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 4 by 4 10 */ 11 #undef __FUNCT__ 12 #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_4" 13 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4(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; 21 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j,*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,x10,x11,x12,x13,x14,x15,x16; 25 MatScalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 26 MatScalar m13,m14,m15,m16; 27 MatScalar *ba = b->a,*aa = a->a; 28 PetscTruth pivotinblocks = b->pivotinblocks; 29 30 PetscFunctionBegin; 31 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 32 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 33 ierr = PetscMalloc(16*(n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 34 35 for (i=0; i<n; i++) { 36 nz = bi[i+1] - bi[i]; 37 ajtmp = bj + bi[i]; 38 for (j=0; j<nz; j++) { 39 x = rtmp+16*ajtmp[j]; 40 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 41 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 42 } 43 /* load in initial (unfactored row) */ 44 idx = r[i]; 45 nz = ai[idx+1] - ai[idx]; 46 ajtmpold = aj + ai[idx]; 47 v = aa + 16*ai[idx]; 48 for (j=0; j<nz; j++) { 49 x = rtmp+16*ic[ajtmpold[j]]; 50 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 51 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 52 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 53 x[14] = v[14]; x[15] = v[15]; 54 v += 16; 55 } 56 row = *ajtmp++; 57 while (row < i) { 58 pc = rtmp + 16*row; 59 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 60 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 61 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 62 p15 = pc[14]; p16 = pc[15]; 63 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 64 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 65 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 66 || p16 != 0.0) { 67 pv = ba + 16*diag_offset[row]; 68 pj = bj + diag_offset[row] + 1; 69 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 70 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 71 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 72 x15 = pv[14]; x16 = pv[15]; 73 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 74 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 75 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 76 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 77 78 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 79 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 80 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 81 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 82 83 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 84 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 85 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 86 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 87 88 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 89 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 90 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 91 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 92 93 nz = bi[row+1] - diag_offset[row] - 1; 94 pv += 16; 95 for (j=0; j<nz; j++) { 96 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 97 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 98 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 99 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 100 x = rtmp + 16*pj[j]; 101 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 102 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 103 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 104 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 105 106 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 107 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 108 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 109 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 110 111 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 112 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 113 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 114 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 115 116 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 117 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 118 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 119 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 120 121 pv += 16; 122 } 123 PetscLogFlops(128*nz+112); 124 } 125 row = *ajtmp++; 126 } 127 /* finished row so stick it into b->a */ 128 pv = ba + 16*bi[i]; 129 pj = bj + bi[i]; 130 nz = bi[i+1] - bi[i]; 131 for (j=0; j<nz; j++) { 132 x = rtmp+16*pj[j]; 133 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 134 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 135 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 136 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 137 pv += 16; 138 } 139 /* invert diagonal block */ 140 w = ba + 16*diag_offset[i]; 141 if (pivotinblocks) { 142 ierr = Kernel_A_gets_inverse_A_4(w);CHKERRQ(ierr); 143 } else { 144 ierr = Kernel_A_gets_inverse_A_4_nopivot(w);CHKERRQ(ierr); 145 } 146 } 147 148 ierr = PetscFree(rtmp);CHKERRQ(ierr); 149 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 150 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 151 C->factor = FACTOR_LU; 152 C->assembled = PETSC_TRUE; 153 PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 154 PetscFunctionReturn(0); 155 } 156