xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact4.c (revision be1d678a52e6eff2808b2fa31ae986cdbf03c9fe)
1*be1d678aSKris Buschelman #define PETSCMAT_DLL
21c2a3de1SBarry Smith 
33a7fca6bSBarry Smith #include "src/mat/impls/sbaij/seq/sbaij.h"
481278733SSatish Balay #include "src/inline/ilu.h"
581278733SSatish Balay 
681278733SSatish Balay /*
781278733SSatish Balay       Version for when blocks are 3 by 3 Using natural ordering
881278733SSatish Balay */
94a2ae208SSatish Balay #undef __FUNCT__
104a2ae208SSatish Balay #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering"
11af281ebdSHong Zhang PetscErrorCode MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering(Mat A,MatFactorInfo *info,Mat *B)
1281278733SSatish Balay {
1381278733SSatish Balay   Mat            C = *B;
1481278733SSatish Balay   Mat_SeqSBAIJ   *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
15dfbe8321SBarry Smith   PetscErrorCode ierr;
1613f74950SBarry Smith   PetscInt       i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
1713f74950SBarry Smith   PetscInt       *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
1881278733SSatish Balay   MatScalar      *ba = b->a,*aa,*ap,*dk,*uik;
1981278733SSatish Balay   MatScalar      *u,*diag,*rtmp,*rtmp_ptr;
2081278733SSatish Balay 
2181278733SSatish Balay   PetscFunctionBegin;
2281278733SSatish Balay 
2381278733SSatish Balay   /* initialization */
2481278733SSatish Balay   ierr = PetscMalloc(9*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
2581278733SSatish Balay   ierr = PetscMemzero(rtmp,9*mbs*sizeof(MatScalar));CHKERRQ(ierr);
2613f74950SBarry Smith   ierr = PetscMalloc(2*mbs*sizeof(PetscInt),&il);CHKERRQ(ierr);
2781278733SSatish Balay   jl   = il + mbs;
2881278733SSatish Balay   for (i=0; i<mbs; i++) {
2981278733SSatish Balay     jl[i] = mbs; il[0] = 0;
3081278733SSatish Balay   }
3181278733SSatish Balay   ierr = PetscMalloc(18*sizeof(MatScalar),&dk);CHKERRQ(ierr);
3281278733SSatish Balay   uik  = dk + 9;
3381278733SSatish Balay 
3481278733SSatish Balay   ai   = a->i; aj = a->j; aa = a->a;
3581278733SSatish Balay 
3681278733SSatish Balay   /* for each row k */
3781278733SSatish Balay   for (k = 0; k<mbs; k++){
3881278733SSatish Balay 
3981278733SSatish Balay     /*initialize k-th row with elements nonzero in row k of A */
4081278733SSatish Balay     jmin = ai[k]; jmax = ai[k+1];
4181278733SSatish Balay     if (jmin < jmax) {
4281278733SSatish Balay       ap = aa + jmin*9;
4381278733SSatish Balay       for (j = jmin; j < jmax; j++){
4481278733SSatish Balay         vj = aj[j];         /* block col. index */
4581278733SSatish Balay         rtmp_ptr = rtmp + vj*9;
4681278733SSatish Balay         for (i=0; i<9; i++) *rtmp_ptr++ = *ap++;
4781278733SSatish Balay       }
4881278733SSatish Balay     }
4981278733SSatish Balay 
5081278733SSatish Balay     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
5181278733SSatish Balay     ierr = PetscMemcpy(dk,rtmp+k*9,9*sizeof(MatScalar));CHKERRQ(ierr);
5281278733SSatish Balay     i = jl[k]; /* first row to be added to k_th row  */
5381278733SSatish Balay 
5481278733SSatish Balay     while (i < mbs){
5581278733SSatish Balay       nexti = jl[i]; /* next row to be added to k_th row */
5681278733SSatish Balay 
5781278733SSatish Balay       /* compute multiplier */
5881278733SSatish Balay       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
5981278733SSatish Balay 
6081278733SSatish Balay       /* uik = -inv(Di)*U_bar(i,k) */
6181278733SSatish Balay       diag = ba + i*9;
6281278733SSatish Balay       u    = ba + ili*9;
6381278733SSatish Balay 
6481278733SSatish Balay       uik[0] = -(diag[0]*u[0] + diag[3]*u[1] + diag[6]*u[2]);
6581278733SSatish Balay       uik[1] = -(diag[1]*u[0] + diag[4]*u[1] + diag[7]*u[2]);
6681278733SSatish Balay       uik[2] = -(diag[2]*u[0] + diag[5]*u[1] + diag[8]*u[2]);
6781278733SSatish Balay 
6881278733SSatish Balay       uik[3] = -(diag[0]*u[3] + diag[3]*u[4] + diag[6]*u[5]);
6981278733SSatish Balay       uik[4] = -(diag[1]*u[3] + diag[4]*u[4] + diag[7]*u[5]);
7081278733SSatish Balay       uik[5] = -(diag[2]*u[3] + diag[5]*u[4] + diag[8]*u[5]);
7181278733SSatish Balay 
7281278733SSatish Balay       uik[6] = -(diag[0]*u[6] + diag[3]*u[7] + diag[6]*u[8]);
7381278733SSatish Balay       uik[7] = -(diag[1]*u[6] + diag[4]*u[7] + diag[7]*u[8]);
7481278733SSatish Balay       uik[8] = -(diag[2]*u[6] + diag[5]*u[7] + diag[8]*u[8]);
7581278733SSatish Balay 
7681278733SSatish Balay       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
7781278733SSatish Balay       dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2];
7881278733SSatish Balay       dk[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2];
