xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact7.c (revision 3ba1676111f5c958fe6c2729b46ca4d523958bb3)
11c2a3de1SBarry Smith 
2c6db04a5SJed Brown #include <../src/mat/impls/sbaij/seq/sbaij.h>
3af0996ceSBarry Smith #include <petsc/private/kernels/blockinvert.h>
481278733SSatish Balay 
581278733SSatish Balay /* Version for when blocks are 5 by 5  */
6d71ae5a4SJacob Faibussowitsch PetscErrorCode MatCholeskyFactorNumeric_SeqSBAIJ_5(Mat C, Mat A, const MatFactorInfo *info)
7d71ae5a4SJacob Faibussowitsch {
881278733SSatish Balay   Mat_SeqSBAIJ   *a = (Mat_SeqSBAIJ *)A->data, *b = (Mat_SeqSBAIJ *)C->data;
981278733SSatish Balay   IS              perm = b->row;
105d0c19d7SBarry Smith   const PetscInt *ai, *aj, *perm_ptr, mbs = a->mbs, *bi = b->i, *bj = b->j;
11c0c53c30SBarry Smith   PetscInt        i, j, *a2anew, k, k1, jmin, jmax, *jl, *il, vj, nexti, ili, ipvt[5];
1281278733SSatish Balay   MatScalar      *ba = b->a, *aa, *ap, *dk, *uik;
13c0c53c30SBarry Smith   MatScalar      *u, *d, *rtmp, *rtmp_ptr, work[25];
14182b8fbaSHong Zhang   PetscReal       shift = info->shiftamount;
15a455e926SHong Zhang   PetscBool       allowzeropivot, zeropivotdetected;
1681278733SSatish Balay 
1781278733SSatish Balay   PetscFunctionBegin;
1881278733SSatish Balay   /* initialization */
190164db54SHong Zhang   allowzeropivot = PetscNot(A->erroriffailure);
209566063dSJacob Faibussowitsch   PetscCall(PetscCalloc1(25 * mbs, &rtmp));
219566063dSJacob Faibussowitsch   PetscCall(PetscMalloc2(mbs, &il, mbs, &jl));
226df5ee2eSHong Zhang   il[0] = 0;
236df5ee2eSHong Zhang   for (i = 0; i < mbs; i++) jl[i] = mbs;
246df5ee2eSHong Zhang 
259566063dSJacob Faibussowitsch   PetscCall(PetscMalloc2(25, &dk, 25, &uik));
269566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(perm, &perm_ptr));
2781278733SSatish Balay 
2881278733SSatish Balay   /* check permutation */
2981278733SSatish Balay   if (!a->permute) {
309371c9d4SSatish Balay     ai = a->i;
319371c9d4SSatish Balay     aj = a->j;
329371c9d4SSatish Balay     aa = a->a;
3381278733SSatish Balay   } else {
349371c9d4SSatish Balay     ai = a->inew;
359371c9d4SSatish Balay     aj = a->jnew;
369566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(25 * ai[mbs], &aa));
379566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(aa, a->a, 25 * ai[mbs]));
389566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(ai[mbs], &a2anew));
399566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(a2anew, a->a2anew, ai[mbs]));
4081278733SSatish Balay 
4181278733SSatish Balay     for (i = 0; i < mbs; i++) {
429371c9d4SSatish Balay       jmin = ai[i];
439371c9d4SSatish Balay       jmax = ai[i + 1];
4481278733SSatish Balay       for (j = jmin; j < jmax; j++) {
4581278733SSatish Balay         while (a2anew[j] != j) {
469371c9d4SSatish Balay           k         = a2anew[j];
479371c9d4SSatish Balay           a2anew[j] = a2anew[k];
489371c9d4SSatish Balay           a2anew[k] = k;
4981278733SSatish Balay           for (k1 = 0; k1 < 25; k1++) {
5081278733SSatish Balay             dk[k1]          = aa[k * 25 + k1];
5181278733SSatish Balay             aa[k * 25 + k1] = aa[j * 25 + k1];
5281278733SSatish Balay             aa[j * 25 + k1] = dk[k1];
5381278733SSatish Balay           }
