1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.20 1996/04/28 02:09:52 bsmith Exp balay $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "baij.h" 9 #include "src/inline/ilu.h" 10 11 /* 12 The symbolic factorization code is identical to that for AIJ format, 13 except for very small changes since this is now a SeqBAIJ datastructure. 14 NOT good code reuse. 15 */ 16 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 17 { 18 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 19 IS isicol; 20 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 21 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 22 int *idnew, idx, row,m,fm, nnz, nzi,len, realloc = 0,nzbd,*im; 23 24 if (a->m != a->n) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must be square"); 25 if (!isrow) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must have row permutation"); 26 if (!iscol) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must have col. permutation"); 27 28 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 29 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 30 31 /* get new row pointers */ 32 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 33 ainew[0] = 0; 34 /* don't know how many column pointers are needed so estimate */ 35 jmax = (int) (f*ai[n] + 1); 36 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 37 /* fill is a linked list of nonzeros in active row */ 38 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 39 im = fill + n + 1; 40 /* idnew is location of diagonal in factor */ 41 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 42 idnew[0] = 0; 43 44 for ( i=0; i<n; i++ ) { 45 /* first copy previous fill into linked list */ 46 nnz = nz = ai[r[i]+1] - ai[r[i]]; 47 ajtmp = aj + ai[r[i]]; 48 fill[n] = n; 49 while (nz--) { 50 fm = n; 51 idx = ic[*ajtmp++]; 52 do { 53 m = fm; 54 fm = fill[m]; 55 } while (fm < idx); 56 fill[m] = idx; 57 fill[idx] = fm; 58 } 59 row = fill[n]; 60 while ( row < i ) { 61 ajtmp = ajnew + idnew[row] + 1; 62 nzbd = 1 + idnew[row] - ainew[row]; 63 nz = im[row] - nzbd; 64 fm = row; 65 while (nz-- > 0) { 66 idx = *ajtmp++; 67 nzbd++; 68 if (idx == i) im[row] = nzbd; 69 do { 70 m = fm; 71 fm = fill[m]; 72 } while (fm < idx); 73 if (fm != idx) { 74 fill[m] = idx; 75 fill[idx] = fm; 76 fm = idx; 77 nnz++; 78 } 79 } 80 row = fill[row]; 81 } 82 /* copy new filled row into permanent storage */ 83 ainew[i+1] = ainew[i] + nnz; 84 if (ainew[i+1] > jmax) { 85 /* allocate a longer ajnew */ 86 int maxadd; 87 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 88 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 89 jmax += maxadd; 90 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 91 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 92 PetscFree(ajnew); 93 ajnew = ajtmp; 94 realloc++; /* count how many times we realloc */ 95 } 96 ajtmp = ajnew + ainew[i]; 97 fm = fill[n]; 98 nzi = 0; 99 im[i] = nnz; 100 while (nnz--) { 101 if (fm < i) nzi++; 102 *ajtmp++ = fm; 103 fm = fill[fm]; 104 } 105 idnew[i] = ainew[i] + nzi; 106 } 107 108 PLogInfo(A, 109 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 110 realloc,f,((double)ainew[n])/((double)ai[i])); 111 112 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 113 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 114 115 PetscFree(fill); 116 117 /* put together the new matrix */ 118 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 119 PLogObjectParent(*B,isicol); 120 ierr = ISDestroy(isicol); CHKERRQ(ierr); 121 b = (Mat_SeqBAIJ *) (*B)->data; 122 PetscFree(b->imax); 123 b->singlemalloc = 0; 124 len = ainew[n]*sizeof(Scalar); 125 /* the next line frees the default space generated by the Create() */ 126 PetscFree(b->a); PetscFree(b->ilen); 127 b->a = (Scalar *) PetscMalloc( len*bs2 ); CHKPTRQ(b->a); 128 b->j = ajnew; 129 b->i = ainew; 130 b->diag = idnew; 131 b->ilen = 0; 132 b->imax = 0; 133 b->row = isrow; 134 b->col = iscol; 135 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 136 CHKPTRQ(b->solve_work); 137 /* In b structure: Free imax, ilen, old a, old j. 138 Allocate idnew, solve_work, new a, new j */ 139 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 140 b->maxnz = b->nz = ainew[n]; 141 142 return 0; 143 } 144 145 /* ----------------------------------------------------------- */ 146 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 147 { 148 Mat C = *B; 149 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 150 IS iscol = b->col, isrow = b->row, isicol; 151 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 152 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j; 153 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = bs*bs,*v_pivots; 154 Scalar *ba = b->a,*aa = a->a; 155 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 156 register int *pj; 157 158 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 159 PLogObjectParent(*B,isicol); 160 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 161 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 162 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 163 164 /* generate work space needed by dense LU factorization */ 165 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 166 CHKPTRQ(v_work); 167 multiplier = v_work + bs; 168 v_pivots = (int *) (multiplier + bs2); 169 170 /* flops in while loop */ 171 bslog = 2*bs*bs2; 172 173 for ( i=0; i<n; i++ ) { 174 nz = bi[i+1] - bi[i]; 175 ajtmp = bj + bi[i]; 176 for ( j=0; j<nz; j++ ) { 177 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 178 } 179 /* load in initial (unfactored row) */ 180 nz = ai[r[i]+1] - ai[r[i]]; 181 ajtmpold = aj + ai[r[i]]; 182 v = aa + bs2*ai[r[i]]; 183 for ( j=0; j<nz; j++ ) { 184 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 185 } 186 row = *ajtmp++; 187 while (row < i) { 188 pc = rtmp + bs2*row; 189 /* if (*pc) { */ 190 pv = ba + bs2*diag_offset[row]; 191 pj = bj + diag_offset[row] + 1; 192 /* 193 BLgemm_("N","N",&bs,&bs,&bs,&one,pc,&bs,pv,&bs,&zero, 194 multiplier,&bs); 195 PetscMemcpy(pc,multiplier,bs2*sizeof(Scalar)); 196 */ 197 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 198 