1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.25 1996/08/04 23:12:52 bsmith Exp curfman $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "baij.h" 9 #include "src/vec/vecimpl.h" 10 #include "src/inline/ilu.h" 11 12 13 /* 14 The symbolic factorization code is identical to that for AIJ format, 15 except for very small changes since this is now a SeqBAIJ datastructure. 16 NOT good code reuse. 17 */ 18 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 19 { 20 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 21 IS isicol; 22 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 23 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 24 int *idnew, idx, row,m,fm, nnz, nzi,len, realloc = 0,nzbd,*im; 25 26 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 27 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 28 29 /* get new row pointers */ 30 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 31 ainew[0] = 0; 32 /* don't know how many column pointers are needed so estimate */ 33 jmax = (int) (f*ai[n] + 1); 34 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 35 /* fill is a linked list of nonzeros in active row */ 36 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 37 im = fill + n + 1; 38 /* idnew is location of diagonal in factor */ 39 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 40 idnew[0] = 0; 41 42 for ( i=0; i<n; i++ ) { 43 /* first copy previous fill into linked list */ 44 nnz = nz = ai[r[i]+1] - ai[r[i]]; 45 ajtmp = aj + ai[r[i]]; 46 fill[n] = n; 47 while (nz--) { 48 fm = n; 49 idx = ic[*ajtmp++]; 50 do { 51 m = fm; 52 fm = fill[m]; 53 } while (fm < idx); 54 fill[m] = idx; 55 fill[idx] = fm; 56 } 57 row = fill[n]; 58 while ( row < i ) { 59 ajtmp = ajnew + idnew[row] + 1; 60 nzbd = 1 + idnew[row] - ainew[row]; 61 nz = im[row] - nzbd; 62 fm = row; 63 while (nz-- > 0) { 64 idx = *ajtmp++; 65 nzbd++; 66 if (idx == i) im[row] = nzbd; 67 do { 68 m = fm; 69 fm = fill[m]; 70 } while (fm < idx); 71 if (fm != idx) { 72 fill[m] = idx; 73 fill[idx] = fm; 74 fm = idx; 75 nnz++; 76 } 77 } 78 row = fill[row]; 79 } 80 /* copy new filled row into permanent storage */ 81 ainew[i+1] = ainew[i] + nnz; 82 if (ainew[i+1] > jmax) { 83 /* allocate a longer ajnew */ 84 int maxadd; 85 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 86 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 87 jmax += maxadd; 88 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 89 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 90 PetscFree(ajnew); 91 ajnew = ajtmp; 92 realloc++; /* count how many times we realloc */ 93 } 94 ajtmp = ajnew + ainew[i]; 95 fm = fill[n]; 96 nzi = 0; 97 im[i] = nnz; 98 while (nnz--) { 99 if (fm < i) nzi++; 100 *ajtmp++ = fm; 101 fm = fill[fm]; 102 } 103 idnew[i] = ainew[i] + nzi; 104 } 105 106 PLogInfo(A, 107 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 108 realloc,f,((double)ainew[n])/((double)ai[i])); 109 110 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 111 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 112 113 PetscFree(fill); 114 115 /* put together the new matrix */ 116 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 117 PLogObjectParent(*B,isicol); 118 ierr = ISDestroy(isicol); CHKERRQ(ierr); 119 b = (Mat_SeqBAIJ *) (*B)->data; 120 PetscFree(b->imax); 121 b->singlemalloc = 0; 122 len = ainew[n]*sizeof(Scalar); 123 /* the next line frees the default space generated by the Create() */ 124 PetscFree(b->a); PetscFree(b->ilen); 125 b->a = (Scalar *) PetscMalloc( len*bs2 ); CHKPTRQ(b->a); 126 b->j = ajnew; 127 b->i = ainew; 128 b->diag = idnew; 129 b->ilen = 0; 130 b->imax = 0; 131 b->row = isrow; 132 b->col = iscol; 133 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 134 CHKPTRQ(b->solve_work); 135 /* In b structure: Free imax, ilen, old a, old j. 136 Allocate idnew, solve_work, new a, new j */ 137 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 138 b->maxnz = b->nz = ainew[n]; 139 140 return 0; 141 } 142 143 /* ----------------------------------------------------------- */ 144 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 145 { 146 Mat C = *B; 147 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 148 IS iscol = b->col, isrow = b->row, isicol; 149 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 150 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j; 151 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = bs*bs,*v_pivots; 152 Scalar *ba = b->a,*aa = a->a; 153 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 154 register int *pj; 155 156 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 157 PLogObjectParent(*B,isicol); 158 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 159 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 160 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 161 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 162 /* generate work space needed by dense LU factorization */ 163 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 164 CHKPTRQ(v_work); 165 multiplier = v_work + bs; 166 v_pivots = (int *) (multiplier + bs2); 167 168 /* flops in while loop */ 169 bslog = 2*bs*bs2; 170 171 for ( i=0; i<n; i++ ) { 172 nz = bi[i+1] - bi[i]; 173 ajtmp = bj + bi[i]; 174 for ( j=0; j<nz; j++ ) { 175 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 176 } 177 /* load in initial (unfactored row) */ 178 nz = ai[r[i]+1] - ai[r[i]]; 179 ajtmpold = aj + ai[r[i]]; 180 v = aa + bs2*ai[r[i]]; 181 for ( j=0; j<nz; j++ ) { 182 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 183 } 184 row = *ajtmp++; 185 while (row < i) { 186 pc = rtmp + bs2*row; 187 /* if (*pc) { */ 188 pv = ba + bs2*diag_offset[row]; 189 pj = bj + diag_offset[row] + 1; 190 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 191 nz = bi[row+1] - diag_offset[row] - 1; 192 pv += bs2; 193 for (j=0; j<nz; j++) { 194 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 195 } 196 PLogFlops(bslog*(nz+1)-bs); 197 /* } */ 198 row = *ajtmp++; 199 } 200 /* finished row so