1428ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 1429ab27746eSHong Zhang 1430ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 1431ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 1432ab27746eSHong Zhang if (jmin < jmax) { 1433ab27746eSHong Zhang ap = aa + jmin*36; 1434ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 1435ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 1436ab27746eSHong Zhang wp = w + vj*36; 1437ab27746eSHong Zhang for (i=0; i<36; i++) *wp++ = *ap++; 143849b5e25fSSatish Balay } 143949b5e25fSSatish Balay } 1440ab27746eSHong Zhang 1441ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1442ab27746eSHong Zhang ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 1443ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1444ab27746eSHong Zhang 1445ab27746eSHong Zhang while (i < mbs){ 1446ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1447ab27746eSHong Zhang 1448ab27746eSHong Zhang /* compute multiplier */ 1449ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1450ab27746eSHong Zhang 1451ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1452ab27746eSHong Zhang d = ba + i*36; 1453ab27746eSHong Zhang u = ba + ili*36; 1454ab27746eSHong Zhang 1455ab27746eSHong 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]); 1456ab27746eSHong 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]); 1457ab27746eSHong 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]); 1458ab27746eSHong 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]); 1459ab27746eSHong 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]); 1460ab27746eSHong 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]); 1461ab27746eSHong Zhang 1462ab27746eSHong 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]); 1463ab27746eSHong 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]); 1464ab27746eSHong 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]); 1465ab27746eSHong 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]); 1466ab27746eSHong 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]); 1467ab27746eSHong 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]); 1468ab27746eSHong Zhang 1469ab27746eSHong 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]); 1470ab27746eSHong 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]); 1471ab27746eSHong 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]); 1472ab27746eSHong 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]); 1473ab27746eSHong 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]); 1474ab27746eSHong 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]); 1475ab27746eSHong Zhang 1476ab27746eSHong 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]); 1477ab27746eSHong 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]); 1478ab27746eSHong 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]); 1479ab27746eSHong 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]); 1480ab27746eSHong 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]); 1481ab27746eSHong 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]); 1482ab27746eSHong Zhang 1483ab27746eSHong 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]); 1484ab27746eSHong 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]); 1485ab27746eSHong 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]); 1486ab27746eSHong 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]); 1487ab27746eSHong 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]); 1488ab27746eSHong 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]); 1489ab27746eSHong Zhang 1490ab27746eSHong 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]); 1491ab27746eSHong 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]); 1492ab27746eSHong 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]); 1493ab27746eSHong 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]); 1494ab27746eSHong 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]); 1495ab27746eSHong 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]); 1496ab27746eSHong Zhang 1497ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1498ab27746eSHong 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]; 1499ab27746eSHong 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]; 1500ab27746eSHong 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]; 1501ab27746eSHong 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]; 1502ab27746eSHong 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]; 1503ab27746eSHong 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]; 1504ab27746eSHong Zhang 1505ab27746eSHong 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]; 1506ab27746eSHong 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]; 1507ab27746eSHong 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]; 1508ab27746eSHong 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]; 1509ab27746eSHong 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]; 1510ab27746eSHong 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]; 1511ab27746eSHong Zhang 1512ab27746eSHong 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]; 1513ab27746eSHong 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]; 1514ab27746eSHong 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]; 1515ab27746eSHong 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]; 1516ab27746eSHong 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]; 1517ab27746eSHong 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]; 1518ab27746eSHong Zhang 1519ab27746eSHong 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]; 1520ab27746eSHong 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]; 1521ab27746eSHong 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]; 1522ab27746eSHong 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]; 1523ab27746eSHong 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]; 1524ab27746eSHong 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]; 1525ab27746eSHong Zhang 1526ab27746eSHong 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]; 1527ab27746eSHong 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]; 1528ab27746eSHong 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]; 1529ab27746eSHong 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]; 1530ab27746eSHong 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]; 1531ab27746eSHong 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]; 1532ab27746eSHong Zhang 1533ab27746eSHong 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]; 1534ab27746eSHong 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]; 1535ab27746eSHong 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]; 1536ab27746eSHong 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]; 1537ab27746eSHong 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]; 1538ab27746eSHong 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]; 1539ab27746eSHong Zhang 1540ab27746eSHong Zhang /* update -U(i,k) */ 1541ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 1542ab27746eSHong Zhang 1543ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 1544ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1545ab27746eSHong Zhang if (jmin < jmax){ 1546ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 1547ab27746eSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 1548ab27746eSHong Zhang wp = w + bj[j]*36; 1549ab27746eSHong Zhang u = ba + j*36; 1550ab27746eSHong 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]; 1551ab27746eSHong 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]; 1552ab27746eSHong 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]; 1553ab27746eSHong 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]; 1554ab27746eSHong 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]; 1555ab27746eSHong 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]; 1556ab27746eSHong Zhang 1557ab27746eSHong 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]; 1558ab27746eSHong 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]; 1559ab27746eSHong 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]; 1560ab27746eSHong 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]; 1561ab27746eSHong 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]; 1562ab27746eSHong 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]; 1563ab27746eSHong Zhang 1564ab27746eSHong 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]; 1565ab27746eSHong 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]; 1566ab27746eSHong 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]; 1567ab27746eSHong 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]; 1568ab27746eSHong 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]; 1569ab27746eSHong 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]; 1570ab27746eSHong Zhang 1571ab27746eSHong 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]; 1572ab27746eSHong 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]; 1573ab27746eSHong 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]; 1574ab27746eSHong 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]; 1575ab27746eSHong 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]; 1576ab27746eSHong 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]; 1577ab27746eSHong Zhang 1578ab27746eSHong 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]; 1579ab27746eSHong 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]; 1580ab27746eSHong 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]; 1581ab27746eSHong 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]; 1582ab27746eSHong 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]; 1583ab27746eSHong 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]; 1584ab27746eSHong Zhang 1585ab27746eSHong 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]; 1586ab27746eSHong 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]; 1587ab27746eSHong 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]; 1588ab27746eSHong 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]; 1589ab27746eSHong 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]; 1590ab27746eSHong 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]; 159149b5e25fSSatish Balay } 1592ab27746eSHong Zhang 1593ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 1594ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1595ab27746eSHong Zhang j = bj[jmin]; 1596ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 159749b5e25fSSatish Balay } 1598ab27746eSHong Zhang i = nexti; 159949b5e25fSSatish Balay } 1600ab27746eSHong Zhang 1601ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 1602ab27746eSHong Zhang 160349b5e25fSSatish Balay /* invert diagonal block */ 1604ab27746eSHong Zhang d = ba+k*36; 1605ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 1606ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 1607ab27746eSHong Zhang 1608ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1609ab27746eSHong Zhang if (jmin < jmax) { 1610ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 1611ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 1612ab27746eSHong Zhang u = ba + j*36; 1613ab27746eSHong Zhang wp = w + vj*36; 1614ab27746eSHong Zhang for (k1=0; k1<36; k1++){ 1615ab27746eSHong Zhang *u++ = *wp; 1616ab27746eSHong Zhang *wp++ = 0.0; 1617ab27746eSHong Zhang } 161849b5e25fSSatish Balay } 161949b5e25fSSatish Balay 1620ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1621ab27746eSHong Zhang il[k] = jmin; 1622ab27746eSHong Zhang i = bj[jmin]; 1623ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 1624ab27746eSHong Zhang } 1625ab27746eSHong Zhang } 1626ab27746eSHong Zhang 1627ab27746eSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 1628ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1629ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1630ab27746eSHong Zhang 1631ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 163249b5e25fSSatish Balay C->assembled = PETSC_TRUE; 1633ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 1634*b0a32e0cSBarry Smith PetscLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 163549b5e25fSSatish Balay PetscFunctionReturn(0); 163649b5e25fSSatish Balay } 163749b5e25fSSatish Balay 1638b59abe41SHong Zhang /* Version for when blocks are 5 by 5 */ 1639b59abe41SHong Zhang #undef __FUNC__ 1640b59abe41SHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_5" 1641b59abe41SHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_5(Mat A,Mat *B) 1642b59abe41SHong Zhang { 1643b59abe41SHong Zhang Mat C = *B; 1644b59abe41SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1645b59abe41SHong Zhang IS perm = b->row; 1646b59abe41SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1647b59abe41SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1648b59abe41SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1649b59abe41SHong Zhang MatScalar *u,*d,*rtmp,*rtmp_ptr; 1650b59abe41SHong Zhang 1651b59abe41SHong Zhang PetscFunctionBegin; 1652b59abe41SHong Zhang /* initialization */ 1653*b0a32e0cSBarry Smith ierr = PetscMalloc(25*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 1654b59abe41SHong Zhang ierr = PetscMemzero(rtmp,25*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1655*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 165628de702eSHong Zhang jl = il + mbs; 1657b59abe41SHong Zhang for (i=0; i<mbs; i++) { 1658b59abe41SHong Zhang jl[i] = mbs; il[0] = 0; 1659b59abe41SHong Zhang } 1660*b0a32e0cSBarry Smith ierr = PetscMalloc(50*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 166128de702eSHong Zhang uik = dk + 25; 1662b59abe41SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 1663b59abe41SHong Zhang 1664b59abe41SHong Zhang /* check permutation */ 1665b59abe41SHong Zhang if (!a->permute){ 1666b59abe41SHong Zhang ai = a->i; aj = a->j; aa = a->a; 1667b59abe41SHong Zhang } else { 1668b59abe41SHong Zhang ai = a->inew; aj = a->jnew; 1669*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(25*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 1670b59abe41SHong Zhang ierr = PetscMemcpy(aa,a->a,25*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 1671*b0a32e0cSBarry Smith a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 1672b59abe41SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 1673b59abe41SHong Zhang 1674b59abe41SHong Zhang for (i=0; i<mbs; i++){ 1675b59abe41SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 1676b59abe41SHong Zhang for (j=jmin; j<jmax; j++){ 1677b59abe41SHong Zhang while (a2anew[j] != j){ 1678b59abe41SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 1679b59abe41SHong Zhang for (k1=0; k1<25; k1++){ 1680b59abe41SHong Zhang dk[k1] = aa[k*25+k1]; 1681b59abe41SHong Zhang aa[k*25+k1] = aa[j*25+k1]; 1682b59abe41SHong Zhang aa[j*25+k1] = dk[k1]; 1683b59abe41SHong Zhang } 1684b59abe41SHong Zhang } 1685b59abe41SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 1686b59abe41SHong Zhang if (i > aj[j]){ 1687b59abe41SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 1688b59abe41SHong Zhang ap = aa + j*25; /* ptr to the beginning of j-th block of aa */ 1689b59abe41SHong Zhang for (k=0; k<25; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 1690b59abe41SHong Zhang for (k=0; k<5; k++){ /* j-th block of aa <- dk^T */ 1691b59abe41SHong Zhang for (k1=0; k1<5; k1++) *ap++ = dk[k + 5*k1]; 1692b59abe41SHong Zhang } 1693b59abe41SHong Zhang } 1694b59abe41SHong Zhang } 1695b59abe41SHong Zhang } 1696b59abe41SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 1697b59abe41SHong Zhang } 1698b59abe41SHong Zhang 1699b59abe41SHong Zhang /* for each row k */ 1700b59abe41SHong Zhang for (k = 0; k<mbs; k++){ 1701b59abe41SHong Zhang 1702b59abe41SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 1703b59abe41SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 1704b59abe41SHong Zhang if (jmin < jmax) { 1705b59abe41SHong Zhang ap = aa + jmin*25; 1706b59abe41SHong Zhang for (j = jmin; j < jmax; j++){ 1707b59abe41SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 1708b59abe41SHong Zhang rtmp_ptr = rtmp + vj*25; 1709b59abe41SHong Zhang for (i=0; i<25; i++) *rtmp_ptr++ = *ap++; 1710b59abe41SHong Zhang } 1711b59abe41SHong Zhang } 1712b59abe41SHong Zhang 1713b59abe41SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1714b59abe41SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*25,25*sizeof(MatScalar));CHKERRQ(ierr); 1715b59abe41SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1716b59abe41SHong Zhang 1717b59abe41SHong Zhang while (i < mbs){ 1718b59abe41SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1719b59abe41SHong Zhang 1720b59abe41SHong Zhang /* compute multiplier */ 1721b59abe41SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1722b59abe41SHong Zhang 1723b59abe41SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1724b59abe41SHong Zhang d = ba + i*25; 1725b59abe41SHong Zhang u = ba + ili*25; 1726b59abe41SHong Zhang 1727b59abe41SHong Zhang uik[0] = -(d[0]*u[0] + d[5]*u[1] + d[10]*u[2] + d[15]*u[3] + d[20]*u[4]); 1728b59abe41SHong Zhang uik[1] = -(d[1]*u[0] + d[6]*u[1] + d[11]*u[2] + d[16]*u[3] + d[21]*u[4]); 1729b59abe41SHong Zhang uik[2] = -(d[2]*u[0] + d[7]*u[1] + d[12]*u[2] + d[17]*u[3] + d[22]*u[4]); 1730b59abe41SHong Zhang uik[3] = -(d[3]*u[0] + d[8]*u[1] + d[13]*u[2] + d[18]*u[3] + d[23]*u[4]); 1731b59abe41SHong Zhang uik[4] = -(d[4]*u[0] + d[9]*u[1] + d[14]*u[2] + d[19]*u[3] + d[24]*u[4]); 1732b59abe41SHong Zhang 1733b59abe41SHong Zhang uik[5] = -(d[0]*u[5] + d[5]*u[6] + d[10]*u[7] + d[15]*u[8] + d[20]*u[9]); 1734b59abe41SHong Zhang uik[6] = -(d[1]*u[5] + d[6]*u[6] + d[11]*u[7] + d[16]*u[8] + d[21]*u[9]); 1735b59abe41SHong Zhang uik[7] = -(d[2]*u[5] + d[7]*u[6] + d[12]*u[7] + d[17]*u[8] + d[22]*u[9]); 1736b59abe41SHong Zhang uik[8] = -(d[3]*u[5] + d[8]*u[6] + d[13]*u[7] + d[18]*u[8] + d[23]*u[9]); 1737b59abe41SHong Zhang uik[9] = -(d[4]*u[5] + d[9]*u[6] + d[14]*u[7] + d[19]*u[8] + d[24]*u[9]); 1738b59abe41SHong Zhang 1739b59abe41SHong Zhang uik[10]= -(d[0]*u[10] + d[5]*u[11] + d[10]*u[12] + d[15]*u[13] + d[20]*u[14]); 1740b59abe41SHong Zhang uik[11]= -(d[1]*u[10] + d[6]*u[11] + d[11]*u[12] + d[16]*u[13] + d[21]*u[14]); 1741b59abe41SHong Zhang uik[12]= -(d[2]*u[10] + d[7]*u[11] + d[12]*u[12] + d[17]*u[13] + d[22]*u[14]); 1742b59abe41SHong Zhang uik[13]= -(d[3]*u[10] + d[8]*u[11] + d[13]*u[12] + d[18]*u[13] + d[23]*u[14]); 1743b59abe41SHong Zhang uik[14]= -(d[4]*u[10] + d[9]*u[11] + d[14]*u[12] + d[19]*u[13] + d[24]*u[14]); 1744b59abe41SHong Zhang 1745b59abe41SHong Zhang uik[15]= -(d[0]*u[15] + d[5]*u[16] + d[10]*u[17] + d[15]*u[18] + d[20]*u[19]); 1746b59abe41SHong Zhang uik[16]= -(d[1]*u[15] + d[6]*u[16] + d[11]*u[17] + d[16]*u[18] + d[21]*u[19]); 1747b59abe41SHong Zhang uik[17]= -(d[2]*u[15] + d[7]*u[16] + d[12]*u[17] + d[17]*u[18] + d[22]*u[19]); 1748b59abe41SHong Zhang uik[18]= -(d[3]*u[15] + d[8]*u[16] + d[13]*u[17] + d[18]*u[18] + d[23]*u[19]); 1749b59abe41SHong Zhang uik[19]= -(d[4]*u[15] + d[9]*u[16] + d[14]*u[17] + d[19]*u[18] + d[24]*u[19]); 1750b59abe41SHong Zhang 1751b59abe41SHong Zhang uik[20]= -(d[0]*u[20] + d[5]*u[21] + d[10]*u[22] + d[15]*u[23] + d[20]*u[24]); 1752b59abe41SHong Zhang uik[21]= -(d[1]*u[20] + d[6]*u[21] + d[11]*u[22] + d[16]*u[23] + d[21]*u[24]); 1753b59abe41SHong Zhang uik[22]= -(d[2]*u[20] + d[7]*u[21] + d[12]*u[22] + d[17]*u[23] + d[22]*u[24]); 1754b59abe41SHong Zhang uik[23]= -(d[3]*u[20] + d[8]*u[21] + d[13]*u[22] + d[18]*u[23] + d[23]*u[24]); 1755b59abe41SHong Zhang uik[24]= -(d[4]*u[20] + d[9]*u[21] + d[14]*u[22] + d[19]*u[23] + d[24]*u[24]); 1756b59abe41SHong Zhang 1757b59abe41SHong Zhang 1758936bb232SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1759936bb232SHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1760936bb232SHong Zhang dk[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1761936bb232SHong Zhang dk[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1762936bb232SHong Zhang dk[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1763936bb232SHong Zhang dk[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1764b59abe41SHong Zhang 1765936bb232SHong Zhang dk[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1766936bb232SHong Zhang dk[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1767936bb232SHong Zhang dk[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1768936bb232SHong Zhang dk[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1769936bb232SHong Zhang dk[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1770b59abe41SHong Zhang 1771936bb232SHong Zhang dk[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1772936bb232SHong Zhang dk[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1773936bb232SHong Zhang dk[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1774936bb232SHong Zhang dk[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1775936bb232SHong Zhang dk[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1776b59abe41SHong Zhang 1777936bb232SHong Zhang dk[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1778936bb232SHong Zhang dk[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1779936bb232SHong Zhang dk[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1780936bb232SHong Zhang dk[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1781936bb232SHong Zhang dk[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1782936bb232SHong Zhang 1783936bb232SHong Zhang dk[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1784936bb232SHong Zhang dk[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1785936bb232SHong Zhang dk[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1786936bb232SHong Zhang dk[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 17874f078a5fSHong Zhang dk[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1788b59abe41SHong Zhang 1789b59abe41SHong Zhang /* update -U(i,k) */ 1790b59abe41SHong Zhang ierr = PetscMemcpy(ba+ili*25,uik,25*sizeof(MatScalar));CHKERRQ(ierr); 1791b59abe41SHong Zhang 1792b59abe41SHong Zhang /* add multiple of row i to k-th row ... */ 1793b59abe41SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1794b59abe41SHong Zhang if (jmin < jmax){ 1795b59abe41SHong Zhang for (j=jmin; j<jmax; j++) { 1796b59abe41SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 1797b59abe41SHong Zhang rtmp_ptr = rtmp + bj[j]*25; 1798b59abe41SHong Zhang u = ba + j*25; 1799936bb232SHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1800936bb232SHong Zhang rtmp_ptr[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1801936bb232SHong Zhang rtmp_ptr[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1802936bb232SHong Zhang rtmp_ptr[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1803936bb232SHong Zhang rtmp_ptr[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1804b59abe41SHong Zhang 1805936bb232SHong Zhang rtmp_ptr[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1806936bb232SHong Zhang rtmp_ptr[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1807936bb232SHong Zhang rtmp_ptr[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1808936bb232SHong Zhang rtmp_ptr[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1809936bb232SHong Zhang rtmp_ptr[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1810b59abe41SHong Zhang 1811936bb232SHong Zhang rtmp_ptr[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1812936bb232SHong Zhang rtmp_ptr[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1813936bb232SHong Zhang rtmp_ptr[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1814936bb232SHong Zhang rtmp_ptr[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1815936bb232SHong Zhang rtmp_ptr[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1816b59abe41SHong Zhang 1817936bb232SHong Zhang rtmp_ptr[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1818936bb232SHong Zhang rtmp_ptr[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1819936bb232SHong Zhang rtmp_ptr[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1820936bb232SHong Zhang rtmp_ptr[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1821936bb232SHong Zhang rtmp_ptr[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1822936bb232SHong Zhang 1823936bb232SHong Zhang rtmp_ptr[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1824936bb232SHong Zhang rtmp_ptr[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1825936bb232SHong Zhang rtmp_ptr[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1826936bb232SHong Zhang rtmp_ptr[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 18274f078a5fSHong Zhang rtmp_ptr[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1828b59abe41SHong Zhang } 1829b59abe41SHong Zhang 1830b59abe41SHong Zhang /* ... add i to row list for next nonzero entry */ 1831b59abe41SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1832b59abe41SHong Zhang j = bj[jmin]; 1833b59abe41SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 1834b59abe41SHong Zhang } 1835b59abe41SHong Zhang i = nexti; 1836b59abe41SHong Zhang } 1837b59abe41SHong Zhang 1838b59abe41SHong Zhang /* save nonzero entries in k-th row of U ... */ 1839b59abe41SHong Zhang 1840b59abe41SHong Zhang /* invert diagonal block */ 1841b59abe41SHong Zhang d = ba+k*25; 1842b59abe41SHong Zhang ierr = PetscMemcpy(d,dk,25*sizeof(MatScalar));CHKERRQ(ierr); 1843b59abe41SHong Zhang ierr = Kernel_A_gets_inverse_A_5(d);CHKERRQ(ierr); 1844b59abe41SHong Zhang 1845b59abe41SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1846b59abe41SHong Zhang if (jmin < jmax) { 1847b59abe41SHong Zhang for (j=jmin; j<jmax; j++){ 1848b59abe41SHong Zhang vj = bj[j]; /* block col. index of U */ 1849b59abe41SHong Zhang u = ba + j*25; 1850b59abe41SHong Zhang rtmp_ptr = rtmp + vj*25; 1851b59abe41SHong Zhang for (k1=0; k1<25; k1++){ 1852b59abe41SHong Zhang *u++ = *rtmp_ptr; 1853b59abe41SHong Zhang *rtmp_ptr++ = 0.0; 1854b59abe41SHong Zhang } 1855b59abe41SHong Zhang } 1856b59abe41SHong Zhang 1857b59abe41SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1858b59abe41SHong Zhang il[k] = jmin; 1859b59abe41SHong Zhang i = bj[jmin]; 1860b59abe41SHong Zhang jl[k] = jl[i]; jl[i] = k; 1861b59abe41SHong Zhang } 1862b59abe41SHong Zhang } 1863b59abe41SHong Zhang 1864b59abe41SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 1865b59abe41SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1866b59abe41SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1867b59abe41SHong Zhang if (a->permute){ 1868b59abe41SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 1869b59abe41SHong Zhang } 1870b59abe41SHong Zhang 1871b59abe41SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 1872b59abe41SHong Zhang C->factor = FACTOR_CHOLESKY; 1873b59abe41SHong Zhang C->assembled = PETSC_TRUE; 