1 /*$Id: baijfact13.c,v 1.2 2001/01/15 21:45:50 bsmith Exp balay $*/ 2 /* 3 Factorization code for BAIJ format. 4 */ 5 #include "src/mat/impls/baij/seq/baij.h" 6 #include "src/vec/vecimpl.h" 7 #include "src/inline/ilu.h" 8 9 /* ------------------------------------------------------------*/ 10 /* 11 Version for when blocks are 3 by 3 12 */ 13 #undef __FUNC__ 14 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 15 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 16 { 17 Mat C = *B; 18 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data; 19 IS isrow = b->row,isicol = b->icol; 20 int *r,*ic,ierr,i,j,n = a->mbs,*bi = b->i,*bj = b->j; 21 int *ajtmpold,*ajtmp,nz,row,*ai=a->i,*aj=a->j; 22 int *diag_offset = b->diag,idx,*pj; 23 MatScalar *pv,*v,*rtmp,*pc,*w,*x; 24 MatScalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 25 MatScalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 26 MatScalar *ba = b->a,*aa = a->a; 27 28 PetscFunctionBegin; 29 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 30 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 31 ierr = PetscMalloc(9*(n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 32 33 for (i=0; i<n; i++) { 34 nz = bi[i+1] - bi[i]; 35 ajtmp = bj + bi[i]; 36 for (j=0; j<nz; j++) { 37 x = rtmp + 9*ajtmp[j]; 38 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = 0.0; 39 } 40 /* load in initial (unfactored row) */ 41 idx = r[i]; 42 nz = ai[idx+1] - ai[idx]; 43 ajtmpold = aj + ai[idx]; 44 v = aa + 9*ai[idx]; 45 for (j=0; j<nz; j++) { 46 x = rtmp + 9*ic[ajtmpold[j]]; 47 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 48 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 49 v += 9; 50 } 51 row = *ajtmp++; 52 while (row < i) { 53 pc = rtmp + 9*row; 54 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 55 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 56 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 57 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 58 pv = ba + 9*diag_offset[row]; 59 pj = bj + diag_offset[row] + 1; 60 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 61 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 62 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 63 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 64 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 65 66 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 67 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 68 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 69 70 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 71 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 72 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 73 nz = bi[row+1] - diag_offset[row] - 1; 74 pv += 9; 75 for (j=0; j<nz; j++) { 76 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 77 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 78 x = rtmp + 9*pj[j]; 79 x[0] -= m1*x1 + m4*x2 + m7*x3; 80 x[1] -= m2*x1 + m5*x2 + m8*x3; 81 x[2] -= m3*x1 + m6*x2 + m9*x3; 82 83 x[3] -= m1*x4 + m4*x5 + m7*x6; 84 x[4] -= m2*x4 + m5*x5 + m8*x6; 85 x[5] -= m3*x4 + m6*x5 + m9*x6; 86 87 x[6] -= m1*x7 + m4*x8 + m7*x9; 88 x[7] -= m2*x7 + m5*x8 + m8*x9; 89 x[8] -= m3*x7 + m6*x8 + m9*x9; 90 pv += 9; 91 } 92 PetscLogFlops(54*nz+36); 93 } 94 row = *ajtmp++; 95 } 96 /* finished row so stick it into b->a */ 97 pv = ba + 9*bi[i]; 98 pj = bj + bi[i]; 99 nz = bi[i+1] - bi[i]; 100 for (j=0; j<nz; j++) { 101 x = rtmp + 9*pj[j]; 102 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 103 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 104 pv += 9; 105 } 106 /* invert diagonal block */ 107 w = ba + 9*diag_offset[i]; 108 ierr = Kernel_A_gets_inverse_A_3(w);CHKERRQ(ierr); 109 } 110 111 ierr = PetscFree(rtmp);CHKERRQ(ierr); 112 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 113 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 114 C->factor = FACTOR_LU; 115 C->assembled = PETSC_TRUE; 116 PetscLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 117 PetscFunctionReturn(0); 118 } 119