5481278733SSatish Balay         }
5581278733SSatish Balay         /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */
5681278733SSatish Balay         if (i > aj[j]) {
5781278733SSatish Balay           /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */
5881278733SSatish Balay           ap = aa + j * 25;                       /* ptr to the beginning of j-th block of aa */
5981278733SSatish Balay           for (k = 0; k < 25; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */
6081278733SSatish Balay           for (k = 0; k < 5; k++) {               /* j-th block of aa <- dk^T */
6181278733SSatish Balay             for (k1 = 0; k1 < 5; k1++) *ap++ = dk[k + 5 * k1];
6281278733SSatish Balay           }
6381278733SSatish Balay         }
6481278733SSatish Balay       }
6581278733SSatish Balay     }
669566063dSJacob Faibussowitsch     PetscCall(PetscFree(a2anew));
6781278733SSatish Balay   }
6881278733SSatish Balay 
6981278733SSatish Balay   /* for each row k */
7081278733SSatish Balay   for (k = 0; k < mbs; k++) {
7181278733SSatish Balay     /*initialize k-th row with elements nonzero in row perm(k) of A */
729371c9d4SSatish Balay     jmin = ai[perm_ptr[k]];
739371c9d4SSatish Balay     jmax = ai[perm_ptr[k] + 1];
7481278733SSatish Balay     if (jmin < jmax) {
7581278733SSatish Balay       ap = aa + jmin * 25;
7681278733SSatish Balay       for (j = jmin; j < jmax; j++) {
7781278733SSatish Balay         vj       = perm_ptr[aj[j]]; /* block col. index */
7881278733SSatish Balay         rtmp_ptr = rtmp + vj * 25;
7981278733SSatish Balay         for (i = 0; i < 25; i++) *rtmp_ptr++ = *ap++;
8081278733SSatish Balay       }
8181278733SSatish Balay     }
8281278733SSatish Balay 
8381278733SSatish Balay     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
849566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(dk, rtmp + k * 25, 25));
8581278733SSatish Balay     i = jl[k]; /* first row to be added to k_th row  */
8681278733SSatish Balay 
8781278733SSatish Balay     while (i < mbs) {
8881278733SSatish Balay       nexti = jl[i]; /* next row to be added to k_th row */
8981278733SSatish Balay 
9081278733SSatish Balay       /* compute multiplier */
9181278733SSatish Balay       ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */
9281278733SSatish Balay 
9381278733SSatish Balay       /* uik = -inv(Di)*U_bar(i,k) */
9481278733SSatish Balay       d = ba + i * 25;
9581278733SSatish Balay       u = ba + ili * 25;
9681278733SSatish Balay 
9781278733SSatish Balay       uik[0] = -(d[0] * u[0] + d[5] * u[1] + d[10] * u[2] + d[15] * u[3] + d[20] * u[4]);
9881278733SSatish Balay       uik[1] = -(d[1] * u[0] + d[6] * u[1] + d[11] * u[2] + d[16] * u[3] + d[21] * u[4]);
9981278733SSatish Balay       uik[2] = -(d[2] * u[0] + d[7] * u[1] + d[12] * u[2] + d[17] * u[3] + d[22] * u[4]);
10081278733SSatish Balay       uik[3] = -(d[3] * u[0] + d[8] * u[1] + d[13] * u[2] + d[18] * u[3] + d[23] * u[4]);
10181278733SSatish Balay       uik[4] = -(d[4] * u[0] + d[9] * u[1] + d[14] * u[2] + d[19] * u[3] + d[24] * u[4]);
10281278733SSatish Balay 
10381278733SSatish Balay       uik[5] = -(d[0] * u[5] + d[5] * u[6] + d[10] * u[7] + d[15] * u[8] + d[20] * u[9]);
10481278733SSatish Balay       uik[6] = -(d[1] * u[5] + d[6] * u[6] + d[11] * u[7] + d[16] * u[8] + d[21] * u[9]);