1874b59abe41SHong Zhang C->preallocated = PETSC_TRUE; 1875*b0a32e0cSBarry Smith PetscLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 1876b59abe41SHong Zhang PetscFunctionReturn(0); 1877b59abe41SHong Zhang } 187849b5e25fSSatish Balay 187949b5e25fSSatish Balay /* 188049b5e25fSSatish Balay Version for when blocks are 5 by 5 Using natural ordering 188149b5e25fSSatish Balay */ 188249b5e25fSSatish Balay #undef __FUNC__ 18836f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering" 18846f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering(Mat A,Mat *B) 188549b5e25fSSatish Balay { 188649b5e25fSSatish Balay Mat C = *B; 1887ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1888ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 188971149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1890ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1891ab27746eSHong Zhang MatScalar *u,*d,*rtmp,*rtmp_ptr; 189249b5e25fSSatish Balay 189349b5e25fSSatish Balay PetscFunctionBegin; 1894671cb588SHong Zhang 1895ab27746eSHong Zhang /* initialization */ 1896*b0a32e0cSBarry Smith ierr = PetscMalloc(25*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 1897ab27746eSHong Zhang ierr = PetscMemzero(rtmp,25*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1898*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 189928de702eSHong Zhang jl = il + mbs; 1900ab27746eSHong Zhang for (i=0; i<mbs; i++) { 1901ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 190249b5e25fSSatish Balay } 1903*b0a32e0cSBarry Smith ierr = PetscMalloc(50*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 190428de702eSHong Zhang uik = dk + 25; 1905ab27746eSHong Zhang 1906ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 1907ab27746eSHong Zhang 1908ab27746eSHong Zhang /* for each row k */ 1909ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 1910ab27746eSHong Zhang 1911ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 1912ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 1913ab27746eSHong Zhang if (jmin < jmax) { 1914ab27746eSHong Zhang ap = aa + jmin*25; 1915ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 1916ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 1917ab27746eSHong Zhang rtmp_ptr = rtmp + vj*25; 1918ab27746eSHong Zhang for (i=0; i<25; i++) *rtmp_ptr++ = *ap++; 191949b5e25fSSatish Balay } 192049b5e25fSSatish Balay } 1921ab27746eSHong Zhang 1922ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1923ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*25,25*sizeof(MatScalar));CHKERRQ(ierr); 1924ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1925ab27746eSHong Zhang 1926ab27746eSHong Zhang while (i < mbs){ 1927ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1928ab27746eSHong Zhang 1929ab27746eSHong Zhang /* compute multiplier */ 1930ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1931ab27746eSHong Zhang 1932ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1933ab27746eSHong Zhang d = ba + i*25; 1934ab27746eSHong Zhang u = ba + ili*25; 1935ab27746eSHong Zhang 1936ab27746eSHong Zhang uik[0] = -(d[0]*u[0] + d[5]*u[1] + d[10]*u[2] + d[15]*u[3] + d[20]*u[4]); 1937ab27746eSHong Zhang uik[1] = -(d[1]*u[0] + d[6]*u[1] + d[11]*u[2] + d[16]*u[3] + d[21]*u[4]); 1938ab27746eSHong Zhang uik[2] = -(d[2]*u[0] + d[7]*u[1] + d[12]*u[2] + d[17]*u[3] + d[22]*u[4]); 1939ab27746eSHong Zhang uik[3] = -(d[3]*u[0] + d[8]*u[1] + d[13]*u[2] + d[18]*u[3] + d[23]*u[4]); 1940ab27746eSHong Zhang uik[4] = -(d[4]*u[0] + d[9]*u[1] + d[14]*u[2] + d[19]*u[3] + d[24]*u[4]); 1941ab27746eSHong Zhang 1942ab27746eSHong Zhang uik[5] = -(d[0]*u[5] + d[5]*u[6] + d[10]*u[7] + d[15]*u[8] + d[20]*u[9]); 1943ab27746eSHong Zhang uik[6] = -(d[1]*u[5] + d[6]*u[6] + d[11]*u[7] + d[16]*u[8] + d[21]*u[9]); 1944ab27746eSHong Zhang uik[7] = -(d[2]*u[5] + d[7]*u[6] + d[12]*u[7] + d[17]*u[8] + d[22]*u[9]); 1945ab27746eSHong Zhang uik[8] = -(d[3]*u[5] + d[8]*u[6] + d[13]*u[7] + d[18]*u[8] + d[23]*u[9]); 1946ab27746eSHong Zhang uik[9] = -(d[4]*u[5] + d[9]*u[6] + d[14]*u[7] + d[19]*u[8] + d[24]*u[9]); 1947ab27746eSHong Zhang 1948ab27746eSHong Zhang uik[10]= -(d[0]*u[10] + d[5]*u[11] + d[10]*u[12] + d[15]*u[13] + d[20]*u[14]); 1949ab27746eSHong Zhang uik[11]= -(d[1]*u[10] + d[6]*u[11] + d[11]*u[12] + d[16]*u[13] + d[21]*u[14]); 1950ab27746eSHong Zhang uik[12]= -(d[2]*u[10] + d[7]*u[11] + d[12]*u[12] + d[17]*u[13] + d[22]*u[14]); 1951ab27746eSHong Zhang uik[13]= -(d[3]*u[10] + d[8]*u[11] + d[13]*u[12] + d[18]*u[13] + d[23]*u[14]); 1952ab27746eSHong Zhang uik[14]= -(d[4]*u[10] + d[9]*u[11] + d[14]*u[12] + d[19]*u[13] + d[24]*u[14]); 1953ab27746eSHong Zhang 1954ab27746eSHong Zhang uik[15]= -(d[0]*u[15] + d[5]*u[16] + d[10]*u[17] + d[15]*u[18] + d[20]*u[19]); 1955ab27746eSHong Zhang uik[16]= -(d[1]*u[15] + d[6]*u[16] + d[11]*u[17] + d[16]*u[18] + d[21]*u[19]); 1956ab27746eSHong Zhang uik[17]= -(d[2]*u[15] + d[7]*u[16] + d[12]*u[17] + d[17]*u[18] + d[22]*u[19]); 1957ab27746eSHong Zhang uik[18]= -(d[3]*u[15] + d[8]*u[16] + d[13]*u[17] + d[18]*u[18] + d[23]*u[19]); 1958ab27746eSHong Zhang uik[19]= -(d[4]*u[15] + d[9]*u[16] + d[14]*u[17] + d[19]*u[18] + d[24]*u[19]); 1959ab27746eSHong Zhang 1960ab27746eSHong Zhang uik[20]= -(d[0]*u[20] + d[5]*u[21] + d[10]*u[22] + d[15]*u[23] + d[20]*u[24]); 1961ab27746eSHong Zhang uik[21]= -(d[1]*u[20] + d[6]*u[21] + d[11]*u[22] + d[16]*u[23] + d[21]*u[24]); 1962ab27746eSHong Zhang uik[22]= -(d[2]*u[20] + d[7]*u[21] + d[12]*u[22] + d[17]*u[23] + d[22]*u[24]); 1963ab27746eSHong Zhang uik[23]= -(d[3]*u[20] + d[8]*u[21] + d[13]*u[22] + d[18]*u[23] + d[23]*u[24]); 1964ab27746eSHong Zhang uik[24]= -(d[4]*u[20] + d[9]*u[21] + d[14]*u[22] + d[19]*u[23] + d[24]*u[24]); 1965ab27746eSHong Zhang 1966ab27746eSHong Zhang 1967ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1968ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1969ab27746eSHong Zhang dk[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1970ab27746eSHong Zhang dk[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1971ab27746eSHong Zhang dk[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1972ab27746eSHong Zhang dk[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1973ab27746eSHong Zhang 1974ab27746eSHong Zhang dk[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1975ab27746eSHong Zhang dk[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1976ab27746eSHong Zhang dk[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1977ab27746eSHong Zhang dk[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1978ab27746eSHong Zhang dk[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1979ab27746eSHong Zhang 1980ab27746eSHong Zhang dk[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1981ab27746eSHong Zhang dk[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1982ab27746eSHong Zhang dk[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1983ab27746eSHong Zhang dk[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1984ab27746eSHong Zhang dk[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1985ab27746eSHong Zhang 1986ab27746eSHong Zhang dk[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1987ab27746eSHong Zhang dk[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1988ab27746eSHong Zhang dk[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1989ab27746eSHong Zhang dk[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1990ab27746eSHong Zhang dk[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1991ab27746eSHong Zhang 1992ab27746eSHong Zhang dk[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1993ab27746eSHong Zhang dk[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1994ab27746eSHong Zhang dk[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1995ab27746eSHong Zhang dk[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 1996ab27746eSHong Zhang dk[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1997ab27746eSHong Zhang 1998ab27746eSHong Zhang /* update -U(i,k) */ 1999ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*25,uik,25*sizeof(MatScalar));CHKERRQ(ierr); 2000ab27746eSHong Zhang 2001ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2002ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2003ab27746eSHong Zhang if (jmin < jmax){ 2004ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2005ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2006ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*25; 2007ab27746eSHong Zhang u = ba + j*25; 2008ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 2009ab27746eSHong Zhang rtmp_ptr[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 2010ab27746eSHong Zhang rtmp_ptr[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 2011ab27746eSHong Zhang rtmp_ptr[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 2012ab27746eSHong Zhang rtmp_ptr[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 2013ab27746eSHong Zhang 2014ab27746eSHong Zhang rtmp_ptr[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 2015ab27746eSHong Zhang rtmp_ptr[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 2016ab27746eSHong Zhang rtmp_ptr[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 2017ab27746eSHong Zhang rtmp_ptr[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 2018ab27746eSHong Zhang rtmp_ptr[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 2019ab27746eSHong Zhang 2020ab27746eSHong Zhang rtmp_ptr[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 2021ab27746eSHong Zhang rtmp_ptr[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 2022ab27746eSHong Zhang rtmp_ptr[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 2023ab27746eSHong Zhang rtmp_ptr[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 2024ab27746eSHong Zhang rtmp_ptr[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 2025ab27746eSHong Zhang 2026ab27746eSHong Zhang rtmp_ptr[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 2027ab27746eSHong Zhang rtmp_ptr[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 2028ab27746eSHong Zhang rtmp_ptr[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 2029ab27746eSHong Zhang rtmp_ptr[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 2030ab27746eSHong Zhang rtmp_ptr[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 2031ab27746eSHong Zhang 2032ab27746eSHong Zhang rtmp_ptr[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 2033ab27746eSHong Zhang rtmp_ptr[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 2034ab27746eSHong Zhang rtmp_ptr[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 2035ab27746eSHong Zhang rtmp_ptr[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 2036ab27746eSHong Zhang rtmp_ptr[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 203749b5e25fSSatish Balay } 2038ab27746eSHong Zhang 2039ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2040ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2041ab27746eSHong Zhang j = bj[jmin]; 2042ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 204349b5e25fSSatish Balay } 2044ab27746eSHong Zhang i = nexti; 204549b5e25fSSatish Balay } 2046ab27746eSHong Zhang 2047ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2048ab27746eSHong Zhang 204949b5e25fSSatish Balay /* invert diagonal block */ 2050ab27746eSHong Zhang d = ba+k*25; 2051ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,25*sizeof(MatScalar));CHKERRQ(ierr); 2052ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_5(d);CHKERRQ(ierr); 2053ab27746eSHong Zhang 2054ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2055ab27746eSHong Zhang if (jmin < jmax) { 2056ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2057ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2058ab27746eSHong Zhang u = ba + j*25; 2059ab27746eSHong Zhang rtmp_ptr = rtmp + vj*25; 2060ab27746eSHong Zhang for (k1=0; k1<25; k1++){ 2061ab27746eSHong Zhang *u++ = *rtmp_ptr; 2062ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2063ab27746eSHong Zhang } 2064ab27746eSHong Zhang } 2065ab27746eSHong Zhang 2066ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2067ab27746eSHong Zhang il[k] = jmin; 2068ab27746eSHong Zhang i = bj[jmin]; 2069ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2070ab27746eSHong Zhang } 207149b5e25fSSatish Balay } 207249b5e25fSSatish Balay 207349b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2074ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2075ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2076ab27746eSHong Zhang 2077ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 207849b5e25fSSatish Balay C->assembled = PETSC_TRUE; 2079ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 2080*b0a32e0cSBarry Smith PetscLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 208149b5e25fSSatish Balay PetscFunctionReturn(0); 208249b5e25fSSatish Balay } 208349b5e25fSSatish Balay 208449b5e25fSSatish Balay /* 208549b5e25fSSatish Balay Version for when blocks are 4 by 4 Using natural ordering 208649b5e25fSSatish Balay */ 208749b5e25fSSatish Balay #undef __FUNC__ 20886f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering" 20896f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering(Mat A,Mat *B) 209049b5e25fSSatish Balay { 209149b5e25fSSatish Balay Mat C = *B; 2092ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2093ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 