stick it into b->a */ 201 pv = ba + bs2*bi[i]; 202 pj = bj + bi[i]; 203 nz = bi[i+1] - bi[i]; 204 for ( j=0; j<nz; j++ ) { 205 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 206 } 207 diag = diag_offset[i] - bi[i]; 208 /* invert diagonal block */ 209 w = pv + bs2*diag; 210 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 211 } 212 213 PetscFree(rtmp); PetscFree(v_work); 214 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 215 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 216 ierr = ISDestroy(isicol); CHKERRQ(ierr); 217 C->factor = FACTOR_LU; 218 C->assembled = PETSC_TRUE; 219 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 220 return 0; 221 } 222 /* ------------------------------------------------------------*/ 223 /* 224 Version for when blocks are 5 by 5 225 */ 226 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 227 { 228 Mat C = *B; 229 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 230 IS iscol = b->col, isrow = b->row, isicol; 231 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 232 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 233 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 234 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 235 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 236 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 237 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 238 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 239 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 240 Scalar *ba = b->a,*aa = a->a; 241 register int *pj; 242 243 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 244 PLogObjectParent(*B,isicol); 245 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 246 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 247 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 248 249 for ( i=0; i<n; i++ ) { 250 nz = bi[i+1] - bi[i]; 251 ajtmp = bj + bi[i]; 252 for ( j=0; j<nz; j++ ) { 253 x = rtmp+25*ajtmp[j]; 254 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 255 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 256 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 257 } 258 /* load in initial (unfactored row) */ 259 idx = r[i]; 260 nz = ai[idx+1] - ai[idx]; 261 ajtmpold = aj + ai[idx]; 262 v = aa + 25*ai[idx]; 263 for ( j=0; j<nz; j++ ) { 264 x = rtmp+25*ic[ajtmpold[j]]; 265 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 266 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 267 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 268 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 269 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 270 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 271 v += 25; 272 } 273 row = *ajtmp++; 274 while (row < i) { 275 pc = rtmp + 25*row; 276 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 277 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 278 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 279 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 280 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 281 p25 = pc[24]; 282 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 283 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 284 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 285 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 286 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 287 p24 != 0.0 || p25 != 0.0) { 288 pv = ba + 25*diag_offset[row]; 289 pj = bj + diag_offset[row] + 1; 290 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 291 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 292 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 293 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 294 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 295 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 296 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 297 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 298 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 299 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 300 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 301 302 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 303 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 304 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 305 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 306 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 307 308 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 309 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 310 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 311 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 312 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 313 314 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 315 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 316 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 317 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 318 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 319 320 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 321 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 322 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 323 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 324 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 325 326 nz = bi[row+1] - diag_offset[row] - 1; 327 pv += 25; 328 for (j=0; j<nz; j++) { 329 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 330 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 331 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 332 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 333 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 334 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 335 x = rtmp + 25*pj[j]; 336 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 337 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 338 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 339 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 