nz = bi[row+1] - diag_offset[row] - 1; 199 pv += bs2; 200 for (j=0; j<nz; j++) { 201 /* 202 BLgemm_("N","N",&bs,&bs,&bs,&mone,pc,&bs,pv+bs2*j,&bs, 203 &one,rtmp+bs2*pj[j],&bs); 204 */ 205 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 206 } 207 PLogFlops(bslog*(nz+1)-bs); 208 /* } */ 209 row = *ajtmp++; 210 } 211 /* finished row so stick it into b->a */ 212 pv = ba + bs2*bi[i]; 213 pj = bj + bi[i]; 214 nz = bi[i+1] - bi[i]; 215 for ( j=0; j<nz; j++ ) { 216 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 217 } 218 diag = diag_offset[i] - bi[i]; 219 /* invert diagonal block */ 220 w = pv + bs2*diag; 221 /* 222 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 223 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 224 */ 225 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 226 } 227 228 PetscFree(rtmp); PetscFree(v_work); 229 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 230 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 231 ierr = ISDestroy(isicol); CHKERRQ(ierr); 232 C->factor = FACTOR_LU; 233 C->assembled = PETSC_TRUE; 234 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 235 return 0; 236 } 237 /* ------------------------------------------------------------*/ 238 /* 239 Version for when blocks are 5 by 5 240 */ 241 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 242 { 243 Mat C = *B; 244 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 245 IS iscol = b->col, isrow = b->row, isicol; 246 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 247 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 248 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 249 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 250 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 251 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 252 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 253 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 254 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 255 Scalar *ba = b->a,*aa = a->a; 256 register int *pj; 257 258 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 259 PLogObjectParent(*B,isicol); 260 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 261 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 262 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 263 264 for ( i=0; i<n; i++ ) { 265 nz = bi[i+1] - bi[i]; 266 ajtmp = bj + bi[i]; 267 for ( j=0; j<nz; j++ ) { 268 x = rtmp+25*ajtmp[j]; 269 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 270 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 271 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 272 } 273 /* load in initial (unfactored row) */ 274 idx = r[i]; 275 nz = ai[idx+1] - ai[idx]; 276 ajtmpold = aj + ai[idx]; 277 v = aa + 25*ai[idx]; 278 for ( j=0; j<nz; j++ ) { 279 x = rtmp+25*ic[ajtmpold[j]]; 280 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 281 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 282 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 283 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 284 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 285 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 286 v += 25; 287 } 288 row = *ajtmp++; 289 while (row < i) { 290 pc = rtmp + 25*row; 291 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 292 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 293 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 294 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 295 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 296 p25 = pc[24]; 297 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 298 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 299 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 300 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 301 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 302 p24 != 0.0 || p25 != 0.0) { 303 pv = ba + 25*diag_offset[row]; 304 pj = bj + diag_offset[row] + 1; 305 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 306 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 307 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 308 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 309 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 310 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 311 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 312 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 313 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 314 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 315 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 316 317 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 318 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 319 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 320 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 321 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 322 323 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 324 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 325 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 326 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 327 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 328 329 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 330 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 331 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 332 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 333 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 334 335 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 336 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 337 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 338 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 339 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 340 341 nz = bi[row+1] - diag_offset[row] - 1; 342 pv += 25; 343 for (j=0; j<nz; j++) { 344 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 345 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 346 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 347 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 