xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact5.c (revision 3a7fca6b988f1faea172e46abdf950477dbfaf76)
1*3a7fca6bSBarry Smith /*$Id: sbaijfact5.c,v 1.3 2001/03/23 23:22:21 balay Exp bsmith $*/
2*3a7fca6bSBarry Smith #include "src/mat/impls/sbaij/seq/sbaij.h"
381278733SSatish Balay #include "src/inline/ilu.h"
481278733SSatish Balay 
581278733SSatish Balay /*
681278733SSatish Balay       Version for when blocks are 4 by 4 Using natural ordering
781278733SSatish Balay */
84a2ae208SSatish Balay #undef __FUNCT__
94a2ae208SSatish Balay #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering"
1081278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering(Mat A,Mat *B)
1181278733SSatish Balay {
1281278733SSatish Balay   Mat                C = *B;
1381278733SSatish Balay   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
1481278733SSatish Balay   int                ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
1581278733SSatish Balay   int                *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
1681278733SSatish Balay   MatScalar          *ba = b->a,*aa,*ap,*dk,*uik;
1781278733SSatish Balay   MatScalar          *u,*diag,*rtmp,*rtmp_ptr;
1881278733SSatish Balay 
1981278733SSatish Balay   PetscFunctionBegin;
2081278733SSatish Balay 
2181278733SSatish Balay   /* initialization */
2281278733SSatish Balay   ierr = PetscMalloc(16*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
2381278733SSatish Balay   ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr);
2481278733SSatish Balay   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
2581278733SSatish Balay   jl   = il + mbs;
2681278733SSatish Balay   for (i=0; i<mbs; i++) {
2781278733SSatish Balay     jl[i] = mbs; il[0] = 0;
2881278733SSatish Balay   }
2981278733SSatish Balay   ierr = PetscMalloc(32*sizeof(MatScalar),&dk);CHKERRQ(ierr);
3081278733SSatish Balay   uik  = dk + 16;
3181278733SSatish Balay 
3281278733SSatish Balay   ai = a->i; aj = a->j; aa = a->a;
3381278733SSatish Balay 
3481278733SSatish Balay   /* for each row k */
3581278733SSatish Balay   for (k = 0; k<mbs; k++){
3681278733SSatish Balay 
3781278733SSatish Balay     /*initialize k-th row with elements nonzero in row k of A */
3881278733SSatish Balay     jmin = ai[k]; jmax = ai[k+1];
3981278733SSatish Balay     if (jmin < jmax) {
4081278733SSatish Balay       ap = aa + jmin*16;
4181278733SSatish Balay       for (j = jmin; j < jmax; j++){
4281278733SSatish Balay         vj = aj[j];         /* block col. index */
4381278733SSatish Balay         rtmp_ptr = rtmp + vj*16;
4481278733SSatish Balay         for (i=0; i<16; i++) *rtmp_ptr++ = *ap++;
4581278733SSatish Balay       }
4681278733SSatish Balay     }
4781278733SSatish Balay 
4881278733SSatish Balay     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
4981278733SSatish Balay     ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr);
5081278733SSatish Balay     i = jl[k]; /* first row to be added to k_th row  */
5181278733SSatish Balay 
5281278733SSatish Balay     while (i < mbs){
5381278733SSatish Balay       nexti = jl[i]; /* next row to be added to k_th row */
5481278733SSatish Balay 
5581278733SSatish Balay       /* compute multiplier */
5681278733SSatish Balay       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
5781278733SSatish Balay 
5881278733SSatish Balay       /* uik = -inv(Di)*U_bar(i,k) */
5981278733SSatish Balay       diag = ba + i*16;
6081278733SSatish Balay       u    = ba + ili*16;
6181278733SSatish Balay 
6281278733SSatish Balay       uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]);
6381278733SSatish Balay       uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]);
6481278733SSatish Balay       uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]);
6581278733SSatish Balay       uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]);
6681278733SSatish Balay 
6781278733SSatish Balay       uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]);
6881278733SSatish Balay       uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]);
6981278733SSatish Balay       uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]);
7081278733SSatish Balay       uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]);
7181278733SSatish Balay 
7281278733SSatish Balay       uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]);
7381278733SSatish Balay       uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]);
7481278733SSatish Balay       uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]);
7581278733SSatish Balay       uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]);
7681278733SSatish Balay 
7781278733SSatish Balay       uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]);
7881278733SSatish Balay       uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]);
7981278733SSatish Balay       uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]);
8081278733SSatish Balay       uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]);
8181278733SSatish Balay 
8281278733SSatish Balay       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
8381278733SSatish Balay       dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3];
8481278733SSatish Balay       dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3];
8581278733SSatish Balay       dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3];
8681278733SSatish Balay       dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3];
8781278733SSatish Balay 
8881278733SSatish Balay       dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7];
