1 #include <../src/mat/impls/baij/seq/baij.h> 2 #include <petsc/private/kernels/blockinvert.h> 3 4 /* ----------------------------------------------------------- */ 5 PetscErrorCode MatSolveTranspose_SeqBAIJ_N_inplace(Mat A, Vec bb, Vec xx) { 6 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 7 IS iscol = a->col, isrow = a->row; 8 const PetscInt *r, *c, *rout, *cout, *ai = a->i, *aj = a->j, *vi; 9 PetscInt i, nz, j; 10 const PetscInt n = a->mbs, bs = A->rmap->bs, bs2 = a->bs2; 11 const MatScalar *aa = a->a, *v; 12 PetscScalar *x, *t, *ls; 13 const PetscScalar *b; 14 15 PetscFunctionBegin; 16 PetscCall(VecGetArrayRead(bb, &b)); 17 PetscCall(VecGetArray(xx, &x)); 18 t = a->solve_work; 19 20 PetscCall(ISGetIndices(isrow, &rout)); 21 r = rout; 22 PetscCall(ISGetIndices(iscol, &cout)); 23 c = cout; 24 25 /* copy the b into temp work space according to permutation */ 26 for (i = 0; i < n; i++) { 27 for (j = 0; j < bs; j++) { t[i * bs + j] = b[c[i] * bs + j]; } 28 } 29 30 /* forward solve the upper triangular transpose */ 31 ls = a->solve_work + A->cmap->n; 32 for (i = 0; i < n; i++) { 33 PetscCall(PetscArraycpy(ls, t + i * bs, bs)); 34 PetscKernel_w_gets_transA_times_v(bs, ls, aa + bs2 * a->diag[i], t + i * bs); 35 v = aa + bs2 * (a->diag[i] + 1); 36 vi = aj + a->diag[i] + 1; 37 nz = ai[i + 1] - a->diag[i] - 1; 38 while (nz--) { 39 PetscKernel_v_gets_v_minus_transA_times_w(bs, t + bs * (*vi++), v, t + i * bs); 40 v += bs2; 41 } 42 } 43 44 /* backward solve the lower triangular transpose */ 45 for (i = n - 1; i >= 0; i--) { 46 v = aa + bs2 * ai[i]; 47 vi = aj + ai[i]; 48 nz = a->diag[i] - ai[i]; 49 while (nz--) { 50 PetscKernel_v_gets_v_minus_transA_times_w(bs, t + bs * (*vi++), v, t + i * bs); 51 v += bs2; 52 } 53 } 54 55 /* copy t into x according to permutation */ 56 for (i = 0; i < n; i++) { 57 for (j = 0; j < bs; j++) { x[bs * r[i] + j] = t[bs * i + j]; } 58 } 59 60 PetscCall(ISRestoreIndices(isrow, &rout)); 61 PetscCall(ISRestoreIndices(iscol, &cout)); 62 PetscCall(VecRestoreArrayRead(bb, &b)); 63 PetscCall(VecRestoreArray(xx, &x)); 64 PetscCall(PetscLogFlops(2.0 * (a->bs2) * (a->nz) - A->rmap->bs * A->cmap->n)); 65 PetscFunctionReturn(0); 66 } 67 68 PetscErrorCode MatSolveTranspose_SeqBAIJ_N(Mat A, Vec bb, Vec xx) { 69 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 70 IS iscol = a->col, isrow = a->row; 71 const PetscInt *r, *c, *rout, *cout; 72 const PetscInt n = a->mbs, *ai = a->i, *aj = a->j, *vi, *diag = a->diag; 73 PetscInt i, j, nz; 74 const PetscInt bs = A->rmap->bs, bs2 = a->bs2; 75 const MatScalar *aa = a->a, *v; 76 PetscScalar *x, *t, *ls; 77 const PetscScalar *b; 78 79 PetscFunctionBegin; 80 PetscCall(VecGetArrayRead(bb, &b)); 81 PetscCall(VecGetArray(xx, &x)); 82 t = a->solve_work; 83 84 PetscCall(ISGetIndices(isrow, &rout)); 85 r = rout; 86 PetscCall(ISGetIndices(iscol, &cout)); 87 c = cout; 88 89 /* copy the b into temp work space according to permutation */ 90 for (i = 0; i < n; i++) { 91 for (j = 0; j < bs; j++) { t[i * bs + j] = b[c[i] * bs + j]; } 92 } 93 94 /* forward solve the upper triangular transpose */ 95 ls = a->solve_work + A->cmap->n; 96 for (i = 0; i < n; i++) { 97 PetscCall(PetscArraycpy(ls, t + i * bs, bs)); 98 PetscKernel_w_gets_transA_times_v(bs, ls, aa + bs2 * diag[i], t + i * bs); 99 v = aa + bs2 * (diag[i] - 1); 100 vi = aj + diag[i] - 1; 101 nz = diag[i] - diag[i + 1] - 1; 102 for (j = 0; j > -nz; j--) { 103 PetscKernel_v_gets_v_minus_transA_times_w(bs, t + bs * (vi[j]), v, t + i * bs); 104 v -= bs2; 105 } 106 } 107 108 /* backward solve the lower triangular transpose */ 109 for (i = n - 1; i >= 0; i--) { 110 v = aa + bs2 * ai[i]; 111 vi = aj + ai[i]; 112 nz = ai[i + 1] - ai[i]; 113 for (j = 0; j < nz; j++) { 114 PetscKernel_v_gets_v_minus_transA_times_w(bs, t + bs * (vi[j]), v, t + i * bs); 115 v += bs2; 116 } 117 } 118 119 /* copy t into x according to permutation */ 120 for (i = 0; i < n; i++) { 121 for (j = 0; j < bs; j++) { x[bs * r[i] + j] = t[bs * i + j]; } 122 } 123 124 PetscCall(ISRestoreIndices(isrow, &rout)); 125 PetscCall(ISRestoreIndices(iscol, &cout)); 126 PetscCall(VecRestoreArrayRead(bb, &b)); 127 PetscCall(VecRestoreArray(xx, &x)); 128 PetscCall(PetscLogFlops(2.0 * (a->bs2) * (a->nz) - A->rmap->bs * A->cmap->n)); 129 PetscFunctionReturn(0); 130 } 131