348 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 349 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 350 x = rtmp + 25*pj[j]; 351 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 352 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 353 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 354 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 355 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 356 357 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 358 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 359 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 360 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 361 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 362 363 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 364 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 365 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 366 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 367 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 368 369 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 370 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 371 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 372 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 373 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 374 375 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 376 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 377 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 378 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 379 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 380 381 pv += 25; 382 } 383 PLogFlops(250*nz+225); 384 } 385 row = *ajtmp++; 386 } 387 /* finished row so stick it into b->a */ 388 pv = ba + 25*bi[i]; 389 pj = bj + bi[i]; 390 nz = bi[i+1] - bi[i]; 391 for ( j=0; j<nz; j++ ) { 392 x = rtmp+25*pj[j]; 393 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 394 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 395 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 396 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 397 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 398 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 399 pv += 25; 400 } 401 /* invert diagonal block */ 402 w = ba + 25*diag_offset[i]; 403 /* 404 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 405 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 406 */ 407 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 408 } 409 410 PetscFree(rtmp); 411 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 412 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 413 ierr = ISDestroy(isicol); CHKERRQ(ierr); 414 C->factor = FACTOR_LU; 415 C->assembled = PETSC_TRUE; 416 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 417 return 0; 418 } 419 420 /* ------------------------------------------------------------*/ 421 /* 422 Version for when blocks are 4 by 4 423 */ 424 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 425 { 426 Mat C = *B; 427 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 428 IS iscol = b->col, isrow = b->row, isicol; 429 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 430 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 431 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 432 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 433 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 434 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 435 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 436 Scalar m13,m14,m15,m16; 437 Scalar *ba = b->a,*aa = a->a; 438 register int *pj; 439 440 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 441 PLogObjectParent(*B,isicol); 442 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 443 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 444 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 445 446 for ( i=0; i<n; i++ ) { 447 nz = bi[i+1] - bi[i]; 448 ajtmp = bj + bi[i]; 449 for ( j=0; j<nz; j++ ) { 450 x = rtmp+16*ajtmp[j]; 451 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 452 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 453 } 454 /* load in initial (unfactored row) */ 455 idx = r[i]; 456 nz = ai[idx+1] - ai[idx]; 457 ajtmpold = aj + ai[idx]; 458 v = aa + 16*ai[idx]; 459 for ( j=0; j<nz; j++ ) { 460 x = rtmp+16*ic[ajtmpold[j]]; 461 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 462 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 463 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 464 x[14] = v[14]; x[15] = v[15]; 465 v += 16; 466 } 467 row = *ajtmp++; 468 while (row < i) { 469 pc = rtmp + 16*row; 470 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 471 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 472 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 473 p15 = pc[14]; p16 = pc[15]; 474 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 475 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 476 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 477 || p16 != 0.0) { 478 pv = ba + 16*diag_offset[row]; 479 pj = bj + diag_offset[row] + 1; 480 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 481 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 482 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 483 x15 = pv[14]; x16 = pv[15]; 484 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 485 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 486 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 487 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 488 489 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 490 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 491 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 492 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 493 494 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 495 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 496 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 497 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 498 499 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 500 