149b5e25fSSatish Balay /* Using Modified Sparse Row (MSR) storage. 249b5e25fSSatish Balay See page 85, "Iterative Methods ..." by Saad. */ 349b5e25fSSatish Balay 4*71149678SHong Zhang /*$Id: sbaijfact.c,v 1.45 2000/11/03 18:39:54 hzhang Exp hzhang $*/ 549b5e25fSSatish Balay /* 6effa298cSHong Zhang Symbolic U^T*D*U factorization for SBAIJ format. Modified from SSF of YSMP. 749b5e25fSSatish Balay */ 849b5e25fSSatish Balay #include "sbaij.h" 949b5e25fSSatish Balay #include "src/mat/impls/baij/seq/baij.h" 1049b5e25fSSatish Balay #include "src/vec/vecimpl.h" 1149b5e25fSSatish Balay #include "src/inline/ilu.h" 1249a6740bSHong Zhang #include "include/petscis.h" 1349b5e25fSSatish Balay 1449b5e25fSSatish Balay #undef __FUNC__ 156f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorSymbolic_SeqSBAIJ" 1677f73120SHong Zhang int MatCholeskyFactorSymbolic_SeqSBAIJ(Mat A,IS perm,PetscReal f,Mat *B) 1749b5e25fSSatish Balay { 1849b5e25fSSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b; 19cb718733SHong Zhang int *rip,ierr,i,mbs = a->mbs,*ai,*aj; 20066653e3SSatish Balay int *jutmp,bs = a->bs,bs2=a->bs2; 21066653e3SSatish Balay int m,nzi,realloc = 0; 22066653e3SSatish Balay int *jl,*q,jumin,jmin,jmax,juptr,nzk,qm,*iu,*ju,k,j,vj,umax,maxadd; 23cb718733SHong Zhang /* PetscTruth ident; */ 2449b5e25fSSatish Balay 2549b5e25fSSatish Balay PetscFunctionBegin; 2677f73120SHong Zhang PetscValidHeaderSpecific(perm,IS_COOKIE); 2729bbc08cSBarry Smith if (A->M != A->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 282d007305SHong Zhang 29cb718733SHong Zhang /* check whether perm is the identity mapping */ 30cb718733SHong Zhang /* 31cb718733SHong Zhang ierr = ISView(perm, VIEWER_STDOUT_SELF);CHKERRA(ierr); 32cb718733SHong Zhang ierr = ISIdentity(perm,&ident);CHKERRQ(ierr); 33cb718733SHong Zhang printf("ident = %d\n", ident); 34cb718733SHong Zhang */ 35cb718733SHong Zhang ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 36cb718733SHong Zhang for (i=0; i<mbs; i++){ 37cb718733SHong Zhang if (rip[i] != i){ 38cb718733SHong Zhang a->permute = PETSC_TRUE; 39cb718733SHong Zhang /* printf("non-trivial perm\n"); */ 40cb718733SHong Zhang break; 4149a6740bSHong Zhang } 42cb718733SHong Zhang } 432d007305SHong Zhang 44cb718733SHong Zhang if (!a->permute){ /* without permutation */ 452d007305SHong Zhang ai = a->i; aj = a->j; 462d007305SHong Zhang } else { /* non-trivial permutation */ 472d007305SHong Zhang ierr = MatReorderingSeqSBAIJ(A, perm);CHKERRA(ierr); 482d007305SHong Zhang ai = a->inew; aj = a->jnew; 492d007305SHong Zhang } 502d007305SHong Zhang 5149b5e25fSSatish Balay /* initialization */ 5249b5e25fSSatish Balay /* Don't know how many column pointers are needed so estimate. 5349b5e25fSSatish Balay Use Modified Sparse Row storage for u and ju, see Sasd pp.85 */ 549f9cb213SHong Zhang iu = (int*)PetscMalloc((mbs+1)*sizeof(int));CHKPTRQ(iu); 5549b5e25fSSatish Balay umax = (int)(f*ai[mbs] + 1); umax += mbs + 1; 569f9cb213SHong Zhang ju = (int*)PetscMalloc(umax*sizeof(int));CHKPTRQ(ju); 5749b5e25fSSatish Balay iu[0] = mbs+1; 5849b5e25fSSatish Balay juptr = mbs; 5949b5e25fSSatish Balay jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 6049b5e25fSSatish Balay q = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(q); 6149b5e25fSSatish Balay for (i=0; i<mbs; i++){ 6249b5e25fSSatish Balay jl[i] = mbs; q[i] = 0; 6349b5e25fSSatish Balay } 6449b5e25fSSatish Balay 6549b5e25fSSatish Balay /* for each row k */ 6649b5e25fSSatish Balay for (k=0; k<mbs; k++){ 6749b5e25fSSatish Balay nzk = 0; /* num. of nz blocks in k-th block row with diagonal block excluded */ 6849b5e25fSSatish Balay q[k] = mbs; 6949b5e25fSSatish Balay /* initialize nonzero structure of k-th row to row rip[k] of A */ 7049b5e25fSSatish Balay jmin = ai[rip[k]]; 7149b5e25fSSatish Balay jmax = ai[rip[k]+1]; 7249b5e25fSSatish Balay for (j=jmin; j<jmax; j++){ 73cb718733SHong Zhang vj = rip[aj[j]]; /* col. value */ 7449b5e25fSSatish Balay if(vj > k){ 7549b5e25fSSatish Balay qm = k; 7649b5e25fSSatish Balay do { 7749b5e25fSSatish Balay m = qm; qm = q[m]; 7849b5e25fSSatish Balay } while(qm < vj); 7949b5e25fSSatish Balay if (qm == vj) { 8049b5e25fSSatish Balay printf(" error: duplicate entry in A\n"); break; 8149b5e25fSSatish Balay } 8249b5e25fSSatish Balay nzk++; 8349b5e25fSSatish Balay q[m] = vj; 8449b5e25fSSatish Balay q[vj] = qm; 8549b5e25fSSatish Balay } /* if(vj > k) */ 8649b5e25fSSatish Balay } /* for (j=jmin; j<jmax; j++) */ 8749b5e25fSSatish Balay 8849b5e25fSSatish Balay /* modify nonzero structure of k-th row by computing fill-in 8949b5e25fSSatish Balay for each row i to be merged in */ 9049b5e25fSSatish Balay i = k; 9149b5e25fSSatish Balay i = jl[i]; /* next pivot row (== mbs for symbolic factorization) */ 9249b5e25fSSatish Balay /* printf(" next pivot row i=%d\n",i); */ 9349b5e25fSSatish Balay while (i < mbs){ 9449b5e25fSSatish Balay /* merge row i into k-th row */ 9549b5e25fSSatish Balay nzi = iu[i+1] - (iu[i]+1); 9649b5e25fSSatish Balay jmin = iu[i] + 1; jmax = iu[i] + nzi; 9749b5e25fSSatish Balay qm = k; 9849b5e25fSSatish Balay for (j=jmin; j<jmax+1; j++){ 9949b5e25fSSatish Balay vj = ju[j]; 10049b5e25fSSatish Balay do { 10149b5e25fSSatish Balay m = qm; qm = q[m]; 10249b5e25fSSatish Balay } while (qm < vj); 10349b5e25fSSatish Balay if (qm != vj){ 10449b5e25fSSatish Balay nzk++; q[m] = vj; q[vj] = qm; qm = vj; 10549b5e25fSSatish Balay } 10649b5e25fSSatish Balay } 10749b5e25fSSatish Balay i = jl[i]; /* next pivot row */ 10849b5e25fSSatish Balay } 10949b5e25fSSatish Balay 11049b5e25fSSatish Balay /* add k to row list for first nonzero element in k-th row */ 11149b5e25fSSatish Balay if (nzk > 0){ 11249b5e25fSSatish Balay i = q[k]; /* col value of first nonzero element in U(k, k+1:mbs-1) */ 11349b5e25fSSatish Balay jl[k] = jl[i]; jl[i] = k; 11449b5e25fSSatish Balay } 11549b5e25fSSatish Balay iu[k+1] = iu[k] + nzk; /* printf(" iu[%d]=%d, umax=%d\n", k+1, iu[k+1],umax);*/ 11649b5e25fSSatish Balay 11749b5e25fSSatish Balay /* allocate more space to ju if needed */ 11849b5e25fSSatish Balay if (iu[k+1] > umax) { printf("allocate more space, iu[%d]=%d > umax=%d\n",k+1, iu[k+1],umax); 11949b5e25fSSatish Balay /* estimate how much additional space we will need */ 12049b5e25fSSatish Balay /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */ 12149b5e25fSSatish Balay /* just double the memory each time */ 12249b5e25fSSatish Balay maxadd = umax; 12349b5e25fSSatish Balay if (maxadd < nzk) maxadd = (mbs-k)*(nzk+1)/2; 12449b5e25fSSatish Balay umax += maxadd; 12549b5e25fSSatish Balay 1269f9cb213SHong Zhang /* allocate a longer ju */ 12749b5e25fSSatish Balay jutmp = (int*)PetscMalloc(umax*sizeof(int));CHKPTRQ(jutmp); 12849b5e25fSSatish Balay ierr = PetscMemcpy(jutmp,ju,iu[k]*sizeof(int));CHKERRQ(ierr); 12949b5e25fSSatish Balay ierr = PetscFree(ju);CHKERRQ(ierr); 1309f9cb213SHong Zhang ju = jutmp; 13149b5e25fSSatish Balay realloc++; /* count how many times we realloc */ 13249b5e25fSSatish Balay } 13349b5e25fSSatish Balay 13449b5e25fSSatish Balay /* save nonzero structure of k-th row in ju */ 13549b5e25fSSatish Balay i=k; 13649b5e25fSSatish Balay jumin = juptr + 1; juptr += nzk; 13749b5e25fSSatish Balay for (j=jumin; j<juptr+1; j++){ 13849b5e25fSSatish Balay i=q[i]; 13949b5e25fSSatish Balay ju[j]=i; 14049b5e25fSSatish Balay } 14177f73120SHong Zhang } 14249b5e25fSSatish Balay 14349b5e25fSSatish Balay if (ai[mbs] != 0) { 14449b5e25fSSatish Balay PetscReal af = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]); 1456f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Reallocs %d Fill ratio:given %g needed %g\n",realloc,f,af); 1466f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Run with -pc_lu_fill %g or use \n",af); 1476f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:PCLUSetFill(pc,%g);\n",af); 1486f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:for best performance.\n"); 14949b5e25fSSatish Balay } else { 1506f9a564bSHong Zhang PLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Empty matrix.\n"); 15149b5e25fSSatish Balay } 15249b5e25fSSatish Balay 15377f73120SHong Zhang ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 15449b5e25fSSatish Balay ierr = PetscFree(q);CHKERRQ(ierr); 15549b5e25fSSatish Balay ierr = PetscFree(jl);CHKERRQ(ierr); 15649b5e25fSSatish Balay 15749b5e25fSSatish Balay /* put together the new matrix */ 15849b5e25fSSatish Balay ierr = MatCreateSeqSBAIJ(A->comm,bs,bs*mbs,bs*mbs,0,PETSC_NULL,B);CHKERRQ(ierr); 1593d840bdaSHong Zhang /* PLogObjectParent(*B,iperm); */ 16049b5e25fSSatish Balay b = (Mat_SeqSBAIJ*)(*B)->data; 16149b5e25fSSatish Balay ierr = PetscFree(b->imax);CHKERRQ(ierr); 16249b5e25fSSatish Balay b->singlemalloc = PETSC_FALSE; 16349b5e25fSSatish Balay /* the next line frees the default space generated by the Create() */ 16449b5e25fSSatish Balay ierr = PetscFree(b->a);CHKERRQ(ierr); 16549b5e25fSSatish Balay ierr = PetscFree(b->ilen);CHKERRQ(ierr); 16649b5e25fSSatish Balay b->a = (MatScalar*)PetscMalloc((iu[mbs]+1)*sizeof(MatScalar)*bs2);CHKPTRQ(b->a); 16749b5e25fSSatish Balay b->j = ju; 16849b5e25fSSatish Balay b->i = iu; 16949b5e25fSSatish Balay b->diag = 0; 17049b5e25fSSatish Balay b->ilen = 0; 17149b5e25fSSatish Balay b->imax = 0; 17277f73120SHong Zhang b->row = perm; 17377f73120SHong Zhang ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 174cb718733SHong Zhang b->icol = perm; 175cb718733SHong Zhang ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 17649b5e25fSSatish Balay b->solve_work = (Scalar*)PetscMalloc((bs*mbs+bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 17749b5e25fSSatish Balay /* In b structure: Free imax, ilen, old a, old j. 17849b5e25fSSatish Balay Allocate idnew, solve_work, new a, new j */ 17949b5e25fSSatish Balay PLogObjectMemory(*B,(iu[mbs]-mbs)*(sizeof(int)+sizeof(MatScalar))); 18049b5e25fSSatish Balay b->s_maxnz = b->s_nz = iu[mbs]; 18149b5e25fSSatish Balay 1825c0bcdfcSHong Zhang (*B)->factor = FACTOR_CHOLESKY; 18349b5e25fSSatish Balay (*B)->info.factor_mallocs = realloc; 18449b5e25fSSatish Balay (*B)->info.fill_ratio_given = f; 18549b5e25fSSatish Balay if (ai[mbs] != 0) { 18649b5e25fSSatish Balay (*B)->info.fill_ratio_needed = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]); 18749b5e25fSSatish Balay } else { 18849b5e25fSSatish Balay (*B)->info.fill_ratio_needed = 0.0; 18949b5e25fSSatish Balay } 1900aa0ce06SHong Zhang 19149b5e25fSSatish Balay PetscFunctionReturn(0); 19249b5e25fSSatish Balay } 19349b5e25fSSatish Balay 19449b5e25fSSatish Balay #undef __FUNC__ 1956f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_N" 1966f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_N(Mat A,Mat *B) 19749b5e25fSSatish Balay { 19849b5e25fSSatish Balay Mat C = *B; 1994c16a6a6SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2004c16a6a6SHong Zhang IS perm = b->row; 2014c16a6a6SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2024c16a6a6SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2034c16a6a6SHong Zhang int bs=a->bs,bs2 = a->bs2; 2044c16a6a6SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2054c16a6a6SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 2064c16a6a6SHong Zhang MatScalar *W,*work; 2074c16a6a6SHong Zhang int *pivots; 2084c16a6a6SHong Zhang 2094c16a6a6SHong Zhang PetscFunctionBegin; 2109706f043SHong Zhang 2114c16a6a6SHong Zhang /* initialization */ 2124c16a6a6SHong Zhang rtmp = (MatScalar*)PetscMalloc(bs2*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2134c16a6a6SHong Zhang ierr = PetscMemzero(rtmp,bs2*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2144c16a6a6SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2154c16a6a6SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2164c16a6a6SHong Zhang for (i=0; i<mbs; i++) { 2174c16a6a6SHong Zhang jl[i] = mbs; il[0] = 0; 2184c16a6a6SHong Zhang } 2194c16a6a6SHong Zhang dk = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(dk); 2204c16a6a6SHong Zhang uik = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(uik); 2214c16a6a6SHong Zhang W = (MatScalar*)PetscMalloc(bs2*sizeof(MatScalar));CHKPTRQ(W); 2224c16a6a6SHong Zhang work = (MatScalar*)PetscMalloc(bs*sizeof(MatScalar));CHKPTRQ(work); 2234c16a6a6SHong Zhang pivots= (int*)PetscMalloc(bs*sizeof(int));CHKPTRQ(pivots); 2244c16a6a6SHong Zhang 2254c16a6a6SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2264c16a6a6SHong Zhang 2274c16a6a6SHong Zhang /* check permutation */ 2284c16a6a6SHong Zhang if (!a->permute){ 2294c16a6a6SHong Zhang ai = a->i; aj = a->j; aa = a->a; 2304c16a6a6SHong Zhang } else { 2314c16a6a6SHong Zhang ai = a->inew; aj = a->jnew; 2324c16a6a6SHong Zhang aa = (MatScalar*)PetscMalloc(bs2*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2334c16a6a6SHong Zhang ierr = PetscMemcpy(aa,a->a,bs2*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2344c16a6a6SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2354c16a6a6SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2364c16a6a6SHong Zhang 2374c16a6a6SHong Zhang for (i=0; i<mbs; i++){ 2384c16a6a6SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2394c16a6a6SHong Zhang for (j=jmin; j<jmax; j++){ 2404c16a6a6SHong Zhang while (a2anew[j] != j){ 2414c16a6a6SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2424c16a6a6SHong Zhang for (k1=0; k1<bs2; k1++){ 2434c16a6a6SHong Zhang dk[k1] = aa[k*bs2+k1]; 2444c16a6a6SHong Zhang aa[k*bs2+k1] = aa[j*bs2+k1]; 2454c16a6a6SHong Zhang aa[j*bs2+k1] = dk[k1]; 2464c16a6a6SHong Zhang } 2474c16a6a6SHong Zhang } 2484c16a6a6SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2494c16a6a6SHong Zhang if (i > aj[j]){ 2504c16a6a6SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2514c16a6a6SHong Zhang ap = aa + j*bs2; /* ptr to the beginning of j-th block of aa */ 2524c16a6a6SHong Zhang for (k=0; k<bs2; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2534c16a6a6SHong Zhang for (k=0; k<bs; k++){ /* j-th block of aa <- dk^T */ 2544c16a6a6SHong Zhang for (k1=0; k1<bs; k1++) *ap++ = dk[k + bs*k1]; 2554c16a6a6SHong Zhang } 2564c16a6a6SHong Zhang } 2574c16a6a6SHong Zhang } 2584c16a6a6SHong Zhang } 2594c16a6a6SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2604c16a6a6SHong Zhang } 2614c16a6a6SHong Zhang 2624c16a6a6SHong Zhang /* for each row k */ 2634c16a6a6SHong Zhang for (k = 0; k<mbs; k++){ 2644c16a6a6SHong Zhang 2654c16a6a6SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2664c16a6a6SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 2674c16a6a6SHong Zhang if (jmin < jmax) { 2684c16a6a6SHong Zhang ap = aa + jmin*bs2; 2694c16a6a6SHong Zhang for (j = jmin; j < jmax; j++){ 2704c16a6a6SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2714c16a6a6SHong Zhang rtmp_ptr = rtmp + vj*bs2; 2724c16a6a6SHong Zhang for (i=0; i<bs2; i++) *rtmp_ptr++ = *ap++; 2734c16a6a6SHong Zhang } 2744c16a6a6SHong Zhang } 2754c16a6a6SHong Zhang 2764c16a6a6SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2774c16a6a6SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2784c16a6a6SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2794c16a6a6SHong Zhang 2804c16a6a6SHong Zhang while (i < mbs){ 2814c16a6a6SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2824c16a6a6SHong Zhang 2834c16a6a6SHong Zhang /* compute multiplier */ 2844c16a6a6SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2854c16a6a6SHong Zhang 2864c16a6a6SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2874c16a6a6SHong Zhang diag = ba + i*bs2; 2884c16a6a6SHong Zhang u = ba + ili*bs2; 2894c16a6a6SHong Zhang ierr = PetscMemzero(uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2904c16a6a6SHong Zhang Kernel_A_gets_A_minus_B_times_C(bs,uik,diag,u); 2914c16a6a6SHong Zhang 2924c16a6a6SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2934c16a6a6SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,dk,uik,u); 2944c16a6a6SHong Zhang 2954c16a6a6SHong Zhang /* update -U(i,k) */ 2964c16a6a6SHong Zhang ierr = PetscMemcpy(ba+ili*bs2,uik,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2974c16a6a6SHong Zhang 2984c16a6a6SHong Zhang /* add multiple of row i to k-th row ... */ 2994c16a6a6SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 3004c16a6a6SHong Zhang if (jmin < jmax){ 3014c16a6a6SHong Zhang for (j=jmin; j<jmax; j++) { 3024c16a6a6SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 3034c16a6a6SHong Zhang rtmp_ptr = rtmp + bj[j]*bs2; 3044c16a6a6SHong Zhang u = ba + j*bs2; 3054c16a6a6SHong Zhang Kernel_A_gets_A_plus_Btranspose_times_C(bs,rtmp_ptr,uik,u); 3064c16a6a6SHong Zhang } 3074c16a6a6SHong Zhang 3084c16a6a6SHong Zhang /* ... add i to row list for next nonzero entry */ 3094c16a6a6SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 3104c16a6a6SHong Zhang j = bj[jmin]; 3114c16a6a6SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 3124c16a6a6SHong Zhang } 3134c16a6a6SHong Zhang i = nexti; 3144c16a6a6SHong Zhang } 3154c16a6a6SHong Zhang 3164c16a6a6SHong Zhang /* save nonzero entries in k-th row of U ... */ 3174c16a6a6SHong Zhang 3184c16a6a6SHong Zhang /* invert diagonal block */ 3194c16a6a6SHong Zhang diag = ba+k*bs2; 3204c16a6a6SHong Zhang ierr = PetscMemcpy(diag,dk,bs2*sizeof(MatScalar));CHKERRQ(ierr); 3214c16a6a6SHong Zhang Kernel_A_gets_inverse_A(bs,diag,pivots,work); 3224c16a6a6SHong Zhang 3234c16a6a6SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 3244c16a6a6SHong Zhang if (jmin < jmax) { 3254c16a6a6SHong Zhang for (j=jmin; j<jmax; j++){ 3264c16a6a6SHong Zhang vj = bj[j]; /* block col. index of U */ 3274c16a6a6SHong Zhang u = ba + j*bs2; 3284c16a6a6SHong Zhang rtmp_ptr = rtmp + vj*bs2; 3294c16a6a6SHong Zhang for (k1=0; k1<bs2; k1++){ 3304c16a6a6SHong Zhang *u++ = *rtmp_ptr; 3314c16a6a6SHong Zhang *rtmp_ptr++ = 0.0; 3324c16a6a6SHong Zhang } 3334c16a6a6SHong Zhang } 3344c16a6a6SHong Zhang 3354c16a6a6SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 3364c16a6a6SHong Zhang il[k] = jmin; 3374c16a6a6SHong Zhang i = bj[jmin]; 3384c16a6a6SHong Zhang jl[k] = jl[i]; jl[i] = k; 3394c16a6a6SHong Zhang } 3404c16a6a6SHong Zhang } 3414c16a6a6SHong Zhang 3424c16a6a6SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 3434c16a6a6SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 3444c16a6a6SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 3454c16a6a6SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 3464c16a6a6SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 3474c16a6a6SHong Zhang ierr = PetscFree(W);CHKERRQ(ierr); 3484c16a6a6SHong Zhang ierr = PetscFree(work);CHKERRQ(ierr); 3494c16a6a6SHong Zhang ierr = PetscFree(pivots);CHKERRQ(ierr); 3504c16a6a6SHong Zhang if (a->permute){ 3514c16a6a6SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 3524c16a6a6SHong Zhang } 3534c16a6a6SHong Zhang 3544c16a6a6SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 3554c16a6a6SHong Zhang C->factor = FACTOR_CHOLESKY; 3564c16a6a6SHong Zhang C->assembled = PETSC_TRUE; 3574c16a6a6SHong Zhang C->preallocated = PETSC_TRUE; 3584c16a6a6SHong Zhang PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 3594c16a6a6SHong Zhang PetscFunctionReturn(0); 3604c16a6a6SHong Zhang } 361d4edadd4SHong Zhang 3620aa0ce06SHong Zhang /* Version for when blocks are 7 by 7 */ 36349b5e25fSSatish Balay #undef __FUNC__ 3646f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_7" 3656f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_7(Mat A,Mat *B) 36649b5e25fSSatish Balay { 36749b5e25fSSatish Balay Mat C = *B; 3680aa0ce06SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 3690aa0ce06SHong Zhang IS perm = b->row; 3700aa0ce06SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 3710aa0ce06SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 3720aa0ce06SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 3730aa0ce06SHong Zhang MatScalar *u,*d,*w,*wp; 37449b5e25fSSatish Balay 37549b5e25fSSatish Balay PetscFunctionBegin; 3760aa0ce06SHong Zhang /* initialization */ 3770aa0ce06SHong Zhang w = (MatScalar*)PetscMalloc(49*mbs*sizeof(MatScalar));CHKPTRQ(w); 3780aa0ce06SHong Zhang ierr = PetscMemzero(w,49*mbs*sizeof(MatScalar));CHKERRQ(ierr); 3790aa0ce06SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 3800aa0ce06SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 3810aa0ce06SHong Zhang for (i=0; i<mbs; i++) { 3820aa0ce06SHong Zhang jl[i] = mbs; il[0] = 0; 38349b5e25fSSatish Balay } 3840aa0ce06SHong Zhang dk = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(dk); 3850aa0ce06SHong Zhang uik = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(uik); 3860aa0ce06SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 3870aa0ce06SHong Zhang 3880aa0ce06SHong Zhang /* check permutation */ 3890aa0ce06SHong Zhang if (!a->permute){ 3900aa0ce06SHong Zhang ai = a->i; aj = a->j; aa = a->a; 3910aa0ce06SHong Zhang } else { 3920aa0ce06SHong Zhang ai = a->inew; aj = a->jnew; 3930aa0ce06SHong Zhang aa = (MatScalar*)PetscMalloc(49*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 3940aa0ce06SHong Zhang ierr = PetscMemcpy(aa,a->a,49*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 3950aa0ce06SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 3960aa0ce06SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 3970aa0ce06SHong Zhang 3980aa0ce06SHong Zhang for (i=0; i<mbs; i++){ 3990aa0ce06SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 4000aa0ce06SHong Zhang for (j=jmin; j<jmax; j++){ 4010aa0ce06SHong Zhang while (a2anew[j] != j){ 4020aa0ce06SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 4030aa0ce06SHong Zhang for (k1=0; k1<49; k1++){ 4040aa0ce06SHong Zhang dk[k1] = aa[k*49+k1]; 4050aa0ce06SHong Zhang aa[k*49+k1] = aa[j*49+k1]; 4060aa0ce06SHong Zhang aa[j*49+k1] = dk[k1]; 40749b5e25fSSatish Balay } 40849b5e25fSSatish Balay } 4090aa0ce06SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 4100aa0ce06SHong Zhang if (i > aj[j]){ 4110aa0ce06SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 4120aa0ce06SHong Zhang ap = aa + j*49; /* ptr to the beginning of j-th block of aa */ 4130aa0ce06SHong Zhang for (k=0; k<49; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 4140aa0ce06SHong Zhang for (k=0; k<7; k++){ /* j-th