xref: /petsc/src/mat/impls/baij/seq/baijsolvtran4.c (revision 31d78bcd2b98084dc1368b20eb1129c8b9fb39fe)
12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith #include <petsc/private/kernels/blockinvert.h>
32c733ed4SBarry Smith 
MatSolveTranspose_SeqBAIJ_4_inplace(Mat A,Vec bb,Vec xx)4d71ae5a4SJacob Faibussowitsch PetscErrorCode MatSolveTranspose_SeqBAIJ_4_inplace(Mat A, Vec bb, Vec xx)
5d71ae5a4SJacob Faibussowitsch {
62c733ed4SBarry Smith   Mat_SeqBAIJ       *a     = (Mat_SeqBAIJ *)A->data;
72c733ed4SBarry Smith   IS                 iscol = a->col, isrow = a->row;
82c733ed4SBarry Smith   const PetscInt    *r, *c, *rout, *cout;
92c733ed4SBarry Smith   const PetscInt    *diag = a->diag, n = a->mbs, *vi, *ai = a->i, *aj = a->j;
102c733ed4SBarry Smith   PetscInt           i, nz, idx, idt, ii, ic, ir, oidx;
112c733ed4SBarry Smith   const MatScalar   *aa = a->a, *v;
122c733ed4SBarry Smith   PetscScalar        s1, s2, s3, s4, x1, x2, x3, x4, *x, *t;
132c733ed4SBarry Smith   const PetscScalar *b;
142c733ed4SBarry Smith 
152c733ed4SBarry Smith   PetscFunctionBegin;
169566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb, &b));
179566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx, &x));
182c733ed4SBarry Smith   t = a->solve_work;
192c733ed4SBarry Smith 
209371c9d4SSatish Balay   PetscCall(ISGetIndices(isrow, &rout));
219371c9d4SSatish Balay   r = rout;
229371c9d4SSatish Balay   PetscCall(ISGetIndices(iscol, &cout));
239371c9d4SSatish Balay   c = cout;
242c733ed4SBarry Smith 
252c733ed4SBarry Smith   /* copy the b into temp work space according to permutation */
262c733ed4SBarry Smith   ii = 0;
272c733ed4SBarry Smith   for (i = 0; i < n; i++) {
282c733ed4SBarry Smith     ic        = 4 * c[i];
292c733ed4SBarry Smith     t[ii]     = b[ic];
302c733ed4SBarry Smith     t[ii + 1] = b[ic + 1];
312c733ed4SBarry Smith     t[ii + 2] = b[ic + 2];
322c733ed4SBarry Smith     t[ii + 3] = b[ic + 3];
332c733ed4SBarry Smith     ii += 4;
342c733ed4SBarry Smith   }
352c733ed4SBarry Smith 
362c733ed4SBarry Smith   /* forward solve the U^T */
372c733ed4SBarry Smith   idx = 0;
382c733ed4SBarry Smith   for (i = 0; i < n; i++) {
392c733ed4SBarry Smith     v = aa + 16 * diag[i];
402c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
419371c9d4SSatish Balay     x1 = t[idx];
429371c9d4SSatish Balay     x2 = t[1 + idx];
439371c9d4SSatish Balay     x3 = t[2 + idx];
449371c9d4SSatish Balay     x4 = t[3 + idx];
452c733ed4SBarry Smith     s1 = v[0] * x1 + v[1] * x2 + v[2] * x3 + v[3] * x4;
462c733ed4SBarry Smith     s2 = v[4] * x1 + v[5] * x2 + v[6] * x3 + v[7] * x4;
472c733ed4SBarry Smith     s3 = v[8] * x1 + v[9] * x2 + v[10] * x3 + v[11] * x4;
482c733ed4SBarry Smith     s4 = v[12] * x1 + v[13] * x2 + v[14] * x3 + v[15] * x4;
492c733ed4SBarry Smith     v += 16;
502c733ed4SBarry Smith 
512c733ed4SBarry Smith     vi = aj + diag[i] + 1;
522c733ed4SBarry Smith     nz = ai[i + 1] - diag[i] - 1;
532c733ed4SBarry Smith     while (nz--) {
542c733ed4SBarry Smith       oidx = 4 * (*vi++);
552c733ed4SBarry Smith       t[oidx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4;
562c733ed4SBarry Smith       t[oidx + 1] -= v[4] * s1 + v[5] * s2 + v[6] * s3 + v[7] * s4;
572c733ed4SBarry Smith       t[oidx + 2] -= v[8] * s1 + v[9] * s2 + v[10] * s3 + v[11] * s4;
582c733ed4SBarry Smith       t[oidx + 3] -= v[12] * s1 + v[13] * s2 + v[14] * s3 + v[15] * s4;
592c733ed4SBarry Smith       v += 16;
602c733ed4SBarry Smith     }
619371c9d4SSatish Balay     t[idx]     = s1;
629371c9d4SSatish Balay     t[1 + idx] = s2;
639371c9d4SSatish Balay     t[2 + idx] = s3;
649371c9d4SSatish Balay     t[3 + idx] = s4;
652c733ed4SBarry Smith     idx += 4;
662c733ed4SBarry Smith   }
672c733ed4SBarry Smith   /* backward solve the L^T */
682c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
