1 #define PETSCMAT_DLL 2 3 #include "src/mat/impls/sbaij/seq/sbaij.h" 4 #include "src/inline/ilu.h" 5 6 /* Version for when blocks are 7 by 7 */ 7 #undef __FUNCT__ 8 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_7" 9 PetscErrorCode MatCholeskyFactorNumeric_SeqSBAIJ_7(Mat A,MatFactorInfo *info,Mat *B) 10 { 11 Mat C = *B; 12 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 13 IS perm = b->row; 14 PetscErrorCode ierr; 15 PetscInt *perm_ptr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 16 PetscInt *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 17 MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 18 MatScalar *u,*d,*w,*wp,u0,u1,u2,u3,u4,u5,u6,u7,u8,u9,u10,u11,u12; 19 MatScalar u13,u14,u15,u16,u17,u18,u19,u20,u21,u22,u23,u24,u25,u26,u27; 20 MatScalar u28,u29,u30,u31,u32,u33,u34,u35,u36,u37,u38,u39,u40,u41; 21 MatScalar u42,u43,u44,u45,u46,u47,u48; 22 PetscReal shift = info->shiftinblocks; 23 24 PetscFunctionBegin; 25 /* initialization */ 26 ierr = PetscMalloc(49*mbs*sizeof(MatScalar),&w);CHKERRQ(ierr); 27 ierr = PetscMemzero(w,49*mbs*sizeof(MatScalar));CHKERRQ(ierr); 28 ierr = PetscMalloc(2*mbs*sizeof(PetscInt),&il);CHKERRQ(ierr); 29 jl = il + mbs; 30 for (i=0; i<mbs; i++) { 31 jl[i] = mbs; il[0] = 0; 32 } 33 ierr = PetscMalloc(98*sizeof(MatScalar),&dk);CHKERRQ(ierr); 34 uik = dk + 49; 35 ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 36 37 /* check permutation */ 38 if (!a->permute){ 39 ai = a->i; aj = a->j; aa = a->a; 40 } else { 41 ai = a->inew; aj = a->jnew; 42 ierr = PetscMalloc(49*ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr); 43 ierr = PetscMemcpy(aa,a->a,49*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 44 ierr = PetscMalloc(ai[mbs]*sizeof(PetscInt),&a2anew);CHKERRQ(ierr); 45 ierr = PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(PetscInt));CHKERRQ(ierr); 46 47 for (i=0; i<mbs; i++){ 48 jmin = ai[i]; jmax = ai[i+1]; 49 for (j=jmin; j<jmax; j++){ 50 while (a2anew[j] != j){ 51 k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 52 for (k1=0; k1<49; k1++){ 53 dk[k1] = aa[k*49+k1]; 54 aa[k*49+k1] = aa[j*49+k1]; 55 aa[j*49+k1] = dk[k1]; 56 } 57 } 58 /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 59 if (i > aj[j]){ 60 /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 61 ap = aa + j*49; /* ptr to the beginning of j-th block of aa */ 62 for (k=0; k<49; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 63 for (k=0; k<7; k++){ /* j-th block of aa <- dk^T */ 64 for (k1=0; k1<7; k1++) *ap++ = dk[k + 7*k1]; 65 } 66 } 67 } 68 } 69 ierr = PetscFree(a2anew);CHKERRQ(ierr); 70 } 71 72 /* for each row k */ 73 for (k = 