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 501 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 502 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 503 504 nz = bi[row+1] - diag_offset[row] - 1; 505 pv += 16; 506 for (j=0; j<nz; j++) { 507 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 508 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 509 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 510 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 511 x = rtmp + 16*pj[j]; 512 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 513 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 514 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 515 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 516 517 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 518 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 519 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 520 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 521 522 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 523 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 524 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 525 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 526 527 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 528 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 529 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 530 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 531 532 pv += 16; 533 } 534 PLogFlops(128*nz+112); 535 } 536 row = *ajtmp++; 537 } 538 /* finished row so stick it into b->a */ 539 pv = ba + 16*bi[i]; 540 pj = bj + bi[i]; 541 nz = bi[i+1] - bi[i]; 542 for ( j=0; j<nz; j++ ) { 543 x = rtmp+16*pj[j]; 544 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 545 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 546 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 547 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 548 pv += 16; 549 } 550 /* invert diagonal block */ 551 w = ba + 16*diag_offset[i]; 552 /* 553 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 554 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 555 */ 556 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 557 } 558 559 PetscFree(rtmp); 560 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 561 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 562 ierr = ISDestroy(isicol); CHKERRQ(ierr); 563 C->factor = FACTOR_LU; 564 C->assembled = PETSC_TRUE; 565 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 566 return 0; 567 } 568 /* ------------------------------------------------------------*/ 569 /* 570 Version for when blocks are 3 by 3 571 */ 572 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 573 { 574 Mat C = *B; 575 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 576 IS iscol = b->col, isrow = b->row, isicol; 577 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 578 int *ajtmpold, *ajtmp, nz, row, v_pivots[3],*ai=a->i,*aj=a->j; 579 int *diag_offset = b->diag,bs = 3,idx; 580 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 581 Scalar p1,p2,p3,p4,v_work[3],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 582 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 583 Scalar *ba = b->a,*aa = a->a; 584 register int *pj; 585 586 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 587 PLogObjectParent(*B,isicol); 588 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 589 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 590 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 591 592 for ( i=0; i<n; i++ ) { 593 nz = bi[i+1] - bi[i]; 594 ajtmp = bj + bi[i]; 595 for ( j=0; j<nz; j++ ) { 596 x = rtmp + 9*ajtmp[j]; 597 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 598 } 599 /* load in initial (unfactored row) */ 600 idx = r[i]; 601 nz = ai[idx+1] - ai[idx]; 602 ajtmpold = aj + ai[idx]; 603 v = aa + 9*ai[idx]; 604 for ( j=0; j<nz; j++ ) { 605 x = rtmp + 9*ic[ajtmpold[j]]; 606 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 607 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 608 v += 9; 609 } 610 row = *ajtmp++; 611 while (row < i) { 612 pc = rtmp + 9*row; 613 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 614 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 615 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 616 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 617 pv = ba + 9*diag_offset[row]; 618 pj = bj + diag_offset[row] + 1; 619 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 620 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 621 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 622 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 623 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 624 625 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 626 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 627 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 628 629 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 630 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 631 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 632 nz = bi[row+1] - diag_offset[row] - 1; 633 pv += 9; 634 for (j=0; j<nz; j++) { 635 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 636 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 637 x = rtmp + 9*pj[j]; 638 x[0] -= m1*x1 + m4*x2 + m7*x3; 639 x[1] -= m2*x1 + m5*x2 + m8*x3; 640 x[2] -= m3*x1 + m6*x2 + m9*x3; 641 642 x[3] -= m1*x4 + m4*x5 + m7*x6; 643 x[4] -= m2*x4 + m5*x5 + m8*x6; 644 x[5] -= m3*x4 + m6*x5 + m9*x6; 645 646 x[6] -= m1*x7 + m4*x8 + m7*x9; 647 x[7] -= m2*x7 + m5*x8 + m8*x9; 648 x[8] -= m3*x7 + m6*x8 + m9*x9; 649 pv += 9; 650 } 651 PLogFlops(54*nz+36); 652 } 653 row = *ajtmp++; 654 } 655 /* finished row so stick it into b->a */ 656 pv = ba + 9*bi[i]; 657 pj = bj + bi[i]; 658 nz = bi[i+1] - bi[i]; 659 for ( j=0; j<nz; j++ ) { 660 x = rtmp + 9*pj[j]; 661 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 662 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 663 pv += 9; 664 } 665 /* invert diagonal block */ 666 w = ba + 9*diag_offset[i]; 667 