692c733ed4SBarry Smith     v   = aa + 16 * diag[i] - 16;
702c733ed4SBarry Smith     vi  = aj + diag[i] - 1;
712c733ed4SBarry Smith     nz  = diag[i] - ai[i];
722c733ed4SBarry Smith     idt = 4 * i;
739371c9d4SSatish Balay     s1  = t[idt];
749371c9d4SSatish Balay     s2  = t[1 + idt];
759371c9d4SSatish Balay     s3  = t[2 + idt];
769371c9d4SSatish Balay     s4  = t[3 + idt];
772c733ed4SBarry Smith     while (nz--) {
782c733ed4SBarry Smith       idx = 4 * (*vi--);
792c733ed4SBarry Smith       t[idx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4;
802c733ed4SBarry Smith       t[idx + 1] -= v[4] * s1 + v[5] * s2 + v[6] * s3 + v[7] * s4;
812c733ed4SBarry Smith       t[idx + 2] -= v[8] * s1 + v[9] * s2 + v[10] * s3 + v[11] * s4;
822c733ed4SBarry Smith       t[idx + 3] -= v[12] * s1 + v[13] * s2 + v[14] * s3 + v[15] * s4;
832c733ed4SBarry Smith       v -= 16;
842c733ed4SBarry Smith     }
852c733ed4SBarry Smith   }
862c733ed4SBarry Smith 
872c733ed4SBarry Smith   /* copy t into x according to permutation */
882c733ed4SBarry Smith   ii = 0;
892c733ed4SBarry Smith   for (i = 0; i < n; i++) {
902c733ed4SBarry Smith     ir        = 4 * r[i];
912c733ed4SBarry Smith     x[ir]     = t[ii];
922c733ed4SBarry Smith     x[ir + 1] = t[ii + 1];
932c733ed4SBarry Smith     x[ir + 2] = t[ii + 2];
942c733ed4SBarry Smith     x[ir + 3] = t[ii + 3];
952c733ed4SBarry Smith     ii += 4;
962c733ed4SBarry Smith   }
972c733ed4SBarry Smith 
989566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow, &rout));
999566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol, &cout));
1009566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb, &b));
1019566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
1029566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * 16 * (a->nz) - 4.0 * A->cmap->n));
103*3ba16761SJacob Faibussowitsch   PetscFunctionReturn(PETSC_SUCCESS);
1042c733ed4SBarry Smith }
1052c733ed4SBarry Smith 
MatSolveTranspose_SeqBAIJ_4(Mat A,Vec bb,Vec xx)106d71ae5a4SJacob Faibussowitsch PetscErrorCode MatSolveTranspose_SeqBAIJ_4(Mat A, Vec bb, Vec xx)
107d71ae5a4SJacob Faibussowitsch {
1082c733ed4SBarry Smith   Mat_SeqBAIJ       *a     = (Mat_SeqBAIJ *)A->data;
1092c733ed4SBarry Smith   IS                 iscol = a->col, isrow = a->row;
1102c733ed4SBarry Smith   const PetscInt     n = a->mbs, *vi, *ai = a->i, *aj = a->j, *diag = a->diag;
1112c733ed4SBarry Smith   const PetscInt    *r, *c, *rout, *cout;
1122c733ed4SBarry Smith   PetscInt           nz, idx, idt, j, i, oidx, ii, ic, ir;
1132c733ed4SBarry Smith   const PetscInt     bs = A->rmap->bs, bs2 = a->bs2;
1142c733ed4SBarry Smith   const MatScalar   *aa = a->a, *v;
1152c733ed4SBarry Smith   PetscScalar        s1, s2, s3, s4, x1, x2, x3, x4, *x, *t;
1162c733ed4SBarry Smith   const PetscScalar *b;
1172c733ed4SBarry Smith 
1182c733ed4SBarry Smith   PetscFunctionBegin;
1199566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb, &b));
1209566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx, &x));
1212c733ed4SBarry Smith   t = a->solve_work;
1222c733ed4SBarry Smith 
1239371c9d4SSatish Balay   PetscCall(ISGetIndices(isrow, &rout));
1249371c9d4SSatish Balay   r = rout;
1259371c9d4SSatish Balay   PetscCall(ISGetIndices(iscol, &cout));
1269371c9d4SSatish Balay   c = cout;
1272c733ed4SBarry Smith 
1282c733ed4SBarry Smith   /* copy b into temp work space according to permutation */
1292c733ed4SBarry Smith   for (i = 0; i < n; i++) {
1309371c9d4SSatish Balay     ii        = bs * i;
1319371c9d4SSatish Balay     ic        = bs * c[i];
1329371c9d4SSatish Balay     t[ii]     = b[ic];
1339371c9d4SSatish Balay     t[ii + 1] = b[ic + 1];
1349371c9d4SSatish Balay     t[ii + 2] = b[ic + 2];
1359371c9d4SSatish Balay     t[ii + 3] = b[ic + 3];
1362c733ed4SBarry Smith   }