340 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 341 342 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 343 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 344 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 345 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 346 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 347 348 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 349 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 350 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 351 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 352 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 353 354 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 355 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 356 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 357 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 358 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 359 360 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 361 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 362 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 363 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 364 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 365 366 pv += 25; 367 } 368 PLogFlops(250*nz+225); 369 } 370 row = *ajtmp++; 371 } 372 /* finished row so stick it into b->a */ 373 pv = ba + 25*bi[i]; 374 pj = bj + bi[i]; 375 nz = bi[i+1] - bi[i]; 376 for ( j=0; j<nz; j++ ) { 377 x = rtmp+25*pj[j]; 378 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 379 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 380 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 381 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 382 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 383 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 384 pv += 25; 385 } 386 /* invert diagonal block */ 387 w = ba + 25*diag_offset[i]; 388 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 389 } 390 391 PetscFree(rtmp); 392 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 393 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 394 ierr = ISDestroy(isicol); CHKERRQ(ierr); 395 C->factor = FACTOR_LU; 396 C->assembled = PETSC_TRUE; 397 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 398 return 0; 399 } 400 401 /* ------------------------------------------------------------*/ 402 /* 403 Version for when blocks are 4 by 4 404 */ 405 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 406 { 407 Mat C = *B; 408 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 409 IS iscol = b->col, isrow = b->row, isicol; 410 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 411 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 412 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 413 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 414 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 415 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 416 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 417 Scalar m13,m14,m15,m16; 418 Scalar *ba = b->a,*aa = a->a; 419 register int *pj; 420 421 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 422 PLogObjectParent(*B,isicol); 423 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 424 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 425 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 426 427 for ( i=0; i<n; i++ ) { 428 nz = bi[i+1] - bi[i]; 429 ajtmp = bj + bi[i]; 430 for ( j=0; j<nz; j++ ) { 431 x = rtmp+16*ajtmp[j]; 432 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 433 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 434 } 435 /* load in initial (unfactored row) */ 436 idx = r[i]; 437 nz = ai[idx+1] - ai[idx]; 438 ajtmpold = aj + ai[idx]; 439 v = aa + 16*ai[idx]; 440 for ( j=0; j<nz; j++ ) { 441 x = rtmp+16*ic[ajtmpold[j]]; 442 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 443 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 444 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 445 x[14] = v[14]; x[15] = v[15]; 446 v += 16; 447 } 448 row = *ajtmp++; 449 while (row < i) { 450 pc = rtmp + 16*row; 451 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 452 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 453 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 454 p15 = pc[14]; p16 = pc[15]; 455 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 456 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 457 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 458 || p16 != 0.0) { 459 pv = ba + 16*diag_offset[row]; 460 pj = bj + diag_offset[row] + 1; 461 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 462 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 463 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 464 x15 = pv[14]; x16 = pv[15]; 465 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 466 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 467 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 468 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 469 470 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 471 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 472 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 473 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 474 475 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 476 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 477 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 478 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 479 480 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 481 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 482 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 483 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 484 485 nz = bi[row+1] - diag_offset[row] - 1; 486 pv += 16; 487 for (j=0; j<nz; j++) { 488 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 489 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 490 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 