block of aa <- dk^T */ 4150aa0ce06SHong Zhang for (k1=0; k1<7; k1++) *ap++ = dk[k + 7*k1]; 41649b5e25fSSatish Balay } 41749b5e25fSSatish Balay } 41849b5e25fSSatish Balay } 4190aa0ce06SHong Zhang } 4200aa0ce06SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 4210aa0ce06SHong Zhang } 4220aa0ce06SHong Zhang 4230aa0ce06SHong Zhang /* for each row k */ 4240aa0ce06SHong Zhang for (k = 0; k<mbs; k++){ 4250aa0ce06SHong Zhang 4260aa0ce06SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 4270aa0ce06SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 4280aa0ce06SHong Zhang if (jmin < jmax) { 4290aa0ce06SHong Zhang ap = aa + jmin*49; 4300aa0ce06SHong Zhang for (j = jmin; j < jmax; j++){ 4310aa0ce06SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 4320aa0ce06SHong Zhang wp = w + vj*49; 4330aa0ce06SHong Zhang for (i=0; i<49; i++) *wp++ = *ap++; 4340aa0ce06SHong Zhang } 4350aa0ce06SHong Zhang } 4360aa0ce06SHong Zhang 4370aa0ce06SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 4380aa0ce06SHong Zhang ierr = PetscMemcpy(dk,w+k*49,49*sizeof(MatScalar));CHKERRQ(ierr); 4390aa0ce06SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 4400aa0ce06SHong Zhang 4410aa0ce06SHong Zhang while (i < mbs){ 4420aa0ce06SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 4430aa0ce06SHong Zhang 4440aa0ce06SHong Zhang /* compute multiplier */ 4450aa0ce06SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 4460aa0ce06SHong Zhang 4470aa0ce06SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 4480aa0ce06SHong Zhang d = ba + i*49; 4490aa0ce06SHong Zhang u = ba + ili*49; 4500aa0ce06SHong Zhang 4510aa0ce06SHong 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]); 4520aa0ce06SHong 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]); 4530aa0ce06SHong 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]); 4540aa0ce06SHong 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]); 4550aa0ce06SHong 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]); 4560aa0ce06SHong 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]); 4570aa0ce06SHong 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]); 4580aa0ce06SHong Zhang 4590aa0ce06SHong 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]); 4600aa0ce06SHong 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]); 4610aa0ce06SHong 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]); 4620aa0ce06SHong 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]); 4630aa0ce06SHong 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]); 4640aa0ce06SHong 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]); 4650aa0ce06SHong 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]); 4660aa0ce06SHong Zhang 4670aa0ce06SHong 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]); 4680aa0ce06SHong 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]); 4690aa0ce06SHong 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]); 4700aa0ce06SHong 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]); 4710aa0ce06SHong 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]); 4720aa0ce06SHong 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]); 4730aa0ce06SHong 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]); 4740aa0ce06SHong Zhang 4750aa0ce06SHong 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]); 4760aa0ce06SHong 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]); 4770aa0ce06SHong 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]); 4780aa0ce06SHong 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]); 4790aa0ce06SHong 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]); 4800aa0ce06SHong 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]); 4810aa0ce06SHong 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]); 4820aa0ce06SHong Zhang 4830aa0ce06SHong 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]); 4840aa0ce06SHong 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]); 4850aa0ce06SHong 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]); 4860aa0ce06SHong 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]); 4870aa0ce06SHong 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]); 4880aa0ce06SHong 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]); 4890aa0ce06SHong 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]); 4900aa0ce06SHong Zhang 4910aa0ce06SHong 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]); 4920aa0ce06SHong 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]); 4930aa0ce06SHong 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]); 4940aa0ce06SHong 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]); 4950aa0ce06SHong 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]); 4960aa0ce06SHong 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]); 4970aa0ce06SHong 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]); 4980aa0ce06SHong Zhang 4990aa0ce06SHong 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]); 5000aa0ce06SHong 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]); 5010aa0ce06SHong 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]); 5020aa0ce06SHong 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]); 5030aa0ce06SHong 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]); 5040aa0ce06SHong 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]); 5050aa0ce06SHong 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]); 5060aa0ce06SHong Zhang 5070aa0ce06SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 5080aa0ce06SHong 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]; 5090aa0ce06SHong 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]; 5100aa0ce06SHong 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]; 5110aa0ce06SHong 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]; 5120aa0ce06SHong 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]; 5130aa0ce06SHong 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]; 5140aa0ce06SHong 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]; 5150aa0ce06SHong Zhang 5160aa0ce06SHong 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]; 5170aa0ce06SHong 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]; 5180aa0ce06SHong 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]; 5190aa0ce06SHong 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]; 5200aa0ce06SHong 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]; 5210aa0ce06SHong 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]; 5220aa0ce06SHong 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]; 5230aa0ce06SHong Zhang 5240aa0ce06SHong 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]; 5250aa0ce06SHong 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]; 5260aa0ce06SHong 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]; 5270aa0ce06SHong 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]; 5280aa0ce06SHong 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]; 5290aa0ce06SHong 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]; 5300aa0ce06SHong 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]; 5310aa0ce06SHong Zhang 5320aa0ce06SHong 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]; 5330aa0ce06SHong 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]; 5340aa0ce06SHong 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]; 5350aa0ce06SHong 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]; 5360aa0ce06SHong 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]; 5370aa0ce06SHong 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]; 5380aa0ce06SHong 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]; 5390aa0ce06SHong Zhang 5400aa0ce06SHong 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]; 5410aa0ce06SHong 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]; 5420aa0ce06SHong 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]; 5430aa0ce06SHong 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]; 5440aa0ce06SHong 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]; 5450aa0ce06SHong 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]; 5460aa0ce06SHong 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]; 5470aa0ce06SHong Zhang 5480aa0ce06SHong 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]; 5490aa0ce06SHong 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]; 5500aa0ce06SHong 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]; 5510aa0ce06SHong 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]; 5520aa0ce06SHong 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]; 5530aa0ce06SHong 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]; 5540aa0ce06SHong 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]; 5550aa0ce06SHong Zhang 5560aa0ce06SHong 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]; 5570aa0ce06SHong 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]; 5580aa0ce06SHong 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]; 5590aa0ce06SHong 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]; 5600aa0ce06SHong 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]; 5610aa0ce06SHong 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]; 5620aa0ce06SHong 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]; 5630aa0ce06SHong Zhang 5640aa0ce06SHong Zhang /* update -U(i,k) */ 5650aa0ce06SHong Zhang ierr = PetscMemcpy(ba+ili*49,uik,49*sizeof(MatScalar));CHKERRQ(ierr); 5660aa0ce06SHong Zhang 5670aa0ce06SHong Zhang /* add multiple of row i to k-th row ... */ 5680aa0ce06SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 5690aa0ce06SHong Zhang if (jmin < jmax){ 5700aa0ce06SHong Zhang for (j=jmin; j<jmax; j++) { 5710aa0ce06SHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 5720aa0ce06SHong Zhang wp = w + bj[j]*49; 5730aa0ce06SHong Zhang u = ba + j*49; 5740aa0ce06SHong Zhang 5750aa0ce06SHong 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]; 5760aa0ce06SHong 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]; 5770aa0ce06SHong 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]; 5780aa0ce06SHong 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]; 5790aa0ce06SHong 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]; 5800aa0ce06SHong 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]; 5810aa0ce06SHong 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]; 5820aa0ce06SHong Zhang 5830aa0ce06SHong 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]; 5840aa0ce06SHong 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]; 5850aa0ce06SHong 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]; 5860aa0ce06SHong 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]; 5870aa0ce06SHong 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]; 5880aa0ce06SHong 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]; 5890aa0ce06SHong 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]; 5900aa0ce06SHong Zhang 5910aa0ce06SHong 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]; 5920aa0ce06SHong 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]; 5930aa0ce06SHong 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]; 5940aa0ce06SHong 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]; 5950aa0ce06SHong 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]; 5960aa0ce06SHong 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]; 5970aa0ce06SHong 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]; 5980aa0ce06SHong Zhang 5990aa0ce06SHong 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]; 6000aa0ce06SHong 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]; 6010aa0ce06SHong 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]; 6020aa0ce06SHong 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]; 6030aa0ce06SHong 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]; 6040aa0ce06SHong 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]; 6050aa0ce06SHong 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]; 6060aa0ce06SHong Zhang 6070aa0ce06SHong 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]; 6080aa0ce06SHong 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]; 6090aa0ce06SHong 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]; 6100aa0ce06SHong 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]; 6110aa0ce06SHong 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]; 6120aa0ce06SHong 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]; 6130aa0ce06SHong 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]; 6140aa0ce06SHong Zhang 6150aa0ce06SHong 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]; 6160aa0ce06SHong 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]; 6170aa0ce06SHong 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]; 6180aa0ce06SHong 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]; 6190aa0ce06SHong 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]; 6200aa0ce06SHong 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]; 6210aa0ce06SHong 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]; 6220aa0ce06SHong Zhang 6230aa0ce06SHong 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]; 6240aa0ce06SHong 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]; 6250aa0ce06SHong 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]; 6260aa0ce06SHong 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]; 6270aa0ce06SHong 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]; 6280aa0ce06SHong 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]; 6290aa0ce06SHong 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]; 6300aa0ce06SHong Zhang } 6310aa0ce06SHong Zhang 6320aa0ce06SHong Zhang /* ... add i to row list for next nonzero entry */ 6330aa0ce06SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 6340aa0ce06SHong Zhang j = bj[jmin]; 6350aa0ce06SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 6360aa0ce06SHong Zhang } 6370aa0ce06SHong Zhang i = nexti; 6380aa0ce06SHong Zhang } 6390aa0ce06SHong Zhang 6400aa0ce06SHong Zhang /* save nonzero entries in k-th row of U ... */ 6410aa0ce06SHong Zhang 64249b5e25fSSatish Balay /* invert diagonal block */ 6430aa0ce06SHong Zhang d = ba+k*49; 6440aa0ce06SHong Zhang ierr = PetscMemcpy(d,dk,49*sizeof(MatScalar));CHKERRQ(ierr); 6450aa0ce06SHong Zhang ierr = Kernel_A_gets_inverse_A_7(d);CHKERRQ(ierr); 6460aa0ce06SHong Zhang 6470aa0ce06SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 6480aa0ce06SHong Zhang if (jmin < jmax) { 6490aa0ce06SHong Zhang for (j=jmin; j<jmax; j++){ 6500aa0ce06SHong Zhang vj = bj[j]; /* block col. index of U */ 6510aa0ce06SHong Zhang u = ba + j*49; 6520aa0ce06SHong Zhang wp = w + vj*49; 6530aa0ce06SHong Zhang for (k1=0; k1<49; k1++){ 6540aa0ce06SHong Zhang *u++ = *wp; 6550aa0ce06SHong Zhang *wp++ = 0.0; 6560aa0ce06SHong Zhang } 65749b5e25fSSatish Balay } 65849b5e25fSSatish Balay 6590aa0ce06SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 6600aa0ce06SHong Zhang il[k] = jmin; 6610aa0ce06SHong Zhang i = bj[jmin]; 6620aa0ce06SHong Zhang jl[k] = jl[i]; jl[i] = k; 6630aa0ce06SHong Zhang } 6640aa0ce06SHong Zhang } 6650aa0ce06SHong Zhang 6660aa0ce06SHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 6670aa0ce06SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 6680aa0ce06SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 6690aa0ce06SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 6700aa0ce06SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 6710aa0ce06SHong Zhang if (a->permute){ 6720aa0ce06SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 6730aa0ce06SHong Zhang } 6740aa0ce06SHong Zhang 6750aa0ce06SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 6760aa0ce06SHong Zhang C->factor = FACTOR_CHOLESKY; 67749b5e25fSSatish Balay C->assembled = PETSC_TRUE; 6780aa0ce06SHong Zhang C->preallocated = PETSC_TRUE; 67949b5e25fSSatish Balay PLogFlops(1.3333*343*b->mbs); /* from inverting diagonal blocks */ 68049b5e25fSSatish Balay PetscFunctionReturn(0); 68149b5e25fSSatish Balay } 68249b5e25fSSatish Balay 68349b5e25fSSatish Balay /* 68449b5e25fSSatish Balay Version for when blocks are 7 by 7 Using natural ordering 68549b5e25fSSatish Balay */ 68649b5e25fSSatish Balay #undef __FUNC__ 6876f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering" 6886f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering(Mat A,Mat *B) 68949b5e25fSSatish Balay { 69049b5e25fSSatish Balay Mat C = *B; 691ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 692ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 693*71149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 694ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 695ab27746eSHong Zhang MatScalar *u,*d,*w,*wp; 69649b5e25fSSatish Balay 69749b5e25fSSatish Balay PetscFunctionBegin; 698ab27746eSHong Zhang /* initialization */ 699ab27746eSHong Zhang w = (MatScalar*)PetscMalloc(49*mbs*sizeof(MatScalar));CHKPTRQ(w); 700ab27746eSHong Zhang ierr = PetscMemzero(w,49*mbs*sizeof(MatScalar));CHKERRQ(ierr); 701ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 702ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 703ab27746eSHong Zhang for (i=0; i<mbs; i++) { 704ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 70549b5e25fSSatish Balay } 706ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(dk); 707ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(49*sizeof(MatScalar));CHKPTRQ(uik); 708ab27746eSHong Zhang 709ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 710ab27746eSHong Zhang 711ab27746eSHong Zhang /* for each row k */ 712ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 713ab27746eSHong Zhang 714ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 715ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 716ab27746eSHong Zhang if (jmin < jmax) { 717ab27746eSHong Zhang ap = aa + jmin*49; 718ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 719ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 720ab27746eSHong Zhang wp = w + vj*49; 721ab27746eSHong Zhang for (i=0; i<49; i++) *wp++ = *ap++; 72249b5e25fSSatish Balay } 72349b5e25fSSatish Balay } 724ab27746eSHong Zhang 725ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 726ab27746eSHong Zhang ierr = PetscMemcpy(dk,w+k*49,49*sizeof(MatScalar));CHKERRQ(ierr); 727ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 728ab27746eSHong Zhang 729ab27746eSHong Zhang while (i < mbs){ 730ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 731ab27746eSHong Zhang 732ab27746eSHong Zhang /* compute multiplier */ 733ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 734ab27746eSHong Zhang 735ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 736ab27746eSHong Zhang d = ba + i*49; 737ab27746eSHong Zhang u = ba + ili*49; 738ab27746eSHong Zhang 739ab27746eSHong 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]); 740ab27746eSHong 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]); 741ab27746eSHong 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]); 742ab27746eSHong 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]); 743ab27746eSHong 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]); 744ab27746eSHong 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]); 745ab27746eSHong 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]); 746ab27746eSHong Zhang 747ab27746eSHong 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]); 748ab27746eSHong 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]); 749ab27746eSHong 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]); 750ab27746eSHong 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]); 751ab27746eSHong 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]); 752ab27746eSHong 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]); 753ab27746eSHong 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]); 754ab27746eSHong Zhang 755ab27746eSHong 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]); 756ab27746eSHong 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]); 757ab27746eSHong 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]); 758ab27746eSHong 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]); 759ab27746eSHong 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]); 760ab27746eSHong 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]); 761ab27746eSHong 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]); 762ab27746eSHong Zhang 763ab27746eSHong 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]); 764ab27746eSHong 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]); 765ab27746eSHong 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]); 766ab27746eSHong 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]); 767ab27746eSHong 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]); 768ab27746eSHong 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]); 769ab27746eSHong 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]); 770ab27746eSHong Zhang 771ab27746eSHong 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]); 772ab27746eSHong 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]); 773ab27746eSHong 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]); 774ab27746eSHong 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]); 775ab27746eSHong 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]); 776ab27746eSHong 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]); 777ab27746eSHong 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]); 778ab27746eSHong Zhang 779ab27746eSHong 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]); 780ab27746eSHong 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]); 781ab27746eSHong 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]); 782ab27746eSHong 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]); 783ab27746eSHong 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]); 784ab27746eSHong 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]); 785ab27746eSHong 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]); 786ab27746eSHong Zhang 787ab27746eSHong 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]); 788ab27746eSHong 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]); 