7981278733SSatish Balay       dk[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2];
8081278733SSatish Balay 
8181278733SSatish Balay       dk[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5];
8281278733SSatish Balay       dk[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5];
8381278733SSatish Balay       dk[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5];
8481278733SSatish Balay 
8581278733SSatish Balay       dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8];
8681278733SSatish Balay       dk[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8];
8781278733SSatish Balay       dk[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8];
8881278733SSatish Balay 
8981278733SSatish Balay       /* update -U(i,k) */
9081278733SSatish Balay       ierr = PetscMemcpy(ba+ili*9,uik,9*sizeof(MatScalar));CHKERRQ(ierr);
9181278733SSatish Balay 
9281278733SSatish Balay       /* add multiple of row i to k-th row ... */
9381278733SSatish Balay       jmin = ili + 1; jmax = bi[i+1];
9481278733SSatish Balay       if (jmin < jmax){
9581278733SSatish Balay         for (j=jmin; j<jmax; j++) {
9681278733SSatish Balay           /* rtmp += -U(i,k)^T * U_bar(i,j) */
9781278733SSatish Balay           rtmp_ptr = rtmp + bj[j]*9;
9881278733SSatish Balay           u = ba + j*9;
9981278733SSatish Balay           rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2];
10081278733SSatish Balay           rtmp_ptr[1] += uik[3]*u[0] + uik[4]*u[1] + uik[5]*u[2];
10181278733SSatish Balay           rtmp_ptr[2] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2];
10281278733SSatish Balay 
10381278733SSatish Balay           rtmp_ptr[3] += uik[0]*u[3] + uik[1]*u[4] + uik[2]*u[5];
10481278733SSatish Balay           rtmp_ptr[4] += uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5];
10581278733SSatish Balay           rtmp_ptr[5] += uik[6]*u[3] + uik[7]*u[4] + uik[8]*u[5];
10681278733SSatish Balay 
10781278733SSatish Balay           rtmp_ptr[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8];
10881278733SSatish Balay           rtmp_ptr[7] += uik[3]*u[6] + uik[4]*u[7] + uik[5]*u[8];
10981278733SSatish Balay           rtmp_ptr[8] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8];
11081278733SSatish Balay         }
11181278733SSatish Balay 
11281278733SSatish Balay         /* ... add i to row list for next nonzero entry */
11381278733SSatish Balay         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
11481278733SSatish Balay         j     = bj[jmin];
11581278733SSatish Balay         jl[i] = jl[j]; jl[j] = i; /* update jl */
11681278733SSatish Balay       }
11781278733SSatish Balay       i = nexti;
11881278733SSatish Balay     }
11981278733SSatish Balay 
12081278733SSatish Balay     /* save nonzero entries in k-th row of U ... */
12181278733SSatish Balay 
12281278733SSatish Balay     /* invert diagonal block */
12381278733SSatish Balay     diag = ba+k*9;
12481278733SSatish Balay     ierr = PetscMemcpy(diag,dk,9*sizeof(MatScalar));CHKERRQ(ierr);
12581278733SSatish Balay     ierr = Kernel_A_gets_inverse_A_3(diag);CHKERRQ(ierr);
12681278733SSatish Balay 
12781278733SSatish Balay     jmin = bi[k]; jmax = bi[k+1];
12881278733SSatish Balay     if (jmin < jmax) {
12981278733SSatish Balay       for (j=jmin; j<jmax; j++){
13081278733SSatish Balay          vj = bj[j];           /* block col. index of U */
13181278733SSatish Balay          u   = ba + j*9;
13281278733SSatish Balay          rtmp_ptr = rtmp + vj*9;
13381278733SSatish Balay          for (k1=0; k1<9; k1++){
13481278733SSatish Balay            *u++        = *rtmp_ptr;
13581278733SSatish Balay            *rtmp_ptr++ = 0.0;
13681278733SSatish Balay          }
13781278733SSatish Balay       }
13881278733SSatish Balay 
13981278733SSatish Balay       /* ... add k to row list for first nonzero entry in k-th row */
14081278733SSatish Balay       il[k] = jmin;
14181278733SSatish Balay       i     = bj[jmin];
14281278733SSatish Balay       jl[k] = jl[i]; jl[i] = k;
14381278733SSatish Balay     }
14481278733SSatish Balay   }
14581278733SSatish Balay 
14681278733SSatish Balay   ierr = PetscFree(rtmp);CHKERRQ(ierr);
14781278733SSatish Balay   ierr = PetscFree(il);CHKERRQ(ierr);
14881278733SSatish Balay   ierr = PetscFree(dk);CHKERRQ(ierr);
14981278733SSatish Balay 
15081278733SSatish Balay   C->factor    = FACTOR_CHOLESKY;
15181278733SSatish Balay   C->assembled = PETSC_TRUE;
15281278733SSatish Balay   C->preallocated = PETSC_TRUE;
153efee365bSSatish Balay   ierr = PetscLogFlops(1.3333*27*b->mbs);CHKERRQ(ierr); /* from inverting diagonal blocks */
15481278733SSatish Balay   PetscFunctionReturn(0);
15581278733SSatish Balay }
156