0; k<mbs; k++){ 74 75 /*initialize k-th row with elements nonzero in row perm(k) of A */ 76 jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 77 if (jmin < jmax) { 78 ap = aa + jmin*49; 79 for (j = jmin; j < jmax; j++){ 80 vj = perm_ptr[aj[j]]; /* block col. index */ 81 wp = w + vj*49; 82 for (i=0; i<49; i++) *wp++ = *ap++; 83 } 84 } 85 86 /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 87 ierr = PetscMemcpy(dk,w+k*49,49*sizeof(MatScalar));CHKERRQ(ierr); 88 i = jl[k]; /* first row to be added to k_th row */ 89 90 while (i < mbs){ 91 nexti = jl[i]; /* next row to be added to k_th row */ 92 93 /* compute multiplier */ 94 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 95 96 /* uik = -inv(Di)*U_bar(i,k) */ 97 d = ba + i*49; 98 u = ba + ili*49; 99 100 u0 = u[0]; u1 = u[1]; u2 = u[2]; u3 = u[3]; u4 = u[4]; u5 = u[5]; u6 = u[6]; 101 u7 = u[7]; u8 = u[8]; u9 = u[9]; u10 = u[10]; u11 = u[11]; u12 = u[12]; u13 = u[13]; 102 u14 = u[14]; u15 = u[15]; u16 = u[16]; u17 = u[17]; u18 = u[18]; u19 = u[19]; u20 = u[20]; 103 u21 = u[21]; u22 = u[22]; u23 = u[23]; u24 = u[24]; u25 = u[25]; u26 = u[26]; u27 = u[27]; 104 u28 = u[28]; u29 = u[29]; u30 = u[30]; u31 = u[31]; u32 = u[32]; u33 = u[33]; u34 = u[34]; 105 u35 = u[35]; u36 = u[36]; u37 = u[37]; u38 = u[38]; u39 = u[39]; u40 = u[40]; u41 = u[41]; u42 = u[42]; 106 u43 = u[43]; u44 = u[44]; u45 = u[45]; u46 = u[46]; u47 = u[47]; u48 = u[48]; 107 108 uik[0] = -(d[0]*u0 + d[7]*u1+ d[14]*u2+ d[21]*u3+ d[28]*u4+ d[35]*u5+ d[42]*u6); 109 uik[1] = -(d[1]*u0 + d[8]*u1+ d[15]*u2+ d[22]*u3+ d[29]*u4+ d[36]*u5+ d[43]*u6); 110 uik[2] = -(d[2]*u0 + d[9]*u1+ d[16]*u2+ d[23]*u3+ d[30]*u4+ d[37]*u5+ d[44]*u6); 111 uik[3] = -(d[3]*u0+ d[10]*u1+ d[17]*u2+ d[24]*u3+ d[31]*u4+ d[38]*u5+ d[45]*u6); 112 uik[4] = -(d[4]*u0+ d[11]*u1+ d[18]*u2+ d[25]*u3+ d[32]*u4+ d[39]*u5+ d[46]*u6); 113 uik[5] = -(d[5]*u0+ d[12]*u1+ d[19]*u2+ d[26]*u3+ d[33]*u4+ d[40]*u5+ d[47]*u6); 114 uik[6] = -(d[6]*u0+ d[13]*u1+ d[20]*u2+ d[27]*u3+ d[34]*u4+ d[41]*u5+ d[48]*u6); 115 116 uik[7] = -(d[0]*u7 + d[7]*u8+ d[14]*u9+ d[21]*u10+ d[28]*u11+ d[35]*u12+ d[42]*u13); 117 uik[8] = -(d[1]*u7 + d[8]*u8+ d[15]*u9+ d[22]*u10+ d[29]*u11+ d[36]*u12+ d[43]*u13); 118 uik[9] = -(d[2]*u7 + d[9]*u8+ d[16]*u9+ d[23]*u10+ d[30]*u11+ d[37]*u12+ d[44]*u13); 119 uik[10]= -(d[3]*u7+ d[10]*u8+ d[17]*u9+ d[24]*u10+ d[31]*u11+ d[38]*u12+ d[45]*u13); 120 uik[11]= -(d[4]*u7+ d[11]*u8+ d[18]*u9+ d[25]*u10+ d[32]*u11+ d[39]*u12+ d[46]*u13); 121 uik[12]= -(d[5]*u7+ d[12]*u8+ d[19]*u9+ d[26]*u10+ d[33]*u11+ d[40]*u12+ d[47]*u13); 122 uik[13]= -(d[6]*u7+ d[13]*u8+ d[20]*u9+ d[27]*u10+ d[34]*u11+ d[41]*u12+ d[48]*u13); 