789ab27746eSHong 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]); 790ab27746eSHong 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]); 791ab27746eSHong 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]); 792ab27746eSHong 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]); 793ab27746eSHong 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]); 794ab27746eSHong Zhang 795ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 796ab27746eSHong 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]; 797ab27746eSHong 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]; 798ab27746eSHong 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]; 799ab27746eSHong 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]; 800ab27746eSHong 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]; 801ab27746eSHong 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]; 802ab27746eSHong 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]; 803ab27746eSHong Zhang 804ab27746eSHong 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]; 805ab27746eSHong 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]; 806ab27746eSHong 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]; 807ab27746eSHong 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]; 808ab27746eSHong 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]; 809ab27746eSHong 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]; 810ab27746eSHong 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]; 811ab27746eSHong Zhang 812ab27746eSHong 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]; 813ab27746eSHong 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]; 814ab27746eSHong 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]; 815ab27746eSHong 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]; 816ab27746eSHong 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]; 817ab27746eSHong 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]; 818ab27746eSHong 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]; 819ab27746eSHong Zhang 820ab27746eSHong 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]; 821ab27746eSHong 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]; 822ab27746eSHong 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]; 823ab27746eSHong 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]; 824ab27746eSHong 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]; 825ab27746eSHong 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]; 826ab27746eSHong 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]; 827ab27746eSHong Zhang 828ab27746eSHong 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]; 829ab27746eSHong 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]; 830ab27746eSHong 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]; 831ab27746eSHong 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]; 832ab27746eSHong 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]; 833ab27746eSHong 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]; 834ab27746eSHong 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]; 835ab27746eSHong Zhang 836ab27746eSHong 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]; 837ab27746eSHong 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]; 838ab27746eSHong 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]; 839ab27746eSHong 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]; 840ab27746eSHong 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]; 841ab27746eSHong 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]; 842ab27746eSHong 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]; 843ab27746eSHong Zhang 844ab27746eSHong 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]; 845ab27746eSHong 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]; 846ab27746eSHong 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]; 847ab27746eSHong 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]; 848ab27746eSHong 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]; 849ab27746eSHong 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]; 850ab27746eSHong 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]; 851ab27746eSHong Zhang 852ab27746eSHong Zhang /* update -U(i,k) */ 853ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*49,uik,49*sizeof(MatScalar));CHKERRQ(ierr); 854ab27746eSHong Zhang 855ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 856ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 857ab27746eSHong Zhang if (jmin < jmax){ 858ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 859ab27746eSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 860ab27746eSHong Zhang wp = w + bj[j]*49; 861ab27746eSHong Zhang u = ba + j*49; 862ab27746eSHong Zhang 863ab27746eSHong 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]; 864ab27746eSHong 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]; 865ab27746eSHong 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]; 866ab27746eSHong 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]; 867ab27746eSHong 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]; 868ab27746eSHong 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]; 869ab27746eSHong 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]; 870ab27746eSHong Zhang 871ab27746eSHong 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]; 872ab27746eSHong 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]; 873ab27746eSHong 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]; 874ab27746eSHong 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]; 875ab27746eSHong 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]; 876ab27746eSHong 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]; 877ab27746eSHong 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]; 878ab27746eSHong Zhang 879ab27746eSHong 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]; 880ab27746eSHong 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]; 881ab27746eSHong 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]; 882ab27746eSHong 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]; 883ab27746eSHong 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]; 884ab27746eSHong 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]; 885ab27746eSHong 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]; 886ab27746eSHong Zhang 887ab27746eSHong 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]; 888ab27746eSHong 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]; 889ab27746eSHong 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]; 890ab27746eSHong 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]; 891ab27746eSHong 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]; 892ab27746eSHong 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]; 893ab27746eSHong 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]; 894ab27746eSHong Zhang 895ab27746eSHong 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]; 896ab27746eSHong 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]; 897ab27746eSHong 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]; 898ab27746eSHong 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]; 899ab27746eSHong 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]; 900ab27746eSHong 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]; 901ab27746eSHong 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]; 902ab27746eSHong Zhang 903ab27746eSHong 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]; 904ab27746eSHong 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]; 905ab27746eSHong 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]; 906ab27746eSHong 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]; 907ab27746eSHong 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]; 908ab27746eSHong 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]; 909ab27746eSHong 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]; 910ab27746eSHong Zhang 911ab27746eSHong 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]; 912ab27746eSHong 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]; 913ab27746eSHong 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]; 914ab27746eSHong 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]; 915ab27746eSHong 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]; 916ab27746eSHong 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]; 917ab27746eSHong 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]; 91849b5e25fSSatish Balay } 919ab27746eSHong Zhang 920ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 921ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 922ab27746eSHong Zhang j = bj[jmin]; 923ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 92449b5e25fSSatish Balay } 925ab27746eSHong Zhang i = nexti; 92649b5e25fSSatish Balay } 927ab27746eSHong Zhang 928ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 929ab27746eSHong Zhang 93049b5e25fSSatish Balay /* invert diagonal block */ 931ab27746eSHong Zhang d = ba+k*49; 932ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,49*sizeof(MatScalar));CHKERRQ(ierr); 933ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_7(d);CHKERRQ(ierr); 934ab27746eSHong Zhang 935ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 936ab27746eSHong Zhang if (jmin < jmax) { 937ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 938ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 939ab27746eSHong Zhang u = ba + j*49; 940ab27746eSHong Zhang wp = w + vj*49; 941ab27746eSHong Zhang for (k1=0; k1<49; k1++){ 942ab27746eSHong Zhang *u++ = *wp; 943ab27746eSHong Zhang *wp++ = 0.0; 944ab27746eSHong Zhang } 94549b5e25fSSatish Balay } 94649b5e25fSSatish Balay 947ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 948ab27746eSHong Zhang il[k] = jmin; 949ab27746eSHong Zhang i = bj[jmin]; 950ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 951ab27746eSHong Zhang } 952ab27746eSHong Zhang } 953ab27746eSHong Zhang 954ab27746eSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 955ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 956ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 957ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 958ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 959ab27746eSHong Zhang 960ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 96149b5e25fSSatish Balay C->assembled = PETSC_TRUE; 962ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 96349b5e25fSSatish Balay PLogFlops(1.3333*343*b->mbs); /* from inverting diagonal blocks */ 96449b5e25fSSatish Balay PetscFunctionReturn(0); 96549b5e25fSSatish Balay } 96649b5e25fSSatish Balay 9679b0e81eeSHong Zhang /* Version for when blocks are 6 by 6 */ 96849b5e25fSSatish Balay #undef __FUNC__ 9696f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6" 9706f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_6(Mat A,Mat *B) 97149b5e25fSSatish Balay { 97249b5e25fSSatish Balay Mat C = *B; 9739b0e81eeSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 9749b0e81eeSHong Zhang IS perm = b->row; 9759b0e81eeSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 9769b0e81eeSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 9779b0e81eeSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 9789b0e81eeSHong Zhang MatScalar *u,*d,*w,*wp; 97949b5e25fSSatish Balay 98049b5e25fSSatish Balay PetscFunctionBegin; 9819b0e81eeSHong Zhang /* initialization */ 9829b0e81eeSHong Zhang w = (MatScalar*)PetscMalloc(36*mbs*sizeof(MatScalar));CHKPTRQ(w); 9839b0e81eeSHong Zhang ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 9849b0e81eeSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 9859b0e81eeSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 9869b0e81eeSHong Zhang for (i=0; i<mbs; i++) { 9879b0e81eeSHong Zhang jl[i] = mbs; il[0] = 0; 98849b5e25fSSatish Balay } 9899b0e81eeSHong Zhang dk = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(dk); 9909b0e81eeSHong Zhang uik = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(uik); 9919b0e81eeSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 9929b0e81eeSHong Zhang 9939b0e81eeSHong Zhang /* check permutation */ 9949b0e81eeSHong Zhang if (!a->permute){ 9959b0e81eeSHong Zhang ai = a->i; aj = a->j; aa = a->a; 9969b0e81eeSHong Zhang } else { 9979b0e81eeSHong Zhang ai = a->inew; aj = a->jnew; 9989b0e81eeSHong Zhang aa = (MatScalar*)PetscMalloc(36*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 9999b0e81eeSHong Zhang ierr = PetscMemcpy(aa,a->a,36*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 10009b0e81eeSHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 10019b0e81eeSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 10029b0e81eeSHong Zhang 10039b0e81eeSHong Zhang for (i=0; i<mbs; i++){ 10049b0e81eeSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 10059b0e81eeSHong Zhang for (j=jmin; j<jmax; j++){ 10069b0e81eeSHong Zhang while (a2anew[j] != j){ 10079b0e81eeSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 10089b0e81eeSHong Zhang for (k1=0; k1<36; k1++){ 10099b0e81eeSHong Zhang dk[k1] = aa[k*36+k1]; 10109b0e81eeSHong Zhang aa[k*36+k1] = aa[j*36+k1]; 10119b0e81eeSHong Zhang aa[j*36+k1] = dk[k1]; 101249b5e25fSSatish Balay } 101349b5e25fSSatish Balay } 10149b0e81eeSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 10159b0e81eeSHong Zhang if (i > aj[j]){ 10169b0e81eeSHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 10179b0e81eeSHong Zhang ap = aa + j*36; /* ptr to the beginning of j-th block of aa */ 10189b0e81eeSHong Zhang for (k=0; k<36; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 10199b0e81eeSHong Zhang for (k=0; k<6; k++){ /* j-th block of aa <- dk^T */ 10209b0e81eeSHong Zhang for (k1=0; k1<6; k1++) *ap++ = dk[k + 6*k1]; 102149b5e25fSSatish Balay } 102249b5e25fSSatish Balay } 102349b5e25fSSatish Balay } 10249b0e81eeSHong Zhang } 10259b0e81eeSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 10269b0e81eeSHong Zhang } 10279b0e81eeSHong Zhang 10289b0e81eeSHong Zhang /* for each row k */ 10299b0e81eeSHong Zhang for (k = 0; k<mbs; k++){ 10309b0e81eeSHong Zhang 10319b0e81eeSHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 10329b0e81eeSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 10339b0e81eeSHong Zhang if (jmin < jmax) { 10349b0e81eeSHong Zhang ap = aa + jmin*36; 10359b0e81eeSHong Zhang for (j = jmin; j < jmax; j++){ 10369b0e81eeSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 10379b0e81eeSHong Zhang wp = w + vj*36; 10389b0e81eeSHong Zhang for (i=0; i<36; i++) *wp++ = *ap++; 10399b0e81eeSHong Zhang } 10409b0e81eeSHong Zhang } 10419b0e81eeSHong Zhang 10429b0e81eeSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 10439b0e81eeSHong Zhang ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 10449b0e81eeSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 10459b0e81eeSHong Zhang 10469b0e81eeSHong Zhang while (i < mbs){ 10479b0e81eeSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 10489b0e81eeSHong Zhang 10499b0e81eeSHong Zhang /* compute multiplier */ 10509b0e81eeSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 10519b0e81eeSHong Zhang 10529b0e81eeSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 10539b0e81eeSHong Zhang d = ba + i*36; 10549b0e81eeSHong Zhang u = ba + ili*36; 10559b0e81eeSHong Zhang 10569b0e81eeSHong 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]); 10579b0e81eeSHong 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]); 10589b0e81eeSHong 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]); 10599b0e81eeSHong 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]); 10609b0e81eeSHong 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]); 10619b0e81eeSHong 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]); 10629b0e81eeSHong Zhang 10639b0e81eeSHong 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]); 10649b0e81eeSHong 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]); 10659b0e81eeSHong 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]); 10669b0e81eeSHong 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]); 10679b0e81eeSHong 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]); 10689b0e81eeSHong 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]); 10699b0e81eeSHong Zhang 10709b0e81eeSHong 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]); 10719b0e81eeSHong 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]); 10729b0e81eeSHong 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]); 10739b0e81eeSHong 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]); 10749b0e81eeSHong 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]); 10759b0e81eeSHong 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]); 10769b0e81eeSHong Zhang 10779b0e81eeSHong 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]); 10789b0e81eeSHong 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]); 10799b0e81eeSHong 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]); 10809b0e81eeSHong 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]); 10819b0e81eeSHong 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]); 10829b0e81eeSHong 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]); 10839b0e81eeSHong Zhang 10849b0e81eeSHong 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]); 10859b0e81eeSHong 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]); 10869b0e81eeSHong 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]); 10879b0e81eeSHong 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]); 10889b0e81eeSHong 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]); 10899b0e81eeSHong 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]); 10909b0e81eeSHong Zhang 10919b0e81eeSHong 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]); 10929b0e81eeSHong 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]); 10939b0e81eeSHong 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]); 10949b0e81eeSHong 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]); 10959b0e81eeSHong 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]); 10969b0e81eeSHong 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]); 10979b0e81eeSHong Zhang 10989b0e81eeSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 10999b0e81eeSHong 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]; 11009b0e81eeSHong 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]; 11019b0e81eeSHong 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]; 11029b0e81eeSHong 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]; 11039b0e81eeSHong 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]; 11049b0e81eeSHong 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]; 11059b0e81eeSHong Zhang 11069b0e81eeSHong 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]; 11079b0e81eeSHong 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]; 11089b0e81eeSHong 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]; 11099b0e81eeSHong 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]; 11109b0e81eeSHong 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]; 11119b0e81eeSHong 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]; 11129b0e81eeSHong Zhang 11139b0e81eeSHong 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]; 11149b0e81eeSHong 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]; 11159b0e81eeSHong 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]; 11169b0e81eeSHong 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]; 11179b0e81eeSHong 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]; 11189b0e81eeSHong 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]; 11199b0e81eeSHong Zhang 11209b0e81eeSHong 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]; 11219b0e81eeSHong 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]; 11229b0e81eeSHong 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]; 11239b0e81eeSHong 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]; 11249b0e81eeSHong 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]; 11259b0e81eeSHong 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]; 11269b0e81eeSHong Zhang 11279b0e81eeSHong 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]; 11289b0e81eeSHong 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]; 11299b0e81eeSHong 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]; 11309b0e81eeSHong 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]; 11319b0e81eeSHong 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]; 11329b0e81eeSHong 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]; 11339b0e81eeSHong Zhang 11349b0e81eeSHong 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]; 11359b0e81eeSHong 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]; 11369b0e81eeSHong 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]; 11379b0e81eeSHong 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]; 11389b0e81eeSHong 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]; 11399b0e81eeSHong 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]; 11409b0e81eeSHong Zhang 11419b0e81eeSHong Zhang /* update -U(i,k) */ 11429b0e81eeSHong Zhang ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 11439b0e81eeSHong Zhang 11449b0e81eeSHong Zhang /* add multiple of row i to k-th row ... */ 11459b0e81eeSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 11469b0e81eeSHong Zhang if (jmin < jmax){ 11479b0e81eeSHong Zhang for (j=jmin; j<jmax; j++) { 11489b0e81eeSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 11499b0e81eeSHong Zhang wp = w + bj[j]*36; 11509b0e81eeSHong Zhang u = ba + j*36; 11519b0e81eeSHong 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]; 11529b0e81eeSHong 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]; 11539b0e81eeSHong 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]; 11549b0e81eeSHong 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]; 11559b0e81eeSHong 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]; 11569b0e81eeSHong 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]; 11579b0e81eeSHong Zhang 11589b0e81eeSHong 