12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith
MatSolveTranspose_SeqBAIJ_7_NaturalOrdering_inplace(Mat A,Vec bb,Vec xx)3d71ae5a4SJacob Faibussowitsch PetscErrorCode MatSolveTranspose_SeqBAIJ_7_NaturalOrdering_inplace(Mat A, Vec bb, Vec xx)
4d71ae5a4SJacob Faibussowitsch {
52c733ed4SBarry Smith Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
62c733ed4SBarry Smith const PetscInt *diag = a->diag, n = a->mbs, *vi, *ai = a->i, *aj = a->j;
72c733ed4SBarry Smith PetscInt i, nz, idx, idt, oidx;
82c733ed4SBarry Smith const MatScalar *aa = a->a, *v;
92c733ed4SBarry Smith PetscScalar s1, s2, s3, s4, s5, s6, s7, x1, x2, x3, x4, x5, x6, x7, *x;
102c733ed4SBarry Smith
112c733ed4SBarry Smith PetscFunctionBegin;
129566063dSJacob Faibussowitsch PetscCall(VecCopy(bb, xx));
139566063dSJacob Faibussowitsch PetscCall(VecGetArray(xx, &x));
142c733ed4SBarry Smith
152c733ed4SBarry Smith /* forward solve the U^T */
162c733ed4SBarry Smith idx = 0;
172c733ed4SBarry Smith for (i = 0; i < n; i++) {
182c733ed4SBarry Smith v = aa + 49 * diag[i];
192c733ed4SBarry Smith /* multiply by the inverse of the block diagonal */
209371c9d4SSatish Balay x1 = x[idx];
219371c9d4SSatish Balay x2 = x[1 + idx];
229371c9d4SSatish Balay x3 = x[2 + idx];
239371c9d4SSatish Balay x4 = x[3 + idx];
249371c9d4SSatish Balay x5 = x[4 + idx];
259371c9d4SSatish Balay x6 = x[5 + idx];
269371c9d4SSatish Balay x7 = x[6 + idx];
272c733ed4SBarry Smith s1 = v[0] * x1 + v[1] * x2 + v[2] * x3 + v[3] * x4 + v[4] * x5 + v[5] * x6 + v[6] * x7;
282c733ed4SBarry Smith s2 = v[7] * x1 + v[8] * x2 + v[9] * x3 + v[10] * x4 + v[11] * x5 + v[12] * x6 + v[13] * x7;
292c733ed4SBarry Smith s3 = v[14] * x1 + v[15] * x2 + v[16] * x3 + v[17] * x4 + v[18] * x5 + v[19] * x6 + v[20] * x7;
302c733ed4SBarry Smith s4 = v[21] * x1 + v[22] * x2 + v[23] * x3 + v[24] * x4 + v[25] * x5 + v[26] * x6 + v[27] * x7;
312c733ed4SBarry Smith s5 = v[28] * x1 + v[29] * x2 + v[30] * x3 + v[31] * x4 + v[32] * x5 + v[33] * x6 + v[34] * x7;
322c733ed4SBarry Smith s6 = v[35] * x1 + v[36] * x2 + v[37] * x3 + v[38] * x4 + v[39] * x5 + v[40] * x6 + v[41] * x7;
332c733ed4SBarry Smith s7 = v[42] * x1 + v[43] * x2 + v[44] * x3 + v[45] * x4 + v[46] * x5 + v[47] * x6 + v[48] * x7;
342c733ed4SBarry Smith v += 49;
352c733ed4SBarry Smith
362c733ed4SBarry Smith vi = aj + diag[i] + 1;
372c733ed4SBarry Smith nz = ai[i + 1] - diag[i] - 1;
382c733ed4SBarry Smith while (nz--) {
392c733ed4SBarry Smith oidx = 7 * (*vi++);
402c733ed4SBarry Smith x[oidx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4 + v[4] * s5 + v[5] * s6 + v[6] * s7;
412c733ed4SBarry Smith x[oidx + 1] -= v[7] * s1 + v[8] * s2 + v[9] * s3 + v[10] * s4 + v[11] * s5 + v[12] * s6 + v[13] * s7;
422c733ed4SBarry Smith x[oidx + 2] -= v[14] * s1 + v[15] * s2 + v[16] * s3 + v[17] * s4 + v[18] * s5 + v[19] * s6 + v[20] * s7;