491 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 492 x = rtmp + 16*pj[j]; 493 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 494 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 495 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 496 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 497 498 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 499 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 500 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 501 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 502 503 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 504 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 505 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 506 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 507 508 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 509 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 510 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 511 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 512 513 pv += 16; 514 } 515 PLogFlops(128*nz+112); 516 } 517 row = *ajtmp++; 518 } 519 /* finished row so stick it into b->a */ 520 pv = ba + 16*bi[i]; 521 pj = bj + bi[i]; 522 nz = bi[i+1] - bi[i]; 523 for ( j=0; j<nz; j++ ) { 524 x = rtmp+16*pj[j]; 525 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 526 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 527 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 528 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 529 pv += 16; 530 } 531 /* invert diagonal block */ 532 w = ba + 16*diag_offset[i]; 533 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 534 } 535 536 PetscFree(rtmp); 537 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 538 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 539 ierr = ISDestroy(isicol); CHKERRQ(ierr); 540 C->factor = FACTOR_LU; 541 C->assembled = PETSC_TRUE; 542 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 543 return 0; 544 } 545 /* ------------------------------------------------------------*/ 546 /* 547 Version for when blocks are 3 by 3 548 */ 549 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 550 { 551 Mat C = *B; 552 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 553 IS iscol = b->col, isrow = b->row, isicol; 554 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 555 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 556 int *diag_offset = b->diag,idx; 557 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 558 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 559 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 560 Scalar *ba = b->a,*aa = a->a; 561 register int *pj; 562 563 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 564 PLogObjectParent(*B,isicol); 565 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 566 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 567 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 568 569 for ( i=0; i<n; i++ ) { 570 nz = bi[i+1] - bi[i]; 571 ajtmp = bj + bi[i]; 572 for ( j=0; j<nz; j++ ) { 573 x = rtmp + 9*ajtmp[j]; 574 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 575 } 576 /* load in initial (unfactored row) */ 577 idx = r[i]; 578 nz = ai[idx+1] - ai[idx]; 579 ajtmpold = aj + ai[idx]; 580 v = aa + 9*ai[idx]; 581 for ( j=0; j<nz; j++ ) { 582 x = rtmp + 9*ic[ajtmpold[j]]; 583 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 584 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 585 v += 9; 586 } 587 row = *ajtmp++; 588 while (row < i) { 589 pc = rtmp + 9*row; 590 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 591 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 592 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 593 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 594 pv = ba + 9*diag_offset[row]; 595 pj = bj + diag_offset[row] + 1; 596 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 597 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 598 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 599 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 600 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 601 602 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 603 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 604 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 605 606 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 607 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 608 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 609 nz = bi[row+1] - diag_offset[row] - 1; 610 pv += 9; 611 for (j=0; j<nz; j++) { 612 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 613 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 614 x = rtmp + 9*pj[j]; 615 x[0] -= m1*x1 + m4*x2 + m7*x3; 616 x[1] -= m2*x1 + m5*x2 + m8*x3; 617 x[2] -= m3*x1 + m6*x2 + m9*x3; 618 619 x[3] -= m1*x4 + m4*x5 + m7*x6; 620 x[4] -= m2*x4 + m5*x5 + m8*x6; 621 x[5] -= m3*x4 + m6*x5 + m9*x6; 622 623 x[6] -= m1*x7 + m4*x8 + m7*x9; 624 x[7] -= m2*x7 + m5*x8 + m8*x9; 625 x[8] -= m3*x7 + m6*x8 + m9*x9; 626 pv += 9; 627 } 628 PLogFlops(54*nz+36); 629 } 630 row = *ajtmp++; 631 } 632 /* finished row so stick it into b->a */ 633 pv = ba + 9*bi[i]; 634 pj = bj + bi[i]; 635 nz = bi[i+1] - bi[i]; 636 for ( j=0; j<nz; j++ ) { 637 x = rtmp + 9*pj[j]; 638 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 639 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 640 pv += 9; 641 } 642 /* invert diagonal block */ 643 w = ba + 9*diag_offset[i]; 644 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 645 } 646 647 PetscFree(rtmp); 648 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 649 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 650 ierr = ISDestroy(isicol); CHKERRQ(ierr); 651 C->factor = FACTOR_LU; 652 C->assembled = PETSC_TRUE; 653 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 654 return 0; 655 } 656 657 /* ------------------------------------------------------------*/ 658 /* 659 Version for when