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]; 11599b0e81eeSHong 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]; 11609b0e81eeSHong 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]; 11619b0e81eeSHong 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]; 11629b0e81eeSHong 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]; 11639b0e81eeSHong 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]; 11649b0e81eeSHong Zhang 11659b0e81eeSHong 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]; 11669b0e81eeSHong 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]; 11679b0e81eeSHong 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]; 11689b0e81eeSHong 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]; 11699b0e81eeSHong 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]; 11709b0e81eeSHong 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]; 11719b0e81eeSHong Zhang 11729b0e81eeSHong 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]; 11739b0e81eeSHong 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]; 11749b0e81eeSHong 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]; 11759b0e81eeSHong 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]; 11769b0e81eeSHong 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]; 11779b0e81eeSHong 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]; 11789b0e81eeSHong Zhang 11799b0e81eeSHong 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]; 11809b0e81eeSHong 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]; 11819b0e81eeSHong 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]; 11829b0e81eeSHong 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]; 11839b0e81eeSHong 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]; 11849b0e81eeSHong 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]; 11859b0e81eeSHong Zhang 11869b0e81eeSHong 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]; 11879b0e81eeSHong 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]; 11889b0e81eeSHong 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]; 11899b0e81eeSHong 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]; 11909b0e81eeSHong 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]; 11919b0e81eeSHong 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]; 11929b0e81eeSHong Zhang } 11939b0e81eeSHong Zhang 11949b0e81eeSHong Zhang /* ... add i to row list for next nonzero entry */ 11959b0e81eeSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 11969b0e81eeSHong Zhang j = bj[jmin]; 11979b0e81eeSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 11989b0e81eeSHong Zhang } 11999b0e81eeSHong Zhang i = nexti; 12009b0e81eeSHong Zhang } 12019b0e81eeSHong Zhang 12029b0e81eeSHong Zhang /* save nonzero entries in k-th row of U ... */ 12039b0e81eeSHong Zhang 120449b5e25fSSatish Balay /* invert diagonal block */ 12059b0e81eeSHong Zhang d = ba+k*36; 12069b0e81eeSHong Zhang ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 12079b0e81eeSHong Zhang ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 12089b0e81eeSHong Zhang 12099b0e81eeSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 12109b0e81eeSHong Zhang if (jmin < jmax) { 12119b0e81eeSHong Zhang for (j=jmin; j<jmax; j++){ 12129b0e81eeSHong Zhang vj = bj[j]; /* block col. index of U */ 12139b0e81eeSHong Zhang u = ba + j*36; 12149b0e81eeSHong Zhang wp = w + vj*36; 12159b0e81eeSHong Zhang for (k1=0; k1<36; k1++){ 12169b0e81eeSHong Zhang *u++ = *wp; 12179b0e81eeSHong Zhang *wp++ = 0.0; 12189b0e81eeSHong Zhang } 121949b5e25fSSatish Balay } 122049b5e25fSSatish Balay 12219b0e81eeSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 12229b0e81eeSHong Zhang il[k] = jmin; 12239b0e81eeSHong Zhang i = bj[jmin]; 12249b0e81eeSHong Zhang jl[k] = jl[i]; jl[i] = k; 12259b0e81eeSHong Zhang } 12269b0e81eeSHong Zhang } 12279b0e81eeSHong Zhang 12289b0e81eeSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 12299b0e81eeSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 12309b0e81eeSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 12319b0e81eeSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 12329b0e81eeSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 12339b0e81eeSHong Zhang if (a->permute){ 12349b0e81eeSHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 12359b0e81eeSHong Zhang } 12369b0e81eeSHong Zhang 12379b0e81eeSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 12389b0e81eeSHong Zhang C->factor = FACTOR_CHOLESKY; 123949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 12409b0e81eeSHong Zhang C->preallocated = PETSC_TRUE; 124149b5e25fSSatish Balay PLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 124249b5e25fSSatish Balay PetscFunctionReturn(0); 124349b5e25fSSatish Balay } 12449b0e81eeSHong Zhang 124549b5e25fSSatish Balay /* 124649b5e25fSSatish Balay Version for when blocks are 6 by 6 Using natural ordering 124749b5e25fSSatish Balay */ 124849b5e25fSSatish Balay #undef __FUNC__ 12496f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering" 12506f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering(Mat A,Mat *B) 125149b5e25fSSatish Balay { 125249b5e25fSSatish Balay Mat C = *B; 1253ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1254ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1255*71149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1256ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1257ab27746eSHong Zhang MatScalar *u,*d,*w,*wp; 125849b5e25fSSatish Balay 125949b5e25fSSatish Balay PetscFunctionBegin; 1260ab27746eSHong Zhang /* initialization */ 1261ab27746eSHong Zhang w = (MatScalar*)PetscMalloc(36*mbs*sizeof(MatScalar));CHKPTRQ(w); 1262ab27746eSHong Zhang ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1263ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 1264ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 1265ab27746eSHong Zhang for (i=0; i<mbs; i++) { 1266ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 126749b5e25fSSatish Balay } 1268ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(dk); 1269ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(36*sizeof(MatScalar));CHKPTRQ(uik); 1270ab27746eSHong Zhang 1271ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 1272ab27746eSHong Zhang 1273ab27746eSHong Zhang /* for each row k */ 1274ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 1275ab27746eSHong Zhang 1276ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 1277ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 1278ab27746eSHong Zhang if (jmin < jmax) { 1279ab27746eSHong Zhang ap = aa + jmin*36; 1280ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 1281ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 1282ab27746eSHong Zhang wp = w + vj*36; 1283ab27746eSHong Zhang for (i=0; i<36; i++) *wp++ = *ap++; 128449b5e25fSSatish Balay } 128549b5e25fSSatish Balay } 1286ab27746eSHong Zhang 1287ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1288ab27746eSHong Zhang ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 1289ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1290ab27746eSHong Zhang 1291ab27746eSHong Zhang while (i < mbs){ 1292ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1293ab27746eSHong Zhang 1294ab27746eSHong Zhang /* compute multiplier */ 1295ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1296ab27746eSHong Zhang 1297ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1298ab27746eSHong Zhang d = ba + i*36; 1299ab27746eSHong Zhang u = ba + ili*36; 1300ab27746eSHong Zhang 1301ab27746eSHong 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]); 1302ab27746eSHong 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]); 1303ab27746eSHong 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]); 1304ab27746eSHong 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]); 1305ab27746eSHong 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]); 1306ab27746eSHong 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]); 1307ab27746eSHong Zhang 1308ab27746eSHong 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]); 1309ab27746eSHong 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]); 1310ab27746eSHong 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]); 1311ab27746eSHong 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]); 1312ab27746eSHong 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]); 1313ab27746eSHong 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]); 1314ab27746eSHong Zhang 1315ab27746eSHong 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]); 1316ab27746eSHong 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]); 1317ab27746eSHong 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]); 1318ab27746eSHong 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]); 1319ab27746eSHong 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]); 1320ab27746eSHong 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]); 1321ab27746eSHong Zhang 1322ab27746eSHong 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]); 1323ab27746eSHong 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]); 1324ab27746eSHong 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]); 1325ab27746eSHong 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]); 1326ab27746eSHong 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]); 1327ab27746eSHong 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]); 1328ab27746eSHong Zhang 1329ab27746eSHong 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]); 1330ab27746eSHong 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]); 1331ab27746eSHong 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]); 1332ab27746eSHong 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]); 1333ab27746eSHong 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]); 1334ab27746eSHong 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]); 1335ab27746eSHong Zhang 1336ab27746eSHong 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]); 1337ab27746eSHong 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]); 1338ab27746eSHong 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]); 1339ab27746eSHong 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]); 1340ab27746eSHong 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]); 1341ab27746eSHong 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]); 1342ab27746eSHong Zhang 1343ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1344ab27746eSHong 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]; 1345ab27746eSHong 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]; 1346ab27746eSHong 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]; 1347ab27746eSHong 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]; 1348ab27746eSHong 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]; 1349ab27746eSHong 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]; 1350ab27746eSHong Zhang 1351ab27746eSHong 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]; 1352ab27746eSHong 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]; 1353ab27746eSHong 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]; 1354ab27746eSHong 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]; 1355ab27746eSHong 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]; 1356ab27746eSHong 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]; 1357ab27746eSHong Zhang 1358ab27746eSHong 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]; 1359ab27746eSHong 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]; 1360ab27746eSHong 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]; 1361ab27746eSHong 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]; 1362ab27746eSHong 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]; 1363ab27746eSHong 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]; 1364ab27746eSHong Zhang 1365ab27746eSHong 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]; 1366ab27746eSHong 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]; 1367ab27746eSHong 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]; 1368ab27746eSHong 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]; 1369ab27746eSHong 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]; 1370ab27746eSHong 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]; 1371ab27746eSHong Zhang 1372ab27746eSHong 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]; 1373ab27746eSHong 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]; 1374ab27746eSHong 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]; 1375ab27746eSHong 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]; 1376ab27746eSHong 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]; 1377ab27746eSHong 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]; 1378ab27746eSHong Zhang 1379ab27746eSHong 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]; 1380ab27746eSHong 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]; 1381ab27746eSHong 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]; 1382ab27746eSHong 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]; 1383ab27746eSHong 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]; 1384ab27746eSHong 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]; 1385ab27746eSHong Zhang 1386ab27746eSHong Zhang /* update -U(i,k) */ 1387ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 1388ab27746eSHong Zhang 1389ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 1390ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1391ab27746eSHong Zhang if (jmin < jmax){ 1392ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 1393ab27746eSHong Zhang /* w += -U(i,k)^T * U_bar(i,j) */ 1394ab27746eSHong Zhang wp = w + bj[j]*36; 1395ab27746eSHong Zhang u = ba + j*36; 1396ab27746eSHong 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]; 1397ab27746eSHong 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]; 1398ab27746eSHong 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]; 1399ab27746eSHong 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]; 1400ab27746eSHong 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]; 1401ab27746eSHong 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]; 1402ab27746eSHong Zhang 1403ab27746eSHong 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]; 1404ab27746eSHong 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]; 1405ab27746eSHong 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]; 1406ab27746eSHong 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]; 1407ab27746eSHong 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]; 1408ab27746eSHong 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]; 1409ab27746eSHong Zhang 1410ab27746eSHong 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]; 1411ab27746eSHong 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]; 1412ab27746eSHong 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]; 1413ab27746eSHong 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]; 1414ab27746eSHong 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]; 1415ab27746eSHong 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]; 1416ab27746eSHong Zhang 1417ab27746eSHong 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]; 1418ab27746eSHong 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]; 1419ab27746eSHong 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]; 1420ab27746eSHong 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]; 1421ab27746eSHong 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]; 1422ab27746eSHong 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]; 1423ab27746eSHong Zhang 1424ab27746eSHong 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]; 1425ab27746eSHong 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]; 1426ab27746eSHong 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]; 1427ab27746eSHong 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]; 1428ab27746eSHong 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]; 1429ab27746eSHong 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]; 1430ab27746eSHong Zhang 1431ab27746eSHong 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]; 1432ab27746eSHong 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]; 1433ab27746eSHong 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]; 1434ab27746eSHong 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]; 1435ab27746eSHong 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]; 1436ab27746eSHong 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]; 143749b5e25fSSatish Balay } 1438ab27746eSHong Zhang 1439ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 1440ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1441ab27746eSHong Zhang j = bj[jmin]; 1442ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 144349b5e25fSSatish Balay } 1444ab27746eSHong Zhang i = nexti; 144549b5e25fSSatish Balay } 1446ab27746eSHong Zhang 1447ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 1448ab27746eSHong Zhang 144949b5e25fSSatish Balay /* invert diagonal block */ 1450ab27746eSHong Zhang d = ba+k*36; 1451ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 1452ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 1453ab27746eSHong Zhang 1454ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1455ab27746eSHong Zhang if (jmin < jmax) { 1456ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 1457ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 1458ab27746eSHong Zhang u = ba + j*36; 1459ab27746eSHong Zhang wp = w + vj*36; 1460ab27746eSHong Zhang for (k1=0; k1<36; k1++){ 1461ab27746eSHong Zhang *u++ = *wp; 1462ab27746eSHong Zhang *wp++ = 0.0; 1463ab27746eSHong Zhang } 146449b5e25fSSatish Balay } 146549b5e25fSSatish Balay 1466ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1467ab27746eSHong Zhang il[k] = jmin; 1468ab27746eSHong Zhang i = bj[jmin]; 1469ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 1470ab27746eSHong Zhang } 1471ab27746eSHong Zhang } 1472ab27746eSHong Zhang 1473ab27746eSHong Zhang ierr = PetscFree(w);CHKERRQ(ierr); 1474ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1475ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 1476ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1477ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 1478ab27746eSHong Zhang 1479ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 148049b5e25fSSatish Balay C->assembled = PETSC_TRUE; 1481ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 148249b5e25fSSatish Balay PLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 148349b5e25fSSatish Balay PetscFunctionReturn(0); 148449b5e25fSSatish Balay } 148549b5e25fSSatish Balay 1486b59abe41SHong Zhang /* Version for when blocks are 5 by 5 */ 1487b59abe41SHong Zhang #undef __FUNC__ 1488b59abe41SHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_5" 1489b59abe41SHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_5(Mat A,Mat *B) 1490b59abe41SHong Zhang { 1491b59abe41SHong Zhang Mat C = *B; 1492b59abe41SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1493b59abe41SHong Zhang IS perm = b->row; 1494b59abe41SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1495b59abe41SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1496b59abe41SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1497b59abe41SHong Zhang MatScalar *u,*d,*rtmp,*rtmp_ptr; 1498b59abe41SHong Zhang 1499b59abe41SHong Zhang PetscFunctionBegin; 1500b59abe41SHong Zhang /* initialization */ 1501b59abe41SHong Zhang rtmp = (MatScalar*)PetscMalloc(25*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 1502b59abe41SHong Zhang ierr = PetscMemzero(rtmp,25*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1503b59abe41SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 1504b59abe41SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 1505b59abe41SHong Zhang for (i=0; i<mbs; i++) { 1506b59abe41SHong Zhang jl[i] = mbs; il[0] = 0; 1507b59abe41SHong Zhang } 1508b59abe41SHong Zhang dk = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(dk); 1509b59abe41SHong Zhang uik = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(uik); 1510b59abe41SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 1511b59abe41SHong Zhang 1512b59abe41SHong Zhang /* check permutation */ 1513b59abe41SHong Zhang if (!a->permute){ 1514b59abe41SHong Zhang ai = a->i; aj = a->j; aa = a->a; 1515b59abe41SHong Zhang } else { 1516b59abe41SHong Zhang ai = a->inew; aj = a->jnew; 1517b59abe41SHong Zhang aa = (MatScalar*)PetscMalloc(25*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 1518b59abe41SHong Zhang ierr = PetscMemcpy(aa,a->a,25*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 1519b59abe41SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 1520b59abe41SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 1521b59abe41SHong Zhang 1522b59abe41SHong Zhang for (i=0; i<mbs; i++){ 1523b59abe41SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 1524b59abe41SHong Zhang for (j=jmin; j<jmax; j++){ 1525b59abe41SHong Zhang while (a2anew[j] != j){ 1526b59abe41SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 1527b59abe41SHong Zhang for (k1=0; k1<25; k1++){ 1528b59abe41SHong Zhang dk[k1] = aa[k*25+k1]; 1529b59abe41SHong Zhang aa[k*25+k1] = aa[j*25+k1]; 1530b59abe41SHong Zhang aa[j*25+k1] = dk[k1]; 1531b59abe41SHong Zhang } 1532b59abe41SHong Zhang } 1533b59abe41SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 1534b59abe41SHong Zhang if (i > aj[j]){ 1535b59abe41SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 1536b59abe41SHong Zhang ap = aa + j*25; /* ptr to the beginning of j-th block of aa */ 1537b59abe41SHong Zhang for (k=0; k<25; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 1538b59abe41SHong Zhang for (k=0; k<5; k++){ /* j-th block of aa <- dk^T */ 1539b59abe41SHong Zhang for (k1=0; k1<5; k1++) *ap++ = dk[k + 5*k1]; 1540b59abe41SHong Zhang } 1541b59abe41SHong Zhang } 1542b59abe41SHong Zhang } 1543b59abe41SHong Zhang } 1544b59abe41SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 1545b59abe41SHong Zhang } 1546b59abe41SHong Zhang 