209471149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2095ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2096ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 209749b5e25fSSatish Balay 209849b5e25fSSatish Balay PetscFunctionBegin; 2099671cb588SHong Zhang 2100ab27746eSHong Zhang /* initialization */ 2101*b0a32e0cSBarry Smith ierr = PetscMalloc(16*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 2102ab27746eSHong Zhang ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2103*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 210428de702eSHong Zhang jl = il + mbs; 2105ab27746eSHong Zhang for (i=0; i<mbs; i++) { 2106ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 210749b5e25fSSatish Balay } 2108*b0a32e0cSBarry Smith ierr = PetscMalloc(32*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 210928de702eSHong Zhang uik = dk + 16; 2110ab27746eSHong Zhang 2111ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2112ab27746eSHong Zhang 2113ab27746eSHong Zhang /* for each row k */ 2114ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 2115ab27746eSHong Zhang 2116ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 2117ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 2118ab27746eSHong Zhang if (jmin < jmax) { 2119ab27746eSHong Zhang ap = aa + jmin*16; 2120ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 2121ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 2122ab27746eSHong Zhang rtmp_ptr = rtmp + vj*16; 2123ab27746eSHong Zhang for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 212449b5e25fSSatish Balay } 212549b5e25fSSatish Balay } 2126ab27746eSHong Zhang 2127ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2128ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 2129ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2130ab27746eSHong Zhang 2131ab27746eSHong Zhang while (i < mbs){ 2132ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2133ab27746eSHong Zhang 2134ab27746eSHong Zhang /* compute multiplier */ 2135ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2136ab27746eSHong Zhang 2137ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2138ab27746eSHong Zhang diag = ba + i*16; 2139ab27746eSHong Zhang u = ba + ili*16; 2140ab27746eSHong Zhang 2141ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 2142ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 2143ab27746eSHong Zhang uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 2144ab27746eSHong Zhang uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 2145ab27746eSHong Zhang 2146ab27746eSHong Zhang uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 2147ab27746eSHong Zhang uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 2148ab27746eSHong Zhang uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 2149ab27746eSHong Zhang uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 2150ab27746eSHong Zhang 2151ab27746eSHong Zhang uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 2152ab27746eSHong Zhang uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 2153ab27746eSHong Zhang uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 2154ab27746eSHong Zhang uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 2155ab27746eSHong Zhang 2156ab27746eSHong Zhang uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 2157ab27746eSHong Zhang uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 2158ab27746eSHong Zhang uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 2159ab27746eSHong Zhang uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 2160ab27746eSHong Zhang 2161ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2162ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2163ab27746eSHong Zhang dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2164ab27746eSHong Zhang dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2165ab27746eSHong Zhang dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2166ab27746eSHong Zhang 2167ab27746eSHong Zhang dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2168ab27746eSHong Zhang dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2169ab27746eSHong Zhang dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2170ab27746eSHong Zhang dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2171ab27746eSHong Zhang 2172ab27746eSHong Zhang dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2173ab27746eSHong Zhang dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2174ab27746eSHong Zhang dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2175ab27746eSHong Zhang dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2176ab27746eSHong Zhang 2177ab27746eSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2178ab27746eSHong Zhang dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2179ab27746eSHong Zhang dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2180ab27746eSHong Zhang dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2181ab27746eSHong Zhang 2182ab27746eSHong Zhang /* update -U(i,k) */ 2183ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 2184ab27746eSHong Zhang 2185ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2186ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2187ab27746eSHong Zhang if (jmin < jmax){ 2188ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2189ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2190ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*16; 2191ab27746eSHong Zhang u = ba + j*16; 2192ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2193ab27746eSHong Zhang rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2194ab27746eSHong Zhang rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2195ab27746eSHong Zhang rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2196ab27746eSHong Zhang 2197ab27746eSHong Zhang rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2198ab27746eSHong Zhang rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2199ab27746eSHong Zhang rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2200ab27746eSHong Zhang rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2201ab27746eSHong Zhang 2202ab27746eSHong Zhang rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2203ab27746eSHong Zhang rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2204ab27746eSHong Zhang rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2205ab27746eSHong Zhang rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2206ab27746eSHong Zhang 2207ab27746eSHong Zhang rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2208ab27746eSHong Zhang rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2209ab27746eSHong Zhang rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2210ab27746eSHong Zhang rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 221149b5e25fSSatish Balay } 2212ab27746eSHong Zhang 2213ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2214ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2215ab27746eSHong Zhang j = bj[jmin]; 2216ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 221749b5e25fSSatish Balay } 2218ab27746eSHong Zhang i = nexti; 221949b5e25fSSatish Balay } 2220ab27746eSHong Zhang 2221ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2222ab27746eSHong Zhang 222349b5e25fSSatish Balay /* invert diagonal block */ 2224ab27746eSHong Zhang diag = ba+k*16; 2225ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 2226ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 2227ab27746eSHong Zhang 2228ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2229ab27746eSHong Zhang if (jmin < jmax) { 2230ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2231ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2232ab27746eSHong Zhang u = ba + j*16; 2233ab27746eSHong Zhang rtmp_ptr = rtmp + vj*16; 2234ab27746eSHong Zhang for (k1=0; k1<16; k1++){ 2235ab27746eSHong Zhang *u++ = *rtmp_ptr; 2236ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2237ab27746eSHong Zhang } 2238ab27746eSHong Zhang } 2239ab27746eSHong Zhang 2240ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2241ab27746eSHong Zhang il[k] = jmin; 2242ab27746eSHong Zhang i = bj[jmin]; 2243ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2244ab27746eSHong Zhang } 224549b5e25fSSatish Balay } 224649b5e25fSSatish Balay 224749b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2248ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2249ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2250ab27746eSHong Zhang 2251ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 225249b5e25fSSatish Balay C->assembled = PETSC_TRUE; 22537a05c331SHong Zhang C->preallocated = PETSC_TRUE; 2254*b0a32e0cSBarry Smith PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 225549b5e25fSSatish Balay PetscFunctionReturn(0); 225649b5e25fSSatish Balay } 225749b5e25fSSatish Balay 2258effa298cSHong Zhang /* Version for when blocks are 4 by 4 */ 2259effa298cSHong Zhang #undef __FUNC__ 2260effa298cSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4" 2261effa298cSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_4(Mat A,Mat *B) 2262effa298cSHong Zhang { 2263effa298cSHong Zhang Mat C = *B; 2264effa298cSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2265effa298cSHong Zhang IS perm = b->row; 2266effa298cSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2267effa298cSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2268effa298cSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2269effa298cSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 2270effa298cSHong Zhang 2271effa298cSHong Zhang PetscFunctionBegin; 2272effa298cSHong Zhang /* initialization */ 2273*b0a32e0cSBarry Smith ierr = PetscMalloc(16*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 2274effa298cSHong Zhang ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2275*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 227628de702eSHong Zhang jl = il + mbs; 2277effa298cSHong Zhang for (i=0; i<mbs; i++) { 2278effa298cSHong Zhang jl[i] = mbs; il[0] = 0; 2279effa298cSHong Zhang } 2280*b0a32e0cSBarry Smith ierr = PetscMalloc(32*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 228128de702eSHong Zhang uik = dk + 16; 2282effa298cSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2283effa298cSHong Zhang 2284effa298cSHong Zhang /* check permutation */ 2285effa298cSHong Zhang if (!a->permute){ 2286effa298cSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2287effa298cSHong Zhang } else { 2288effa298cSHong Zhang ai = a->inew; aj = a->jnew; 2289*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(16*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2290effa298cSHong Zhang ierr = PetscMemcpy(aa,a->a,16*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2291*b0a32e0cSBarry Smith a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 2292effa298cSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2293effa298cSHong Zhang 2294effa298cSHong Zhang for (i=0; i<mbs; i++){ 2295effa298cSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2296effa298cSHong Zhang for (j=jmin; j<jmax; j++){ 2297effa298cSHong Zhang while (a2anew[j] != j){ 2298effa298cSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2299effa298cSHong Zhang for (k1=0; k1<16; k1++){ 2300effa298cSHong Zhang dk[k1] = aa[k*16+k1]; 2301effa298cSHong Zhang aa[k*16+k1] = aa[j*16+k1]; 2302effa298cSHong Zhang aa[j*16+k1] = dk[k1]; 2303effa298cSHong Zhang } 2304effa298cSHong Zhang } 2305effa298cSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2306effa298cSHong Zhang if (i > aj[j]){ 2307effa298cSHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2308effa298cSHong Zhang ap = aa + j*16; /* ptr to the beginning of j-th block of aa */ 2309effa298cSHong Zhang for (k=0; k<16; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2310effa298cSHong Zhang for (k=0; k<4; k++){ /* j-th block of aa <- dk^T */ 2311effa298cSHong Zhang for (k1=0; k1<4; k1++) *ap++ = dk[k + 4*k1]; 2312effa298cSHong Zhang } 2313effa298cSHong Zhang } 2314effa298cSHong Zhang } 2315effa298cSHong Zhang } 2316effa298cSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2317effa298cSHong Zhang } 2318effa298cSHong Zhang 2319effa298cSHong Zhang /* for each row k */ 2320effa298cSHong Zhang for (k = 0; k<mbs; k++){ 2321effa298cSHong Zhang 2322effa298cSHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2323effa298cSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 2324effa298cSHong Zhang if (jmin < jmax) { 2325effa298cSHong Zhang ap = aa + jmin*16; 2326effa298cSHong Zhang for (j = jmin; j < jmax; j++){ 2327effa298cSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2328effa298cSHong Zhang rtmp_ptr = rtmp + vj*16; 2329effa298cSHong Zhang for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 2330effa298cSHong Zhang } 2331effa298cSHong Zhang } 2332effa298cSHong Zhang 2333effa298cSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2334effa298cSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 2335effa298cSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2336effa298cSHong Zhang 2337effa298cSHong Zhang while (i < mbs){ 2338effa298cSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2339effa298cSHong Zhang 2340effa298cSHong Zhang /* compute multiplier */ 2341effa298cSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2342effa298cSHong Zhang 2343effa298cSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2344effa298cSHong Zhang diag = ba + i*16; 2345effa298cSHong