blocks are 2 by 2 660 */ 661 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 662 { 663 Mat C = *B; 664 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 665 IS iscol = b->col, isrow = b->row, isicol; 666 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 667 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 668 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 669 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 670 Scalar p1,p2,p3,p4,v_work[2]; 671 Scalar *ba = b->a,*aa = a->a; 672 register int *pj; 673 674 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 675 PLogObjectParent(*B,isicol); 676 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 677 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 678 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 679 680 for ( i=0; i<n; i++ ) { 681 nz = bi[i+1] - bi[i]; 682 ajtmp = bj + bi[i]; 683 for ( j=0; j<nz; j++ ) { 684 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 685 } 686 /* load in initial (unfactored row) */ 687 idx = r[i]; 688 nz = ai[idx+1] - ai[idx]; 689 ajtmpold = aj + ai[idx]; 690 v = aa + 4*ai[idx]; 691 for ( j=0; j<nz; j++ ) { 692 x = rtmp+4*ic[ajtmpold[j]]; 693 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 694 v += 4; 695 } 696 row = *ajtmp++; 697 while (row < i) { 698 pc = rtmp + 4*row; 699 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 700 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 701 pv = ba + 4*diag_offset[row]; 702 pj = bj + diag_offset[row] + 1; 703 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 704 pc[0] = m1 = p1*x1 + p3*x2; 705 pc[1] = m2 = p2*x1 + p4*x2; 706 pc[2] = m3 = p1*x3 + p3*x4; 707 pc[3] = m4 = p2*x3 + p4*x4; 708 nz = bi[row+1] - diag_offset[row] - 1; 709 pv += 4; 710 for (j=0; j<nz; j++) { 711 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 712 x = rtmp + 4*pj[j]; 713 x[0] -= m1*x1 + m3*x2; 714 x[1] -= m2*x1 + m4*x2; 715 x[2] -= m1*x3 + m3*x4; 716 x[3] -= m2*x3 + m4*x4; 717 pv += 4; 718 } 719 PLogFlops(16*nz+12); 720 } 721 row = *ajtmp++; 722 } 723 /* finished row so stick it into b->a */ 724 pv = ba + 4*bi[i]; 725 pj = bj + bi[i]; 726 nz = bi[i+1] - bi[i]; 727 for ( j=0; j<nz; j++ ) { 728 x = rtmp+4*pj[j]; 729 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 730 pv += 4; 731 } 732 /* invert diagonal block */ 733 w = ba + 4*diag_offset[i]; 734 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 735 } 736 737 PetscFree(rtmp); 738 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 739 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 740 ierr = ISDestroy(isicol); CHKERRQ(ierr); 741 C->factor = FACTOR_LU; 742 C->assembled = PETSC_TRUE; 743 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 744 return 0; 745 } 746 747 /* ----------------------------------------------------------- */ 748 /* 749 Version for when blocks are 1 by 1. 750 */ 751 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 752 { 753 Mat C = *B; 754 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 755 IS iscol = b->col, isrow = b->row, isicol; 756 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 757 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 758 int *diag_offset = b->diag,diag; 759 register Scalar *pv,*v,*rtmp,multiplier,*pc; 760 Scalar *ba = b->a,*aa = a->a; 761 register int *pj; 762 763 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 764 PLogObjectParent(*B,isicol); 765 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 766 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 767 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 768 769 for ( i=0; i<n; i++ ) { 770 nz = bi[i+1] - bi[i]; 771 ajtmp = bj + bi[i]; 772 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 773 774 /* load in initial (unfactored row) */ 775 nz = ai[r[i]+1] - ai[r[i]]; 776 ajtmpold = aj + ai[r[i]]; 777 v = aa + ai[r[i]]; 778 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 779 780 row = *ajtmp++; 781 while (row < i) { 782 pc = rtmp + row; 783 if (*pc != 0.0) { 784 pv = ba + diag_offset[row]; 785 pj = bj + diag_offset[row] + 1; 786 multiplier = *pc * *pv++; 787 *pc = multiplier; 788 nz = bi[row+1] - diag_offset[row] - 1; 789 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 790 PLogFlops(1+2*nz); 791 } 792 row = *ajtmp++; 793 } 794 /* finished row so stick it into b->a */ 795 pv = ba + bi[i]; 796 pj = bj + bi[i]; 797 nz = bi[i+1] - bi[i]; 798 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 799 diag = diag_offset[i] - bi[i]; 800 /* check pivot entry for current row */ 801 if (pv[diag] == 0.0) { 802 SETERRQ(1,"MatLUFactorNumeric_SeqAIJ:Zero pivot"); 803 } 804 pv[diag] = 1.0/pv[diag]; 805 } 806 807 PetscFree(rtmp); 808 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 809 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 810 ierr = ISDestroy(isicol); CHKERRQ(ierr); 811 C->factor = FACTOR_LU; 812 C->assembled = PETSC_TRUE; 813 PLogFlops(b->n); 814 return 0; 815 } 816 817 /* ----------------------------------------------------------- */ 818 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 819 { 820 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 821 int ierr; 822 Mat C; 823 824 ierr = MatLUFactorSymbolic_SeqBAIJ(A,row,col,f,&C); CHKERRQ(ierr); 825 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 826 827 /* free all the data structures from mat */ 828 PetscFree(mat->a); 829 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 830 if (mat->diag) PetscFree(mat->diag); 831 if (mat->ilen) PetscFree(mat->ilen); 832 if (mat->imax) PetscFree(mat->imax); 833 if (mat->solve_work) PetscFree(mat->solve_work); 834 if (mat->mult_work) PetscFree(mat->mult_work); 835 PetscFree(mat); 836 837 PetscMemcpy(A,C,sizeof(struct _Mat)); 838 PetscHeaderDestroy(C); 839 return 0; 840 } 841 /* ----------------------------------------------------------- */ 842 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 843 { 844 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 845 IS iscol=a->col,isrow=a->row; 846 