10581278733SSatish Balay       uik[7] = -(d[2] * u[5] + d[7] * u[6] + d[12] * u[7] + d[17] * u[8] + d[22] * u[9]);
10681278733SSatish Balay       uik[8] = -(d[3] * u[5] + d[8] * u[6] + d[13] * u[7] + d[18] * u[8] + d[23] * u[9]);
10781278733SSatish Balay       uik[9] = -(d[4] * u[5] + d[9] * u[6] + d[14] * u[7] + d[19] * u[8] + d[24] * u[9]);
10881278733SSatish Balay 
10981278733SSatish Balay       uik[10] = -(d[0] * u[10] + d[5] * u[11] + d[10] * u[12] + d[15] * u[13] + d[20] * u[14]);
11081278733SSatish Balay       uik[11] = -(d[1] * u[10] + d[6] * u[11] + d[11] * u[12] + d[16] * u[13] + d[21] * u[14]);
11181278733SSatish Balay       uik[12] = -(d[2] * u[10] + d[7] * u[11] + d[12] * u[12] + d[17] * u[13] + d[22] * u[14]);
11281278733SSatish Balay       uik[13] = -(d[3] * u[10] + d[8] * u[11] + d[13] * u[12] + d[18] * u[13] + d[23] * u[14]);
11381278733SSatish Balay       uik[14] = -(d[4] * u[10] + d[9] * u[11] + d[14] * u[12] + d[19] * u[13] + d[24] * u[14]);
11481278733SSatish Balay 
11581278733SSatish Balay       uik[15] = -(d[0] * u[15] + d[5] * u[16] + d[10] * u[17] + d[15] * u[18] + d[20] * u[19]);
11681278733SSatish Balay       uik[16] = -(d[1] * u[15] + d[6] * u[16] + d[11] * u[17] + d[16] * u[18] + d[21] * u[19]);
11781278733SSatish Balay       uik[17] = -(d[2] * u[15] + d[7] * u[16] + d[12] * u[17] + d[17] * u[18] + d[22] * u[19]);
11881278733SSatish Balay       uik[18] = -(d[3] * u[15] + d[8] * u[16] + d[13] * u[17] + d[18] * u[18] + d[23] * u[19]);
11981278733SSatish Balay       uik[19] = -(d[4] * u[15] + d[9] * u[16] + d[14] * u[17] + d[19] * u[18] + d[24] * u[19]);
12081278733SSatish Balay 
12181278733SSatish Balay       uik[20] = -(d[0] * u[20] + d[5] * u[21] + d[10] * u[22] + d[15] * u[23] + d[20] * u[24]);
12281278733SSatish Balay       uik[21] = -(d[1] * u[20] + d[6] * u[21] + d[11] * u[22] + d[16] * u[23] + d[21] * u[24]);
12381278733SSatish Balay       uik[22] = -(d[2] * u[20] + d[7] * u[21] + d[12] * u[22] + d[17] * u[23] + d[22] * u[24]);
12481278733SSatish Balay       uik[23] = -(d[3] * u[20] + d[8] * u[21] + d[13] * u[22] + d[18] * u[23] + d[23] * u[24]);
12581278733SSatish Balay       uik[24] = -(d[4] * u[20] + d[9] * u[21] + d[14] * u[22] + d[19] * u[23] + d[24] * u[24]);
12681278733SSatish Balay 
12781278733SSatish Balay       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
12881278733SSatish Balay       dk[0] += uik[0] * u[0] + uik[1] * u[1] + uik[2] * u[2] + uik[3] * u[3] + uik[4] * u[4];
12981278733SSatish Balay       dk[1] += uik[5] * u[0] + uik[6] * u[1] + uik[7] * u[2] + uik[8] * u[3] + uik[9] * u[4];
13081278733SSatish Balay       dk[2] += uik[10] * u[0] + uik[11] * u[1] + uik[12] * u[2] + uik[13] * u[3] + uik[14] * u[4];
13181278733SSatish Balay       dk[3] += uik[15] * u[0] + uik[16] * u[1] + uik[17] * u[2] + uik[18] * u[3] + uik[19] * u[4];
13281278733SSatish Balay       dk[4] += uik[20] * u[0] + uik[21] * u[1] + uik[22] * u[2] + uik[23] * u[3] + uik[24] * u[4];
13381278733SSatish Balay 
13481278733SSatish Balay       dk[5] += uik[0] * u[5] + uik[1] * u[6] + uik[2] * u[7] + uik[3] * u[8] + uik[4] * u[9];
13581278733SSatish Balay       dk[6] += uik[5] * u[5] + uik[6] * u[6] + uik[7] * u[7] + uik[8] * u[8] + uik[9] * u[9];
