xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact5.c (revision dc0b31ed242eb6637dca0468fbfa12899dc55c4b)
1be1d678aSKris Buschelman #define PETSCMAT_DLL
21c2a3de1SBarry Smith 
37c4f633dSBarry Smith #include "../src/mat/impls/sbaij/seq/sbaij.h"
47c4f633dSBarry Smith #include "../src/inline/ilu.h"
581278733SSatish Balay 
681278733SSatish Balay /*
781278733SSatish Balay       Version for when blocks are 4 by 4 Using natural ordering
881278733SSatish Balay */
94a2ae208SSatish Balay #undef __FUNCT__
104a2ae208SSatish Balay #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering"
110481f469SBarry Smith PetscErrorCode MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering(Mat C,Mat A,const MatFactorInfo *info)
1281278733SSatish Balay {
1381278733SSatish Balay   Mat_SeqSBAIJ   *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
14dfbe8321SBarry Smith   PetscErrorCode ierr;
1513f74950SBarry Smith   PetscInt       i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
1613f74950SBarry Smith   PetscInt       *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
1781278733SSatish Balay   MatScalar      *ba = b->a,*aa,*ap,*dk,*uik;
1881278733SSatish Balay   MatScalar      *u,*diag,*rtmp,*rtmp_ptr;
19bcd9e38bSBarry Smith   PetscTruth     pivotinblocks = b->pivotinblocks;
2062bba022SBarry Smith   PetscReal      shift = info->shiftinblocks;
2181278733SSatish Balay 
2281278733SSatish Balay   PetscFunctionBegin;
2381278733SSatish Balay 
2481278733SSatish Balay   /* initialization */
2581278733SSatish Balay   ierr = PetscMalloc(16*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
2681278733SSatish Balay   ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr);
2713f74950SBarry Smith   ierr = PetscMalloc(2*mbs*sizeof(PetscInt),&il);CHKERRQ(ierr);
2881278733SSatish Balay   jl   = il + mbs;
2981278733SSatish Balay   for (i=0; i<mbs; i++) {
3081278733SSatish Balay     jl[i] = mbs; il[0] = 0;
3181278733SSatish Balay   }
3281278733SSatish Balay   ierr = PetscMalloc(32*sizeof(MatScalar),&dk);CHKERRQ(ierr);
3381278733SSatish Balay   uik  = dk + 16;
3481278733SSatish Balay 
3581278733SSatish Balay   ai = a->i; aj = a->j; aa = a->a;
3681278733SSatish Balay 
3781278733SSatish Balay   /* for each row k */
3881278733SSatish Balay   for (k = 0; k<mbs; k++){
3981278733SSatish Balay 
4081278733SSatish Balay     /*initialize k-th row with elements nonzero in row k of A */
4181278733SSatish Balay     jmin = ai[k]; jmax = ai[k+1];
4281278733SSatish Balay     if (jmin < jmax) {
4381278733SSatish Balay       ap = aa + jmin*16;
4481278733SSatish Balay       for (j = jmin; j < jmax; j++){
4581278733SSatish Balay         vj = aj[j];         /* block col. index */
4681278733SSatish Balay         rtmp_ptr = rtmp + vj*16;
4781278733SSatish Balay         for (i=0; i<16; i++) *rtmp_ptr++ = *ap++;
4881278733SSatish Balay       }
4981278733SSatish Balay     }
5081278733SSatish Balay 
5181278733SSatish Balay     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
5281278733SSatish Balay     ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr);
5381278733SSatish Balay     i = jl[k]; /* first row to be added to k_th row  */
5481278733SSatish Balay 
5581278733SSatish Balay     while (i < mbs){
5681278733SSatish Balay       nexti = jl[i]; /* next row to be added to k_th row */
5781278733SSatish Balay 
5881278733SSatish Balay       /* compute multiplier */
5981278733SSatish Balay       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
6081278733SSatish Balay 
6181278733SSatish Balay       /* uik = -inv(Di)*U_bar(i,k) */
6281278733SSatish Balay       diag = ba + i*16;
6381278733SSatish Balay       u    = ba + ili*16;
6481278733SSatish Balay 
6581278733SSatish Balay       uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]);
6681278733SSatish Balay       uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]);
6781278733SSatish Balay       uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]);
6881278733SSatish Balay       uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]);
6981278733SSatish Balay 
7081278733SSatish Balay       uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]);
7181278733SSatish Balay       uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]);
7281278733SSatish Balay       uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]);
7381278733SSatish Balay       uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]);
7481278733SSatish Balay 
7581278733SSatish Balay       uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]);
7681278733SSatish Balay       uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]);
7781278733SSatish Balay       uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]);
7881278733SSatish Balay       uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]);
7981278733SSatish Balay 
8081278733SSatish Balay       uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]);
8181278733SSatish Balay       uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]);
8281278733SSatish Balay       uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]);
8381278733SSatish Balay       uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]);
8481278733SSatish Balay 
8581278733SSatish Balay       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
8681278733SSatish Balay       dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3];
8781278733SSatish Balay       dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3];
8881278733SSatish Balay       dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3];
8981278733SSatish Balay       dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3];
9081278733SSatish Balay 
9181278733SSatish Balay       dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7];
9281278733SSatish Balay       dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7];
9381278733SSatish Balay       dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7];
9481278733SSatish Balay       dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7];
9581278733SSatish Balay 