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 847 int nz,bs=a->bs,bs2=a->bs2; 848 Scalar *aa=a->a,*sum; 849 register Scalar *x,*b,*lsum,*tmp,*v; 850 851 VecGetArray_Fast(bb,b); 852 VecGetArray_Fast(xx,x); 853 tmp = a->solve_work; 854 855 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 856 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 857 858 /* forward solve the lower triangular */ 859 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 860 for ( i=1; i<n; i++ ) { 861 v = aa + bs2*ai[i]; 862 vi = aj + ai[i]; 863 nz = a->diag[i] - ai[i]; 864 sum = tmp + bs*i; 865 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 866 while (nz--) { 867 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 868 v += bs2; 869 } 870 } 871 /* backward solve the upper triangular */ 872 lsum = a->solve_work + a->n; 873 for ( i=n-1; i>=0; i-- ){ 874 v = aa + bs2*(a->diag[i] + 1); 875 vi = aj + a->diag[i] + 1; 876 nz = ai[i+1] - a->diag[i] - 1; 877 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 878 while (nz--) { 879 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 880 v += bs2; 881 } 882 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 883 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 884 } 885 886 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 887 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 888 VecRestoreArray_Fast(bb,b); 889 VecRestoreArray_Fast(xx,x); 890 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 891 return 0; 892 } 893 894 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 895 { 896 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 897 IS iscol=a->col,isrow=a->row; 898 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 899 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 900 register Scalar *x,*b,*tmp,*v; 901 902 VecGetArray_Fast(bb,b); 903 VecGetArray_Fast(xx,x); 904 tmp = a->solve_work; 905 906 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 907 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 908 909 /* forward solve the lower triangular */ 910 idx = 5*(*r++); 911 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 912 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 913 for ( i=1; i<n; i++ ) { 914 v = aa + 25*ai[i]; 915 vi = aj + ai[i]; 916 nz = a->diag[i] - ai[i]; 917 idx = 5*(*r++); 918 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 919 sum5 = b[4+idx]; 920 while (nz--) { 921 idx = 5*(*vi++); 922 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 923 x4 = tmp[3+idx];x5 = tmp[4+idx]; 924 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 925 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 926 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 927 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 928 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 929 v += 25; 930 } 931 idx = 5*i; 932 tmp[idx] = sum1;tmp[1+idx] = sum2; 933 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 934 } 935 /* backward solve the upper triangular */ 936 for ( i=n-1; i>=0; i-- ){ 937 v = aa + 25*a->diag[i] + 25; 938 vi = aj + a->diag[i] + 1; 939 nz = ai[i+1] - a->diag[i] - 1; 940 idt = 5*i; 941 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 942 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 943 while (nz--) { 944 idx = 5*(*vi++); 945 x1 = tmp[idx]; x2 = tmp[1+idx]; 946 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 947 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 948 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 949 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 950 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 951 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 952 v += 25; 953 } 954 idc = 5*(*c--); 955 v = aa + 25*a->diag[i]; 956 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 957 v[15]*sum4+v[20]*sum5; 958 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 959 v[16]*sum4+v[21]*sum5; 960 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 961 v[17]*sum4+v[22]*sum5; 962 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 963 v[18]*sum4+v[23]*sum5; 964 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 965 v[19]*sum4+v[24]*sum5; 966 } 967 968 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 969 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 970 VecRestoreArray_Fast(bb,b); 971 VecRestoreArray_Fast(xx,x); 972 PLogFlops(2*25*(a->nz) - a->n); 973 return 0; 974 } 975 976 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 977 { 978 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 979 IS iscol=a->col,isrow=a->row; 980 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 981 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 982 register Scalar *x,*b,*tmp,*v; 983 984 VecGetArray_Fast(bb,b); 985 VecGetArray_Fast(xx,x); 986 tmp = a->solve_work; 987 988 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 989 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 990 991 /* forward solve the lower triangular */ 992 idx = 4*(*r++); 993 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 994 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 995 for ( i=1; i<n; i++ ) { 996 v = aa + 16*ai[i]; 997 vi = aj + ai[i]; 998 nz = a->diag[i] - ai[i]; 999 idx = 4*(*r++); 1000 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1001 while (nz--) { 1002 idx = 4*(*vi++); 1003 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1004 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1005 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1006 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1007 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1008 v += 16; 1009 } 1010 idx = 4*i; 1011 tmp[idx] = sum1;tmp[1+idx] = sum2; 1012 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1013 } 1014 /* backward solve the upper triangular */ 1015 for ( i=n-1; i>=0; i-- ){ 1016 v = aa + 16*a->diag[i] + 16; 1017 vi = aj + a->diag[i] + 1; 1018 nz = ai[i+1] - a->diag[i] - 1; 1019 idt = 4*i; 1020 