13681278733SSatish Balay       dk[7] += uik[10] * u[5] + uik[11] * u[6] + uik[12] * u[7] + uik[13] * u[8] + uik[14] * u[9];
13781278733SSatish Balay       dk[8] += uik[15] * u[5] + uik[16] * u[6] + uik[17] * u[7] + uik[18] * u[8] + uik[19] * u[9];
13881278733SSatish Balay       dk[9] += uik[20] * u[5] + uik[21] * u[6] + uik[22] * u[7] + uik[23] * u[8] + uik[24] * u[9];
13981278733SSatish Balay 
14081278733SSatish Balay       dk[10] += uik[0] * u[10] + uik[1] * u[11] + uik[2] * u[12] + uik[3] * u[13] + uik[4] * u[14];
14181278733SSatish Balay       dk[11] += uik[5] * u[10] + uik[6] * u[11] + uik[7] * u[12] + uik[8] * u[13] + uik[9] * u[14];
14281278733SSatish Balay       dk[12] += uik[10] * u[10] + uik[11] * u[11] + uik[12] * u[12] + uik[13] * u[13] + uik[14] * u[14];
14381278733SSatish Balay       dk[13] += uik[15] * u[10] + uik[16] * u[11] + uik[17] * u[12] + uik[18] * u[13] + uik[19] * u[14];
14481278733SSatish Balay       dk[14] += uik[20] * u[10] + uik[21] * u[11] + uik[22] * u[12] + uik[23] * u[13] + uik[24] * u[14];
14581278733SSatish Balay 
14681278733SSatish Balay       dk[15] += uik[0] * u[15] + uik[1] * u[16] + uik[2] * u[17] + uik[3] * u[18] + uik[4] * u[19];
14781278733SSatish Balay       dk[16] += uik[5] * u[15] + uik[6] * u[16] + uik[7] * u[17] + uik[8] * u[18] + uik[9] * u[19];
14881278733SSatish Balay       dk[17] += uik[10] * u[15] + uik[11] * u[16] + uik[12] * u[17] + uik[13] * u[18] + uik[14] * u[19];
14981278733SSatish Balay       dk[18] += uik[15] * u[15] + uik[16] * u[16] + uik[17] * u[17] + uik[18] * u[18] + uik[19] * u[19];
15081278733SSatish Balay       dk[19] += uik[20] * u[15] + uik[21] * u[16] + uik[22] * u[17] + uik[23] * u[18] + uik[24] * u[19];
15181278733SSatish Balay 
15281278733SSatish Balay       dk[20] += uik[0] * u[20] + uik[1] * u[21] + uik[2] * u[22] + uik[3] * u[23] + uik[4] * u[24];
15381278733SSatish Balay       dk[21] += uik[5] * u[20] + uik[6] * u[21] + uik[7] * u[22] + uik[8] * u[23] + uik[9] * u[24];
15481278733SSatish Balay       dk[22] += uik[10] * u[20] + uik[11] * u[21] + uik[12] * u[22] + uik[13] * u[23] + uik[14] * u[24];
15581278733SSatish Balay       dk[23] += uik[15] * u[20] + uik[16] * u[21] + uik[17] * u[22] + uik[18] * u[23] + uik[19] * u[24];
15681278733SSatish Balay       dk[24] += uik[20] * u[20] + uik[21] * u[21] + uik[22] * u[22] + uik[23] * u[23] + uik[24] * u[24];
15781278733SSatish Balay 
1589566063dSJacob Faibussowitsch       PetscCall(PetscLogFlops(125.0 * 4.0));
159187a9f4bSHong Zhang 
16081278733SSatish Balay       /* update -U(i,k) */
1619566063dSJacob Faibussowitsch       PetscCall(PetscArraycpy(ba + ili * 25, uik, 25));
16281278733SSatish Balay 
16381278733SSatish Balay       /* add multiple of row i to k-th row ... */
1649371c9d4SSatish Balay       jmin = ili + 1;
1659371c9d4SSatish Balay       jmax = bi[i + 1];
16681278733SSatish Balay       if (jmin < jmax) {
16781278733SSatish Balay         for (j = jmin; j < jmax; j++) {
16881278733SSatish Balay           /* rtmp += -U(i,k)^T * U_bar(i,j) */
16981278733SSatish Balay           rtmp_ptr = rtmp + bj[j] * 25;
17081278733SSatish Balay           u        = ba + j * 25;
17181278733SSatish Balay           rtmp_ptr[0] += uik[0] * u[0] + uik[1] * u[1] + uik[2] * u[2] + uik[3] * u[3] + uik[4] * u[4];