123 124 uik[14]= -(d[0]*u14 + d[7]*u15+ d[14]*u16+ d[21]*u17+ d[28]*u18+ d[35]*u19+ d[42]*u20); 125 uik[15]= -(d[1]*u14 + d[8]*u15+ d[15]*u16+ d[22]*u17+ d[29]*u18+ d[36]*u19+ d[43]*u20); 126 uik[16]= -(d[2]*u14 + d[9]*u15+ d[16]*u16+ d[23]*u17+ d[30]*u18+ d[37]*u19+ d[44]*u20); 127 uik[17]= -(d[3]*u14+ d[10]*u15+ d[17]*u16+ d[24]*u17+ d[31]*u18+ d[38]*u19+ d[45]*u20); 128 uik[18]= -(d[4]*u14+ d[11]*u15+ d[18]*u16+ d[25]*u17+ d[32]*u18+ d[39]*u19+ d[46]*u20); 129 uik[19]= -(d[5]*u14+ d[12]*u15+ d[19]*u16+ d[26]*u17+ d[33]*u18+ d[40]*u19+ d[47]*u20); 130 uik[20]= -(d[6]*u14+ d[13]*u15+ d[20]*u16+ d[27]*u17+ d[34]*u18+ d[41]*u19+ d[48]*u20); 131 132 uik[21]= -(d[0]*u21 + d[7]*u22+ d[14]*u23+ d[21]*u24+ d[28]*u25+ d[35]*u26+ d[42]*u27); 133 uik[22]= -(d[1]*u21 + d[8]*u22+ d[15]*u23+ d[22]*u24+ d[29]*u25+ d[36]*u26+ d[43]*u27); 134 uik[23]= -(d[2]*u21 + d[9]*u22+ d[16]*u23+ d[23]*u24+ d[30]*u25+ d[37]*u26+ d[44]*u27); 135 uik[24]= -(d[3]*u21+ d[10]*u22+ d[17]*u23+ d[24]*u24+ d[31]*u25+ d[38]*u26+ d[45]*u27); 136 uik[25]= -(d[4]*u21+ d[11]*u22+ d[18]*u23+ d[25]*u24+ d[32]*u25+ d[39]*u26+ d[46]*u27); 137 uik[26]= -(d[5]*u21+ d[12]*u22+ d[19]*u23+ d[26]*u24+ d[33]*u25+ d[40]*u26+ d[47]*u27); 138 uik[27]= -(d[6]*u21+ d[13]*u22+ d[20]*u23+ d[27]*u24+ d[34]*u25+ d[41]*u26+ d[48]*u27); 139 140 uik[28]= -(d[0]*u28 + d[7]*u29+ d[14]*u30+ d[21]*u31+ d[28]*u32+ d[35]*u33+ d[42]*u34); 141 uik[29]= -(d[1]*u28 + d[8]*u29+ d[15]*u30+ d[22]*u31+ d[29]*u32+ d[36]*u33+ d[43]*u34); 142 uik[30]= -(d[2]*u28 + d[9]*u29+ d[16]*u30+ d[23]*u31+ d[30]*u32+ d[37]*u33+ d[44]*u34); 143 uik[31]= -(d[3]*u28+ d[10]*u29+ d[17]*u30+ d[24]*u31+ d[31]*u32+ d[38]*u33+ d[45]*u34); 144 uik[32]= -(d[4]*u28+ d[11]*u29+ d[18]*u30+ d[25]*u31+ d[32]*u32+ d[39]*u33+ d[46]*u34); 145 uik[33]= -(d[5]*u28+ d[12]*u29+ d[19]*u30+ d[26]*u31+ d[33]*u32+ d[40]*u33+ d[47]*u34); 146 uik[34]= -(d[6]*u28+ d[13]*u29+ d[20]*u30+ d[27]*u31+ d[34]*u32+ d[41]*u33+ d[48]*u34); 147 148 uik[35]= -(d[0]*u35 + d[7]*u36+ d[14]*u37+ d[21]*u38+ d[28]*u39+ d[35]*u40+ d[42]*u41); 149 uik[36]= -(d[1]*u35 + d[8]*u36+ d[15]*u37+ d[22]*u38+ d[29]*u39+ d[36]*u40+ d[43]*u41); 150 uik[37]= -(d[2]*u35 + d[9]*u36+ d[16]*u37+ d[23]*u38+ d[30]*u39+ d[37]*u40+ d[44]*u41); 151 uik[38]= -(d[3]*u35+ d[10]*u36+ d[17]*u37+ d[24]*u38+ d[31]*u39+ d[38]*u40+ d[45]*u41); 152 uik[39]= -(d[4]*u35+ d[11]*u36+ d[18]*u37+ d[25]*u38+ d[32]*u39+ d[39]*u40+ d[46]*u41); 153 uik[40]= -(d[5]*u35+ d[12]*u36+ d[19]*u37+ d[26]*u38+ d[33]*u39+ d[40]*u40+ d[47]*u41); 154 uik[41]= -(d[6]*u35+ d[13]*u36+ d[20]*u37+ d[27]*u38+ d[34]*u39+ d[41]*u40+ d[48]*u41); 155 156 uik[42]= -(d[0]*u42 + d[7]*u43+ d[14]*u44+ d[21]*u45+ d[28]*u46+ d[35]*u47+ d[42]*u48); 157 uik[43]= -(d[1]*u42 + d[8]*u43+ d[15]*u44+ d[22]*u45+ d[29]*u46+ d[36]*u47+ d[43]*u48); 158 uik[44]= -(d[2]*u42 + d[9]*u43+ d[16]*u44+ d[23]*u45+ d[30]*u46+ d[37]*u47+ d[44]*u48); 159 uik[45]= -(d[3]*u42+ d[10]*u43+ d[17]*u44+ d[24]*u45+ d[31]*u46+ d[38]*u47+ d[45]*u48); 160 uik[46]= -(d[4]*u42+ d[11]*u43+ d[18]*u44+ d[25]*u45+ d[32]*u46+ d[39]*u47+ d[46]*u48); 161 uik[47]= -(d[5]*u42+ d[12]*u43+ d[19]*u44+ d[26]*u45+ d[33]*u46+ d[40]*u47+ d[47]*u48); 162 uik[48]= -(d[6]*u42+ d[13]*u43+ d[20]*u44+ d[27]*u45+ d[34]*u46+ d[41]*u47+ d[48]*u48); 163 164 /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 165 dk[0]+= uik[0]*u0 + uik[1]*u1 + uik[2]*u2 + uik[3]*u3 + uik[4]*u4 + uik[5]*u5 + uik[6]*u6; 166 dk[1]+= uik[7]*u0 + uik[8]*u1 + uik[9]*u2+ uik[10]*u3+ uik[11]*u4+ uik[12]*u5+ uik[13]*u6; 167 dk[2]+= uik[14]*u0+ uik[15]*u1+ uik[16]*u2+ uik[17]*u3+ uik[18]*u4+ uik[19]*u5+ uik[20]*u6; 168 dk[3]+= uik[21]*u0+ uik[22]*u1+ uik[23]*u2+ uik[24]*u3+ uik[25]*u4+ uik[26]*u5+ uik[27]*u6; 169 dk[4]+= uik[28]*u0+ uik[29]*u1+ uik[30]*u2+ uik[31]*u3+ uik[32]*u4+ uik[33]*u5+ uik[34]*u6; 170 dk[5]+= uik[35]*u0+ uik[36]*u1+ uik[37]*u2+ uik[38]*u3+ uik[39]*u4+ uik[40]*u5+ uik[41]*u6; 171 dk[6]+= uik[42]*u0+ uik[43]*u1+ uik[44]*u2+ uik[45]*u3+ uik[46]*u4+ uik[47]*u5+ uik[48]*u6; 172 173 dk[7]+= uik[0]*u7 + uik[1]*u8 + uik[2]*u9 + uik[3]*u10 + uik[4]*u11 + uik[5]*u12 + uik[6]*u13; 174 dk[8]+= uik[7]*u7 + uik[8]*u8 + uik[9]*u9+ uik[10]*u10+ uik[11]*u11+ uik[12]*u12+ uik[13]*u13; 175 dk[9]+= uik[14]*u7+ uik[15]*u8+ uik[16]*u9+ uik[17]*u10+ uik[18]*u11+ uik[19]*u12+ uik[20]*u13; 176 dk[10]+=uik[21]*u7+ uik[22]*u8+ uik[23]*u9+ uik[24]*u10+ uik[25]*u11+ uik[26]*u12+ uik[27]*u13; 177 dk[11]+=uik[28]*u7+ uik[29]*u8+ uik[30]*u9+ uik[31]*u10+ uik[32]*u11+ uik[33]*u12+ uik[34]*u13; 178 dk[12]+=uik[35]*u7+ uik[36]*u8+ uik[37]*u9+ uik[38]*u10+ uik[39]*u11+ uik[40]*u12+ uik[41]*u13; 179 dk[13]+=uik[42]*u7+ uik[43]*u8+ uik[44]*u9+ uik[45]*u10+ uik[46]*u11+ uik[47]*u12+ uik[48]*u13; 180 181 dk[14]+= uik[0]*u14 + uik[1]*u15 + uik[2]*u16 + uik[3]*u17 + uik[4]*u18 + uik[5]*u19 + uik[6]*u20; 182 dk[15]+= uik[7]*u14 + uik[8]*u15 + uik[9]*u16+ uik[10]*u17+ uik[11]*u18+ uik[12]*u19+ uik[13]*u20; 183 dk[16]+= uik[14]*u14+ uik[15]*u15+ uik[16]*u16+ uik[17]*u17+ uik[18]*u18+ uik[19]*u19+ uik[20]*u20; 184 dk[17]+= uik[21]*u14+ uik[22]*u15+ uik[23]*u16+ uik[24]*u17+ uik[25]*u18+ uik[26]*u19+ uik[27]*u20; 185 dk[18]+= uik[28]*u14+ uik[29]*u15+ uik[30]*u16+ uik[31]*u17+ uik[32]*u18+ uik[33]*u19+ uik[34]*u20; 186 dk[19]+= uik[35]*u14+ uik[36]*u15+ uik[37]*u16+ uik[38]*u17+ uik[39]*u18+ uik[40]*u19+ uik[41]*u20; 187 dk[20]+= uik[42]*u14+ uik[43]*u15+ uik[44]*u16+ uik[45]*u17+ uik[46]*u18+ uik[47]*u19+ uik[48]*u20; 188 189 dk[21]+= uik[0]*u21 + uik[1]*u22 + uik[2]*u23 + uik[3]*u24 + uik[4]*u25 + uik[5]*u26 + uik[6]*u27; 190 dk[22]+= uik[7]*u21 + uik[8]*u22 + uik[9]*u23+ uik[10]*u24+ uik[11]*u25+ uik[12]*u26+ uik[13]*u27; 191 dk[23]+= uik[14]*u21+ uik[15]*u22+ uik[16]*u23+ uik[17]*u24+ uik[18]*u25+ uik[19]*u26+ uik[20]*u27; 192 dk[24]+= uik[21]*u21+ uik[22]*u22+ uik[23]*u23+ uik[24]*u24+ uik[25]*u25+ uik[26]*u26+ uik[27]*u27; 193 dk[25]+= uik[28]*u21+ uik[29]*u22+ uik[30]*u23+ uik[31]*u24+ uik[32]*u25+ uik[33]*u26+ uik[34]*u27; 194 dk[26]+= uik[35]*u21+ uik[36]*u22+ uik[37]*u23+ uik[38]*u24+ uik[39]*u25+ uik[40]*u26+ uik[41]*u27; 195 dk[27]+= uik[42]*u21+ uik[43]*u22+ uik[44]*u23+ uik[45]*u24+ uik[46]*u25+ uik[47]*u26+ uik[48]*u27; 196 197 dk[28]+= uik[0]*u28 + uik[1]*u29 + uik[2]*u30 + uik[3]*u31 + uik[4]*u32 + uik[5]*u33 + uik[6]*u34; 198 dk[29]+= uik[7]*u28 + uik[8]*u29 + uik[9]*u30+ uik[10]*u31+ uik[11]*u32+ uik[12]*u33+ uik[13]*u34; 199 dk[30]+= uik[14]*u28+ uik[15]*u29+ uik[16]*u30+ uik[17]*u31+ uik[18]*u32+ uik[19]*u33+ uik[20]*u34; 200 dk[31]+= uik[21]*u28+ uik[22]*u29+ uik[23]*u30+ uik[24]*u31+ uik[25]*u32+ uik[26]*u33+ uik[27]*u34; 201 dk[32]+= uik[28]*u28+ uik[29]*u29+ uik[30]*u30+ uik[31]*u31+ uik[32]*u32+ uik[33]*u33+ uik[34]*u34; 202 dk[33]+= uik[35]*u28+ uik[36]*u29+ uik[37]*u30+ uik[38]*u31+ uik[39]*u32+ uik[40]*u33+ uik[41]*u34; 203 dk[34]+= uik[42]*u28+ uik[43]*u29+ uik[44]*u30+ uik[45]*u31+ uik[46]*u32+ uik[47]*u33+ uik[48]*u34; 204 205 dk[35]+= uik[0]*u35 + uik[1]*u36 + uik[2]*u37 + uik[3]*u38 + uik[4]*u39 + uik[5]*u40 + uik[6]*u41; 206 dk[36]+= uik[7]*u35 + uik[8]*u36 + uik[9]*u37+ uik[10]*u38+ uik[11]*u39+ uik[12]*u40+ uik[13]*u41; 207 dk[37]+= uik[14]*u35+ uik[15]*u36+ uik[16]*u37+ uik[17]*u38+ uik[18]*u39+ uik[19]*u40+ uik[20]*u41; 208 dk[38]+= uik[21]*u35+ uik[22]*u36+ uik[23]*u37+ uik[24]*u38+ uik[25]*u39+ uik[26]*u40+ uik[27]*u41; 209 dk[39]+= uik[28]*u35+ uik[29]*u36+ uik[30]*u37+ uik[31]*u38+ uik[32]*u39+ uik[33]*u40+ uik[34]*u41; 210 dk[40]+= uik[35]*u35+ uik[36]*u36+ uik[37]*u37+ uik[38]*u38+ uik[39]*u39+ uik[40]*u40+ uik[41]*u41; 211 dk[41]+= uik[42]*u35+ uik[43]*u36+ uik[44]*u37+ uik[45]*u38+ uik[46]*u39+ uik[47]*u40+ uik[48]*u41; 212 213 dk[42]+= uik[0]*u42 + uik[1]*u43 + uik[2]*u44 + uik[3]*u45 + uik[4]*u46 + uik[5]*u47 + uik[6]*u48; 214 dk[43]+= uik[7]*u42 + uik[8]*u43 + uik[9]*u44+ uik[10]*u45+ uik[11]*u46+ uik[12]*u47+ uik[13]*u48; 215 dk[44]+= uik[14]*u42+ uik[15]*u43+ uik[16]*u44+ uik[17]*u45+ uik[18]*u46+ uik[19]*u47+ uik[20]*u48; 216 