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1021 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1022 while (nz--) { 1023 idx = 4*(*vi++); 1024 x1 = tmp[idx]; x2 = tmp[1+idx]; 1025 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1026 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1027 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1028 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1029 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1030 v += 16; 1031 } 1032 idc = 4*(*c--); 1033 v = aa + 16*a->diag[i]; 1034 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1035 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1036 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1037 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1038 } 1039 1040 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1041 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1042 VecRestoreArray_Fast(bb,b); 1043 VecRestoreArray_Fast(xx,x); 1044 PLogFlops(2*16*(a->nz) - a->n); 1045 return 0; 1046 } 1047 1048 1049 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1050 { 1051 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1052 IS iscol=a->col,isrow=a->row; 1053 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 1054 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1055 register Scalar *x,*b,*tmp,*v; 1056 1057 VecGetArray_Fast(bb,b); 1058 VecGetArray_Fast(xx,x); 1059 tmp = a->solve_work; 1060 1061 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1062 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1063 1064 /* forward solve the lower triangular */ 1065 idx = 3*(*r++); 1066 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1067 for ( i=1; i<n; i++ ) { 1068 v = aa + 9*ai[i]; 1069 vi = aj + ai[i]; 1070 nz = a->diag[i] - ai[i]; 1071 idx = 3*(*r++); 1072 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1073 while (nz--) { 1074 idx = 3*(*vi++); 1075 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1076 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1077 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1078 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1079 v += 9; 1080 } 1081 idx = 3*i; 1082 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1083 } 1084 /* backward solve the upper triangular */ 1085 for ( i=n-1; i>=0; i-- ){ 1086 v = aa + 9*a->diag[i] + 9; 1087 vi = aj + a->diag[i] + 1; 1088 nz = ai[i+1] - a->diag[i] - 1; 1089 idt = 3*i; 1090 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1091 while (nz--) { 1092 idx = 3*(*vi++); 1093 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1094 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1095 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1096 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1097 v += 9; 1098 } 1099 idc = 3*(*c--); 1100 v = aa + 9*a->diag[i]; 1101 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1102 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1103 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1104 } 1105 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1106 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1107 VecRestoreArray_Fast(bb,b); 1108 VecRestoreArray_Fast(xx,x); 1109 PLogFlops(2*9*(a->nz) - a->n); 1110 return 0; 1111 } 1112 1113 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1114 { 1115 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1116 IS iscol=a->col,isrow=a->row; 1117 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 1118 Scalar *aa=a->a,sum1,sum2,x1,x2; 1119 register Scalar *x,*b,*tmp,*v; 1120 1121 VecGetArray_Fast(bb,b); 1122 VecGetArray_Fast(xx,x); 1123 tmp = a->solve_work; 1124 1125 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1126 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1127 1128 /* forward solve the lower triangular */ 1129 idx = 2*(*r++); 1130 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1131 for ( i=1; i<n; i++ ) { 1132 v = aa + 4*ai[i]; 1133 vi = aj + ai[i]; 1134 nz = a->diag[i] - ai[i]; 1135 idx = 2*(*r++); 1136 sum1 = b[idx]; sum2 = b[1+idx]; 1137 while (nz--) { 1138 idx = 2*(*vi++); 1139 x1 = tmp[idx]; x2 = tmp[1+idx]; 1140 sum1 -= v[0]*x1 + v[2]*x2; 1141 sum2 -= v[1]*x1 + v[3]*x2; 1142 v += 4; 1143 } 1144 idx = 2*i; 1145 tmp[idx] = sum1; tmp[1+idx] = sum2; 1146 } 1147 /* backward solve the upper triangular */ 1148 for ( i=n-1; i>=0; i-- ){ 1149 v = aa + 4*a->diag[i] + 4; 1150 vi = aj + a->diag[i] + 1; 1151 nz = ai[i+1] - a->diag[i] - 1; 1152 idt = 2*i; 1153 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1154 while (nz--) { 1155 idx = 2*(*vi++); 1156 x1 = tmp[idx]; x2 = tmp[1+idx]; 1157 sum1 -= v[0]*x1 + v[2]*x2; 1158 sum2 -= v[1]*x1 + v[3]*x2; 1159 v += 4; 1160 } 1161 idc = 2*(*c--); 1162 v = aa + 4*a->diag[i]; 1163 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1164 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1165 } 1166 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1167 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1168 VecRestoreArray_Fast(bb,b); 1169 VecRestoreArray_Fast(xx,x); 1170 PLogFlops(2*4*(a->nz) - a->n); 1171 return 0; 1172 } 1173 1174 1175 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1176 { 1177 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1178 IS iscol=a->col,isrow=a->row; 1179 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz; 1180 Scalar *aa=a->a,sum1; 1181 register Scalar *x,*b,*tmp,*v; 1182 1183 VecGetArray_Fast(bb,b); 1184 VecGetArray_Fast(xx,x); 1185 tmp = a->solve_work; 1186 1187 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1188 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1189 1190 /* forward solve the lower triangular */ 1191 tmp[0] = b[*r++]; 1192 for ( i=1; i<n; i++ ) { 1193 v = aa + ai[i]; 1194 vi = aj + ai[i]; 1195 nz = a->diag[i] - ai[i]; 1196 sum1 = b[*r++]; 1197 while (nz--) { 1198 sum1 -= (*v++)*tmp[*vi++]; 1199 } 1200 tmp[i] = sum1; 1201 } 1202 /* backward solve the upper triangular */ 1203 for ( i=n-1; i>=0; i-- ){ 1204 v = aa + a->diag[i] + 1; 1205 vi = aj + a->diag[i] + 1; 1206 nz = ai[i+1] - a->diag[i] - 1; 1207 sum1 = tmp[i]; 1208 while (nz--) { 1209 sum1 -= (*v++)*tmp[*vi++]; 1210 } 1211 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1212 } 1213 1214 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1215 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1216 VecRestoreArray_Fast(bb,b); 1217 VecRestoreArray_Fast(xx,x); 1218 PLogFlops(2*1*(a->nz) - a->n); 1219 return 0; 1220 } 1221 1222 /* ----------------------------------------------------------------*/ 1223 /* 1224 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1225 except that the data structure of Mat_SeqAIJ is slightly different. 