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