Zhang u = ba + ili*16; 2346effa298cSHong Zhang 2347effa298cSHong Zhang uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 2348effa298cSHong Zhang uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 2349effa298cSHong Zhang uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 2350effa298cSHong Zhang uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 2351effa298cSHong Zhang 2352effa298cSHong Zhang uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 2353effa298cSHong Zhang uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 2354effa298cSHong Zhang uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 2355effa298cSHong Zhang uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 2356effa298cSHong Zhang 2357effa298cSHong Zhang uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 2358effa298cSHong Zhang uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 2359effa298cSHong Zhang uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 2360effa298cSHong Zhang uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 2361effa298cSHong Zhang 2362effa298cSHong Zhang uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 2363effa298cSHong Zhang uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 2364effa298cSHong Zhang uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 2365effa298cSHong Zhang uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 2366effa298cSHong Zhang 2367effa298cSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2368effa298cSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2369effa298cSHong Zhang dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2370effa298cSHong Zhang dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2371effa298cSHong Zhang dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2372effa298cSHong Zhang 2373effa298cSHong Zhang dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2374effa298cSHong Zhang dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2375effa298cSHong Zhang dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2376effa298cSHong Zhang dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2377effa298cSHong Zhang 2378effa298cSHong Zhang dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2379effa298cSHong Zhang dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2380effa298cSHong Zhang dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2381effa298cSHong Zhang dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2382effa298cSHong Zhang 2383effa298cSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2384effa298cSHong Zhang dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2385effa298cSHong Zhang dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2386effa298cSHong Zhang dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2387effa298cSHong Zhang 2388effa298cSHong Zhang /* update -U(i,k) */ 2389effa298cSHong Zhang ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 2390effa298cSHong Zhang 2391effa298cSHong Zhang /* add multiple of row i to k-th row ... */ 2392effa298cSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2393effa298cSHong Zhang if (jmin < jmax){ 2394effa298cSHong Zhang for (j=jmin; j<jmax; j++) { 2395effa298cSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2396effa298cSHong Zhang rtmp_ptr = rtmp + bj[j]*16; 2397effa298cSHong Zhang u = ba + j*16; 2398effa298cSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2399effa298cSHong Zhang rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2400effa298cSHong Zhang rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2401effa298cSHong Zhang rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2402effa298cSHong Zhang 2403effa298cSHong Zhang rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2404effa298cSHong Zhang rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2405effa298cSHong Zhang rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2406effa298cSHong Zhang rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2407effa298cSHong Zhang 2408effa298cSHong Zhang rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2409effa298cSHong Zhang rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2410effa298cSHong Zhang rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2411effa298cSHong Zhang rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2412effa298cSHong Zhang 2413effa298cSHong Zhang rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2414effa298cSHong Zhang rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2415effa298cSHong Zhang rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2416effa298cSHong Zhang rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2417effa298cSHong Zhang } 2418effa298cSHong Zhang 2419effa298cSHong Zhang /* ... add i to row list for next nonzero entry */ 2420effa298cSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2421effa298cSHong Zhang j = bj[jmin]; 2422effa298cSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 2423effa298cSHong Zhang } 2424effa298cSHong Zhang i = nexti; 2425effa298cSHong Zhang } 2426effa298cSHong Zhang 2427effa298cSHong Zhang /* save nonzero entries in k-th row of U ... */ 2428effa298cSHong Zhang 2429effa298cSHong Zhang /* invert diagonal block */ 2430effa298cSHong Zhang diag = ba+k*16; 2431effa298cSHong Zhang ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 2432effa298cSHong Zhang ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 2433effa298cSHong Zhang 2434effa298cSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2435effa298cSHong Zhang if (jmin < jmax) { 2436effa298cSHong Zhang for (j=jmin; j<jmax; j++){ 2437effa298cSHong Zhang vj = bj[j]; /* block col. index of U */ 2438effa298cSHong Zhang u = ba + j*16; 2439effa298cSHong Zhang rtmp_ptr = rtmp + vj*16; 2440effa298cSHong Zhang for (k1=0; k1<16; k1++){ 2441effa298cSHong Zhang *u++ = *rtmp_ptr; 2442effa298cSHong Zhang *rtmp_ptr++ = 0.0; 2443effa298cSHong Zhang } 2444effa298cSHong Zhang } 2445effa298cSHong Zhang 2446effa298cSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2447effa298cSHong Zhang il[k] = jmin; 2448effa298cSHong Zhang i = bj[jmin]; 2449effa298cSHong Zhang jl[k] = jl[i]; jl[i] = k; 2450effa298cSHong Zhang } 2451effa298cSHong Zhang } 2452effa298cSHong Zhang 2453effa298cSHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 2454effa298cSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2455effa298cSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2456effa298cSHong Zhang if (a->permute){ 2457effa298cSHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 2458effa298cSHong Zhang } 2459effa298cSHong Zhang 2460effa298cSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 2461effa298cSHong Zhang C->factor = FACTOR_CHOLESKY; 2462effa298cSHong Zhang C->assembled = PETSC_TRUE; 2463effa298cSHong Zhang C->preallocated = PETSC_TRUE; 2464*b0a32e0cSBarry Smith PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 2465effa298cSHong Zhang PetscFunctionReturn(0); 2466effa298cSHong Zhang } 2467effa298cSHong Zhang 2468a1723e09SHong Zhang /* Version for when blocks are 3 by 3 */ 246949b5e25fSSatish Balay #undef __FUNC__ 24706f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_3" 24716f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_3(Mat A,Mat *B) 247249b5e25fSSatish Balay { 247349b5e25fSSatish Balay Mat C = *B; 2474a1723e09SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2475a1723e09SHong Zhang IS perm = b->row; 2476a1723e09SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2477a1723e09SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2478a1723e09SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2479a1723e09SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 248049b5e25fSSatish Balay 248149b5e25fSSatish Balay PetscFunctionBegin; 2482671cb588SHong Zhang 2483a1723e09SHong Zhang /* initialization */ 2484*b0a32e0cSBarry Smith ierr = PetscMalloc(9*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 2485a1723e09SHong Zhang ierr = PetscMemzero(rtmp,9*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2486*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 248728de702eSHong Zhang jl = il + mbs; 2488a1723e09SHong Zhang for (i=0; i<mbs; i++) { 2489a1723e09SHong Zhang jl[i] = mbs; il[0] = 0; 2490a1723e09SHong Zhang } 2491*b0a32e0cSBarry Smith ierr = PetscMalloc(18*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 249228de702eSHong Zhang uik = dk + 9; 2493a1723e09SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 249449b5e25fSSatish Balay 2495a1723e09SHong Zhang /* check permutation */ 2496a1723e09SHong Zhang if (!a->permute){ 2497a1723e09SHong Zhang ai = a->i; aj = a->j; aa = a->a; 2498a1723e09SHong Zhang } else { 2499a1723e09SHong Zhang ai = a->inew; aj = a->jnew; 2500*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(9*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2501a1723e09SHong Zhang ierr = PetscMemcpy(aa,a->a,9*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2502*b0a32e0cSBarry Smith a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 2503a1723e09SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 250449b5e25fSSatish Balay 2505a1723e09SHong Zhang for (i=0; i<mbs; i++){ 2506a1723e09SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2507a1723e09SHong Zhang for (j=jmin; j<jmax; j++){ 2508a1723e09SHong Zhang while (a2anew[j] != j){ 2509a1723e09SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2510a1723e09SHong Zhang for (k1=0; k1<9; k1++){ 2511a1723e09SHong Zhang dk[k1] = aa[k*9+k1]; 2512a1723e09SHong Zhang aa[k*9+k1] = aa[j*9+k1]; 2513a1723e09SHong Zhang aa[j*9+k1] = dk[k1]; 2514a1723e09SHong Zhang } 2515a1723e09SHong Zhang } 2516a1723e09SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2517a1723e09SHong Zhang if (i > aj[j]){ 251847c40530SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2519a1723e09SHong Zhang ap = aa + j*9; /* ptr to the beginning of j-th block of aa */ 2520a1723e09SHong Zhang for (k=0; k<9; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2521a1723e09SHong Zhang for (k=0; k<3; k++){ /* j-th block of aa <- dk^T */ 252247c40530SHong Zhang for (k1=0; k1<3; k1++) *ap++ = dk[k + 3*k1]; 2523a1723e09SHong Zhang } 2524a1723e09SHong Zhang } 2525a1723e09SHong Zhang } 2526a1723e09SHong Zhang } 2527a1723e09SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2528a1723e09SHong Zhang } 252949b5e25fSSatish Balay 2530a1723e09SHong Zhang /* for each row k */ 2531a1723e09SHong Zhang for (k = 0; k<mbs; k++){ 253249b5e25fSSatish Balay 2533a1723e09SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2534a1723e09SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 2535a1723e09SHong Zhang if (jmin < jmax) { 2536a1723e09SHong Zhang ap = aa + jmin*9; 2537a1723e09SHong Zhang for (j = jmin; j < jmax; j++){ 2538a1723e09SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2539a1723e09SHong Zhang rtmp_ptr = rtmp + vj*9; 2540a1723e09SHong Zhang for (i=0; i<9; i++) *rtmp_ptr++ = *ap++; 2541a1723e09SHong Zhang } 2542a1723e09SHong Zhang } 254349b5e25fSSatish Balay 2544a1723e09SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2545a1723e09SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*9,9*sizeof(MatScalar));CHKERRQ(ierr); 2546a1723e09SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2547a1723e09SHong Zhang 2548a1723e09SHong Zhang while (i < mbs){ 2549a1723e09SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2550a1723e09SHong Zhang 2551a1723e09SHong Zhang /* compute multiplier */ 2552a1723e09SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2553a1723e09SHong Zhang 2554a1723e09SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2555a1723e09SHong Zhang diag = ba + i*9; 2556a1723e09SHong Zhang u = ba + ili*9; 2557a1723e09SHong Zhang 2558a1723e09SHong Zhang uik[0] = -(diag[0]*u[0] + diag[3]*u[1] + diag[6]*u[2]); 2559a1723e09SHong Zhang uik[1] = -(diag[1]*u[0] + diag[4]*u[1] + diag[7]*u[2]); 2560a1723e09SHong Zhang uik[2] = -(diag[2]*u[0] + diag[5]*u[1] + diag[8]*u[2]); 2561a1723e09SHong Zhang 2562a1723e09SHong Zhang uik[3] = -(diag[0]*u[3] + diag[3]*u[4] + diag[6]*u[5]); 2563a1723e09SHong Zhang uik[4] = -(diag[1]*u[3] + diag[4]*u[4] + diag[7]*u[5]); 2564a1723e09SHong Zhang uik[5] = -(diag[2]*u[3] + diag[5]*u[4] + diag[8]*u[5]); 2565a1723e09SHong Zhang 2566a1723e09SHong Zhang uik[6] = -(diag[0]*u[6] + diag[3]*u[7] + diag[6]*u[8]); 2567a1723e09SHong Zhang uik[7] = -(diag[1]*u[6] + diag[4]*u[7] + diag[7]*u[8]); 2568a1723e09SHong Zhang uik[8] = -(diag[2]*u[6] + diag[5]*u[7] + diag[8]*u[8]); 2569a1723e09SHong Zhang 2570a1723e09SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2571a1723e09SHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2572a1723e09SHong Zhang dk[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2573a1723e09SHong Zhang dk[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2574a1723e09SHong Zhang 2575a1723e09SHong Zhang dk[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2576a1723e09SHong Zhang dk[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2577a1723e09SHong Zhang dk[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2578a1723e09SHong Zhang 2579a1723e09SHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2580a1723e09SHong Zhang dk[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2581a1723e09SHong Zhang dk[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 2582a1723e09SHong Zhang 2583a1723e09SHong Zhang /* update -U(i,k) */ 2584a1723e09SHong Zhang ierr = PetscMemcpy(ba+ili*9,uik,9*sizeof(MatScalar));CHKERRQ(ierr); 2585a1723e09SHong Zhang 2586a1723e09SHong Zhang /* add multiple of row i to k-th row ... */ 2587a1723e09SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2588a1723e09SHong Zhang if (jmin < jmax){ 2589a1723e09SHong Zhang for (j=jmin; j<jmax; j++) { 2590a1723e09SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2591a1723e09SHong Zhang rtmp_ptr = rtmp + bj[j]*9; 2592a1723e09SHong Zhang u = ba + j*9; 2593a1723e09SHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2594a1723e09SHong Zhang rtmp_ptr[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2595a1723e09SHong Zhang rtmp_ptr[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2596a1723e09SHong Zhang 2597a1723e09SHong Zhang rtmp_ptr[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2598a1723e09SHong Zhang rtmp_ptr[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2599a1723e09SHong Zhang rtmp_ptr[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2600a1723e09SHong Zhang 2601a1723e09SHong Zhang rtmp_ptr[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2602a1723e09SHong Zhang rtmp_ptr[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2603a1723e09SHong Zhang rtmp_ptr[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 260449b5e25fSSatish Balay } 2605a1723e09SHong Zhang 2606a1723e09SHong Zhang /* ... add i to row list for next nonzero entry */ 2607a1723e09SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2608a1723e09SHong Zhang j = bj[jmin]; 2609a1723e09SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 261049b5e25fSSatish Balay } 2611a1723e09SHong Zhang i = nexti; 261249b5e25fSSatish Balay } 2613a1723e09SHong Zhang 2614a1723e09SHong Zhang /* save nonzero entries in k-th row of U ... */ 2615a1723e09SHong Zhang 261649b5e25fSSatish Balay /* invert diagonal block */ 2617a1723e09SHong Zhang diag = ba+k*9; 2618a1723e09SHong Zhang ierr = PetscMemcpy(diag,dk,9*sizeof(MatScalar));CHKERRQ(ierr); 2619a1723e09SHong Zhang ierr = Kernel_A_gets_inverse_A_3(diag);CHKERRQ(ierr); 2620a1723e09SHong Zhang 2621a1723e09SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2622a1723e09SHong Zhang if (jmin < jmax) { 2623a1723e09SHong Zhang for (j=jmin; j<jmax; j++){ 2624a1723e09SHong Zhang vj = bj[j]; /* block col. index of U */ 2625a1723e09SHong Zhang u = ba + j*9; 2626a1723e09SHong Zhang rtmp_ptr = rtmp + vj*9; 2627a1723e09SHong Zhang for (k1=0; k1<9; k1++){ 2628a1723e09SHong Zhang *u++ = *rtmp_ptr; 2629a1723e09SHong Zhang *rtmp_ptr++ = 0.0; 2630a1723e09SHong Zhang } 2631a1723e09SHong Zhang } 2632a1723e09SHong Zhang 2633a1723e09SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2634a1723e09SHong Zhang il[k] = jmin; 2635a1723e09SHong Zhang i = bj[jmin]; 2636a1723e09SHong Zhang jl[k] = jl[i]; jl[i] = k; 2637a1723e09SHong Zhang } 263849b5e25fSSatish Balay } 263949b5e25fSSatish Balay 264049b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2641a1723e09SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2642a1723e09SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2643a1723e09SHong Zhang if (a->permute){ 2644a1723e09SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 2645a1723e09SHong Zhang } 2646a1723e09SHong Zhang 2647a1723e09SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 2648a1723e09SHong Zhang C->factor = FACTOR_CHOLESKY; 264949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 26507a05c331SHong Zhang C->preallocated = PETSC_TRUE; 2651*b0a32e0cSBarry Smith PetscLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 265249b5e25fSSatish Balay PetscFunctionReturn(0); 265349b5e25fSSatish Balay } 2654a1723e09SHong Zhang 265549b5e25fSSatish Balay /* 265649b5e25fSSatish Balay Version for when blocks are 3 by 3 Using natural ordering 265749b5e25fSSatish Balay */ 265849b5e25fSSatish Balay #undef __FUNC__ 26596f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering" 26606f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering(Mat A,Mat *B) 266149b5e25fSSatish Balay { 266249b5e25fSSatish Balay Mat C = *B; 2663ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 266471149678SHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 266571149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2666ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2667ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 266849b5e25fSSatish Balay 266949b5e25fSSatish Balay PetscFunctionBegin; 2670671cb588SHong Zhang 2671ab27746eSHong Zhang /* initialization */ 2672*b0a32e0cSBarry Smith ierr = PetscMalloc(9*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 2673ab27746eSHong Zhang ierr = PetscMemzero(rtmp,9*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2674*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 267528de702eSHong Zhang jl = il + mbs; 2676ab27746eSHong Zhang for (i=0; i<mbs; i++) { 2677ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 267849b5e25fSSatish Balay } 2679*b0a32e0cSBarry Smith ierr = PetscMalloc(18*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 268028de702eSHong Zhang uik = dk + 9; 268149b5e25fSSatish Balay 2682ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 268349b5e25fSSatish Balay 2684ab27746eSHong Zhang /* for each row k */ 2685ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 268649b5e25fSSatish Balay 2687ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 2688ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 2689ab27746eSHong Zhang if (jmin < jmax) { 2690ab27746eSHong Zhang ap = aa + jmin*9; 2691ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 2692ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 2693ab27746eSHong Zhang rtmp_ptr = rtmp + vj*9; 2694ab27746eSHong Zhang for (i=0; i<9; i++) *rtmp_ptr++ = *ap++; 2695ab27746eSHong Zhang } 2696ab27746eSHong Zhang } 269749b5e25fSSatish Balay 2698ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2699ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*9,9*sizeof(MatScalar));CHKERRQ(ierr); 2700ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 270149b5e25fSSatish Balay 2702ab27746eSHong Zhang while (i < mbs){ 2703ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2704ab27746eSHong Zhang 2705ab27746eSHong Zhang /* compute multiplier */ 2706ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2707ab27746eSHong Zhang 2708ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2709ab27746eSHong Zhang diag = ba + i*9; 2710ab27746eSHong Zhang u = ba + ili*9; 2711ab27746eSHong Zhang 2712ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[3]*u[1] + diag[6]*u[2]); 2713ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[4]*u[1] + diag[7]*u[2]); 2714ab27746eSHong Zhang uik[2] = -(diag[2]*u[0] + diag[5]*u[1] + diag[8]*u[2]); 2715ab27746eSHong Zhang 2716ab27746eSHong Zhang uik[3] = -(diag[0]*u[3] + diag[3]*u[4] + diag[6]*u[5]); 2717ab27746eSHong Zhang uik[4] = -(diag[1]*u[3] + diag[4]*u[4] + diag[7]*u[5]); 2718ab27746eSHong Zhang uik[5] = -(diag[2]*u[3] + diag[5]*u[4] + diag[8]*u[5]); 2719ab27746eSHong Zhang 2720ab27746eSHong Zhang uik[6] = -(diag[0]*u[6] + diag[3]*u[7] + diag[6]*u[8]); 2721ab27746eSHong Zhang uik[7] = -(diag[1]*u[6] + diag[4]*u[7] + diag[7]*u[8]); 2722ab27746eSHong Zhang uik[8] = -(diag[2]*u[6] + diag[5]*u[7] + diag[8]*u[8]); 2723ab27746eSHong Zhang 2724ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2725ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2726ab27746eSHong Zhang dk[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2727ab27746eSHong Zhang dk[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2728ab27746eSHong Zhang 2729ab27746eSHong Zhang dk[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2730ab27746eSHong Zhang dk[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2731ab27746eSHong Zhang dk[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2732ab27746eSHong Zhang 2733ab27746eSHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2734ab27746eSHong Zhang dk[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2735ab27746eSHong Zhang dk[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 2736ab27746eSHong Zhang 2737ab27746eSHong Zhang /* update -U(i,k) */ 2738ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*9,uik,9*sizeof(MatScalar));CHKERRQ(ierr); 2739ab27746eSHong Zhang 2740ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2741ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2742ab27746eSHong Zhang if (jmin < jmax){ 2743ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2744ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2745ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*9; 2746ab27746eSHong Zhang u = ba + j*9; 2747ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2748ab27746eSHong Zhang rtmp_ptr[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2749ab27746eSHong Zhang rtmp_ptr[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2750ab27746eSHong Zhang 2751ab27746eSHong Zhang rtmp_ptr[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2752ab27746eSHong Zhang rtmp_ptr[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2753ab27746eSHong Zhang rtmp_ptr[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2754ab27746eSHong Zhang 2755ab27746eSHong Zhang rtmp_ptr[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2756ab27746eSHong Zhang rtmp_ptr[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2757ab27746eSHong Zhang rtmp_ptr[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 275849b5e25fSSatish Balay } 2759ab27746eSHong Zhang 2760ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2761ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2762ab27746eSHong Zhang j = bj[jmin]; 2763ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 276449b5e25fSSatish Balay } 2765ab27746eSHong Zhang i = nexti; 276649b5e25fSSatish Balay } 2767ab27746eSHong Zhang 2768ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2769ab27746eSHong Zhang 277049b5e25fSSatish Balay /* invert diagonal block */ 2771ab27746eSHong Zhang diag = ba+k*9; 2772ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,9*sizeof(MatScalar));CHKERRQ(ierr); 2773ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_3(diag);CHKERRQ(ierr); 2774ab27746eSHong Zhang 2775ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2776ab27746eSHong Zhang if (jmin < jmax) { 2777ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2778ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2779ab27746eSHong Zhang u = ba + j*9; 2780ab27746eSHong Zhang rtmp_ptr = rtmp + vj*9; 2781ab27746eSHong Zhang for (k1=0; k1<9; k1++){ 2782ab27746eSHong Zhang *u++ = *rtmp_ptr; 2783ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2784ab27746eSHong Zhang } 2785ab27746eSHong Zhang } 2786ab27746eSHong Zhang 2787ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2788ab27746eSHong Zhang il[k] = jmin; 2789ab27746eSHong Zhang i = bj[jmin]; 2790ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2791ab27746eSHong Zhang } 279249b5e25fSSatish Balay } 279349b5e25fSSatish Balay 279449b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2795ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2796ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2797ab27746eSHong Zhang 2798effa298cSHong Zhang C->factor = FACTOR_CHOLESKY; 279949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 28005c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 2801*b0a32e0cSBarry Smith PetscLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 280249b5e25fSSatish Balay PetscFunctionReturn(0); 280349b5e25fSSatish Balay } 280449b5e25fSSatish Balay 280549b5e25fSSatish Balay /* 2806fcf159c0SHong Zhang Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP. 2807cc0c071aSHong Zhang Version for blocks 2 by 2. 280849b5e25fSSatish Balay */ 280949b5e25fSSatish Balay #undef __FUNC__ 28106f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_2" 28116f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_2(Mat A,Mat *B) 281249b5e25fSSatish Balay { 281349b5e25fSSatish Balay Mat C = *B; 281491602c66SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2815cc0c071aSHong Zhang IS perm = b->row; 2816cc0c071aSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2817cc0c071aSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2818a1723e09SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2819cc0c071aSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 282049b5e25fSSatish Balay 282149b5e25fSSatish Balay PetscFunctionBegin; 2822671cb588SHong Zhang 282391602c66SHong Zhang /* initialization */ 282491602c66SHong Zhang /* il and jl record the first nonzero element in each row of the accessing 282591602c66SHong Zhang window U(0:k, k:mbs-1). 