9681278733SSatish Balay       dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11];
9781278733SSatish Balay       dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11];
9881278733SSatish Balay       dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11];
9981278733SSatish Balay       dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11];
10081278733SSatish Balay 
10181278733SSatish Balay       dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15];
10281278733SSatish Balay       dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15];
10381278733SSatish Balay       dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15];
10481278733SSatish Balay       dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15];
10581278733SSatish Balay 
106*dc0b31edSSatish Balay       ierr = PetscLogFlops(64.0*4.0);CHKERRQ(ierr);
107187a9f4bSHong Zhang 
10881278733SSatish Balay       /* update -U(i,k) */
10981278733SSatish Balay       ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr);
11081278733SSatish Balay 
11181278733SSatish Balay       /* add multiple of row i to k-th row ... */
11281278733SSatish Balay       jmin = ili + 1; jmax = bi[i+1];
11381278733SSatish Balay       if (jmin < jmax){
11481278733SSatish Balay         for (j=jmin; j<jmax; j++) {
11581278733SSatish Balay           /* rtmp += -U(i,k)^T * U_bar(i,j) */
11681278733SSatish Balay           rtmp_ptr = rtmp + bj[j]*16;
11781278733SSatish Balay           u = ba + j*16;
11881278733SSatish Balay           rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3];
11981278733SSatish Balay           rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3];
12081278733SSatish Balay           rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3];
12181278733SSatish Balay           rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3];
12281278733SSatish Balay 
12381278733SSatish Balay           rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7];
12481278733SSatish Balay           rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7];
12581278733SSatish Balay           rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7];
12681278733SSatish Balay           rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7];
12781278733SSatish Balay 
12881278733SSatish Balay           rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11];
12981278733SSatish Balay           rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11];
13081278733SSatish Balay           rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11];
13181278733SSatish Balay           rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11];
13281278733SSatish Balay 
13381278733SSatish Balay           rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15];
13481278733SSatish Balay           rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15];
13581278733SSatish Balay           rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15];
13681278733SSatish Balay           rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15];
13781278733SSatish Balay         }
138*dc0b31edSSatish Balay         ierr = PetscLogFlops(2.0*64.0*(jmax-jmin));CHKERRQ(ierr);
13981278733SSatish Balay 
14081278733SSatish Balay         /* ... add i to row list for next nonzero entry */
14181278733SSatish Balay         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
14281278733SSatish Balay         j     = bj[jmin];
14381278733SSatish Balay         jl[i] = jl[j]; jl[j] = i; /* update jl */
14481278733SSatish Balay       }
14581278733SSatish Balay       i = nexti;
14681278733SSatish Balay     }
14781278733SSatish Balay 
14881278733SSatish Balay     /* save nonzero entries in k-th row of U ... */
14981278733SSatish Balay 
15081278733SSatish Balay     /* invert diagonal block */
15181278733SSatish Balay     diag = ba+k*16;
15281278733SSatish Balay     ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr);
153bcd9e38bSBarry Smith     if (pivotinblocks) {
15462bba022SBarry Smith       ierr = Kernel_A_gets_inverse_A_4(diag,shift);CHKERRQ(ierr);
155bcd9e38bSBarry Smith     } else {
156bcd9e38bSBarry Smith       ierr = Kernel_A_gets_inverse_A_4_nopivot(diag);CHKERRQ(ierr);
157bcd9e38bSBarry Smith     }
15881278733SSatish Balay 
15981278733SSatish Balay     jmin = bi[k]; jmax = bi[k+1];
16081278733SSatish Balay     if (jmin < jmax) {
16181278733SSatish Balay       for (j=jmin; j<jmax; j++){
16281278733SSatish Balay          vj = bj[j];           /* block col. index of U */
16381278733SSatish Balay          u   = ba + j*16;
16481278733SSatish Balay          rtmp_ptr = rtmp + vj*16;
16581278733SSatish Balay          for (k1=0; k1<16; k1++){
16681278733SSatish Balay            *u++        = *rtmp_ptr;
16781278733SSatish Balay            *rtmp_ptr++ = 0.0;
16881278733SSatish Balay          }
16981278733SSatish Balay       }
17081278733SSatish Balay 
17181278733SSatish Balay       /* ... add k to row list for first nonzero entry in k-th row */
17281278733SSatish Balay       il[k] = jmin;
17381278733SSatish Balay       i     = bj[jmin];
17481278733SSatish Balay       jl[k] = jl[i]; jl[i] = k;
17581278733SSatish Balay     }
17681278733SSatish Balay   }
17781278733SSatish Balay 
17881278733SSatish Balay   ierr = PetscFree(rtmp);CHKERRQ(ierr);
17981278733SSatish Balay   ierr = PetscFree(il);CHKERRQ(ierr);
18081278733SSatish Balay   ierr = PetscFree(dk);CHKERRQ(ierr);
18181278733SSatish Balay 
182db4efbfdSBarry Smith   C->ops->solve          = MatSolve_SeqSBAIJ_4_NaturalOrdering;
183db4efbfdSBarry Smith   C->ops->solvetranspose = MatSolve_SeqSBAIJ_4_NaturalOrdering;
184db4efbfdSBarry Smith   C->ops->forwardsolve   = MatForwardSolve_SeqSBAIJ_4_NaturalOrdering;
185db4efbfdSBarry Smith   C->ops->backwardsolve  = MatBackwardSolve_SeqSBAIJ_4_NaturalOrdering;
186db4efbfdSBarry Smith 
18781278733SSatish Balay   C->assembled = PETSC_TRUE;
18881278733SSatish Balay   C->preallocated = PETSC_TRUE;
189efee365bSSatish Balay   ierr = PetscLogFlops(1.3333*64*b->mbs);CHKERRQ(ierr); /* from inverting diagonal blocks */
19081278733SSatish Balay   PetscFunctionReturn(0);
19181278733SSatish Balay }
192