17281278733SSatish Balay           rtmp_ptr[1] += uik[5] * u[0] + uik[6] * u[1] + uik[7] * u[2] + uik[8] * u[3] + uik[9] * u[4];
17381278733SSatish Balay           rtmp_ptr[2] += uik[10] * u[0] + uik[11] * u[1] + uik[12] * u[2] + uik[13] * u[3] + uik[14] * u[4];
17481278733SSatish Balay           rtmp_ptr[3] += uik[15] * u[0] + uik[16] * u[1] + uik[17] * u[2] + uik[18] * u[3] + uik[19] * u[4];
17581278733SSatish Balay           rtmp_ptr[4] += uik[20] * u[0] + uik[21] * u[1] + uik[22] * u[2] + uik[23] * u[3] + uik[24] * u[4];
17681278733SSatish Balay 
17781278733SSatish Balay           rtmp_ptr[5] += uik[0] * u[5] + uik[1] * u[6] + uik[2] * u[7] + uik[3] * u[8] + uik[4] * u[9];
17881278733SSatish Balay           rtmp_ptr[6] += uik[5] * u[5] + uik[6] * u[6] + uik[7] * u[7] + uik[8] * u[8] + uik[9] * u[9];
17981278733SSatish Balay           rtmp_ptr[7] += uik[10] * u[5] + uik[11] * u[6] + uik[12] * u[7] + uik[13] * u[8] + uik[14] * u[9];
18081278733SSatish Balay           rtmp_ptr[8] += uik[15] * u[5] + uik[16] * u[6] + uik[17] * u[7] + uik[18] * u[8] + uik[19] * u[9];
18181278733SSatish Balay           rtmp_ptr[9] += uik[20] * u[5] + uik[21] * u[6] + uik[22] * u[7] + uik[23] * u[8] + uik[24] * u[9];
18281278733SSatish Balay 
18381278733SSatish Balay           rtmp_ptr[10] += uik[0] * u[10] + uik[1] * u[11] + uik[2] * u[12] + uik[3] * u[13] + uik[4] * u[14];
18481278733SSatish Balay           rtmp_ptr[11] += uik[5] * u[10] + uik[6] * u[11] + uik[7] * u[12] + uik[8] * u[13] + uik[9] * u[14];
18581278733SSatish Balay           rtmp_ptr[12] += uik[10] * u[10] + uik[11] * u[11] + uik[12] * u[12] + uik[13] * u[13] + uik[14] * u[14];
18681278733SSatish Balay           rtmp_ptr[13] += uik[15] * u[10] + uik[16] * u[11] + uik[17] * u[12] + uik[18] * u[13] + uik[19] * u[14];
18781278733SSatish Balay           rtmp_ptr[14] += uik[20] * u[10] + uik[21] * u[11] + uik[22] * u[12] + uik[23] * u[13] + uik[24] * u[14];
18881278733SSatish Balay 
18981278733SSatish Balay           rtmp_ptr[15] += uik[0] * u[15] + uik[1] * u[16] + uik[2] * u[17] + uik[3] * u[18] + uik[4] * u[19];
19081278733SSatish Balay           rtmp_ptr[16] += uik[5] * u[15] + uik[6] * u[16] + uik[7] * u[17] + uik[8] * u[18] + uik[9] * u[19];
19181278733SSatish Balay           rtmp_ptr[17] += uik[10] * u[15] + uik[11] * u[16] + uik[12] * u[17] + uik[13] * u[18] + uik[14] * u[19];
19281278733SSatish Balay           rtmp_ptr[18] += uik[15] * u[15] + uik[16] * u[16] + uik[17] * u[17] + uik[18] * u[18] + uik[19] * u[19];
19381278733SSatish Balay           rtmp_ptr[19] += uik[20] * u[15] + uik[21] * u[16] + uik[22] * u[17] + uik[23] * u[18] + uik[24] * u[19];
19481278733SSatish Balay 
19581278733SSatish Balay           rtmp_ptr[20] += uik[0] * u[20] + uik[1] * u[21] + uik[2] * u[22] + uik[3] * u[23] + uik[4] * u[24];
19681278733SSatish Balay           rtmp_ptr[21] += uik[5] * u[20] + uik[6] * u[21] + uik[7] * u[22] + uik[8] * u[23] + uik[9] * u[24];
19781278733SSatish Balay           rtmp_ptr[22] += uik[10] * u[20] + uik[11] * u[21] + uik[12] * u[22] + uik[13] * u[23] + uik[14] * u[24];
19881278733SSatish Balay           rtmp_ptr[23] += uik[15] * u[20] + uik[16] * u[21] + uik[17] * u[22] + uik[18] * u[23] + uik[19] * u[24];