282691602c66SHong Zhang jl: list of rows to be added to uneliminated rows 282791602c66SHong Zhang i>= k: jl(i) is the first row to be added to row i 282891602c66SHong Zhang i< k: jl(i) is the row following row i in some list of rows 282991602c66SHong Zhang jl(i) = mbs indicates the end of a list 283091602c66SHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 283191602c66SHong Zhang row i of U */ 2832*b0a32e0cSBarry Smith ierr = PetscMalloc(4*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 2833cc0c071aSHong Zhang ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2834*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 283528de702eSHong Zhang jl = il + mbs; 283691602c66SHong Zhang for (i=0; i<mbs; i++) { 28373845f261SHong Zhang jl[i] = mbs; il[0] = 0; 283891602c66SHong Zhang } 2839*b0a32e0cSBarry Smith ierr = PetscMalloc(8*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 284028de702eSHong Zhang uik = dk + 4; 2841cc0c071aSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2842cc0c071aSHong Zhang 2843cc0c071aSHong Zhang /* check permutation */ 2844cc0c071aSHong Zhang if (!a->permute){ 2845cc0c071aSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2846cc0c071aSHong Zhang } else { 2847cc0c071aSHong Zhang ai = a->inew; aj = a->jnew; 2848*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(4*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2849cc0c071aSHong Zhang ierr = PetscMemcpy(aa,a->a,4*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2850*b0a32e0cSBarry Smith a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 2851cc0c071aSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2852cc0c071aSHong Zhang 2853cc0c071aSHong Zhang for (i=0; i<mbs; i++){ 2854cc0c071aSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2855cc0c071aSHong Zhang for (j=jmin; j<jmax; j++){ 2856cc0c071aSHong Zhang while (a2anew[j] != j){ 2857cc0c071aSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2858cc0c071aSHong Zhang for (k1=0; k1<4; k1++){ 2859cc0c071aSHong Zhang dk[k1] = aa[k*4+k1]; 2860cc0c071aSHong Zhang aa[k*4+k1] = aa[j*4+k1]; 2861cc0c071aSHong Zhang aa[j*4+k1] = dk[k1]; 2862cc0c071aSHong Zhang } 2863cc0c071aSHong Zhang } 2864cc0c071aSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2865cc0c071aSHong Zhang if (i > aj[j]){ 2866a1723e09SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2867cc0c071aSHong Zhang ap = aa + j*4; /* ptr to the beginning of the block */ 2868cc0c071aSHong Zhang dk[1] = ap[1]; /* swap ap[1] and ap[2] */ 2869cc0c071aSHong Zhang ap[1] = ap[2]; 2870cc0c071aSHong Zhang ap[2] = dk[1]; 2871cc0c071aSHong Zhang } 2872cc0c071aSHong Zhang } 2873cc0c071aSHong Zhang } 2874cc0c071aSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2875cc0c071aSHong Zhang } 28763845f261SHong Zhang 287791602c66SHong Zhang /* for each row k */ 287891602c66SHong Zhang for (k = 0; k<mbs; k++){ 287991602c66SHong Zhang 288091602c66SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2881cc0c071aSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 288291602c66SHong Zhang if (jmin < jmax) { 2883cc0c071aSHong Zhang ap = aa + jmin*4; 288491602c66SHong Zhang for (j = jmin; j < jmax; j++){ 2885cc0c071aSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2886cc0c071aSHong Zhang rtmp_ptr = rtmp + vj*4; 2887cc0c071aSHong Zhang for (i=0; i<4; i++) *rtmp_ptr++ = *ap++; 288891602c66SHong Zhang } 288991602c66SHong Zhang } 289091602c66SHong Zhang 289191602c66SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2892cc0c071aSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr); 289391602c66SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 289491602c66SHong Zhang 289591602c66SHong Zhang while (i < mbs){ 289691602c66SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 289791602c66SHong Zhang 28983845f261SHong Zhang /* compute multiplier */ 289991602c66SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 29003845f261SHong Zhang 29013845f261SHong Zhang /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */ 2902cc0c071aSHong Zhang diag = ba + i*4; 2903cc0c071aSHong Zhang u = ba + ili*4; 2904cc0c071aSHong Zhang uik[0] = -(diag[0]*u[0] + diag[2]*u[1]); 2905cc0c071aSHong Zhang uik[1] = -(diag[1]*u[0] + diag[3]*u[1]); 2906cc0c071aSHong Zhang uik[2] = -(diag[0]*u[2] + diag[2]*u[3]); 2907cc0c071aSHong Zhang uik[3] = -(diag[1]*u[2] + diag[3]*u[3]); 29083845f261SHong Zhang 29093845f261SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */ 2910cc0c071aSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1]; 2911cc0c071aSHong Zhang dk[1] += uik[2]*u[0] + uik[3]*u[1]; 2912cc0c071aSHong Zhang dk[2] += uik[0]*u[2] + uik[1]*u[3]; 2913cc0c071aSHong Zhang dk[3] += uik[2]*u[2] + uik[3]*u[3]; 29143845f261SHong Zhang 29153845f261SHong Zhang /* update -U(i,k): ba[ili] = uik */ 2916cc0c071aSHong Zhang ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr); 291791602c66SHong Zhang 291891602c66SHong Zhang /* add multiple of row i to k-th row ... */ 291991602c66SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 292091602c66SHong Zhang if (jmin < jmax){ 29213845f261SHong Zhang for (j=jmin; j<jmax; j++) { 29223845f261SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */ 2923cc0c071aSHong Zhang rtmp_ptr = rtmp + bj[j]*4; 2924cc0c071aSHong Zhang u = ba + j*4; 2925cc0c071aSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1]; 2926cc0c071aSHong Zhang rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1]; 2927cc0c071aSHong Zhang rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3]; 2928cc0c071aSHong Zhang rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3]; 29293845f261SHong Zhang } 29303845f261SHong Zhang 293191602c66SHong Zhang /* ... add i to row list for next nonzero entry */ 293291602c66SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 293391602c66SHong Zhang j = bj[jmin]; 293491602c66SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 293591602c66SHong Zhang } 2936a1723e09SHong Zhang i = nexti; 293791602c66SHong Zhang } 2938cc0c071aSHong Zhang 293991602c66SHong Zhang /* save nonzero entries in k-th row of U ... */ 29403845f261SHong Zhang 2941cc0c071aSHong Zhang /* invert diagonal block */ 2942cc0c071aSHong Zhang diag = ba+k*4; 2943cc0c071aSHong Zhang ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr); 29443845f261SHong Zhang ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr); 29453845f261SHong Zhang 294691602c66SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 294791602c66SHong Zhang if (jmin < jmax) { 294891602c66SHong Zhang for (j=jmin; j<jmax; j++){ 2949cc0c071aSHong Zhang vj = bj[j]; /* block col. index of U */ 2950cc0c071aSHong Zhang u = ba + j*4; 2951cc0c071aSHong Zhang rtmp_ptr = rtmp + vj*4; 2952cc0c071aSHong Zhang for (k1=0; k1<4; k1++){ 2953cc0c071aSHong Zhang *u++ = *rtmp_ptr; 2954cc0c071aSHong Zhang *rtmp_ptr++ = 0.0; 2955cc0c071aSHong Zhang } 2956cc0c071aSHong Zhang } 29573845f261SHong Zhang 295891602c66SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 295991602c66SHong Zhang il[k] = jmin; 296091602c66SHong Zhang i = bj[jmin]; 296191602c66SHong Zhang jl[k] = jl[i]; jl[i] = k; 296291602c66SHong Zhang } 296391602c66SHong Zhang } 29643845f261SHong Zhang 296549b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 296691602c66SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 29673845f261SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 296891602c66SHong Zhang if (a->permute) { 296991602c66SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 297091602c66SHong Zhang } 2971cc0c071aSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 297291602c66SHong Zhang C->factor = FACTOR_CHOLESKY; 297349b5e25fSSatish Balay C->assembled = PETSC_TRUE; 29745c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 2975*b0a32e0cSBarry Smith PetscLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 297649b5e25fSSatish Balay PetscFunctionReturn(0); 297749b5e25fSSatish Balay } 297891602c66SHong Zhang 297949b5e25fSSatish Balay /* 298049b5e25fSSatish Balay Version for when blocks are 2 by 2 Using natural ordering 298149b5e25fSSatish Balay */ 298249b5e25fSSatish Balay #undef __FUNC__ 29836f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering" 29846f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering(Mat A,Mat *B) 298549b5e25fSSatish Balay { 298649b5e25fSSatish Balay Mat C = *B; 2987ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2988ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 298971149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2990ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2991ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 299249b5e25fSSatish Balay 299349b5e25fSSatish Balay PetscFunctionBegin; 2994671cb588SHong Zhang 2995ab27746eSHong Zhang /* initialization */ 2996ab27746eSHong Zhang /* il and jl record the first nonzero element in each row of the accessing 2997ab27746eSHong Zhang window U(0:k, k:mbs-1). 2998ab27746eSHong Zhang jl: list of rows to be added to uneliminated rows 2999ab27746eSHong Zhang i>= k: jl(i) is the first row to be added to row i 3000ab27746eSHong Zhang i< k: jl(i) is the row following row i in some list of rows 3001ab27746eSHong Zhang jl(i) = mbs indicates the end of a list 3002ab27746eSHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 3003ab27746eSHong Zhang row i of U */ 3004*b0a32e0cSBarry Smith ierr = PetscMalloc(4*mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 3005ab27746eSHong Zhang ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr); 3006*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 300728de702eSHong Zhang jl = il + mbs; 3008ab27746eSHong Zhang for (i=0; i<mbs; i++) { 3009ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 301049b5e25fSSatish Balay } 3011*b0a32e0cSBarry Smith ierr = PetscMalloc(8*sizeof(MatScalar),&( dk ));CHKERRQ(ierr); 301228de702eSHong Zhang uik = dk + 4; 3013ab27746eSHong Zhang 3014ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 3015ab27746eSHong Zhang 3016ab27746eSHong Zhang /* for each row k */ 3017ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 3018ab27746eSHong Zhang 3019ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 3020ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 3021ab27746eSHong Zhang if (jmin < jmax) { 3022ab27746eSHong Zhang ap = aa + jmin*4; 3023ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 3024ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 3025ab27746eSHong Zhang rtmp_ptr = rtmp + vj*4; 3026ab27746eSHong Zhang for (i=0; i<4; i++) *rtmp_ptr++ = *ap++; 302749b5e25fSSatish Balay } 302849b5e25fSSatish Balay } 3029ab27746eSHong Zhang 3030ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 3031ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr); 3032ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 3033ab27746eSHong Zhang 3034ab27746eSHong Zhang while (i < mbs){ 3035ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 3036ab27746eSHong Zhang 3037ab27746eSHong Zhang /* compute multiplier */ 3038ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 3039ab27746eSHong Zhang 3040ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */ 3041ab27746eSHong Zhang diag = ba + i*4; 3042ab27746eSHong Zhang u = ba + ili*4; 3043ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[2]*u[1]); 3044ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[3]*u[1]); 3045ab27746eSHong Zhang uik[2] = -(diag[0]*u[2] + diag[2]*u[3]); 3046ab27746eSHong Zhang uik[3] = -(diag[1]*u[2] + diag[3]*u[3]); 3047ab27746eSHong Zhang 3048ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */ 3049ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1]; 3050ab27746eSHong Zhang dk[1] += uik[2]*u[0] + uik[3]*u[1]; 3051ab27746eSHong Zhang dk[2] += uik[0]*u[2] + uik[1]*u[3]; 3052ab27746eSHong Zhang dk[3] += uik[2]*u[2] + uik[3]*u[3]; 3053ab27746eSHong Zhang 3054ab27746eSHong Zhang /* update -U(i,k): ba[ili] = uik */ 3055ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr); 3056ab27746eSHong Zhang 3057ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 3058ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 3059ab27746eSHong Zhang if (jmin < jmax){ 3060ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 3061ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */ 3062ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*4; 3063ab27746eSHong Zhang u = ba + j*4; 3064ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1]; 3065ab27746eSHong Zhang rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1]; 3066ab27746eSHong Zhang rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3]; 3067ab27746eSHong Zhang rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3]; 306849b5e25fSSatish Balay } 3069ab27746eSHong Zhang 3070ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 3071ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 3072ab27746eSHong Zhang j = bj[jmin]; 3073ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 307449b5e25fSSatish Balay } 3075ab27746eSHong Zhang i = nexti; 307649b5e25fSSatish Balay } 3077ab27746eSHong Zhang 3078ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 3079ab27746eSHong Zhang 308049b5e25fSSatish Balay /* invert diagonal block */ 3081ab27746eSHong Zhang diag = ba+k*4; 3082ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr); 3083ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr); 3084ab27746eSHong Zhang 3085ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 3086ab27746eSHong Zhang if (jmin < jmax) { 3087ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 3088ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 3089ab27746eSHong Zhang u = ba + j*4; 3090ab27746eSHong Zhang rtmp_ptr = rtmp + vj*4; 3091ab27746eSHong Zhang for (k1=0; k1<4; k1++){ 3092ab27746eSHong Zhang *u++ = *rtmp_ptr; 3093ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 3094ab27746eSHong Zhang } 3095ab27746eSHong Zhang } 3096ab27746eSHong Zhang 3097ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 3098ab27746eSHong Zhang il[k] = jmin; 3099ab27746eSHong Zhang i = bj[jmin]; 3100ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 3101ab27746eSHong Zhang } 310249b5e25fSSatish Balay } 310349b5e25fSSatish Balay 310449b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 3105ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 3106ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 3107ab27746eSHong Zhang 3108ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 310949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 31105c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 3111*b0a32e0cSBarry Smith PetscLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 311249b5e25fSSatish Balay PetscFunctionReturn(0); 311349b5e25fSSatish Balay } 311449b5e25fSSatish Balay 311549b5e25fSSatish Balay /* 31165c0bcdfcSHong Zhang Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP. 311791602c66SHong Zhang Version for blocks are 1 by 1. 