1226 Not a good example of code reuse. 1227 */ 1228 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1229 Mat *fact) 1230 { 1231 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1232 IS isicol; 1233 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1234 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1235 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1236 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1237 PetscTruth col_identity, row_identity; 1238 1239 /* special case that simply copies fill pattern */ 1240 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1241 if (levels == 0 && row_identity && col_identity) { 1242 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1243 (*fact)->factor = FACTOR_LU; 1244 b = (Mat_SeqBAIJ *) (*fact)->data; 1245 if (!b->diag) { 1246 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1247 } 1248 b->row = isrow; 1249 b->col = iscol; 1250 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1251 return 0; 1252 } 1253 1254 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1255 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1256 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1257 1258 /* get new row pointers */ 1259 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1260 ainew[0] = 0; 1261 /* don't know how many column pointers are needed so estimate */ 1262 jmax = (int) (f*ai[n] + 1); 1263 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1264 /* ajfill is level of fill for each fill entry */ 1265 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1266 /* fill is a linked list of nonzeros in active row */ 1267 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1268 /* im is level for each filled value */ 1269 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1270 /* dloc is location of diagonal in factor */ 1271 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1272 dloc[0] = 0; 1273 for ( prow=0; prow<n; prow++ ) { 1274 /* first copy previous fill into linked list */ 1275 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1276 xi = aj + ai[r[prow]]; 1277 fill[n] = n; 1278 while (nz--) { 1279 fm = n; 1280 idx = ic[*xi++]; 1281 do { 1282 m = fm; 1283 fm = fill[m]; 1284 } while (fm < idx); 1285 fill[m] = idx; 1286 fill[idx] = fm; 1287 im[idx] = 0; 1288 } 1289 nzi = 0; 1290 row = fill[n]; 1291 while ( row < prow ) { 1292 incrlev = im[row] + 1; 1293 nz = dloc[row]; 1294 xi = ajnew + ainew[row] + nz; 1295 flev = ajfill + ainew[row] + nz + 1; 1296 nnz = ainew[row+1] - ainew[row] - nz - 1; 1297 if (*xi++ != row) { 1298 SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:zero pivot"); 1299 } 1300 fm = row; 1301 while (nnz-- > 0) { 1302 idx = *xi++; 1303 if (*flev + incrlev > levels) { 1304 flev++; 1305 continue; 1306 } 1307 do { 1308 m = fm; 1309 fm = fill[m]; 1310 } while (fm < idx); 1311 if (fm != idx) { 1312 im[idx] = *flev + incrlev; 1313 fill[m] = idx; 1314 fill[idx] = fm; 1315 fm = idx; 1316 nzf++; 1317 } 1318 else { 1319 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1320 } 1321 flev++; 1322 } 1323 row = fill[row]; 1324 nzi++; 1325 } 1326 /* copy new filled row into permanent storage */ 1327 ainew[prow+1] = ainew[prow] + nzf; 1328 if (ainew[prow+1] > jmax) { 1329 /* allocate a longer ajnew */ 1330 int maxadd; 1331 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1332 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1333 jmax += maxadd; 1334 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1335 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1336 PetscFree(ajnew); 1337 ajnew = xi; 1338 /* allocate a longer ajfill */ 1339 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1340 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1341 PetscFree(ajfill); 1342 ajfill = xi; 1343 realloc++; 1344 } 1345 xi = ajnew + ainew[prow]; 1346 flev = ajfill + ainew[prow]; 1347 dloc[prow] = nzi; 1348 fm = fill[n]; 1349 while (nzf--) { 1350 *xi++ = fm; 1351 *flev++ = im[fm]; 1352 fm = fill[fm]; 1353 } 1354 } 1355 PetscFree(ajfill); 1356 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1357 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1358 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1359 PetscFree(fill); PetscFree(im); 1360 1361 PLogInfo(A, 1362 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1363 realloc,f,((double)ainew[n])/((double)ai[prow])); 1364 1365 /* put together the new matrix */ 1366 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1367 b = (Mat_SeqBAIJ *) (*fact)->data; 1368 PetscFree(b->imax); 1369 b->singlemalloc = 0; 1370 len = bs2*ainew[n]*sizeof(Scalar); 1371 /* the next line frees the default space generated by the Create() */ 1372 PetscFree(b->a); PetscFree(b->ilen); 1373 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1374 b->j = ajnew; 1375 b->i = ainew; 1376 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1377 b->diag = dloc; 1378 b->ilen = 0; 1379 b->imax = 0; 1380 b->row = isrow; 1381 b->col = iscol; 1382 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1383 CHKPTRQ(b->solve_work); 1384 /* In b structure: Free imax, ilen, old a, old j. 1385 Allocate dloc, solve_work, new a, new j */ 1386 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1387 b->maxnz = b->nz = ainew[n]; 1388 (*fact)->factor = FACTOR_LU; 1389 return 0; 1390 } 1391 1392 1393 1394 1395