1372c733ed4SBarry Smith 
1382c733ed4SBarry Smith   /* forward solve the U^T */
1392c733ed4SBarry Smith   idx = 0;
1402c733ed4SBarry Smith   for (i = 0; i < n; i++) {
1412c733ed4SBarry Smith     v = aa + bs2 * diag[i];
1422c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
1439371c9d4SSatish Balay     x1 = t[idx];
1449371c9d4SSatish Balay     x2 = t[1 + idx];
1459371c9d4SSatish Balay     x3 = t[2 + idx];
1469371c9d4SSatish Balay     x4 = t[3 + idx];
1472c733ed4SBarry Smith     s1 = v[0] * x1 + v[1] * x2 + v[2] * x3 + v[3] * x4;
1482c733ed4SBarry Smith     s2 = v[4] * x1 + v[5] * x2 + v[6] * x3 + v[7] * x4;
1492c733ed4SBarry Smith     s3 = v[8] * x1 + v[9] * x2 + v[10] * x3 + v[11] * x4;
1502c733ed4SBarry Smith     s4 = v[12] * x1 + v[13] * x2 + v[14] * x3 + v[15] * x4;
1512c733ed4SBarry Smith     v -= bs2;
1522c733ed4SBarry Smith 
1532c733ed4SBarry Smith     vi = aj + diag[i] - 1;
1542c733ed4SBarry Smith     nz = diag[i] - diag[i + 1] - 1;
1552c733ed4SBarry Smith     for (j = 0; j > -nz; j--) {
1562c733ed4SBarry Smith       oidx = bs * vi[j];
1572c733ed4SBarry Smith       t[oidx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4;
1582c733ed4SBarry Smith       t[oidx + 1] -= v[4] * s1 + v[5] * s2 + v[6] * s3 + v[7] * s4;
1592c733ed4SBarry Smith       t[oidx + 2] -= v[8] * s1 + v[9] * s2 + v[10] * s3 + v[11] * s4;
1602c733ed4SBarry Smith       t[oidx + 3] -= v[12] * s1 + v[13] * s2 + v[14] * s3 + v[15] * s4;
1612c733ed4SBarry Smith       v -= bs2;
1622c733ed4SBarry Smith     }
1639371c9d4SSatish Balay     t[idx]     = s1;
1649371c9d4SSatish Balay     t[1 + idx] = s2;
1659371c9d4SSatish Balay     t[2 + idx] = s3;
1669371c9d4SSatish Balay     t[3 + idx] = s4;
1672c733ed4SBarry Smith     idx += bs;
1682c733ed4SBarry Smith   }
1692c733ed4SBarry Smith   /* backward solve the L^T */
1702c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
1712c733ed4SBarry Smith     v   = aa + bs2 * ai[i];
1722c733ed4SBarry Smith     vi  = aj + ai[i];
1732c733ed4SBarry Smith     nz  = ai[i + 1] - ai[i];
1742c733ed4SBarry Smith     idt = bs * i;
1759371c9d4SSatish Balay     s1  = t[idt];
1769371c9d4SSatish Balay     s2  = t[1 + idt];
1779371c9d4SSatish Balay     s3  = t[2 + idt];
1789371c9d4SSatish Balay     s4  = t[3 + idt];
1792c733ed4SBarry Smith     for (j = 0; j < nz; j++) {
1802c733ed4SBarry Smith       idx = bs * vi[j];
1812c733ed4SBarry Smith       t[idx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4;
1822c733ed4SBarry Smith       t[idx + 1] -= v[4] * s1 + v[5] * s2 + v[6] * s3 + v[7] * s4;
1832c733ed4SBarry Smith       t[idx + 2] -= v[8] * s1 + v[9] * s2 + v[10] * s3 + v[11] * s4;
1842c733ed4SBarry Smith       t[idx + 3] -= v[12] * s1 + v[13] * s2 + v[14] * s3 + v[15] * s4;
1852c733ed4SBarry Smith       v += bs2;
1862c733ed4SBarry Smith     }
1872c733ed4SBarry Smith   }
1882c733ed4SBarry Smith 
1892c733ed4SBarry Smith   /* copy t into x according to permutation */
1902c733ed4SBarry Smith   for (i = 0; i < n; i++) {
1919371c9d4SSatish Balay     ii        = bs * i;
1929371c9d4SSatish Balay     ir        = bs * r[i];
1939371c9d4SSatish Balay     x[ir]     = t[ii];
1949371c9d4SSatish Balay     x[ir + 1] = t[ii + 1];
1959371c9d4SSatish Balay     x[ir + 2] = t[ii + 2];
1969371c9d4SSatish Balay     x[ir + 3] = t[ii + 3];
1972c733ed4SBarry Smith   }
1982c733ed4SBarry Smith 
1999566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow, &rout));
2009566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol, &cout));
2019566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb, &b));
2029566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
2039566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * bs2 * (a->nz) - bs * A->cmap->n));
204*3ba16761SJacob Faibussowitsch   PetscFunctionReturn(PETSC_SUCCESS);
2052c733ed4SBarry Smith }
206