311849b5e25fSSatish Balay */ 311949b5e25fSSatish Balay #undef __FUNC__ 31206f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_1" 31216f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_1(Mat A,Mat *B) 312249b5e25fSSatish Balay { 312349b5e25fSSatish Balay Mat C = *B; 312449b5e25fSSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 312549b5e25fSSatish Balay IS ip = b->row; 3126351d0355SHong Zhang int *rip,ierr,i,j,mbs = a->mbs,*bi = b->i,*bj = b->j; 3127cb718733SHong Zhang int *ai,*aj,*r; 3128ab27746eSHong Zhang int k,jmin,jmax,*jl,*il,vj,nexti,ili; 3129066653e3SSatish Balay MatScalar *rtmp; 31302d007305SHong Zhang MatScalar *ba = b->a,*aa,ak; 313149b5e25fSSatish Balay MatScalar dk,uikdi; 313249b5e25fSSatish Balay 313349b5e25fSSatish Balay PetscFunctionBegin; 313449b5e25fSSatish Balay ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 3135cb718733SHong Zhang if (!a->permute){ 31362d007305SHong Zhang ai = a->i; aj = a->j; aa = a->a; 31372d007305SHong Zhang } else { 31382d007305SHong Zhang ai = a->inew; aj = a->jnew; 3139*b0a32e0cSBarry Smith aa = (MatScalar*)PetscMalloc(ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 3140cb718733SHong Zhang ierr = PetscMemcpy(aa,a->a,ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 3141*b0a32e0cSBarry Smith r = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKERRQ(ierr); 31422d007305SHong Zhang ierr= PetscMemcpy(r,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 31432d007305SHong Zhang 31442d007305SHong Zhang jmin = ai[0]; jmax = ai[mbs]; 31452d007305SHong Zhang for (j=jmin; j<jmax; j++){ 3146cb718733SHong Zhang while (r[j] != j){ 31472d007305SHong Zhang k = r[j]; r[j] = r[k]; r[k] = k; 31482d007305SHong Zhang ak = aa[k]; aa[k] = aa[j]; aa[j] = ak; 31492d007305SHong Zhang } 31502d007305SHong Zhang } 31512d007305SHong Zhang ierr = PetscFree(r);CHKERRA(ierr); 31522d007305SHong Zhang } 31532d007305SHong Zhang 315491602c66SHong Zhang /* initialization */ 315549b5e25fSSatish Balay /* il and jl record the first nonzero element in each row of the accessing 315649b5e25fSSatish Balay window U(0:k, k:mbs-1). 315749b5e25fSSatish Balay jl: list of rows to be added to uneliminated rows 315849b5e25fSSatish Balay i>= k: jl(i) is the first row to be added to row i 315949b5e25fSSatish Balay i< k: jl(i) is the row following row i in some list of rows 316049b5e25fSSatish Balay jl(i) = mbs indicates the end of a list 316149b5e25fSSatish Balay il(i): points to the first nonzero element in columns k,...,mbs-1 of 316249b5e25fSSatish Balay row i of U */ 3163*b0a32e0cSBarry Smith ierr = PetscMalloc(mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 3164*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 316528de702eSHong Zhang jl = il + mbs; 316649b5e25fSSatish Balay for (i=0; i<mbs; i++) { 316749b5e25fSSatish Balay rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0; 316849b5e25fSSatish Balay } 316949b5e25fSSatish Balay 317091602c66SHong Zhang /* for each row k */ 317149b5e25fSSatish Balay for (k = 0; k<mbs; k++){ 317249b5e25fSSatish Balay 317391602c66SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 317449b5e25fSSatish Balay jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 317549b5e25fSSatish Balay if (jmin < jmax) { 317649b5e25fSSatish Balay for (j = jmin; j < jmax; j++){ 3177351d0355SHong Zhang vj = rip[aj[j]]; 3178ab27746eSHong Zhang rtmp[vj] = aa[j]; 317949b5e25fSSatish Balay } 318049b5e25fSSatish Balay } 318149b5e25fSSatish Balay 318291602c66SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 318349b5e25fSSatish Balay dk = rtmp[k]; 318449b5e25fSSatish Balay i = jl[k]; /* first row to be added to k_th row */ 318549b5e25fSSatish Balay 318649b5e25fSSatish Balay while (i < mbs){ 318749b5e25fSSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 318849b5e25fSSatish Balay 318991602c66SHong Zhang /* compute multiplier, update D(k) and U(i,k) */ 319049b5e25fSSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 319149b5e25fSSatish Balay uikdi = - ba[ili]*ba[i]; 319249b5e25fSSatish Balay dk += uikdi*ba[ili]; 3193658e7b3eSHong Zhang ba[ili] = uikdi; /* -U(i,k) */ 319449b5e25fSSatish Balay 319591602c66SHong Zhang /* add multiple of row i to k-th row ... */ 319649b5e25fSSatish Balay jmin = ili + 1; jmax = bi[i+1]; 319749b5e25fSSatish Balay if (jmin < jmax){ 319849b5e25fSSatish Balay for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 319991602c66SHong Zhang /* ... add i to row list for next nonzero entry */ 320049b5e25fSSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 320149b5e25fSSatish Balay j = bj[jmin]; 320249b5e25fSSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 320349b5e25fSSatish Balay } 3204ab27746eSHong Zhang i = nexti; 320549b5e25fSSatish Balay } 320649b5e25fSSatish Balay 320791602c66SHong Zhang /* check for zero pivot and save diagoanl element */ 320849b5e25fSSatish Balay if (dk == 0.0){ 320929bbc08cSBarry Smith SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot"); 3210671cb588SHong Zhang /* 32118be8c0c7SHong Zhang }else if (PetscRealPart(dk) < 0){ 32128be8c0c7SHong Zhang ierr = PetscPrintf(PETSC_COMM_SELF,"Negative pivot: d[%d] = %g\n",k,dk); 3213671cb588SHong Zhang */ 321449b5e25fSSatish Balay } 321549b5e25fSSatish Balay 321691602c66SHong Zhang /* save nonzero entries in k-th row of U ... */ 3217f3dd7b2dSKris Buschelman ba[k] = 1.0/dk; 321849b5e25fSSatish Balay jmin = bi[k]; jmax = bi[k+1]; 321949b5e25fSSatish Balay if (jmin < jmax) { 322049b5e25fSSatish Balay for (j=jmin; j<jmax; j++){ 3221cc0c071aSHong Zhang vj = bj[j]; ba[j] = rtmp[vj]; rtmp[vj] = 0.0; 322249b5e25fSSatish Balay } 322391602c66SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 322449b5e25fSSatish Balay il[k] = jmin; 322549b5e25fSSatish Balay i = bj[jmin]; 322649b5e25fSSatish Balay jl[k] = jl[i]; jl[i] = k; 322749b5e25fSSatish Balay } 322849b5e25fSSatish Balay } 322949b5e25fSSatish Balay 323049b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 323149b5e25fSSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 3232cb718733SHong Zhang if (a->permute){ 3233cb718733SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 3234cb718733SHong Zhang } 323549b5e25fSSatish Balay 323649b5e25fSSatish Balay ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 32378be8c0c7SHong Zhang C->factor = FACTOR_CHOLESKY; 323849b5e25fSSatish Balay C->assembled = PETSC_TRUE; 32395c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 3240*b0a32e0cSBarry Smith PetscLogFlops(b->mbs); 324149b5e25fSSatish Balay PetscFunctionReturn(0); 324249b5e25fSSatish Balay } 324349b5e25fSSatish Balay 3244671cb588SHong Zhang /* 3245671cb588SHong Zhang Version for when blocks are 1 by 1 Using natural ordering 3246671cb588SHong Zhang */ 3247671cb588SHong Zhang #undef __FUNC__ 3248671cb588SHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering" 3249671cb588SHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering(Mat A,Mat *B) 3250671cb588SHong Zhang { 3251671cb588SHong Zhang Mat C = *B; 3252671cb588SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 3253671cb588SHong Zhang int ierr,i,j,mbs = a->mbs,*bi = b->i,*bj = b->j; 32541ad1de41SHong Zhang int *ai,*aj; 3255671cb588SHong Zhang int k,jmin,jmax,*jl,*il,vj,nexti,ili; 32561ad1de41SHong Zhang MatScalar *rtmp,*ba = b->a,*aa,dk,uikdi; 3257671cb588SHong Zhang 3258671cb588SHong Zhang PetscFunctionBegin; 3259671cb588SHong Zhang 3260671cb588SHong Zhang /* initialization */ 3261671cb588SHong Zhang /* il and jl record the first nonzero element in each row of the accessing 3262671cb588SHong Zhang window U(0:k, k:mbs-1). 3263671cb588SHong Zhang jl: list of rows to be added to uneliminated rows 3264671cb588SHong Zhang i>= k: jl(i) is the first row to be added to row i 3265671cb588SHong Zhang i< k: jl(i) is the row following row i in some list of rows 3266671cb588SHong Zhang jl(i) = mbs indicates the end of a list 3267671cb588SHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 3268671cb588SHong Zhang row i of U */ 3269*b0a32e0cSBarry Smith ierr = PetscMalloc(mbs*sizeof(MatScalar),&( rtmp ));CHKERRQ(ierr); 3270*b0a32e0cSBarry Smith ierr = PetscMalloc(2*mbs*sizeof(int),&( il ));CHKERRQ(ierr); 327128de702eSHong Zhang jl = il + mbs; 3272671cb588SHong Zhang for (i=0; i<mbs; i++) { 3273671cb588SHong Zhang rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0; 3274671cb588SHong Zhang } 3275671cb588SHong Zhang 3276671cb588SHong Zhang ai = a->i; aj = a->j; aa = a->a; 3277671cb588SHong Zhang 3278671cb588SHong Zhang /* for each row k */ 3279671cb588SHong Zhang for (k = 0; k<mbs; k++){ 3280671cb588SHong Zhang 3281671cb588SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 3282671cb588SHong Zhang jmin = ai[k]; jmax = ai[k+1]; 3283671cb588SHong Zhang if (jmin < jmax) { 3284671cb588SHong Zhang for (j = jmin; j < jmax; j++){ 3285671cb588SHong Zhang vj = aj[j]; 3286671cb588SHong Zhang rtmp[vj] = aa[j]; 3287671cb588SHong Zhang } 3288671cb588SHong Zhang } 3289671cb588SHong Zhang 3290671cb588SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 3291671cb588SHong Zhang dk = rtmp[k]; 3292671cb588SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 3293671cb588SHong Zhang 3294671cb588SHong Zhang while (i < mbs){ 3295671cb588SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 3296671cb588SHong Zhang 3297671cb588SHong Zhang /* compute multiplier, update D(k) and U(i,k) */ 3298671cb588SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 3299671cb588SHong Zhang uikdi = - ba[ili]*ba[i]; 3300671cb588SHong Zhang dk += uikdi*ba[ili]; 3301671cb588SHong Zhang ba[ili] = uikdi; /* -U(i,k) */ 3302671cb588SHong Zhang 3303671cb588SHong Zhang /* add multiple of row i to k-th row ... */ 3304671cb588SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 3305671cb588SHong Zhang if (jmin < jmax){ 3306671cb588SHong Zhang for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 3307671cb588SHong Zhang /* ... add i to row list for next nonzero entry */ 3308671cb588SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 3309671cb588SHong Zhang j = bj[jmin]; 3310671cb588SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 3311671cb588SHong Zhang } 3312671cb588SHong Zhang i = nexti; 3313671cb588SHong Zhang } 3314671cb588SHong Zhang 3315671cb588SHong Zhang /* check for zero pivot and save diagoanl element */ 3316671cb588SHong Zhang if (dk == 0.0){ 3317671cb588SHong Zhang SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot"); 3318671cb588SHong Zhang /* 3319671cb588SHong Zhang }else if (PetscRealPart(dk) < 0){ 3320671cb588SHong Zhang ierr = PetscPrintf(PETSC_COMM_SELF,"Negative pivot: d[%d] = %g\n",k,dk); 3321671cb588SHong Zhang */ 3322671cb588SHong Zhang } 3323671cb588SHong Zhang 3324671cb588SHong Zhang /* save nonzero entries in k-th row of U ... */ 3325671cb588SHong Zhang ba[k] = 1.0/dk; 3326671cb588SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 3327671cb588SHong Zhang if (jmin < jmax) { 3328671cb588SHong Zhang for (j=jmin; j<jmax; j++){ 3329671cb588SHong Zhang vj = bj[j]; ba[j] = rtmp[vj]; rtmp[vj] = 0.0; 3330671cb588SHong Zhang } 3331671cb588SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 3332671cb588SHong Zhang il[k] = jmin; 3333671cb588SHong Zhang i = bj[jmin]; 3334671cb588SHong Zhang jl[k] = jl[i]; jl[i] = k; 3335671cb588SHong Zhang } 3336671cb588SHong Zhang } 3337671cb588SHong Zhang 3338671cb588SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 3339671cb588SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 3340671cb588SHong Zhang 3341671cb588SHong Zhang C->factor = FACTOR_CHOLESKY; 3342671cb588SHong Zhang C->assembled = PETSC_TRUE; 3343671cb588SHong Zhang C->preallocated = PETSC_TRUE; 3344*b0a32e0cSBarry Smith PetscLogFlops(b->mbs); 3345671cb588SHong Zhang PetscFunctionReturn(0); 3346671cb588SHong Zhang } 3347671cb588SHong Zhang 334849b5e25fSSatish Balay #undef __FUNC__ 33496f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactor_SeqSBAIJ" 3350b8b23757SHong Zhang int MatCholeskyFactor_SeqSBAIJ(Mat A,IS perm,PetscReal f) 335149b5e25fSSatish Balay { 33524445ddedSHong Zhang int ierr; 335349b5e25fSSatish Balay Mat C; 335449b5e25fSSatish Balay 335549b5e25fSSatish Balay PetscFunctionBegin; 3356b8b23757SHong Zhang ierr = MatCholeskyFactorSymbolic(A,perm,f,&C);CHKERRQ(ierr); 3357a4ada70bSHong Zhang ierr = MatCholeskyFactorNumeric(A,&C);CHKERRQ(ierr); 33584445ddedSHong Zhang ierr = MatHeaderCopy(A,C);CHKERRQ(ierr); 335949b5e25fSSatish Balay PetscFunctionReturn(0); 336049b5e25fSSatish Balay } 336149b5e25fSSatish Balay 336249b5e25fSSatish Balay 3363