1547b59abe41SHong Zhang /* for each row k */ 1548b59abe41SHong Zhang for (k = 0; k<mbs; k++){ 1549b59abe41SHong Zhang 1550b59abe41SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 1551b59abe41SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 1552b59abe41SHong Zhang if (jmin < jmax) { 1553b59abe41SHong Zhang ap = aa + jmin*25; 1554b59abe41SHong Zhang for (j = jmin; j < jmax; j++){ 1555b59abe41SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 1556b59abe41SHong Zhang rtmp_ptr = rtmp + vj*25; 1557b59abe41SHong Zhang for (i=0; i<25; i++) *rtmp_ptr++ = *ap++; 1558b59abe41SHong Zhang } 1559b59abe41SHong Zhang } 1560b59abe41SHong Zhang 1561b59abe41SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1562b59abe41SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*25,25*sizeof(MatScalar));CHKERRQ(ierr); 1563b59abe41SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1564b59abe41SHong Zhang 1565b59abe41SHong Zhang while (i < mbs){ 1566b59abe41SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1567b59abe41SHong Zhang 1568b59abe41SHong Zhang /* compute multiplier */ 1569b59abe41SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1570b59abe41SHong Zhang 1571b59abe41SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1572b59abe41SHong Zhang d = ba + i*25; 1573b59abe41SHong Zhang u = ba + ili*25; 1574b59abe41SHong Zhang 1575b59abe41SHong Zhang uik[0] = -(d[0]*u[0] + d[5]*u[1] + d[10]*u[2] + d[15]*u[3] + d[20]*u[4]); 1576b59abe41SHong Zhang uik[1] = -(d[1]*u[0] + d[6]*u[1] + d[11]*u[2] + d[16]*u[3] + d[21]*u[4]); 1577b59abe41SHong Zhang uik[2] = -(d[2]*u[0] + d[7]*u[1] + d[12]*u[2] + d[17]*u[3] + d[22]*u[4]); 1578b59abe41SHong Zhang uik[3] = -(d[3]*u[0] + d[8]*u[1] + d[13]*u[2] + d[18]*u[3] + d[23]*u[4]); 1579b59abe41SHong Zhang uik[4] = -(d[4]*u[0] + d[9]*u[1] + d[14]*u[2] + d[19]*u[3] + d[24]*u[4]); 1580b59abe41SHong Zhang 1581b59abe41SHong Zhang uik[5] = -(d[0]*u[5] + d[5]*u[6] + d[10]*u[7] + d[15]*u[8] + d[20]*u[9]); 1582b59abe41SHong Zhang uik[6] = -(d[1]*u[5] + d[6]*u[6] + d[11]*u[7] + d[16]*u[8] + d[21]*u[9]); 1583b59abe41SHong Zhang uik[7] = -(d[2]*u[5] + d[7]*u[6] + d[12]*u[7] + d[17]*u[8] + d[22]*u[9]); 1584b59abe41SHong Zhang uik[8] = -(d[3]*u[5] + d[8]*u[6] + d[13]*u[7] + d[18]*u[8] + d[23]*u[9]); 1585b59abe41SHong Zhang uik[9] = -(d[4]*u[5] + d[9]*u[6] + d[14]*u[7] + d[19]*u[8] + d[24]*u[9]); 1586b59abe41SHong Zhang 1587b59abe41SHong Zhang uik[10]= -(d[0]*u[10] + d[5]*u[11] + d[10]*u[12] + d[15]*u[13] + d[20]*u[14]); 1588b59abe41SHong Zhang uik[11]= -(d[1]*u[10] + d[6]*u[11] + d[11]*u[12] + d[16]*u[13] + d[21]*u[14]); 1589b59abe41SHong Zhang uik[12]= -(d[2]*u[10] + d[7]*u[11] + d[12]*u[12] + d[17]*u[13] + d[22]*u[14]); 1590b59abe41SHong Zhang uik[13]= -(d[3]*u[10] + d[8]*u[11] + d[13]*u[12] + d[18]*u[13] + d[23]*u[14]); 1591b59abe41SHong Zhang uik[14]= -(d[4]*u[10] + d[9]*u[11] + d[14]*u[12] + d[19]*u[13] + d[24]*u[14]); 1592b59abe41SHong Zhang 1593b59abe41SHong Zhang uik[15]= -(d[0]*u[15] + d[5]*u[16] + d[10]*u[17] + d[15]*u[18] + d[20]*u[19]); 1594b59abe41SHong Zhang uik[16]= -(d[1]*u[15] + d[6]*u[16] + d[11]*u[17] + d[16]*u[18] + d[21]*u[19]); 1595b59abe41SHong Zhang uik[17]= -(d[2]*u[15] + d[7]*u[16] + d[12]*u[17] + d[17]*u[18] + d[22]*u[19]); 1596b59abe41SHong Zhang uik[18]= -(d[3]*u[15] + d[8]*u[16] + d[13]*u[17] + d[18]*u[18] + d[23]*u[19]); 1597b59abe41SHong Zhang uik[19]= -(d[4]*u[15] + d[9]*u[16] + d[14]*u[17] + d[19]*u[18] + d[24]*u[19]); 1598b59abe41SHong Zhang 1599b59abe41SHong Zhang uik[20]= -(d[0]*u[20] + d[5]*u[21] + d[10]*u[22] + d[15]*u[23] + d[20]*u[24]); 1600b59abe41SHong Zhang uik[21]= -(d[1]*u[20] + d[6]*u[21] + d[11]*u[22] + d[16]*u[23] + d[21]*u[24]); 1601b59abe41SHong Zhang uik[22]= -(d[2]*u[20] + d[7]*u[21] + d[12]*u[22] + d[17]*u[23] + d[22]*u[24]); 1602b59abe41SHong Zhang uik[23]= -(d[3]*u[20] + d[8]*u[21] + d[13]*u[22] + d[18]*u[23] + d[23]*u[24]); 1603b59abe41SHong Zhang uik[24]= -(d[4]*u[20] + d[9]*u[21] + d[14]*u[22] + d[19]*u[23] + d[24]*u[24]); 1604b59abe41SHong Zhang 1605b59abe41SHong Zhang 1606936bb232SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1607936bb232SHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1608936bb232SHong Zhang dk[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1609936bb232SHong Zhang dk[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1610936bb232SHong Zhang dk[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1611936bb232SHong Zhang dk[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1612b59abe41SHong Zhang 1613936bb232SHong Zhang dk[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1614936bb232SHong Zhang dk[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1615936bb232SHong Zhang dk[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1616936bb232SHong Zhang dk[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1617936bb232SHong Zhang dk[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1618b59abe41SHong Zhang 1619936bb232SHong Zhang dk[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1620936bb232SHong Zhang dk[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1621936bb232SHong Zhang dk[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1622936bb232SHong Zhang dk[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1623936bb232SHong Zhang dk[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1624b59abe41SHong Zhang 1625936bb232SHong Zhang dk[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1626936bb232SHong Zhang dk[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1627936bb232SHong Zhang dk[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1628936bb232SHong Zhang dk[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1629936bb232SHong Zhang dk[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1630936bb232SHong Zhang 1631936bb232SHong Zhang dk[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1632936bb232SHong Zhang dk[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1633936bb232SHong Zhang dk[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1634936bb232SHong Zhang dk[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 16354f078a5fSHong Zhang dk[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1636b59abe41SHong Zhang 1637b59abe41SHong Zhang /* update -U(i,k) */ 1638b59abe41SHong Zhang ierr = PetscMemcpy(ba+ili*25,uik,25*sizeof(MatScalar));CHKERRQ(ierr); 1639b59abe41SHong Zhang 1640b59abe41SHong Zhang /* add multiple of row i to k-th row ... */ 1641b59abe41SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1642b59abe41SHong Zhang if (jmin < jmax){ 1643b59abe41SHong Zhang for (j=jmin; j<jmax; j++) { 1644b59abe41SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 1645b59abe41SHong Zhang rtmp_ptr = rtmp + bj[j]*25; 1646b59abe41SHong Zhang u = ba + j*25; 1647936bb232SHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1648936bb232SHong Zhang rtmp_ptr[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1649936bb232SHong Zhang rtmp_ptr[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1650936bb232SHong Zhang rtmp_ptr[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1651936bb232SHong Zhang rtmp_ptr[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1652b59abe41SHong Zhang 1653936bb232SHong Zhang rtmp_ptr[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1654936bb232SHong Zhang rtmp_ptr[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1655936bb232SHong Zhang rtmp_ptr[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1656936bb232SHong Zhang rtmp_ptr[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1657936bb232SHong Zhang rtmp_ptr[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1658b59abe41SHong Zhang 1659936bb232SHong Zhang rtmp_ptr[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1660936bb232SHong Zhang rtmp_ptr[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1661936bb232SHong Zhang rtmp_ptr[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1662936bb232SHong Zhang rtmp_ptr[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1663936bb232SHong Zhang rtmp_ptr[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1664b59abe41SHong Zhang 1665936bb232SHong Zhang rtmp_ptr[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1666936bb232SHong Zhang rtmp_ptr[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1667936bb232SHong Zhang rtmp_ptr[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1668936bb232SHong Zhang rtmp_ptr[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1669936bb232SHong Zhang rtmp_ptr[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1670936bb232SHong Zhang 1671936bb232SHong Zhang rtmp_ptr[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1672936bb232SHong Zhang rtmp_ptr[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1673936bb232SHong Zhang rtmp_ptr[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1674936bb232SHong Zhang rtmp_ptr[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 16754f078a5fSHong Zhang rtmp_ptr[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1676b59abe41SHong Zhang } 1677b59abe41SHong Zhang 1678b59abe41SHong Zhang /* ... add i to row list for next nonzero entry */ 1679b59abe41SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1680b59abe41SHong Zhang j = bj[jmin]; 1681b59abe41SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 1682b59abe41SHong Zhang } 1683b59abe41SHong Zhang i = nexti; 1684b59abe41SHong Zhang } 1685b59abe41SHong Zhang 1686b59abe41SHong Zhang /* save nonzero entries in k-th row of U ... */ 1687b59abe41SHong Zhang 1688b59abe41SHong Zhang /* invert diagonal block */ 1689b59abe41SHong Zhang d = ba+k*25; 1690b59abe41SHong Zhang ierr = PetscMemcpy(d,dk,25*sizeof(MatScalar));CHKERRQ(ierr); 1691b59abe41SHong Zhang ierr = Kernel_A_gets_inverse_A_5(d);CHKERRQ(ierr); 1692b59abe41SHong Zhang 1693b59abe41SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1694b59abe41SHong Zhang if (jmin < jmax) { 1695b59abe41SHong Zhang for (j=jmin; j<jmax; j++){ 1696b59abe41SHong Zhang vj = bj[j]; /* block col. index of U */ 1697b59abe41SHong Zhang u = ba + j*25; 1698b59abe41SHong Zhang rtmp_ptr = rtmp + vj*25; 1699b59abe41SHong Zhang for (k1=0; k1<25; k1++){ 1700b59abe41SHong Zhang *u++ = *rtmp_ptr; 1701b59abe41SHong Zhang *rtmp_ptr++ = 0.0; 1702b59abe41SHong Zhang } 1703b59abe41SHong Zhang } 1704b59abe41SHong Zhang 1705b59abe41SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1706b59abe41SHong Zhang il[k] = jmin; 1707b59abe41SHong Zhang i = bj[jmin]; 1708b59abe41SHong Zhang jl[k] = jl[i]; jl[i] = k; 1709b59abe41SHong Zhang } 1710b59abe41SHong Zhang } 1711b59abe41SHong Zhang 1712b59abe41SHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 1713b59abe41SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1714b59abe41SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 1715b59abe41SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1716b59abe41SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 1717b59abe41SHong Zhang if (a->permute){ 1718b59abe41SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 1719b59abe41SHong Zhang } 1720b59abe41SHong Zhang 1721b59abe41SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 1722b59abe41SHong Zhang C->factor = FACTOR_CHOLESKY; 1723b59abe41SHong Zhang C->assembled = PETSC_TRUE; 1724b59abe41SHong Zhang C->preallocated = PETSC_TRUE; 1725b59abe41SHong Zhang PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 1726b59abe41SHong Zhang PetscFunctionReturn(0); 1727b59abe41SHong Zhang } 172849b5e25fSSatish Balay 172949b5e25fSSatish Balay /* 173049b5e25fSSatish Balay Version for when blocks are 5 by 5 Using natural ordering 173149b5e25fSSatish Balay */ 173249b5e25fSSatish Balay #undef __FUNC__ 17336f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering" 17346f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering(Mat A,Mat *B) 173549b5e25fSSatish Balay { 173649b5e25fSSatish Balay Mat C = *B; 1737ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1738ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1739*71149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1740ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1741ab27746eSHong Zhang MatScalar *u,*d,*rtmp,*rtmp_ptr; 174249b5e25fSSatish Balay 174349b5e25fSSatish Balay PetscFunctionBegin; 1744ab27746eSHong Zhang /* initialization */ 1745ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(25*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 1746ab27746eSHong Zhang ierr = PetscMemzero(rtmp,25*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1747ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 1748ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 1749ab27746eSHong Zhang for (i=0; i<mbs; i++) { 1750ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 175149b5e25fSSatish Balay } 1752ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(dk); 1753ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(25*sizeof(MatScalar));CHKPTRQ(uik); 1754ab27746eSHong Zhang 1755ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 1756ab27746eSHong Zhang 1757ab27746eSHong Zhang /* for each row k */ 1758ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 1759ab27746eSHong Zhang 1760ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 1761ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 1762ab27746eSHong Zhang if (jmin < jmax) { 1763ab27746eSHong Zhang ap = aa + jmin*25; 1764ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 1765ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 1766ab27746eSHong Zhang rtmp_ptr = rtmp + vj*25; 1767ab27746eSHong Zhang for (i=0; i<25; i++) *rtmp_ptr++ = *ap++; 176849b5e25fSSatish Balay } 176949b5e25fSSatish Balay } 1770ab27746eSHong Zhang 1771ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1772ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*25,25*sizeof(MatScalar));CHKERRQ(ierr); 1773ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1774ab27746eSHong Zhang 1775ab27746eSHong Zhang while (i < mbs){ 1776ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1777ab27746eSHong Zhang 1778ab27746eSHong Zhang /* compute multiplier */ 1779ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1780ab27746eSHong Zhang 1781ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1782ab27746eSHong Zhang d = ba + i*25; 1783ab27746eSHong Zhang u = ba + ili*25; 1784ab27746eSHong Zhang 1785ab27746eSHong Zhang uik[0] = -(d[0]*u[0] + d[5]*u[1] + d[10]*u[2] + d[15]*u[3] + d[20]*u[4]); 1786ab27746eSHong Zhang uik[1] = -(d[1]*u[0] + d[6]*u[1] + d[11]*u[2] + d[16]*u[3] + d[21]*u[4]); 1787ab27746eSHong Zhang uik[2] = -(d[2]*u[0] + d[7]*u[1] + d[12]*u[2] + d[17]*u[3] + d[22]*u[4]); 1788ab27746eSHong Zhang uik[3] = -(d[3]*u[0] + d[8]*u[1] + d[13]*u[2] + d[18]*u[3] + d[23]*u[4]); 1789ab27746eSHong Zhang uik[4] = -(d[4]*u[0] + d[9]*u[1] + d[14]*u[2] + d[19]*u[3] + d[24]*u[4]); 1790ab27746eSHong Zhang 1791ab27746eSHong Zhang uik[5] = -(d[0]*u[5] + d[5]*u[6] + d[10]*u[7] + d[15]*u[8] + d[20]*u[9]); 1792ab27746eSHong Zhang uik[6] = -(d[1]*u[5] + d[6]*u[6] + d[11]*u[7] + d[16]*u[8] + d[21]*u[9]); 1793ab27746eSHong Zhang uik[7] = -(d[2]*u[5] + d[7]*u[6] + d[12]*u[7] + d[17]*u[8] + d[22]*u[9]); 1794ab27746eSHong Zhang uik[8] = -(d[3]*u[5] + d[8]*u[6] + d[13]*u[7] + d[18]*u[8] + d[23]*u[9]); 1795ab27746eSHong Zhang uik[9] = -(d[4]*u[5] + d[9]*u[6] + d[14]*u[7] + d[19]*u[8] + d[24]*u[9]); 1796ab27746eSHong Zhang 1797ab27746eSHong Zhang uik[10]= -(d[0]*u[10] + d[5]*u[11] + d[10]*u[12] + d[15]*u[13] + d[20]*u[14]); 1798ab27746eSHong Zhang uik[11]= -(d[1]*u[10] + d[6]*u[11] + d[11]*u[12] + d[16]*u[13] + d[21]*u[14]); 1799ab27746eSHong Zhang uik[12]= -(d[2]*u[10] + d[7]*u[11] + d[12]*u[12] + d[17]*u[13] + d[22]*u[14]); 1800ab27746eSHong Zhang uik[13]= -(d[3]*u[10] + d[8]*u[11] + d[13]*u[12] + d[18]*u[13] + d[23]*u[14]); 1801ab27746eSHong Zhang uik[14]= -(d[4]*u[10] + d[9]*u[11] + d[14]*u[12] + d[19]*u[13] + d[24]*u[14]); 1802ab27746eSHong Zhang 1803ab27746eSHong Zhang uik[15]= -(d[0]*u[15] + d[5]*u[16] + d[10]*u[17] + d[15]*u[18] + d[20]*u[19]); 1804ab27746eSHong Zhang uik[16]= -(d[1]*u[15] + d[6]*u[16] + d[11]*u[17] + d[16]*u[18] + d[21]*u[19]); 1805ab27746eSHong Zhang uik[17]= -(d[2]*u[15] + d[7]*u[16] + d[12]*u[17] + d[17]*u[18] + d[22]*u[19]); 1806ab27746eSHong Zhang uik[18]= -(d[3]*u[15] + d[8]*u[16] + d[13]*u[17] + d[18]*u[18] + d[23]*u[19]); 1807ab27746eSHong Zhang uik[19]= -(d[4]*u[15] + d[9]*u[16] + d[14]*u[17] + d[19]*u[18] + d[24]*u[19]); 1808ab27746eSHong Zhang 1809ab27746eSHong Zhang uik[20]= -(d[0]*u[20] + d[5]*u[21] + d[10]*u[22] + d[15]*u[23] + d[20]*u[24]); 1810ab27746eSHong Zhang uik[21]= -(d[1]*u[20] + d[6]*u[21] + d[11]*u[22] + d[16]*u[23] + d[21]*u[24]); 1811ab27746eSHong Zhang uik[22]= -(d[2]*u[20] + d[7]*u[21] + d[12]*u[22] + d[17]*u[23] + d[22]*u[24]); 1812ab27746eSHong Zhang uik[23]= -(d[3]*u[20] + d[8]*u[21] + d[13]*u[22] + d[18]*u[23] + d[23]*u[24]); 1813ab27746eSHong Zhang uik[24]= -(d[4]*u[20] + d[9]*u[21] + d[14]*u[22] + d[19]*u[23] + d[24]*u[24]); 1814ab27746eSHong Zhang 1815ab27746eSHong Zhang 1816ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 1817ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1818ab27746eSHong Zhang dk[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1819ab27746eSHong Zhang dk[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1820ab27746eSHong Zhang dk[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1821ab27746eSHong Zhang dk[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1822ab27746eSHong Zhang 1823ab27746eSHong Zhang dk[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1824ab27746eSHong Zhang dk[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1825ab27746eSHong Zhang dk[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1826ab27746eSHong Zhang dk[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1827ab27746eSHong Zhang dk[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1828ab27746eSHong Zhang 1829ab27746eSHong Zhang dk[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1830ab27746eSHong Zhang dk[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1831ab27746eSHong Zhang dk[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1832ab27746eSHong Zhang dk[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1833ab27746eSHong Zhang dk[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1834ab27746eSHong Zhang 1835ab27746eSHong Zhang dk[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1836ab27746eSHong Zhang dk[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1837ab27746eSHong Zhang dk[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1838ab27746eSHong Zhang dk[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1839ab27746eSHong Zhang dk[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1840ab27746eSHong Zhang 1841ab27746eSHong Zhang dk[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1842ab27746eSHong Zhang dk[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1843ab27746eSHong Zhang dk[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1844ab27746eSHong Zhang dk[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 1845ab27746eSHong Zhang dk[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 1846ab27746eSHong Zhang 1847ab27746eSHong Zhang /* update -U(i,k) */ 1848ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*25,uik,25*sizeof(MatScalar));CHKERRQ(ierr); 1849ab27746eSHong Zhang 1850ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 1851ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 1852ab27746eSHong Zhang if (jmin < jmax){ 1853ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 1854ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 1855ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*25; 1856ab27746eSHong Zhang u = ba + j*25; 1857ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4]; 1858ab27746eSHong Zhang rtmp_ptr[1] += uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4]; 1859ab27746eSHong Zhang rtmp_ptr[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4]; 1860ab27746eSHong Zhang rtmp_ptr[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4]; 1861ab27746eSHong Zhang rtmp_ptr[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4]; 1862ab27746eSHong Zhang 1863ab27746eSHong Zhang rtmp_ptr[5] += uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9]; 1864ab27746eSHong Zhang rtmp_ptr[6] += uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]; 1865ab27746eSHong Zhang rtmp_ptr[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9]; 1866ab27746eSHong Zhang rtmp_ptr[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9]; 1867ab27746eSHong Zhang rtmp_ptr[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9]; 1868ab27746eSHong Zhang 1869ab27746eSHong Zhang rtmp_ptr[10] += uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14]; 1870ab27746eSHong Zhang rtmp_ptr[11] += uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14]; 1871ab27746eSHong Zhang rtmp_ptr[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]; 1872ab27746eSHong Zhang rtmp_ptr[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14]; 1873ab27746eSHong Zhang rtmp_ptr[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14]; 1874ab27746eSHong Zhang 1875ab27746eSHong Zhang rtmp_ptr[15] += uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19]; 1876ab27746eSHong Zhang rtmp_ptr[16] += uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19]; 1877ab27746eSHong Zhang rtmp_ptr[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19]; 1878ab27746eSHong Zhang rtmp_ptr[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19]; 1879ab27746eSHong Zhang rtmp_ptr[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19]; 1880ab27746eSHong Zhang 1881ab27746eSHong Zhang rtmp_ptr[20] += uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24]; 1882ab27746eSHong Zhang rtmp_ptr[21] += uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24]; 1883ab27746eSHong Zhang rtmp_ptr[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24]; 1884ab27746eSHong Zhang rtmp_ptr[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24]; 1885ab27746eSHong Zhang