1 2 /* 3 Factorization code for BAIJ format. 4 */ 5 #include <../src/mat/impls/baij/seq/baij.h> 6 #include <petsc/private/kernels/blockinvert.h> 7 8 /* ----------------------------------------------------------- */ 9 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_N_inplace(Mat C, Mat A, const MatFactorInfo *info) 10 { 11 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data, *b = (Mat_SeqBAIJ *)C->data; 12 IS isrow = b->row, isicol = b->icol; 13 const PetscInt *r, *ic; 14 PetscInt i, j, n = a->mbs, *bi = b->i, *bj = b->j; 15 PetscInt *ajtmpold, *ajtmp, nz, row, *ai = a->i, *aj = a->j, k, flg; 16 PetscInt *diag_offset = b->diag, diag, bs = A->rmap->bs, bs2 = a->bs2, *pj, *v_pivots; 17 MatScalar *ba = b->a, *aa = a->a, *pv, *v, *rtmp, *multiplier, *v_work, *pc, *w; 18 PetscBool allowzeropivot, zeropivotdetected; 19 20 PetscFunctionBegin; 21 PetscCall(ISGetIndices(isrow, &r)); 22 PetscCall(ISGetIndices(isicol, &ic)); 23 allowzeropivot = PetscNot(A->erroriffailure); 24 25 PetscCall(PetscCalloc1(bs2 * (n + 1), &rtmp)); 26 /* generate work space needed by dense LU factorization */ 27 PetscCall(PetscMalloc3(bs, &v_work, bs2, &multiplier, bs, &v_pivots)); 28 29 for (i = 0; i < n; i++) { 30 nz = bi[i + 1] - bi[i]; 31 ajtmp = bj + bi[i]; 32 for (j = 0; j < nz; j++) PetscCall(PetscArrayzero(rtmp + bs2 * ajtmp[j], bs2)); 33 /* load in initial (unfactored row) */ 34 nz = ai[r[i] + 1] - ai[r[i]]; 35 ajtmpold = aj + ai[r[i]]; 36 v = aa + bs2 * ai[r[i]]; 37 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(rtmp + bs2 * ic[ajtmpold[j]], v + bs2 * j, bs2)); 38 row = *ajtmp++; 39 while (row < i) { 40 pc = rtmp + bs2 * row; 41 /* if (*pc) { */ 42 for (flg = 0, k = 0; k < bs2; k++) { 43 if (pc[k] != 0.0) { 44 flg = 1; 45 break; 46 } 47 } 48 if (flg) { 49 pv = ba + bs2 * diag_offset[row]; 50 pj = bj + diag_offset[row] + 1; 51 PetscKernel_A_gets_A_times_B(bs, pc, pv, multiplier); 52 nz = bi[row + 1] - diag_offset[row] - 1; 53 pv += bs2; 54 for (j = 0; j < nz; j++) PetscKernel_A_gets_A_minus_B_times_C(bs, rtmp + bs2 * pj[j], pc, pv + bs2 * j); 55 PetscCall(PetscLogFlops(2.0 * bs * bs2 * (nz + 1.0) - bs)); 56 } 57 row = *ajtmp++; 58 } 59 /* finished row so stick it into b->a */ 60 pv = ba + bs2 * bi[i]; 61 pj = bj + bi[i]; 62 nz = bi[i + 1] - bi[i]; 63 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(pv + bs2 * j, rtmp + bs2 * pj[j], bs2)); 64 diag = diag_offset[i] - bi[i]; 65 /* invert diagonal block */ 66 w = pv + bs2 * diag; 67 68 PetscCall(PetscKernel_A_gets_inverse_A(bs, w, v_pivots, v_work, allowzeropivot, &zeropivotdetected)); 69 if (zeropivotdetected) C->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 70 } 71 72 PetscCall(PetscFree(rtmp)); 73 PetscCall(PetscFree3(v_work, multiplier, v_pivots)); 74 PetscCall(ISRestoreIndices(isicol, &ic)); 75 PetscCall(ISRestoreIndices(isrow, &r)); 76 77 C->ops->solve = MatSolve_SeqBAIJ_N_inplace; 78 C->ops->solvetranspose = MatSolveTranspose_SeqBAIJ_N_inplace; 79 C->assembled = PETSC_TRUE; 80 81 PetscCall(PetscLogFlops(1.333333333333 * bs * bs2 * b->mbs)); /* from inverting diagonal blocks */ 82 PetscFunctionReturn(0); 83 } 84