432c733ed4SBarry Smith x[oidx + 3] -= v[21] * s1 + v[22] * s2 + v[23] * s3 + v[24] * s4 + v[25] * s5 + v[26] * s6 + v[27] * s7;
442c733ed4SBarry Smith x[oidx + 4] -= v[28] * s1 + v[29] * s2 + v[30] * s3 + v[31] * s4 + v[32] * s5 + v[33] * s6 + v[34] * s7;
452c733ed4SBarry Smith x[oidx + 5] -= v[35] * s1 + v[36] * s2 + v[37] * s3 + v[38] * s4 + v[39] * s5 + v[40] * s6 + v[41] * s7;
462c733ed4SBarry Smith x[oidx + 6] -= v[42] * s1 + v[43] * s2 + v[44] * s3 + v[45] * s4 + v[46] * s5 + v[47] * s6 + v[48] * s7;
472c733ed4SBarry Smith v += 49;
482c733ed4SBarry Smith }
499371c9d4SSatish Balay x[idx] = s1;
509371c9d4SSatish Balay x[1 + idx] = s2;
519371c9d4SSatish Balay x[2 + idx] = s3;
529371c9d4SSatish Balay x[3 + idx] = s4;
539371c9d4SSatish Balay x[4 + idx] = s5;
549371c9d4SSatish Balay x[5 + idx] = s6;
559371c9d4SSatish Balay x[6 + idx] = s7;
562c733ed4SBarry Smith idx += 7;
572c733ed4SBarry Smith }
582c733ed4SBarry Smith /* backward solve the L^T */
592c733ed4SBarry Smith for (i = n - 1; i >= 0; i--) {
602c733ed4SBarry Smith v = aa + 49 * diag[i] - 49;
612c733ed4SBarry Smith vi = aj + diag[i] - 1;
622c733ed4SBarry Smith nz = diag[i] - ai[i];
632c733ed4SBarry Smith idt = 7 * i;
649371c9d4SSatish Balay s1 = x[idt];
659371c9d4SSatish Balay s2 = x[1 + idt];
669371c9d4SSatish Balay s3 = x[2 + idt];
679371c9d4SSatish Balay s4 = x[3 + idt];
689371c9d4SSatish Balay s5 = x[4 + idt];
699371c9d4SSatish Balay s6 = x[5 + idt];
709371c9d4SSatish Balay s7 = x[6 + idt];
712c733ed4SBarry Smith while (nz--) {
722c733ed4SBarry Smith idx = 7 * (*vi--);
732c733ed4SBarry Smith x[idx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4 + v[4] * s5 + v[5] * s6 + v[6] * s7;
742c733ed4SBarry Smith x[idx + 1] -= v[7] * s1 + v[8] * s2 + v[9] * s3 + v[10] * s4 + v[11] * s5 + v[12] * s6 + v[13] * s7;
752c733ed4SBarry Smith x[idx + 2] -= v[14] * s1 + v[15] * s2 + v[16] * s3 + v[17] * s4 + v[18] * s5 + v[19] * s6 + v[20] * s7;
762c733ed4SBarry Smith x[idx + 3] -= v[21] * s1 + v[22] * s2 + v[23] * s3 + v[24] * s4 + v[25] * s5 + v[26] * s6 + v[27] * s7;
772c733ed4SBarry Smith x[idx + 4] -= v[28] * s1 + v[29] * s2 + v[30] * s3 + v[31] * s4 + v[32] * s5 + v[33] * s6 + v[34] * s7;
782c733ed4SBarry Smith x[idx + 5] -= v[35] * s1 + v[36] * s2 + v[37] * s3 + v[38] * s4 + v[39] * s5 + v[40] * s6 + v[41] * s7;
792c733ed4SBarry Smith x[idx + 6] -= v[42] * s1 + v[43] * s2 + v[44] * s3 + v[45] * s4 + v[46] * s5 + v[47] * s6 + v[48] * s7;
802c733ed4SBarry Smith v -= 49;
812c733ed4SBarry Smith }
822c733ed4SBarry Smith }
839566063dSJacob Faibussowitsch PetscCall(VecRestoreArray(xx, &x));
849566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(2.0 * 49 * (a->nz) - 7.0 * A->cmap->n));
85*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS);
862c733ed4SBarry Smith }