rtmp_ptr[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24]; 188649b5e25fSSatish Balay } 1887ab27746eSHong Zhang 1888ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 1889ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 1890ab27746eSHong Zhang j = bj[jmin]; 1891ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 189249b5e25fSSatish Balay } 1893ab27746eSHong Zhang i = nexti; 189449b5e25fSSatish Balay } 1895ab27746eSHong Zhang 1896ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 1897ab27746eSHong Zhang 189849b5e25fSSatish Balay /* invert diagonal block */ 1899ab27746eSHong Zhang d = ba+k*25; 1900ab27746eSHong Zhang ierr = PetscMemcpy(d,dk,25*sizeof(MatScalar));CHKERRQ(ierr); 1901ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_5(d);CHKERRQ(ierr); 1902ab27746eSHong Zhang 1903ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 1904ab27746eSHong Zhang if (jmin < jmax) { 1905ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 1906ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 1907ab27746eSHong Zhang u = ba + j*25; 1908ab27746eSHong Zhang rtmp_ptr = rtmp + vj*25; 1909ab27746eSHong Zhang for (k1=0; k1<25; k1++){ 1910ab27746eSHong Zhang *u++ = *rtmp_ptr; 1911ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 1912ab27746eSHong Zhang } 1913ab27746eSHong Zhang } 1914ab27746eSHong Zhang 1915ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 1916ab27746eSHong Zhang il[k] = jmin; 1917ab27746eSHong Zhang i = bj[jmin]; 1918ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 1919ab27746eSHong Zhang } 192049b5e25fSSatish Balay } 192149b5e25fSSatish Balay 192249b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 1923ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 1924ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 1925ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 1926ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 1927ab27746eSHong Zhang 1928ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 192949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 1930ab27746eSHong Zhang C->preallocated = PETSC_TRUE; 193149b5e25fSSatish Balay PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 193249b5e25fSSatish Balay PetscFunctionReturn(0); 193349b5e25fSSatish Balay } 193449b5e25fSSatish Balay 193549b5e25fSSatish Balay /* 193649b5e25fSSatish Balay Version for when blocks are 4 by 4 Using natural ordering 193749b5e25fSSatish Balay */ 193849b5e25fSSatish Balay #undef __FUNC__ 19396f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering" 19406f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering(Mat A,Mat *B) 194149b5e25fSSatish Balay { 194249b5e25fSSatish Balay Mat C = *B; 1943ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1944ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1945*71149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1946ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 1947ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 194849b5e25fSSatish Balay 194949b5e25fSSatish Balay PetscFunctionBegin; 1950ab27746eSHong Zhang /* initialization */ 1951ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(16*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 1952ab27746eSHong Zhang ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1953ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 1954ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 1955ab27746eSHong Zhang for (i=0; i<mbs; i++) { 1956ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 195749b5e25fSSatish Balay } 1958ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(dk); 1959ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(uik); 1960ab27746eSHong Zhang 1961ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 1962ab27746eSHong Zhang 1963ab27746eSHong Zhang /* for each row k */ 1964ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 1965ab27746eSHong Zhang 1966ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 1967ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 1968ab27746eSHong Zhang if (jmin < jmax) { 1969ab27746eSHong Zhang ap = aa + jmin*16; 1970ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 1971ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 1972ab27746eSHong Zhang rtmp_ptr = rtmp + vj*16; 1973ab27746eSHong Zhang for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 197449b5e25fSSatish Balay } 197549b5e25fSSatish Balay } 1976ab27746eSHong Zhang 1977ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 1978ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 1979ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 1980ab27746eSHong Zhang 1981ab27746eSHong Zhang while (i < mbs){ 1982ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 1983ab27746eSHong Zhang 1984ab27746eSHong Zhang /* compute multiplier */ 1985ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1986ab27746eSHong Zhang 1987ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 1988ab27746eSHong Zhang diag = ba + i*16; 1989ab27746eSHong Zhang u = ba + ili*16; 1990ab27746eSHong Zhang 1991ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 1992ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 1993ab27746eSHong Zhang uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 1994ab27746eSHong Zhang uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 1995ab27746eSHong Zhang 1996ab27746eSHong Zhang uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 1997ab27746eSHong Zhang uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 1998ab27746eSHong Zhang uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 1999ab27746eSHong Zhang uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 2000ab27746eSHong Zhang 2001ab27746eSHong Zhang uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 2002ab27746eSHong Zhang uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 2003ab27746eSHong Zhang uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 2004ab27746eSHong Zhang uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 2005ab27746eSHong Zhang 2006ab27746eSHong Zhang uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 2007ab27746eSHong Zhang uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 2008ab27746eSHong Zhang uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 2009ab27746eSHong Zhang uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 2010ab27746eSHong Zhang 2011ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2012ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2013ab27746eSHong Zhang dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2014ab27746eSHong Zhang dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2015ab27746eSHong Zhang dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2016ab27746eSHong Zhang 2017ab27746eSHong Zhang dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2018ab27746eSHong Zhang dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2019ab27746eSHong Zhang dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2020ab27746eSHong Zhang dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2021ab27746eSHong Zhang 2022ab27746eSHong Zhang dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2023ab27746eSHong Zhang dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2024ab27746eSHong Zhang dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2025ab27746eSHong Zhang dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2026ab27746eSHong Zhang 2027ab27746eSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2028ab27746eSHong Zhang dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2029ab27746eSHong Zhang dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2030ab27746eSHong Zhang dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2031ab27746eSHong Zhang 2032ab27746eSHong Zhang /* update -U(i,k) */ 2033ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 2034ab27746eSHong Zhang 2035ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2036ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2037ab27746eSHong Zhang if (jmin < jmax){ 2038ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2039ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2040ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*16; 2041ab27746eSHong Zhang u = ba + j*16; 2042ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2043ab27746eSHong Zhang rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2044ab27746eSHong Zhang rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2045ab27746eSHong Zhang rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2046ab27746eSHong Zhang 2047ab27746eSHong Zhang rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2048ab27746eSHong Zhang rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2049ab27746eSHong Zhang rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2050ab27746eSHong Zhang rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2051ab27746eSHong Zhang 2052ab27746eSHong Zhang rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2053ab27746eSHong Zhang rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2054ab27746eSHong Zhang rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2055ab27746eSHong Zhang rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2056ab27746eSHong Zhang 2057ab27746eSHong Zhang rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2058ab27746eSHong Zhang rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2059ab27746eSHong Zhang rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2060ab27746eSHong Zhang rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 206149b5e25fSSatish Balay } 2062ab27746eSHong Zhang 2063ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2064ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2065ab27746eSHong Zhang j = bj[jmin]; 2066ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 206749b5e25fSSatish Balay } 2068ab27746eSHong Zhang i = nexti; 206949b5e25fSSatish Balay } 2070ab27746eSHong Zhang 2071ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2072ab27746eSHong Zhang 207349b5e25fSSatish Balay /* invert diagonal block */ 2074ab27746eSHong Zhang diag = ba+k*16; 2075ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 2076ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 2077ab27746eSHong Zhang 2078ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2079ab27746eSHong Zhang if (jmin < jmax) { 2080ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2081ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2082ab27746eSHong Zhang u = ba + j*16; 2083ab27746eSHong Zhang rtmp_ptr = rtmp + vj*16; 2084ab27746eSHong Zhang for (k1=0; k1<16; k1++){ 2085ab27746eSHong Zhang *u++ = *rtmp_ptr; 2086ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2087ab27746eSHong Zhang } 2088ab27746eSHong Zhang } 2089ab27746eSHong Zhang 2090ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2091ab27746eSHong Zhang il[k] = jmin; 2092ab27746eSHong Zhang i = bj[jmin]; 2093ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2094ab27746eSHong Zhang } 209549b5e25fSSatish Balay } 209649b5e25fSSatish Balay 209749b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2098ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2099ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2100ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2101ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2102ab27746eSHong Zhang 2103ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 210449b5e25fSSatish Balay C->assembled = PETSC_TRUE; 21057a05c331SHong Zhang C->preallocated = PETSC_TRUE; 210649b5e25fSSatish Balay PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 210749b5e25fSSatish Balay PetscFunctionReturn(0); 210849b5e25fSSatish Balay } 210949b5e25fSSatish Balay 2110effa298cSHong Zhang /* Version for when blocks are 4 by 4 */ 2111effa298cSHong Zhang #undef __FUNC__ 2112effa298cSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4" 2113effa298cSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_4(Mat A,Mat *B) 2114effa298cSHong Zhang { 2115effa298cSHong Zhang Mat C = *B; 2116effa298cSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2117effa298cSHong Zhang IS perm = b->row; 2118effa298cSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2119effa298cSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2120effa298cSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2121effa298cSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 2122effa298cSHong Zhang 2123effa298cSHong Zhang PetscFunctionBegin; 2124effa298cSHong Zhang /* initialization */ 2125effa298cSHong Zhang rtmp = (MatScalar*)PetscMalloc(16*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2126effa298cSHong Zhang ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2127effa298cSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2128effa298cSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2129effa298cSHong Zhang for (i=0; i<mbs; i++) { 2130effa298cSHong Zhang jl[i] = mbs; il[0] = 0; 2131effa298cSHong Zhang } 2132effa298cSHong Zhang dk = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(dk); 2133effa298cSHong Zhang uik = (MatScalar*)PetscMalloc(16*sizeof(MatScalar));CHKPTRQ(uik); 2134effa298cSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2135effa298cSHong Zhang 2136effa298cSHong Zhang /* check permutation */ 2137effa298cSHong Zhang if (!a->permute){ 2138effa298cSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2139effa298cSHong Zhang } else { 2140effa298cSHong Zhang ai = a->inew; aj = a->jnew; 2141effa298cSHong Zhang aa = (MatScalar*)PetscMalloc(16*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2142effa298cSHong Zhang ierr = PetscMemcpy(aa,a->a,16*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2143effa298cSHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2144effa298cSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2145effa298cSHong Zhang 2146effa298cSHong Zhang for (i=0; i<mbs; i++){ 2147effa298cSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2148effa298cSHong Zhang for (j=jmin; j<jmax; j++){ 2149effa298cSHong Zhang while (a2anew[j] != j){ 2150effa298cSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2151effa298cSHong Zhang for (k1=0; k1<16; k1++){ 2152effa298cSHong Zhang dk[k1] = aa[k*16+k1]; 2153effa298cSHong Zhang aa[k*16+k1] = aa[j*16+k1]; 2154effa298cSHong Zhang aa[j*16+k1] = dk[k1]; 2155effa298cSHong Zhang } 2156effa298cSHong Zhang } 2157effa298cSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2158effa298cSHong Zhang if (i > aj[j]){ 2159effa298cSHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2160effa298cSHong Zhang ap = aa + j*16; /* ptr to the beginning of j-th block of aa */ 2161effa298cSHong Zhang for (k=0; k<16; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2162effa298cSHong Zhang for (k=0; k<4; k++){ /* j-th block of aa <- dk^T */ 2163effa298cSHong Zhang for (k1=0; k1<4; k1++) *ap++ = dk[k + 4*k1]; 2164effa298cSHong Zhang } 2165effa298cSHong Zhang } 2166effa298cSHong Zhang } 2167effa298cSHong Zhang } 2168effa298cSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2169effa298cSHong Zhang } 2170effa298cSHong Zhang 2171effa298cSHong Zhang /* for each row k */ 2172effa298cSHong Zhang for (k = 0; k<mbs; k++){ 2173effa298cSHong Zhang 2174effa298cSHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2175effa298cSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 2176effa298cSHong Zhang if (jmin < jmax) { 2177effa298cSHong Zhang ap = aa + jmin*16; 2178effa298cSHong Zhang for (j = jmin; j < jmax; j++){ 2179effa298cSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2180effa298cSHong Zhang rtmp_ptr = rtmp + vj*16; 2181effa298cSHong Zhang for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 2182effa298cSHong Zhang } 2183effa298cSHong Zhang } 2184effa298cSHong Zhang 2185effa298cSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2186effa298cSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 2187effa298cSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2188effa298cSHong Zhang 2189effa298cSHong Zhang while (i < mbs){ 2190effa298cSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2191effa298cSHong Zhang 2192effa298cSHong Zhang /* compute multiplier */ 2193effa298cSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2194effa298cSHong Zhang 2195effa298cSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2196effa298cSHong Zhang diag = ba + i*16; 2197effa298cSHong Zhang u = ba + ili*16; 2198effa298cSHong Zhang 2199effa298cSHong Zhang uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 2200effa298cSHong Zhang uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 2201effa298cSHong Zhang uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 2202effa298cSHong Zhang uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 2203effa298cSHong Zhang 2204effa298cSHong Zhang uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 2205effa298cSHong Zhang uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 2206effa298cSHong Zhang uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 2207effa298cSHong Zhang uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 2208effa298cSHong Zhang 2209effa298cSHong Zhang uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 2210effa298cSHong Zhang uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 2211effa298cSHong Zhang uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 2212effa298cSHong Zhang uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 2213effa298cSHong Zhang 2214effa298cSHong Zhang uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 2215effa298cSHong Zhang uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 2216effa298cSHong Zhang uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 2217effa298cSHong Zhang uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 2218effa298cSHong Zhang 2219effa298cSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2220effa298cSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2221effa298cSHong Zhang dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2222effa298cSHong Zhang dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2223effa298cSHong Zhang dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2224effa298cSHong Zhang 2225effa298cSHong Zhang dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2226effa298cSHong Zhang dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2227effa298cSHong Zhang dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2228effa298cSHong Zhang dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2229effa298cSHong Zhang 2230effa298cSHong Zhang dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2231effa298cSHong Zhang dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2232effa298cSHong Zhang dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2233effa298cSHong Zhang dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2234effa298cSHong Zhang 2235effa298cSHong Zhang dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2236effa298cSHong Zhang dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2237effa298cSHong Zhang dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2238effa298cSHong Zhang dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2239effa298cSHong Zhang 2240effa298cSHong Zhang /* update -U(i,k) */ 2241effa298cSHong Zhang ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 2242effa298cSHong Zhang 2243effa298cSHong Zhang /* add multiple of row i to k-th row ... */ 2244effa298cSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2245effa298cSHong Zhang if (jmin < jmax){ 2246effa298cSHong Zhang for (j=jmin; j<jmax; j++) { 2247effa298cSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2248effa298cSHong Zhang rtmp_ptr = rtmp + bj[j]*16; 2249effa298cSHong Zhang u = ba + j*16; 2250effa298cSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 2251effa298cSHong Zhang rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 2252effa298cSHong Zhang rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 2253effa298cSHong Zhang rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 2254effa298cSHong Zhang 2255effa298cSHong Zhang rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 2256effa298cSHong Zhang rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 2257effa298cSHong Zhang rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 2258effa298cSHong Zhang rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 2259effa298cSHong Zhang 2260effa298cSHong Zhang rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 2261effa298cSHong Zhang rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 2262effa298cSHong Zhang rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 2263effa298cSHong Zhang rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 2264effa298cSHong Zhang 2265effa298cSHong Zhang rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 2266effa298cSHong Zhang rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 2267effa298cSHong Zhang rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 2268effa298cSHong Zhang rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 2269effa298cSHong Zhang } 2270effa298cSHong Zhang 2271effa298cSHong Zhang /* ... add i to row list for next nonzero entry */ 2272effa298cSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2273effa298cSHong Zhang j = bj[jmin]; 2274effa298cSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 2275effa298cSHong Zhang } 2276effa298cSHong Zhang i = nexti; 2277effa298cSHong Zhang } 2278effa298cSHong Zhang 2279effa298cSHong Zhang /* save nonzero entries in k-th row of U ... */ 2280effa298cSHong Zhang 2281effa298cSHong Zhang /* invert diagonal block */ 2282effa298cSHong Zhang diag = ba+k*16; 2283effa298cSHong Zhang ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 2284effa298cSHong Zhang ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 2285effa298cSHong Zhang 2286effa298cSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2287effa298cSHong Zhang if (jmin < jmax) { 2288effa298cSHong Zhang for (j=jmin; j<jmax; j++){ 2289effa298cSHong Zhang vj = bj[j]; /* block col. index of U */ 2290effa298cSHong