8981278733SSatish Balay       dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7];
9081278733SSatish Balay       dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7];
9181278733SSatish Balay       dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7];
9281278733SSatish Balay 
9381278733SSatish Balay       dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11];
9481278733SSatish Balay       dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11];
9581278733SSatish Balay       dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11];
9681278733SSatish Balay       dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11];
9781278733SSatish Balay 
9881278733SSatish Balay       dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15];
9981278733SSatish Balay       dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15];
10081278733SSatish Balay       dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15];
10181278733SSatish Balay       dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15];
10281278733SSatish Balay 
10381278733SSatish Balay       /* update -U(i,k) */
10481278733SSatish Balay       ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr);
10581278733SSatish Balay 
10681278733SSatish Balay       /* add multiple of row i to k-th row ... */
10781278733SSatish Balay       jmin = ili + 1; jmax = bi[i+1];
10881278733SSatish Balay       if (jmin < jmax){
10981278733SSatish Balay         for (j=jmin; j<jmax; j++) {
11081278733SSatish Balay           /* rtmp += -U(i,k)^T * U_bar(i,j) */
11181278733SSatish Balay           rtmp_ptr = rtmp + bj[j]*16;
11281278733SSatish Balay           u = ba + j*16;
11381278733SSatish Balay           rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3];
11481278733SSatish Balay           rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3];
11581278733SSatish Balay           rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3];
11681278733SSatish Balay           rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3];
11781278733SSatish Balay 
11881278733SSatish Balay           rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7];
11981278733SSatish Balay           rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7];
12081278733SSatish Balay           rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7];
12181278733SSatish Balay           rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7];
12281278733SSatish Balay 
12381278733SSatish Balay           rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11];
12481278733SSatish Balay           rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11];
12581278733SSatish Balay           rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11];
12681278733SSatish Balay           rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11];
12781278733SSatish Balay 
12881278733SSatish Balay           rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15];
12981278733SSatish Balay           rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15];
13081278733SSatish Balay           rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15];
13181278733SSatish Balay           rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15];
13281278733SSatish Balay         }
13381278733SSatish Balay 
13481278733SSatish Balay         /* ... add i to row list for next nonzero entry */
13581278733SSatish Balay         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
13681278733SSatish Balay         j     = bj[jmin];
13781278733SSatish Balay         jl[i] = jl[j]; jl[j] = i; /* update jl */
13881278733SSatish Balay       }
13981278733SSatish Balay       i = nexti;
14081278733SSatish Balay     }
14181278733SSatish Balay 
14281278733SSatish Balay     /* save nonzero entries in k-th row of U ... */
14381278733SSatish Balay 
14481278733SSatish Balay     /* invert diagonal block */
14581278733SSatish Balay     diag = ba+k*16;
14681278733SSatish Balay     ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr);
14781278733SSatish Balay     ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr);
14881278733SSatish Balay 
14981278733SSatish Balay     jmin = bi[k]; jmax = bi[k+1];
15081278733SSatish Balay     if (jmin < jmax) {
15181278733SSatish Balay       for (j=jmin; j<jmax; j++){
15281278733SSatish Balay          vj = bj[j];           /* block col. index of U */
15381278733SSatish Balay          u   = ba + j*16;
15481278733SSatish Balay          rtmp_ptr = rtmp + vj*16;
15581278733SSatish Balay          for (k1=0; k1<16; k1++){
15681278733SSatish Balay            *u++        = *rtmp_ptr;
15781278733SSatish Balay            *rtmp_ptr++ = 0.0;
15881278733SSatish Balay          }
15981278733SSatish Balay       }
16081278733SSatish Balay 
16181278733SSatish Balay       /* ... add k to row list for first nonzero entry in k-th row */
16281278733SSatish Balay       il[k] = jmin;
16381278733SSatish Balay       i     = bj[jmin];
16481278733SSatish Balay       jl[k] = jl[i]; jl[i] = k;
16581278733SSatish Balay     }
16681278733SSatish Balay   }
16781278733SSatish Balay 
16881278733SSatish Balay   ierr = PetscFree(rtmp);CHKERRQ(ierr);
16981278733SSatish Balay   ierr = PetscFree(il);CHKERRQ(ierr);
17081278733SSatish Balay   ierr = PetscFree(dk);CHKERRQ(ierr);
17181278733SSatish Balay 
17281278733SSatish Balay   C->factor    = FACTOR_CHOLESKY;
17381278733SSatish Balay   C->assembled = PETSC_TRUE;
17481278733SSatish Balay   C->preallocated = PETSC_TRUE;
17581278733SSatish Balay   PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */
17681278733SSatish Balay   PetscFunctionReturn(0);
17781278733SSatish Balay }
178