MatSolveTranspose_SeqBAIJ_7_NaturalOrdering(Mat A,Vec bb,Vec xx)87d71ae5a4SJacob Faibussowitsch PetscErrorCode MatSolveTranspose_SeqBAIJ_7_NaturalOrdering(Mat A, Vec bb, Vec xx)
88d71ae5a4SJacob Faibussowitsch {
892c733ed4SBarry Smith Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
902c733ed4SBarry Smith const PetscInt n = a->mbs, *vi, *ai = a->i, *aj = a->j, *diag = a->diag;
912c733ed4SBarry Smith PetscInt nz, idx, idt, j, i, oidx;
922c733ed4SBarry Smith const PetscInt bs = A->rmap->bs, bs2 = a->bs2;
932c733ed4SBarry Smith const MatScalar *aa = a->a, *v;
942c733ed4SBarry Smith PetscScalar s1, s2, s3, s4, s5, s6, s7, x1, x2, x3, x4, x5, x6, x7, *x;
952c733ed4SBarry Smith
962c733ed4SBarry Smith PetscFunctionBegin;
979566063dSJacob Faibussowitsch PetscCall(VecCopy(bb, xx));
989566063dSJacob Faibussowitsch PetscCall(VecGetArray(xx, &x));
992c733ed4SBarry Smith
1002c733ed4SBarry Smith /* forward solve the U^T */
1012c733ed4SBarry Smith idx = 0;
1022c733ed4SBarry Smith for (i = 0; i < n; i++) {
1032c733ed4SBarry Smith v = aa + bs2 * diag[i];
1042c733ed4SBarry Smith /* multiply by the inverse of the block diagonal */
1059371c9d4SSatish Balay x1 = x[idx];
1069371c9d4SSatish Balay x2 = x[1 + idx];
1079371c9d4SSatish Balay x3 = x[2 + idx];
1089371c9d4SSatish Balay x4 = x[3 + idx];
1099371c9d4SSatish Balay x5 = x[4 + idx];
1109371c9d4SSatish Balay x6 = x[5 + idx];
1119371c9d4SSatish Balay x7 = x[6 + idx];
1122c733ed4SBarry Smith s1 = v[0] * x1 + v[1] * x2 + v[2] * x3 + v[3] * x4 + v[4] * x5 + v[5] * x6 + v[6] * x7;
1132c733ed4SBarry Smith s2 = v[7] * x1 + v[8] * x2 + v[9] * x3 + v[10] * x4 + v[11] * x5 + v[12] * x6 + v[13] * x7;
1142c733ed4SBarry Smith s3 = v[14] * x1 + v[15] * x2 + v[16] * x3 + v[17] * x4 + v[18] * x5 + v[19] * x6 + v[20] * x7;
1152c733ed4SBarry Smith s4 = v[21] * x1 + v[22] * x2 + v[23] * x3 + v[24] * x4 + v[25] * x5 + v[26] * x6 + v[27] * x7;
1162c733ed4SBarry Smith s5 = v[28] * x1 + v[29] * x2 + v[30] * x3 + v[31] * x4 + v[32] * x5 + v[33] * x6 + v[34] * x7;
1172c733ed4SBarry Smith s6 = v[35] * x1 + v[36] * x2 + v[37] * x3 + v[38] * x4 + v[39] * x5 + v[40] * x6 + v[41] * x7;
1182c733ed4SBarry Smith s7 = v[42] * x1 + v[43] * x2 + v[44] * x3 + v[45] * x4 + v[46] * x5 + v[47] * x6 + v[48] * x7;
1192c733ed4SBarry Smith v -= bs2;
1202c733ed4SBarry Smith vi = aj + diag[i] - 1;
1212c733ed4SBarry Smith nz = diag[i] - diag[i + 1] - 1;
1222c733ed4SBarry Smith for (j = 0; j > -nz; j--) {
1232c733ed4SBarry Smith oidx = bs * vi[j];
1242c733ed4SBarry Smith x[oidx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4 + v[4] * s5 + v[5] * s6 + v[6] * s7;
1252c733ed4SBarry Smith x[oidx + 1] -= v[7] * s1 + v[8] * s2 + v[9] * s3 + v[10] * s4 + v[11] * s5 + v[12] * s6 + v[13] * s7;
1262c733ed4SBarry Smith x[oidx + 2] -= v[14] * s1 + v[15] * s2 + v[16] * s3 + v[17] * s4 + v[18] * s5 + v[19] * s6 + v[20] * s7;