Zhang u = ba + j*16; 2291effa298cSHong Zhang rtmp_ptr = rtmp + vj*16; 2292effa298cSHong Zhang for (k1=0; k1<16; k1++){ 2293effa298cSHong Zhang *u++ = *rtmp_ptr; 2294effa298cSHong Zhang *rtmp_ptr++ = 0.0; 2295effa298cSHong Zhang } 2296effa298cSHong Zhang } 2297effa298cSHong Zhang 2298effa298cSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2299effa298cSHong Zhang il[k] = jmin; 2300effa298cSHong Zhang i = bj[jmin]; 2301effa298cSHong Zhang jl[k] = jl[i]; jl[i] = k; 2302effa298cSHong Zhang } 2303effa298cSHong Zhang } 2304effa298cSHong Zhang 2305effa298cSHong Zhang ierr = PetscFree(rtmp);CHKERRQ(ierr); 2306effa298cSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2307effa298cSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2308effa298cSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2309effa298cSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2310effa298cSHong Zhang if (a->permute){ 2311effa298cSHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 2312effa298cSHong Zhang } 2313effa298cSHong Zhang 2314effa298cSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 2315effa298cSHong Zhang C->factor = FACTOR_CHOLESKY; 2316effa298cSHong Zhang C->assembled = PETSC_TRUE; 2317effa298cSHong Zhang C->preallocated = PETSC_TRUE; 2318effa298cSHong Zhang PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 2319effa298cSHong Zhang PetscFunctionReturn(0); 2320effa298cSHong Zhang } 2321effa298cSHong Zhang 2322a1723e09SHong Zhang /* Version for when blocks are 3 by 3 */ 232349b5e25fSSatish Balay #undef __FUNC__ 23246f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_3" 23256f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_3(Mat A,Mat *B) 232649b5e25fSSatish Balay { 232749b5e25fSSatish Balay Mat C = *B; 2328a1723e09SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2329a1723e09SHong Zhang IS perm = b->row; 2330a1723e09SHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2331a1723e09SHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2332a1723e09SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2333a1723e09SHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 233449b5e25fSSatish Balay 233549b5e25fSSatish Balay PetscFunctionBegin; 2336a1723e09SHong Zhang /* initialization */ 2337a1723e09SHong Zhang rtmp = (MatScalar*)PetscMalloc(9*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2338a1723e09SHong Zhang ierr = PetscMemzero(rtmp,9*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2339a1723e09SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2340a1723e09SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2341a1723e09SHong Zhang for (i=0; i<mbs; i++) { 2342a1723e09SHong Zhang jl[i] = mbs; il[0] = 0; 2343a1723e09SHong Zhang } 2344a1723e09SHong Zhang dk = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(dk); 2345a1723e09SHong Zhang uik = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(uik); 2346a1723e09SHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 234749b5e25fSSatish Balay 2348a1723e09SHong Zhang /* check permutation */ 2349a1723e09SHong Zhang if (!a->permute){ 2350a1723e09SHong Zhang ai = a->i; aj = a->j; aa = a->a; 2351a1723e09SHong Zhang } else { 2352a1723e09SHong Zhang ai = a->inew; aj = a->jnew; 2353a1723e09SHong Zhang aa = (MatScalar*)PetscMalloc(9*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2354a1723e09SHong Zhang ierr = PetscMemcpy(aa,a->a,9*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2355a1723e09SHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2356a1723e09SHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 235749b5e25fSSatish Balay 2358a1723e09SHong Zhang for (i=0; i<mbs; i++){ 2359a1723e09SHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2360a1723e09SHong Zhang for (j=jmin; j<jmax; j++){ 2361a1723e09SHong Zhang while (a2anew[j] != j){ 2362a1723e09SHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2363a1723e09SHong Zhang for (k1=0; k1<9; k1++){ 2364a1723e09SHong Zhang dk[k1] = aa[k*9+k1]; 2365a1723e09SHong Zhang aa[k*9+k1] = aa[j*9+k1]; 2366a1723e09SHong Zhang aa[j*9+k1] = dk[k1]; 2367a1723e09SHong Zhang } 2368a1723e09SHong Zhang } 2369a1723e09SHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2370a1723e09SHong Zhang if (i > aj[j]){ 237147c40530SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2372a1723e09SHong Zhang ap = aa + j*9; /* ptr to the beginning of j-th block of aa */ 2373a1723e09SHong Zhang for (k=0; k<9; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 2374a1723e09SHong Zhang for (k=0; k<3; k++){ /* j-th block of aa <- dk^T */ 237547c40530SHong Zhang for (k1=0; k1<3; k1++) *ap++ = dk[k + 3*k1]; 2376a1723e09SHong Zhang } 2377a1723e09SHong Zhang } 2378a1723e09SHong Zhang } 2379a1723e09SHong Zhang } 2380a1723e09SHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2381a1723e09SHong Zhang } 238249b5e25fSSatish Balay 2383a1723e09SHong Zhang /* for each row k */ 2384a1723e09SHong Zhang for (k = 0; k<mbs; k++){ 238549b5e25fSSatish Balay 2386a1723e09SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2387a1723e09SHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 2388a1723e09SHong Zhang if (jmin < jmax) { 2389a1723e09SHong Zhang ap = aa + jmin*9; 2390a1723e09SHong Zhang for (j = jmin; j < jmax; j++){ 2391a1723e09SHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2392a1723e09SHong Zhang rtmp_ptr = rtmp + vj*9; 2393a1723e09SHong Zhang for (i=0; i<9; i++) *rtmp_ptr++ = *ap++; 2394a1723e09SHong Zhang } 2395a1723e09SHong Zhang } 239649b5e25fSSatish Balay 2397a1723e09SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2398a1723e09SHong Zhang ierr = PetscMemcpy(dk,rtmp+k*9,9*sizeof(MatScalar));CHKERRQ(ierr); 2399a1723e09SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2400a1723e09SHong Zhang 2401a1723e09SHong Zhang while (i < mbs){ 2402a1723e09SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2403a1723e09SHong Zhang 2404a1723e09SHong Zhang /* compute multiplier */ 2405a1723e09SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2406a1723e09SHong Zhang 2407a1723e09SHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2408a1723e09SHong Zhang diag = ba + i*9; 2409a1723e09SHong Zhang u = ba + ili*9; 2410a1723e09SHong Zhang 2411a1723e09SHong Zhang uik[0] = -(diag[0]*u[0] + diag[3]*u[1] + diag[6]*u[2]); 2412a1723e09SHong Zhang uik[1] = -(diag[1]*u[0] + diag[4]*u[1] + diag[7]*u[2]); 2413a1723e09SHong Zhang uik[2] = -(diag[2]*u[0] + diag[5]*u[1] + diag[8]*u[2]); 2414a1723e09SHong Zhang 2415a1723e09SHong Zhang uik[3] = -(diag[0]*u[3] + diag[3]*u[4] + diag[6]*u[5]); 2416a1723e09SHong Zhang uik[4] = -(diag[1]*u[3] + diag[4]*u[4] + diag[7]*u[5]); 2417a1723e09SHong Zhang uik[5] = -(diag[2]*u[3] + diag[5]*u[4] + diag[8]*u[5]); 2418a1723e09SHong Zhang 2419a1723e09SHong Zhang uik[6] = -(diag[0]*u[6] + diag[3]*u[7] + diag[6]*u[8]); 2420a1723e09SHong Zhang uik[7] = -(diag[1]*u[6] + diag[4]*u[7] + diag[7]*u[8]); 2421a1723e09SHong Zhang uik[8] = -(diag[2]*u[6] + diag[5]*u[7] + diag[8]*u[8]); 2422a1723e09SHong Zhang 2423a1723e09SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2424a1723e09SHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2425a1723e09SHong Zhang dk[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2426a1723e09SHong Zhang dk[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2427a1723e09SHong Zhang 2428a1723e09SHong Zhang dk[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2429a1723e09SHong Zhang dk[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2430a1723e09SHong Zhang dk[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2431a1723e09SHong Zhang 2432a1723e09SHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2433a1723e09SHong Zhang dk[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2434a1723e09SHong Zhang dk[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 2435a1723e09SHong Zhang 2436a1723e09SHong Zhang /* update -U(i,k) */ 2437a1723e09SHong Zhang ierr = PetscMemcpy(ba+ili*9,uik,9*sizeof(MatScalar));CHKERRQ(ierr); 2438a1723e09SHong Zhang 2439a1723e09SHong Zhang /* add multiple of row i to k-th row ... */ 2440a1723e09SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2441a1723e09SHong Zhang if (jmin < jmax){ 2442a1723e09SHong Zhang for (j=jmin; j<jmax; j++) { 2443a1723e09SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2444a1723e09SHong Zhang rtmp_ptr = rtmp + bj[j]*9; 2445a1723e09SHong Zhang u = ba + j*9; 2446a1723e09SHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2447a1723e09SHong Zhang rtmp_ptr[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2448a1723e09SHong Zhang rtmp_ptr[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2449a1723e09SHong Zhang 2450a1723e09SHong Zhang rtmp_ptr[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2451a1723e09SHong Zhang rtmp_ptr[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2452a1723e09SHong Zhang rtmp_ptr[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2453a1723e09SHong Zhang 2454a1723e09SHong Zhang rtmp_ptr[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2455a1723e09SHong Zhang rtmp_ptr[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2456a1723e09SHong Zhang rtmp_ptr[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 245749b5e25fSSatish Balay } 2458a1723e09SHong Zhang 2459a1723e09SHong Zhang /* ... add i to row list for next nonzero entry */ 2460a1723e09SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2461a1723e09SHong Zhang j = bj[jmin]; 2462a1723e09SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 246349b5e25fSSatish Balay } 2464a1723e09SHong Zhang i = nexti; 246549b5e25fSSatish Balay } 2466a1723e09SHong Zhang 2467a1723e09SHong Zhang /* save nonzero entries in k-th row of U ... */ 2468a1723e09SHong Zhang 246949b5e25fSSatish Balay /* invert diagonal block */ 2470a1723e09SHong Zhang diag = ba+k*9; 2471a1723e09SHong Zhang ierr = PetscMemcpy(diag,dk,9*sizeof(MatScalar));CHKERRQ(ierr); 2472a1723e09SHong Zhang ierr = Kernel_A_gets_inverse_A_3(diag);CHKERRQ(ierr); 2473a1723e09SHong Zhang 2474a1723e09SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2475a1723e09SHong Zhang if (jmin < jmax) { 2476a1723e09SHong Zhang for (j=jmin; j<jmax; j++){ 2477a1723e09SHong Zhang vj = bj[j]; /* block col. index of U */ 2478a1723e09SHong Zhang u = ba + j*9; 2479a1723e09SHong Zhang rtmp_ptr = rtmp + vj*9; 2480a1723e09SHong Zhang for (k1=0; k1<9; k1++){ 2481a1723e09SHong Zhang *u++ = *rtmp_ptr; 2482a1723e09SHong Zhang *rtmp_ptr++ = 0.0; 2483a1723e09SHong Zhang } 2484a1723e09SHong Zhang } 2485a1723e09SHong Zhang 2486a1723e09SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2487a1723e09SHong Zhang il[k] = jmin; 2488a1723e09SHong Zhang i = bj[jmin]; 2489a1723e09SHong Zhang jl[k] = jl[i]; jl[i] = k; 2490a1723e09SHong Zhang } 249149b5e25fSSatish Balay } 249249b5e25fSSatish Balay 249349b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2494a1723e09SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2495a1723e09SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2496a1723e09SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2497a1723e09SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2498a1723e09SHong Zhang if (a->permute){ 2499a1723e09SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 2500a1723e09SHong Zhang } 2501a1723e09SHong Zhang 2502a1723e09SHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 2503a1723e09SHong Zhang C->factor = FACTOR_CHOLESKY; 250449b5e25fSSatish Balay C->assembled = PETSC_TRUE; 25057a05c331SHong Zhang C->preallocated = PETSC_TRUE; 2506effa298cSHong Zhang PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 250749b5e25fSSatish Balay PetscFunctionReturn(0); 250849b5e25fSSatish Balay } 2509a1723e09SHong Zhang 251049b5e25fSSatish Balay /* 251149b5e25fSSatish Balay Version for when blocks are 3 by 3 Using natural ordering 251249b5e25fSSatish Balay */ 251349b5e25fSSatish Balay #undef __FUNC__ 25146f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering" 25156f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering(Mat A,Mat *B) 251649b5e25fSSatish Balay { 251749b5e25fSSatish Balay Mat C = *B; 2518ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2519*71149678SHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2520*71149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2521ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2522ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 252349b5e25fSSatish Balay 252449b5e25fSSatish Balay PetscFunctionBegin; 2525ab27746eSHong Zhang /* initialization */ 2526ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(9*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2527ab27746eSHong Zhang ierr = PetscMemzero(rtmp,9*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2528ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2529ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2530ab27746eSHong Zhang for (i=0; i<mbs; i++) { 2531ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 253249b5e25fSSatish Balay } 2533ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(dk); 2534ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(9*sizeof(MatScalar));CHKPTRQ(uik); 253549b5e25fSSatish Balay 2536ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 253749b5e25fSSatish Balay 2538ab27746eSHong Zhang /* for each row k */ 2539ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 254049b5e25fSSatish Balay 2541ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 2542ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 2543ab27746eSHong Zhang if (jmin < jmax) { 2544ab27746eSHong Zhang ap = aa + jmin*9; 2545ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 2546ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 2547ab27746eSHong Zhang rtmp_ptr = rtmp + vj*9; 2548ab27746eSHong Zhang for (i=0; i<9; i++) *rtmp_ptr++ = *ap++; 2549ab27746eSHong Zhang } 2550ab27746eSHong Zhang } 255149b5e25fSSatish Balay 2552ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2553ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*9,9*sizeof(MatScalar));CHKERRQ(ierr); 2554ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 255549b5e25fSSatish Balay 2556ab27746eSHong Zhang while (i < mbs){ 2557ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2558ab27746eSHong Zhang 2559ab27746eSHong Zhang /* compute multiplier */ 2560ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2561ab27746eSHong Zhang 2562ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k) */ 2563ab27746eSHong Zhang diag = ba + i*9; 2564ab27746eSHong Zhang u = ba + ili*9; 2565ab27746eSHong Zhang 2566ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[3]*u[1] + diag[6]*u[2]); 2567ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[4]*u[1] + diag[7]*u[2]); 2568ab27746eSHong Zhang uik[2] = -(diag[2]*u[0] + diag[5]*u[1] + diag[8]*u[2]); 2569ab27746eSHong Zhang 2570ab27746eSHong Zhang uik[3] = -(diag[0]*u[3] + diag[3]*u[4] + diag[6]*u[5]); 2571ab27746eSHong Zhang uik[4] = -(diag[1]*u[3] + diag[4]*u[4] + diag[7]*u[5]); 2572ab27746eSHong Zhang uik[5] = -(diag[2]*u[3] + diag[5]*u[4] + diag[8]*u[5]); 2573ab27746eSHong Zhang 2574ab27746eSHong Zhang uik[6] = -(diag[0]*u[6] + diag[3]*u[7] + diag[6]*u[8]); 2575ab27746eSHong Zhang uik[7] = -(diag[1]*u[6] + diag[4]*u[7] + diag[7]*u[8]); 2576ab27746eSHong Zhang uik[8] = -(diag[2]*u[6] + diag[5]*u[7] + diag[8]*u[8]); 2577ab27746eSHong Zhang 2578ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2579ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2580ab27746eSHong Zhang dk[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2581ab27746eSHong Zhang dk[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2582ab27746eSHong Zhang 2583ab27746eSHong Zhang dk[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2584ab27746eSHong Zhang dk[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2585ab27746eSHong Zhang dk[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2586ab27746eSHong Zhang 2587ab27746eSHong Zhang dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2588ab27746eSHong Zhang dk[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2589ab27746eSHong Zhang dk[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 2590ab27746eSHong Zhang 2591ab27746eSHong Zhang /* update -U(i,k) */ 2592ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*9,uik,9*sizeof(MatScalar));CHKERRQ(ierr); 2593ab27746eSHong Zhang 2594ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2595ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2596ab27746eSHong Zhang if (jmin < jmax){ 2597ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2598ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j) */ 2599ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*9; 2600ab27746eSHong Zhang u = ba + j*9; 2601ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2]; 2602ab27746eSHong Zhang rtmp_ptr[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2]; 2603ab27746eSHong Zhang rtmp_ptr[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2]; 2604ab27746eSHong Zhang 2605ab27746eSHong Zhang rtmp_ptr[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5]; 2606ab27746eSHong Zhang rtmp_ptr[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 2607ab27746eSHong Zhang rtmp_ptr[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5]; 2608ab27746eSHong Zhang 2609ab27746eSHong Zhang rtmp_ptr[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8]; 2610ab27746eSHong Zhang rtmp_ptr[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8]; 2611ab27746eSHong Zhang rtmp_ptr[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8]; 261249b5e25fSSatish Balay } 2613ab27746eSHong Zhang 2614ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2615ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2616ab27746eSHong Zhang j = bj[jmin]; 2617ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 261849b5e25fSSatish Balay } 2619ab27746eSHong Zhang i = nexti; 262049b5e25fSSatish Balay } 2621ab27746eSHong Zhang 2622ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2623ab27746eSHong Zhang 262449b5e25fSSatish Balay /* invert diagonal block */ 2625ab27746eSHong Zhang diag = ba+k*9; 2626ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,9*sizeof(MatScalar));CHKERRQ(ierr); 2627ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_3(diag);CHKERRQ(ierr); 2628ab27746eSHong Zhang 2629ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2630ab27746eSHong Zhang if (jmin < jmax) { 2631ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2632ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2633ab27746eSHong Zhang u = ba + j*9; 2634ab27746eSHong Zhang rtmp_ptr = rtmp + vj*9; 2635ab27746eSHong Zhang for (k1=0; k1<9; k1++){ 2636ab27746eSHong Zhang *u++ = *rtmp_ptr; 2637ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2638ab27746eSHong Zhang } 2639ab27746eSHong Zhang } 2640ab27746eSHong Zhang 2641ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2642ab27746eSHong Zhang il[k] = jmin; 2643ab27746eSHong Zhang i = bj[jmin]; 2644ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2645ab27746eSHong Zhang } 264649b5e25fSSatish Balay } 264749b5e25fSSatish Balay 264849b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2649ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2650ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2651ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2652ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2653ab27746eSHong Zhang 2654effa298cSHong Zhang C->factor = FACTOR_CHOLESKY; 265549b5e25fSSatish Balay C->assembled = PETSC_TRUE; 26565c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 265749b5e25fSSatish Balay PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 265849b5e25fSSatish Balay PetscFunctionReturn(0); 265949b5e25fSSatish Balay } 266049b5e25fSSatish Balay 266149b5e25fSSatish Balay /* 2662fcf159c0SHong Zhang Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP. 2663cc0c071aSHong Zhang Version for blocks 2 by 2. 266449b5e25fSSatish Balay */ 266549b5e25fSSatish Balay #undef __FUNC__ 26666f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_2" 26676f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_2(Mat A,Mat *B) 266849b5e25fSSatish Balay { 266949b5e25fSSatish Balay Mat C = *B; 267091602c66SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2671cc0c071aSHong Zhang IS perm = b->row; 2672cc0c071aSHong Zhang int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2673cc0c071aSHong Zhang int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2674a1723e09SHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2675cc0c071aSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 267649b5e25fSSatish Balay 267749b5e25fSSatish Balay PetscFunctionBegin; 2678103535cfSHong Zhang /* printf("called Numeric_SeqSBAIJ_2\n"); */ 267991602c66SHong Zhang /* initialization */ 268091602c66SHong Zhang /* il and jl record the first nonzero element in each row of the accessing 268191602c66SHong Zhang window U(0:k, k:mbs-1). 