19981278733SSatish Balay           rtmp_ptr[24] += uik[20] * u[20] + uik[21] * u[21] + uik[22] * u[22] + uik[23] * u[23] + uik[24] * u[24];
20081278733SSatish Balay         }
2019566063dSJacob Faibussowitsch         PetscCall(PetscLogFlops(2.0 * 125.0 * (jmax - jmin)));
20281278733SSatish Balay 
20381278733SSatish Balay         /* ... add i to row list for next nonzero entry */
20481278733SSatish Balay         il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */
20581278733SSatish Balay         j     = bj[jmin];
2069371c9d4SSatish Balay         jl[i] = jl[j];
2079371c9d4SSatish Balay         jl[j] = i; /* update jl */
20881278733SSatish Balay       }
20981278733SSatish Balay       i = nexti;
21081278733SSatish Balay     }
21181278733SSatish Balay 
21281278733SSatish Balay     /* save nonzero entries in k-th row of U ... */
21381278733SSatish Balay 
21481278733SSatish Balay     /* invert diagonal block */
21581278733SSatish Balay     d = ba + k * 25;
2169566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(d, dk, 25));
2179566063dSJacob Faibussowitsch     PetscCall(PetscKernel_A_gets_inverse_A_5(d, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
2187b6c816cSBarry Smith     if (zeropivotdetected) C->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
21981278733SSatish Balay 
2209371c9d4SSatish Balay     jmin = bi[k];
2219371c9d4SSatish Balay     jmax = bi[k + 1];
22281278733SSatish Balay     if (jmin < jmax) {
22381278733SSatish Balay       for (j = jmin; j < jmax; j++) {
22481278733SSatish Balay         vj       = bj[j]; /* block col. index of U */
22581278733SSatish Balay         u        = ba + j * 25;
22681278733SSatish Balay         rtmp_ptr = rtmp + vj * 25;
22781278733SSatish Balay         for (k1 = 0; k1 < 25; k1++) {
22881278733SSatish Balay           *u++        = *rtmp_ptr;
22981278733SSatish Balay           *rtmp_ptr++ = 0.0;
23081278733SSatish Balay         }
23181278733SSatish Balay       }
23281278733SSatish Balay 
23381278733SSatish Balay       /* ... add k to row list for first nonzero entry in k-th row */
23481278733SSatish Balay       il[k] = jmin;
23581278733SSatish Balay       i     = bj[jmin];
2369371c9d4SSatish Balay       jl[k] = jl[i];
2379371c9d4SSatish Balay       jl[i] = k;
23881278733SSatish Balay     }
23981278733SSatish Balay   }
24081278733SSatish Balay 
2419566063dSJacob Faibussowitsch   PetscCall(PetscFree(rtmp));
2429566063dSJacob Faibussowitsch   PetscCall(PetscFree2(il, jl));
2439566063dSJacob Faibussowitsch   PetscCall(PetscFree2(dk, uik));
2441baa6e33SBarry Smith   if (a->permute) PetscCall(PetscFree(aa));
24581278733SSatish Balay 
2469566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(perm, &perm_ptr));
24726fbe8dcSKarl Rupp 
2484f79d315SHong Zhang   C->ops->solve          = MatSolve_SeqSBAIJ_5_inplace;
2494f79d315SHong Zhang   C->ops->solvetranspose = MatSolve_SeqSBAIJ_5_inplace;
25081278733SSatish Balay   C->assembled           = PETSC_TRUE;
25181278733SSatish Balay   C->preallocated        = PETSC_TRUE;
25226fbe8dcSKarl Rupp 
2539566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(1.3333 * 125 * b->mbs)); /* from inverting diagonal blocks */
254*3ba16761SJacob Faibussowitsch   PetscFunctionReturn(PETSC_SUCCESS);
25581278733SSatish Balay }
256