dk[45]+= uik[21]*u42+ uik[22]*u43+ uik[23]*u44+ uik[24]*u45+ uik[25]*u46+ uik[26]*u47+ uik[27]*u48; 217 dk[46]+= uik[28]*u42+ uik[29]*u43+ uik[30]*u44+ uik[31]*u45+ uik[32]*u46+ uik[33]*u47+ uik[34]*u48; 218 dk[47]+= uik[35]*u42+ uik[36]*u43+ uik[37]*u44+ uik[38]*u45+ uik[39]*u46+ uik[40]*u47+ uik[41]*u48; 219 dk[48]+= uik[42]*u42+ uik[43]*u43+ uik[44]*u44+ uik[45]*u45+ uik[46]*u46+ uik[47]*u47+ uik[48]*u48; 220 221 ierr = PetscLogFlops(343*4);CHKERRQ(ierr); 222 223 /* update -U(i,k) */ 224 ierr = PetscMemcpy(ba+ili*49,uik,49*sizeof(MatScalar));CHKERRQ(ierr); 225 226 /* add multiple of row i to k-th row ... */ 227 jmin = ili + 1; jmax = bi[i+1]; 228 if (jmin < jmax){ 229 for (j=jmin; j<jmax; j++) { 230 /* w += -U(i,k)^T * U_bar(i,j) */ 231 wp = w + bj[j]*49; 232 u = ba + j*49; 233 234 u0 = u[0]; u1 = u[1]; u2 = u[2]; u3 = u[3]; u4 = u[4]; u5 = u[5]; u6 = u[6]; 235 u7 = u[7]; u8 = u[8]; u9 = u[9]; u10 = u[10]; u11 = u[11]; u12 = u[12]; u13 = u[13]; 236 u14 = u[14]; u15 = u[15]; u16 = u[16]; u17 = u[17]; u18 = u[18]; u19 = u[19]; u20 = u[20]; 237 u21 = u[21]; u22 = u[22]; u23 = u[23]; u24 = u[24]; u25 = u[25]; u26 = u[26]; u27 = u[27]; 238 u28 = u[28]; u29 = u[29]; u30 = u[30]; u31 = u[31]; u32 = u[32]; u33 = u[33]; u34 = u[34]; 239 u35 = u[35]; u36 = u[36]; u37 = u[37]; u38 = u[38]; u39 = u[39]; u40 = u[40]; u41 = u[41]; u42 = u[42]; 240 u43 = u[43]; u44 = u[44]; u45 = u[45]; u46 = u[46]; u47 = u[47]; u48 = u[48]; 241 242 wp[0]+= uik[0]*u0 + uik[1]*u1 + uik[2]*u2 + uik[3]*u3 + uik[4]*u4 + uik[5]*u5 + uik[6]*u6; 243 wp[1]+= uik[7]*u0 + uik[8]*u1 + uik[9]*u2+ uik[10]*u3+ uik[11]*u4+ uik[12]*u5+ uik[13]*u6; 244 wp[2]+= uik[14]*u0+ uik[15]*u1+ uik[16]*u2+ uik[17]*u3+ uik[18]*u4+ uik[19]*u5+ uik[20]*u6; 245 wp[3]+= uik[21]*u0+ uik[22]*u1+ uik[23]*u2+ uik[24]*u3+ uik[25]*u4+ uik[26]*u5+ uik[27]*u6; 246 wp[4]+= uik[28]*u0+ uik[29]*u1+ uik[30]*u2+ uik[31]*u3+ uik[32]*u4+ uik[33]*u5+ uik[34]*u6; 247 wp[5]+= uik[35]*u0+ uik[36]*u1+ uik[37]*u2+ uik[38]*u3+ uik[39]*u4+ uik[40]*u5+ uik[41]*u6; 248 wp[6]+= uik[42]*u0+ uik[43]*u1+ uik[44]*u2+ uik[45]*u3+ uik[46]*u4+ uik[47]*u5+ uik[48]*u6; 249 250 wp[7]+= uik[0]*u7 + uik[1]*u8 + uik[2]*u9 + uik[3]*u10 + uik[4]*u11 + uik[5]*u12 + uik[6]*u13; 251 wp[8]+= uik[7]*u7 + uik[8]*u8 + uik[9]*u9+ uik[10]*u10+ uik[11]*u11+ uik[12]*u12+ uik[13]*u13; 252 wp[9]+= uik[14]*u7+ uik[15]*u8+ uik[16]*u9+ uik[17]*u10+ uik[18]*u11+ uik[19]*u12+ uik[20]*u13; 253 wp[10]+=uik[21]*u7+ uik[22]*u8+ uik[23]*u9+ uik[24]*u10+ uik[25]*u11+ uik[26]*u12+ uik[27]*u13; 254 wp[11]+=uik[28]*u7+ uik[29]*u8+ uik[30]*u9+ uik[31]*u10+ uik[32]*u11+ uik[33]*u12+ uik[34]*u13; 255 wp[12]+=uik[35]*u7+ uik[36]*u8+ uik[37]*u9+ uik[38]*u10+ uik[39]*u11+ uik[40]*u12+ uik[41]*u13; 256 wp[13]+=uik[42]*u7+ uik[43]*u8+ uik[44]*u9+ uik[45]*u10+ uik[46]*u11+ uik[47]*u12+ uik[48]*u13; 257 258 wp[14]+= uik[0]*u14 + uik[1]*u15 + uik[2]*u16 + uik[3]*u17 + uik[4]*u18 + uik[5]*u19 + uik[6]*u20; 259 wp[15]+= uik[7]*u14 + uik[8]*u15 + uik[9]*u16+ uik[10]*u17+ uik[11]*u18+ uik[12]*u19+ uik[13]*u20; 260 wp[16]+= uik[14]*u14+ uik[15]*u15+ uik[16]*u16+ uik[17]*u17+ uik[18]*u18+ uik[19]*u19+ uik[20]*u20; 261 wp[17]+= uik[21]*u14+ uik[22]*u15+ uik[23]*u16+ uik[24]*u17+ uik[25]*u18+ uik[26]*u19+ uik[27]*u20; 262 wp[18]+= uik[28]*u14+ uik[29]*u15+ uik[30]*u16+ uik[31]*u17+ uik[32]*u18+ uik[33]*u19+ uik[34]*u20; 263 wp[19]+= uik[35]*u14+ uik[36]*u15+ uik[37]*u16+ uik[38]*u17+ uik[39]*u18+ uik[40]*u19+ uik[41]*u20; 264 wp[20]+= uik[42]*u14+ uik[43]*u15+ uik[44]*u16+ uik[45]*u17+ uik[46]*u18+ uik[47]*u19+ uik[48]*u20; 265 266 wp[21]+= uik[0]*u21 + uik[1]*u22 + uik[2]*u23 + uik[3]*u24 + uik[4]*u25 + uik[5]*u26 + uik[6]*u27; 267 wp[22]+= uik[7]*u21 + uik[8]*u22 + uik[9]*u23+ uik[10]*u24+ uik[11]*u25+ uik[12]*u26+ uik[13]*u27; 268 wp[23]+= uik[14]*u21+ uik[15]*u22+ uik[16]*u23+ uik[17]*u24+ uik[18]*u25+ uik[19]*u26+ uik[20]*u27; 269 wp[24]+= uik[21]*u21+ uik[22]*u22+ uik[23]*u23+ uik[24]*u24+ uik[25]*u25+ uik[26]*u26+ uik[27]*u27; 270 wp[25]+= uik[28]*u21+ uik[29]*u22+ uik[30]*u23+ uik[31]*u24+ uik[32]*u25+ uik[33]*u26+ uik[34]*u27; 271 wp[26]+= uik[35]*u21+ uik[36]*u22+ uik[37]*u23+ uik[38]*u24+ uik[39]*u25+ uik[40]*u26+ uik[41]*u27; 272 wp[27]+= uik[42]*u21+ uik[43]*u22+ uik[44]*u23+ uik[45]*u24+ uik[46]*u25+ uik[47]*u26+ uik[48]*u27; 273 274 wp[28]+= uik[0]*u28 + uik[1]*u29 + uik[2]*u30 + uik[3]*u31 + uik[4]*u32 + uik[5]*u33 + uik[6]*u34; 275 wp[29]+= uik[7]*u28 + uik[8]*u29 + uik[9]*u30+ uik[10]*u31+ uik[11]*u32+ uik[12]*u33+ uik[13]*u34; 276 wp[30]+= uik[14]*u28+ uik[15]*u29+ uik[16]*u30+ uik[17]*u31+ uik[18]*u32+ uik[19]*u33+ uik[20]*u34; 277 wp[31]+= uik[21]*u28+ uik[22]*u29+ uik[23]*u30+ uik[24]*u31+ uik[25]*u32+ uik[26]*u33+ uik[27]*u34; 278 wp[32]+= uik[28]*u28+ uik[29]*u29+ uik[30]*u30+ uik[31]*u31+ uik[32]*u32+ uik[33]*u33+ uik[34]*u34; 279 wp[33]+= uik[35]*u28+ uik[36]*u29+ uik[37]*u30+ uik[38]*u31+ uik[39]*u32+ uik[40]*u33+ uik[41]*u34; 280 wp[34]+= uik[42]*u28+ uik[43]*u29+ uik[44]*u30+ uik[45]*u31+ uik[46]*u32+ uik[47]*u33+ uik[48]*u34; 281 282 wp[35]+= uik[0]*u35 + uik[1]*u36 + uik[2]*u37 + uik[3]*u38 + uik[4]*u39 + uik[5]*u40 + uik[6]*u41; 283 wp[36]+= uik[7]*u35 + uik[8]*u36 + uik[9]*u37+ uik[10]*u38+ uik[11]*u39+ uik[12]*u40+ uik[13]*u41; 284 wp[37]+= uik[14]*u35+ uik[15]*u36+ uik[16]*u37+ uik[17]*u38+ uik[18]*u39+ uik[19]*u40+ uik[20]*u41; 285 wp[38]+= uik[21]*u35+ uik[22]*u36+ uik[23]*u37+ uik[24]*u38+ uik[25]*u39+ uik[26]*u40+ uik[27]*u41; 286 wp[39]+= uik[28]*u35+ uik[29]*u36+ uik[30]*u37+ uik[31]*u38+ uik[32]*u39+ uik[33]*u40+ uik[34]*u41; 287 wp[40]+= uik[35]*u35+ uik[36]*u36+ uik[37]*u37+ uik[38]*u38+ uik[39]*u39+ uik[40]*u40+ uik[41]*u41; 288 wp[41]+= uik[42]*u35+ uik[43]*u36+ uik[44]*u37+ uik[45]*u38+ uik[46]*u39+ uik[47]*u40+ uik[48]*u41; 289 290 wp[42]+= uik[0]*u42 + uik[1]*u43 + uik[2]*u44 + uik[3]*u45 + uik[4]*u46 + uik[5]*u47 + uik[6]*u48; 291 wp[43]+= uik[7]*u42 + uik[8]*u43 + uik[9]*u44+ uik[10]*u45+ uik[11]*u46+ uik[12]*u47+ uik[13]*u48; 292 wp[44]+= uik[14]*u42+ uik[15]*u43+ uik[16]*u44+ uik[17]*u45+ uik[18]*u46+ uik[19]*u47+ uik[20]*u48; 293 wp[45]+= uik[21]*u42+ uik[22]*u43+ uik[23]*u44+ uik[24]*u45+ uik[25]*u46+ uik[26]*u47+ uik[27]*u48; 294 wp[46]+= uik[28]*u42+ uik[29]*u43+ uik[30]*u44+ uik[31]*u45+ uik[32]*u46+ uik[33]*u47+ uik[34]*u48; 295 wp[47]+= uik[35]*u42+ uik[36]*u43+ uik[37]*u44+ uik[38]*u45+ uik[39]*u46+ uik[40]*u47+ uik[41]*u48; 296 wp[48]+= uik[42]*u42+ uik[43]*u43+ uik[44]*u44+ uik[45]*u45+ uik[46]*u46+ uik[47]*u47+ uik[48]*u48; 297 } 298 ierr = PetscLogFlops(2*343*(jmax-jmin));CHKERRQ(ierr); 299 300 /* ... add i to row list for next nonzero entry */ 301 il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 302 j = bj[jmin]; 303 jl[i] = jl[j]; jl[j] = i; /* update jl */ 304 } 305 i = nexti; 306 } 307 308 /* save nonzero entries in k-th row of U ... */ 309 310 /* invert diagonal block */ 311 d = ba+k*49; 312 ierr = PetscMemcpy(d,dk,49*sizeof(MatScalar));CHKERRQ(ierr); 313 ierr = Kernel_A_gets_inverse_A_7(d,shift);CHKERRQ(ierr); 314 315 jmin = bi[k]; jmax = bi[k+1]; 316 if (jmin < jmax) { 317 for (j=jmin; j<jmax; j++){ 318 vj = bj[j]; /* block col. index of U */ 319 u = ba + j*49; 320 wp = w + vj*49; 321 for (k1=0; k1<49; k1++){ 322 *u++ = *wp; 323 *wp++ = 0.0; 324 } 325 } 326 327 /* ... add k to row list for first nonzero entry in k-th row */ 328 il[k] = jmin; 329 i = bj[jmin]; 330 jl[k] = jl[i]; jl[i] = k; 331 } 332 } 333 334 ierr = PetscFree(w);CHKERRQ(ierr); 335 ierr = PetscFree(il);CHKERRQ(ierr); 336 ierr = PetscFree(dk);CHKERRQ(ierr); 337 if (a->permute){ 338 ierr = PetscFree(aa);CHKERRQ(ierr); 339 } 340 341 ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 342 C->ops->solve = MatSolve_SeqSBAIJ_7; 343 C->ops->solvetranspose = MatSolve_SeqSBAIJ_7; 344 C->assembled = PETSC_TRUE; 345 C->preallocated = PETSC_TRUE; 346 ierr = PetscLogFlops(1.3333*343*b->mbs);CHKERRQ(ierr); /* from inverting diagonal blocks */ 347 PetscFunctionReturn(0); 348 } 349