1272c733ed4SBarry Smith x[oidx + 3] -= v[21] * s1 + v[22] * s2 + v[23] * s3 + v[24] * s4 + v[25] * s5 + v[26] * s6 + v[27] * s7;
1282c733ed4SBarry Smith x[oidx + 4] -= v[28] * s1 + v[29] * s2 + v[30] * s3 + v[31] * s4 + v[32] * s5 + v[33] * s6 + v[34] * s7;
1292c733ed4SBarry Smith x[oidx + 5] -= v[35] * s1 + v[36] * s2 + v[37] * s3 + v[38] * s4 + v[39] * s5 + v[40] * s6 + v[41] * s7;
1302c733ed4SBarry Smith x[oidx + 6] -= v[42] * s1 + v[43] * s2 + v[44] * s3 + v[45] * s4 + v[46] * s5 + v[47] * s6 + v[48] * s7;
1312c733ed4SBarry Smith v -= bs2;
1322c733ed4SBarry Smith }
1339371c9d4SSatish Balay x[idx] = s1;
1349371c9d4SSatish Balay x[1 + idx] = s2;
1359371c9d4SSatish Balay x[2 + idx] = s3;
1369371c9d4SSatish Balay x[3 + idx] = s4;
1379371c9d4SSatish Balay x[4 + idx] = s5;
1389371c9d4SSatish Balay x[5 + idx] = s6;
1399371c9d4SSatish Balay x[6 + idx] = s7;
1402c733ed4SBarry Smith idx += bs;
1412c733ed4SBarry Smith }
1422c733ed4SBarry Smith /* backward solve the L^T */
1432c733ed4SBarry Smith for (i = n - 1; i >= 0; i--) {
1442c733ed4SBarry Smith v = aa + bs2 * ai[i];
1452c733ed4SBarry Smith vi = aj + ai[i];
1462c733ed4SBarry Smith nz = ai[i + 1] - ai[i];
1472c733ed4SBarry Smith idt = bs * i;
1489371c9d4SSatish Balay s1 = x[idt];
1499371c9d4SSatish Balay s2 = x[1 + idt];
1509371c9d4SSatish Balay s3 = x[2 + idt];
1519371c9d4SSatish Balay s4 = x[3 + idt];
1529371c9d4SSatish Balay s5 = x[4 + idt];
1539371c9d4SSatish Balay s6 = x[5 + idt];
1549371c9d4SSatish Balay s7 = x[6 + idt];
1552c733ed4SBarry Smith for (j = 0; j < nz; j++) {
1562c733ed4SBarry Smith idx = bs * vi[j];
1572c733ed4SBarry Smith x[idx] -= v[0] * s1 + v[1] * s2 + v[2] * s3 + v[3] * s4 + v[4] * s5 + v[5] * s6 + v[6] * s7;
1582c733ed4SBarry Smith x[idx + 1] -= v[7] * s1 + v[8] * s2 + v[9] * s3 + v[10] * s4 + v[11] * s5 + v[12] * s6 + v[13] * s7;
1592c733ed4SBarry Smith x[idx + 2] -= v[14] * s1 + v[15] * s2 + v[16] * s3 + v[17] * s4 + v[18] * s5 + v[19] * s6 + v[20] * s7;
1602c733ed4SBarry Smith x[idx + 3] -= v[21] * s1 + v[22] * s2 + v[23] * s3 + v[24] * s4 + v[25] * s5 + v[26] * s6 + v[27] * s7;
1612c733ed4SBarry Smith x[idx + 4] -= v[28] * s1 + v[29] * s2 + v[30] * s3 + v[31] * s4 + v[32] * s5 + v[33] * s6 + v[34] * s7;
1622c733ed4SBarry Smith x[idx + 5] -= v[35] * s1 + v[36] * s2 + v[37] * s3 + v[38] * s4 + v[39] * s5 + v[40] * s6 + v[41] * s7;
1632c733ed4SBarry Smith x[idx + 6] -= v[42] * s1 + v[43] * s2 + v[44] * s3 + v[45] * s4 + v[46] * s5 + v[47] * s6 + v[48] * s7;
1642c733ed4SBarry Smith v += bs2;
1652c733ed4SBarry Smith }
1662c733ed4SBarry Smith }
1679566063dSJacob Faibussowitsch PetscCall(VecRestoreArray(xx, &x));
1689566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(2.0 * bs2 * (a->nz) - bs * A->cmap->n));
169*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS);
1702c733ed4SBarry Smith }
171