268291602c66SHong Zhang jl: list of rows to be added to uneliminated rows 268391602c66SHong Zhang i>= k: jl(i) is the first row to be added to row i 268491602c66SHong Zhang i< k: jl(i) is the row following row i in some list of rows 268591602c66SHong Zhang jl(i) = mbs indicates the end of a list 268691602c66SHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 268791602c66SHong Zhang row i of U */ 2688cc0c071aSHong Zhang rtmp = (MatScalar*)PetscMalloc(4*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2689cc0c071aSHong Zhang ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr); 269091602c66SHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 269191602c66SHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 269291602c66SHong Zhang for (i=0; i<mbs; i++) { 26933845f261SHong Zhang jl[i] = mbs; il[0] = 0; 269491602c66SHong Zhang } 2695cc0c071aSHong Zhang dk = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(dk); 2696cc0c071aSHong Zhang uik = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(uik); 2697cc0c071aSHong Zhang ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 2698cc0c071aSHong Zhang 2699cc0c071aSHong Zhang /* check permutation */ 2700cc0c071aSHong Zhang if (!a->permute){ 2701cc0c071aSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2702cc0c071aSHong Zhang } else { 2703cc0c071aSHong Zhang ai = a->inew; aj = a->jnew; 2704cc0c071aSHong Zhang aa = (MatScalar*)PetscMalloc(4*ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 2705cc0c071aSHong Zhang ierr = PetscMemcpy(aa,a->a,4*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 2706cc0c071aSHong Zhang a2anew = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(a2anew); 2707cc0c071aSHong Zhang ierr= PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 2708cc0c071aSHong Zhang 2709cc0c071aSHong Zhang for (i=0; i<mbs; i++){ 2710cc0c071aSHong Zhang jmin = ai[i]; jmax = ai[i+1]; 2711cc0c071aSHong Zhang for (j=jmin; j<jmax; j++){ 2712cc0c071aSHong Zhang while (a2anew[j] != j){ 2713cc0c071aSHong Zhang k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 2714cc0c071aSHong Zhang for (k1=0; k1<4; k1++){ 2715cc0c071aSHong Zhang dk[k1] = aa[k*4+k1]; 2716cc0c071aSHong Zhang aa[k*4+k1] = aa[j*4+k1]; 2717cc0c071aSHong Zhang aa[j*4+k1] = dk[k1]; 2718cc0c071aSHong Zhang } 2719cc0c071aSHong Zhang } 2720cc0c071aSHong Zhang /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 2721cc0c071aSHong Zhang if (i > aj[j]){ 2722a1723e09SHong Zhang /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 2723cc0c071aSHong Zhang ap = aa + j*4; /* ptr to the beginning of the block */ 2724cc0c071aSHong Zhang dk[1] = ap[1]; /* swap ap[1] and ap[2] */ 2725cc0c071aSHong Zhang ap[1] = ap[2]; 2726cc0c071aSHong Zhang ap[2] = dk[1]; 2727cc0c071aSHong Zhang } 2728cc0c071aSHong Zhang } 2729cc0c071aSHong Zhang } 2730cc0c071aSHong Zhang ierr = PetscFree(a2anew);CHKERRA(ierr); 2731cc0c071aSHong Zhang } 27323845f261SHong Zhang 273391602c66SHong Zhang /* for each row k */ 273491602c66SHong Zhang for (k = 0; k<mbs; k++){ 273591602c66SHong Zhang 273691602c66SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 2737cc0c071aSHong Zhang jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 273891602c66SHong Zhang if (jmin < jmax) { 2739cc0c071aSHong Zhang ap = aa + jmin*4; 274091602c66SHong Zhang for (j = jmin; j < jmax; j++){ 2741cc0c071aSHong Zhang vj = perm_ptr[aj[j]]; /* block col. index */ 2742cc0c071aSHong Zhang rtmp_ptr = rtmp + vj*4; 2743cc0c071aSHong Zhang for (i=0; i<4; i++) *rtmp_ptr++ = *ap++; 274491602c66SHong Zhang } 274591602c66SHong Zhang } 274691602c66SHong Zhang 274791602c66SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2748cc0c071aSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr); 274991602c66SHong Zhang i = jl[k]; /* first row to be added to k_th row */ 275091602c66SHong Zhang 275191602c66SHong Zhang while (i < mbs){ 275291602c66SHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 275391602c66SHong Zhang 27543845f261SHong Zhang /* compute multiplier */ 275591602c66SHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 27563845f261SHong Zhang 27573845f261SHong Zhang /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */ 2758cc0c071aSHong Zhang diag = ba + i*4; 2759cc0c071aSHong Zhang u = ba + ili*4; 2760cc0c071aSHong Zhang uik[0] = -(diag[0]*u[0] + diag[2]*u[1]); 2761cc0c071aSHong Zhang uik[1] = -(diag[1]*u[0] + diag[3]*u[1]); 2762cc0c071aSHong Zhang uik[2] = -(diag[0]*u[2] + diag[2]*u[3]); 2763cc0c071aSHong Zhang uik[3] = -(diag[1]*u[2] + diag[3]*u[3]); 27643845f261SHong Zhang 27653845f261SHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */ 2766cc0c071aSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1]; 2767cc0c071aSHong Zhang dk[1] += uik[2]*u[0] + uik[3]*u[1]; 2768cc0c071aSHong Zhang dk[2] += uik[0]*u[2] + uik[1]*u[3]; 2769cc0c071aSHong Zhang dk[3] += uik[2]*u[2] + uik[3]*u[3]; 27703845f261SHong Zhang 27713845f261SHong Zhang /* update -U(i,k): ba[ili] = uik */ 2772cc0c071aSHong Zhang ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr); 277391602c66SHong Zhang 277491602c66SHong Zhang /* add multiple of row i to k-th row ... */ 277591602c66SHong Zhang jmin = ili + 1; jmax = bi[i+1]; 277691602c66SHong Zhang if (jmin < jmax){ 27773845f261SHong Zhang for (j=jmin; j<jmax; j++) { 27783845f261SHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */ 2779cc0c071aSHong Zhang rtmp_ptr = rtmp + bj[j]*4; 2780cc0c071aSHong Zhang u = ba + j*4; 2781cc0c071aSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1]; 2782cc0c071aSHong Zhang rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1]; 2783cc0c071aSHong Zhang rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3]; 2784cc0c071aSHong Zhang rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3]; 27853845f261SHong Zhang } 27863845f261SHong Zhang 278791602c66SHong Zhang /* ... add i to row list for next nonzero entry */ 278891602c66SHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 278991602c66SHong Zhang j = bj[jmin]; 279091602c66SHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 279191602c66SHong Zhang } 2792a1723e09SHong Zhang i = nexti; 279391602c66SHong Zhang } 2794cc0c071aSHong Zhang 279591602c66SHong Zhang /* save nonzero entries in k-th row of U ... */ 27963845f261SHong Zhang 2797cc0c071aSHong Zhang /* invert diagonal block */ 2798cc0c071aSHong Zhang diag = ba+k*4; 2799cc0c071aSHong Zhang ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr); 28003845f261SHong Zhang ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr); 28013845f261SHong Zhang 280291602c66SHong Zhang jmin = bi[k]; jmax = bi[k+1]; 280391602c66SHong Zhang if (jmin < jmax) { 280491602c66SHong Zhang for (j=jmin; j<jmax; j++){ 2805cc0c071aSHong Zhang vj = bj[j]; /* block col. index of U */ 2806cc0c071aSHong Zhang u = ba + j*4; 2807cc0c071aSHong Zhang rtmp_ptr = rtmp + vj*4; 2808cc0c071aSHong Zhang for (k1=0; k1<4; k1++){ 2809cc0c071aSHong Zhang *u++ = *rtmp_ptr; 2810cc0c071aSHong Zhang *rtmp_ptr++ = 0.0; 2811cc0c071aSHong Zhang } 2812cc0c071aSHong Zhang } 28133845f261SHong Zhang 281491602c66SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 281591602c66SHong Zhang il[k] = jmin; 281691602c66SHong Zhang i = bj[jmin]; 281791602c66SHong Zhang jl[k] = jl[i]; jl[i] = k; 281891602c66SHong Zhang } 281991602c66SHong Zhang } 28203845f261SHong Zhang 282149b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 282291602c66SHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 282391602c66SHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 28243845f261SHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 28253845f261SHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 282691602c66SHong Zhang if (a->permute) { 282791602c66SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 282891602c66SHong Zhang } 2829cc0c071aSHong Zhang ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 283091602c66SHong Zhang C->factor = FACTOR_CHOLESKY; 283149b5e25fSSatish Balay C->assembled = PETSC_TRUE; 28325c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 2833effa298cSHong Zhang PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 283449b5e25fSSatish Balay PetscFunctionReturn(0); 283549b5e25fSSatish Balay } 283691602c66SHong Zhang 283749b5e25fSSatish Balay /* 283849b5e25fSSatish Balay Version for when blocks are 2 by 2 Using natural ordering 283949b5e25fSSatish Balay */ 284049b5e25fSSatish Balay #undef __FUNC__ 28416f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering" 28426f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering(Mat A,Mat *B) 284349b5e25fSSatish Balay { 284449b5e25fSSatish Balay Mat C = *B; 2845ab27746eSHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 2846ab27746eSHong Zhang int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 2847*71149678SHong Zhang int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 2848ab27746eSHong Zhang MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 2849ab27746eSHong Zhang MatScalar *u,*diag,*rtmp,*rtmp_ptr; 285049b5e25fSSatish Balay 285149b5e25fSSatish Balay PetscFunctionBegin; 2852103535cfSHong Zhang /* printf("called Numeric_SeqSBAIJ_2_NaturalOrdering\n"); */ 2853ab27746eSHong Zhang /* initialization */ 2854ab27746eSHong Zhang /* il and jl record the first nonzero element in each row of the accessing 2855ab27746eSHong Zhang window U(0:k, k:mbs-1). 2856ab27746eSHong Zhang jl: list of rows to be added to uneliminated rows 2857ab27746eSHong Zhang i>= k: jl(i) is the first row to be added to row i 2858ab27746eSHong Zhang i< k: jl(i) is the row following row i in some list of rows 2859ab27746eSHong Zhang jl(i) = mbs indicates the end of a list 2860ab27746eSHong Zhang il(i): points to the first nonzero element in columns k,...,mbs-1 of 2861ab27746eSHong Zhang row i of U */ 2862ab27746eSHong Zhang rtmp = (MatScalar*)PetscMalloc(4*mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 2863ab27746eSHong Zhang ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2864ab27746eSHong Zhang il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 2865ab27746eSHong Zhang jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 2866ab27746eSHong Zhang for (i=0; i<mbs; i++) { 2867ab27746eSHong Zhang jl[i] = mbs; il[0] = 0; 286849b5e25fSSatish Balay } 2869ab27746eSHong Zhang dk = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(dk); 2870ab27746eSHong Zhang uik = (MatScalar*)PetscMalloc(4*sizeof(MatScalar));CHKPTRQ(uik); 2871ab27746eSHong Zhang 2872ab27746eSHong Zhang ai = a->i; aj = a->j; aa = a->a; 2873ab27746eSHong Zhang 2874ab27746eSHong Zhang /* for each row k */ 2875ab27746eSHong Zhang for (k = 0; k<mbs; k++){ 2876ab27746eSHong Zhang 2877ab27746eSHong Zhang /*initialize k-th row with elements nonzero in row k of A */ 2878ab27746eSHong Zhang jmin = ai[k]; jmax = ai[k+1]; 2879ab27746eSHong Zhang if (jmin < jmax) { 2880ab27746eSHong Zhang ap = aa + jmin*4; 2881ab27746eSHong Zhang for (j = jmin; j < jmax; j++){ 2882ab27746eSHong Zhang vj = aj[j]; /* block col. index */ 2883ab27746eSHong Zhang rtmp_ptr = rtmp + vj*4; 2884ab27746eSHong Zhang for (i=0; i<4; i++) *rtmp_ptr++ = *ap++; 288549b5e25fSSatish Balay } 288649b5e25fSSatish Balay } 2887ab27746eSHong Zhang 2888ab27746eSHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 2889ab27746eSHong Zhang ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr); 2890ab27746eSHong Zhang i = jl[k]; /* first row to be added to k_th row */ 2891ab27746eSHong Zhang 2892ab27746eSHong Zhang while (i < mbs){ 2893ab27746eSHong Zhang nexti = jl[i]; /* next row to be added to k_th row */ 2894ab27746eSHong Zhang 2895ab27746eSHong Zhang /* compute multiplier */ 2896ab27746eSHong Zhang ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2897ab27746eSHong Zhang 2898ab27746eSHong Zhang /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */ 2899ab27746eSHong Zhang diag = ba + i*4; 2900ab27746eSHong Zhang u = ba + ili*4; 2901ab27746eSHong Zhang uik[0] = -(diag[0]*u[0] + diag[2]*u[1]); 2902ab27746eSHong Zhang uik[1] = -(diag[1]*u[0] + diag[3]*u[1]); 2903ab27746eSHong Zhang uik[2] = -(diag[0]*u[2] + diag[2]*u[3]); 2904ab27746eSHong Zhang uik[3] = -(diag[1]*u[2] + diag[3]*u[3]); 2905ab27746eSHong Zhang 2906ab27746eSHong Zhang /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */ 2907ab27746eSHong Zhang dk[0] += uik[0]*u[0] + uik[1]*u[1]; 2908ab27746eSHong Zhang dk[1] += uik[2]*u[0] + uik[3]*u[1]; 2909ab27746eSHong Zhang dk[2] += uik[0]*u[2] + uik[1]*u[3]; 2910ab27746eSHong Zhang dk[3] += uik[2]*u[2] + uik[3]*u[3]; 2911ab27746eSHong Zhang 2912ab27746eSHong Zhang /* update -U(i,k): ba[ili] = uik */ 2913ab27746eSHong Zhang ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr); 2914ab27746eSHong Zhang 2915ab27746eSHong Zhang /* add multiple of row i to k-th row ... */ 2916ab27746eSHong Zhang jmin = ili + 1; jmax = bi[i+1]; 2917ab27746eSHong Zhang if (jmin < jmax){ 2918ab27746eSHong Zhang for (j=jmin; j<jmax; j++) { 2919ab27746eSHong Zhang /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */ 2920ab27746eSHong Zhang rtmp_ptr = rtmp + bj[j]*4; 2921ab27746eSHong Zhang u = ba + j*4; 2922ab27746eSHong Zhang rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1]; 2923ab27746eSHong Zhang rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1]; 2924ab27746eSHong Zhang rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3]; 2925ab27746eSHong Zhang rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3]; 292649b5e25fSSatish Balay } 2927ab27746eSHong Zhang 2928ab27746eSHong Zhang /* ... add i to row list for next nonzero entry */ 2929ab27746eSHong Zhang il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 2930ab27746eSHong Zhang j = bj[jmin]; 2931ab27746eSHong Zhang jl[i] = jl[j]; jl[j] = i; /* update jl */ 293249b5e25fSSatish Balay } 2933ab27746eSHong Zhang i = nexti; 293449b5e25fSSatish Balay } 2935ab27746eSHong Zhang 2936ab27746eSHong Zhang /* save nonzero entries in k-th row of U ... */ 2937ab27746eSHong Zhang 293849b5e25fSSatish Balay /* invert diagonal block */ 2939ab27746eSHong Zhang diag = ba+k*4; 2940ab27746eSHong Zhang ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr); 2941ab27746eSHong Zhang ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr); 2942ab27746eSHong Zhang 2943ab27746eSHong Zhang jmin = bi[k]; jmax = bi[k+1]; 2944ab27746eSHong Zhang if (jmin < jmax) { 2945ab27746eSHong Zhang for (j=jmin; j<jmax; j++){ 2946ab27746eSHong Zhang vj = bj[j]; /* block col. index of U */ 2947ab27746eSHong Zhang u = ba + j*4; 2948ab27746eSHong Zhang rtmp_ptr = rtmp + vj*4; 2949ab27746eSHong Zhang for (k1=0; k1<4; k1++){ 2950ab27746eSHong Zhang *u++ = *rtmp_ptr; 2951ab27746eSHong Zhang *rtmp_ptr++ = 0.0; 2952ab27746eSHong Zhang } 2953ab27746eSHong Zhang } 2954ab27746eSHong Zhang 2955ab27746eSHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 2956ab27746eSHong Zhang il[k] = jmin; 2957ab27746eSHong Zhang i = bj[jmin]; 2958ab27746eSHong Zhang jl[k] = jl[i]; jl[i] = k; 2959ab27746eSHong Zhang } 296049b5e25fSSatish Balay } 296149b5e25fSSatish Balay 296249b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 2963ab27746eSHong Zhang ierr = PetscFree(il);CHKERRQ(ierr); 2964ab27746eSHong Zhang ierr = PetscFree(jl);CHKERRQ(ierr); 2965ab27746eSHong Zhang ierr = PetscFree(dk);CHKERRQ(ierr); 2966ab27746eSHong Zhang ierr = PetscFree(uik);CHKERRQ(ierr); 2967ab27746eSHong Zhang 2968ab27746eSHong Zhang C->factor = FACTOR_CHOLESKY; 296949b5e25fSSatish Balay C->assembled = PETSC_TRUE; 29705c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 297149b5e25fSSatish Balay PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 297249b5e25fSSatish Balay PetscFunctionReturn(0); 297349b5e25fSSatish Balay } 297449b5e25fSSatish Balay 297549b5e25fSSatish Balay /* 29765c0bcdfcSHong Zhang Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP. 297791602c66SHong Zhang Version for blocks are 1 by 1. 297849b5e25fSSatish Balay */ 297949b5e25fSSatish Balay #undef __FUNC__ 29806f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_1" 29816f9a564bSHong Zhang int MatCholeskyFactorNumeric_SeqSBAIJ_1(Mat A,Mat *B) 298249b5e25fSSatish Balay { 298349b5e25fSSatish Balay Mat C = *B; 298449b5e25fSSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 298549b5e25fSSatish Balay IS ip = b->row; 2986351d0355SHong Zhang int *rip,ierr,i,j,mbs = a->mbs,*bi = b->i,*bj = b->j; 2987cb718733SHong Zhang int *ai,*aj,*r; 2988ab27746eSHong Zhang int k,jmin,jmax,*jl,*il,vj,nexti,ili; 2989066653e3SSatish Balay MatScalar *rtmp; 29902d007305SHong Zhang MatScalar *ba = b->a,*aa,ak; 299149b5e25fSSatish Balay MatScalar dk,uikdi; 299249b5e25fSSatish Balay 299349b5e25fSSatish Balay PetscFunctionBegin; 299449b5e25fSSatish Balay ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 2995cb718733SHong Zhang if (!a->permute){ 29962d007305SHong Zhang ai = a->i; aj = a->j; aa = a->a; 29972d007305SHong Zhang } else { 29982d007305SHong Zhang ai = a->inew; aj = a->jnew; 29992d007305SHong Zhang aa = (MatScalar*)PetscMalloc(ai[mbs]*sizeof(MatScalar));CHKPTRQ(aa); 3000cb718733SHong Zhang ierr = PetscMemcpy(aa,a->a,ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 30012d007305SHong Zhang r = (int*)PetscMalloc(ai[mbs]*sizeof(int));CHKPTRQ(r); 30022d007305SHong Zhang ierr= PetscMemcpy(r,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 30032d007305SHong Zhang 30042d007305SHong Zhang jmin = ai[0]; jmax = ai[mbs]; 30052d007305SHong Zhang for (j=jmin; j<jmax; j++){ 3006cb718733SHong Zhang while (r[j] != j){ 30072d007305SHong Zhang k = r[j]; r[j] = r[k]; r[k] = k; 30082d007305SHong Zhang ak = aa[k]; aa[k] = aa[j]; aa[j] = ak; 30092d007305SHong Zhang } 30102d007305SHong Zhang } 30112d007305SHong Zhang ierr = PetscFree(r);CHKERRA(ierr); 30122d007305SHong Zhang } 30132d007305SHong Zhang 301491602c66SHong Zhang /* initialization */ 301549b5e25fSSatish Balay /* il and jl record the first nonzero element in each row of the accessing 301649b5e25fSSatish Balay window U(0:k, k:mbs-1). 301749b5e25fSSatish Balay jl: list of rows to be added to uneliminated rows 301849b5e25fSSatish Balay i>= k: jl(i) is the first row to be added to row i 301949b5e25fSSatish Balay i< k: jl(i) is the row following row i in some list of rows 302049b5e25fSSatish Balay jl(i) = mbs indicates the end of a list 302149b5e25fSSatish Balay il(i): points to the first nonzero element in columns k,...,mbs-1 of 302249b5e25fSSatish Balay row i of U */ 302349b5e25fSSatish Balay rtmp = (MatScalar*)PetscMalloc(mbs*sizeof(MatScalar));CHKPTRQ(rtmp); 302449b5e25fSSatish Balay il = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(il); 302549b5e25fSSatish Balay jl = (int*)PetscMalloc(mbs*sizeof(int));CHKPTRQ(jl); 302649b5e25fSSatish Balay for (i=0; i<mbs; i++) { 302749b5e25fSSatish Balay rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0; 302849b5e25fSSatish Balay } 302949b5e25fSSatish Balay 303091602c66SHong Zhang /* for each row k */ 303149b5e25fSSatish Balay for (k = 0; k<mbs; k++){ 303249b5e25fSSatish Balay 303391602c66SHong Zhang /*initialize k-th row with elements nonzero in row perm(k) of A */ 303449b5e25fSSatish Balay jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 303549b5e25fSSatish Balay if (jmin < jmax) { 303649b5e25fSSatish Balay for (j = jmin; j < jmax; j++){ 3037351d0355SHong Zhang vj = rip[aj[j]]; 3038ab27746eSHong Zhang rtmp[vj] = aa[j]; 303949b5e25fSSatish Balay } 304049b5e25fSSatish Balay } 304149b5e25fSSatish Balay 304291602c66SHong Zhang /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 304349b5e25fSSatish Balay dk = rtmp[k]; 304449b5e25fSSatish Balay i = jl[k]; /* first row to be added to k_th row */ 304549b5e25fSSatish Balay 304649b5e25fSSatish Balay while (i < mbs){ 304749b5e25fSSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 304849b5e25fSSatish Balay 304991602c66SHong Zhang /* compute multiplier, update D(k) and U(i,k) */ 305049b5e25fSSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 305149b5e25fSSatish Balay uikdi = - ba[ili]*ba[i]; 305249b5e25fSSatish Balay dk += uikdi*ba[ili]; 3053658e7b3eSHong Zhang ba[ili] = uikdi; /* -U(i,k) */ 305449b5e25fSSatish Balay 305591602c66SHong Zhang /* add multiple of row i to k-th row ... */ 305649b5e25fSSatish Balay jmin = ili + 1; jmax = bi[i+1]; 305749b5e25fSSatish Balay if (jmin < jmax){ 305849b5e25fSSatish Balay for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 305991602c66SHong Zhang /* ... add i to row list for next nonzero entry */ 306049b5e25fSSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 306149b5e25fSSatish Balay j = bj[jmin]; 306249b5e25fSSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 306349b5e25fSSatish Balay } 3064ab27746eSHong Zhang i = nexti; 306549b5e25fSSatish Balay } 306649b5e25fSSatish Balay 306791602c66SHong Zhang /* check for zero pivot and save diagoanl element */ 306849b5e25fSSatish Balay if (dk == 0.0){ 306929bbc08cSBarry Smith SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot"); 30708be8c0c7SHong Zhang }else if (PetscRealPart(dk) < 0){ 30718be8c0c7SHong Zhang ierr = PetscPrintf(PETSC_COMM_SELF,"Negative pivot: d[%d] = %g\n",k,dk); 307249b5e25fSSatish Balay } 307349b5e25fSSatish Balay 307491602c66SHong Zhang /* save nonzero entries in k-th row of U ... */ 3075f3dd7b2dSKris Buschelman ba[k] = 1.0/dk; 307649b5e25fSSatish Balay jmin = bi[k]; jmax = bi[k+1]; 307749b5e25fSSatish Balay if (jmin < jmax) { 307849b5e25fSSatish Balay for (j=jmin; j<jmax; j++){ 3079cc0c071aSHong Zhang vj = bj[j]; ba[j] = rtmp[vj]; rtmp[vj] = 0.0; 308049b5e25fSSatish Balay } 308191602c66SHong Zhang /* ... add k to row list for first nonzero entry in k-th row */ 308249b5e25fSSatish Balay il[k] = jmin; 308349b5e25fSSatish Balay i = bj[jmin]; 308449b5e25fSSatish Balay jl[k] = jl[i]; jl[i] = k; 308549b5e25fSSatish Balay } 308649b5e25fSSatish Balay } 308749b5e25fSSatish Balay 308849b5e25fSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 308949b5e25fSSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 309049b5e25fSSatish Balay ierr = PetscFree(jl);CHKERRQ(ierr); 3091cb718733SHong Zhang if (a->permute){ 3092cb718733SHong Zhang ierr = PetscFree(aa);CHKERRQ(ierr); 3093cb718733SHong Zhang } 309449b5e25fSSatish Balay 309549b5e25fSSatish Balay ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 30968be8c0c7SHong Zhang C->factor = FACTOR_CHOLESKY; 309749b5e25fSSatish Balay C->assembled = PETSC_TRUE; 30985c0bcdfcSHong Zhang C->preallocated = PETSC_TRUE; 309949b5e25fSSatish Balay PLogFlops(b->mbs); 310049b5e25fSSatish Balay PetscFunctionReturn(0); 310149b5e25fSSatish Balay } 310249b5e25fSSatish Balay 310349b5e25fSSatish Balay #undef __FUNC__ 31046f9a564bSHong Zhang #define __FUNC__ "MatCholeskyFactor_SeqSBAIJ" 3105b8b23757SHong Zhang int MatCholeskyFactor_SeqSBAIJ(Mat A,IS perm,PetscReal f) 310649b5e25fSSatish Balay { 31074445ddedSHong Zhang int ierr; 310849b5e25fSSatish Balay Mat C; 310949b5e25fSSatish Balay 311049b5e25fSSatish Balay PetscFunctionBegin; 3111b8b23757SHong Zhang ierr = MatCholeskyFactorSymbolic(A,perm,f,&C);CHKERRQ(ierr); 3112a4ada70bSHong Zhang ierr = MatCholeskyFactorNumeric(A,&C);CHKERRQ(ierr); 31134445ddedSHong Zhang ierr = MatHeaderCopy(A,C);CHKERRQ(ierr); 311449b5e25fSSatish Balay PetscFunctionReturn(0); 311549b5e25fSSatish Balay } 311649b5e25fSSatish Balay 311749b5e25fSSatish Balay 3118