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 668 /* Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); */ 669 } 670 671 PetscFree(rtmp); 672 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 673 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 674 ierr = ISDestroy(isicol); CHKERRQ(ierr); 675 C->factor = FACTOR_LU; 676 C->assembled = PETSC_TRUE; 677 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 678 return 0; 679 } 680 681 /* ------------------------------------------------------------*/ 682 /* 683 Version for when blocks are 2 by 2 684 */ 685 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 686 { 687 Mat C = *B; 688 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 689 IS iscol = b->col, isrow = b->row, isicol; 690 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 691 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 692 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 693 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 694 Scalar p1,p2,p3,p4,v_work[2]; 695 Scalar *ba = b->a,*aa = a->a; 696 register int *pj; 697 698 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 699 PLogObjectParent(*B,isicol); 700 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 701 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 702 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 703 704 for ( i=0; i<n; i++ ) { 705 nz = bi[i+1] - bi[i]; 706 ajtmp = bj + bi[i]; 707 for ( j=0; j<nz; j++ ) { 708 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 709 } 710 /* load in initial (unfactored row) */ 711 idx = r[i]; 712 nz = ai[idx+1] - ai[idx]; 713 ajtmpold = aj + ai[idx]; 714 v = aa + 4*ai[idx]; 715 for ( j=0; j<nz; j++ ) { 716 x = rtmp+4*ic[ajtmpold[j]]; 717 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 718 v += 4; 719 } 720 row = *ajtmp++; 721 while (row < i) { 722 pc = rtmp + 4*row; 723 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 724 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 725 pv = ba + 4*diag_offset[row]; 726 pj = bj + diag_offset[row] + 1; 727 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 728 pc[0] = m1 = p1*x1 + p3*x2; 729 pc[1] = m2 = p2*x1 + p4*x2; 730 pc[2] = m3 = p1*x3 + p3*x4; 731 pc[3] = m4 = p2*x3 + p4*x4; 732 nz = bi[row+1] - diag_offset[row] - 1; 733 pv += 4; 734 for (j=0; j<nz; j++) { 735 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 736 x = rtmp + 4*pj[j]; 737 x[0] -= m1*x1 + m3*x2; 738 x[1] -= m2*x1 + m4*x2; 739 x[2] -= m1*x3 + m3*x4; 740 x[3] -= m2*x3 + m4*x4; 741 pv += 4; 742 } 743 PLogFlops(16*nz+12); 744 } 745 row = *ajtmp++; 746 } 747 /* finished row so stick it into b->a */ 748 pv = ba + 4*bi[i]; 749 pj = bj + bi[i]; 750 nz = bi[i+1] - bi[i]; 751 for ( j=0; j<nz; j++ ) { 752 x = rtmp+4*pj[j]; 753 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 754 pv += 4; 755 } 756 /* invert diagonal block */ 757 w = ba + 4*diag_offset[i]; 758 /* 759 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 760 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 761 */ 762 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 763 } 764 765 PetscFree(rtmp); 766 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 767 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 768 ierr = ISDestroy(isicol); CHKERRQ(ierr); 769 C->factor = FACTOR_LU; 770 C->assembled = PETSC_TRUE; 771 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 772 return 0; 773 } 774 775 /* ----------------------------------------------------------- */ 776 /* 777 Version for when blocks are 1 by 1. 778 */ 779 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 780 { 781 Mat C = *B; 782 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 783 IS iscol = b->col, isrow = b->row, isicol; 784 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 785 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 786 int *diag_offset = b->diag,diag; 787 register Scalar *pv,*v,*rtmp,multiplier,*pc; 788 Scalar *ba = b->a,*aa = a->a; 789 register int *pj; 790 791 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 792 PLogObjectParent(*B,isicol); 793 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 794 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 795 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 796 797 for ( i=0; i<n; i++ ) { 798 nz = bi[i+1] - bi[i]; 799 ajtmp = bj + bi[i]; 800 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 801 802 /* load in initial (unfactored row) */ 803 nz = ai[r[i]+1] - ai[r[i]]; 804 ajtmpold = aj + ai[r[i]]; 805 v = aa + ai[r[i]]; 806 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 807 808 row = *ajtmp++; 809 while (row < i) { 810 pc = rtmp + row; 811 if (*pc != 0.0) { 812 pv = ba + diag_offset[row]; 813 pj = bj + diag_offset[row] + 1; 814 multiplier = *pc * *pv++; 815 *pc = multiplier; 816 nz = bi[row+1] - diag_offset[row] - 1; 817 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 818 PLogFlops(1+2*nz); 819 } 820 row = *ajtmp++; 821 } 822 /* finished row so stick it into b->a */ 823 pv = ba + bi[i]; 824 pj = bj + bi[i]; 825 nz = bi[i+1] - bi[i]; 826 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 827 diag = diag_offset[i] - bi[i]; 828 /* check pivot entry for current row */ 829 if (pv[diag] == 0.0) { 830 SETERRQ(1,"MatLUFactorNumeric_SeqAIJ:Zero pivot"); 831 } 832 pv[diag] = 1.0/pv[diag]; 833 } 834 835 PetscFree(rtmp); 836 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 837 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 838 ierr = ISDestroy(isicol); CHKERRQ(ierr); 839 C->factor = FACTOR_LU; 840 C->assembled = PETSC_TRUE; 841 PLogFlops(b->n); 842 return 0; 843 } 844 845 /* ----------------------------------------------------------- */ 846 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 847 { 848 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 849 int ierr; 850 Mat C; 851 852 ierr = MatLUFactorSymbolic_SeqBAIJ(A,row,col,f,&C); CHKERRQ(ierr); 853 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 854 855 /* free all the data structures from mat */ 856 PetscFree(mat->a); 857 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 858 if (mat->diag) PetscFree(mat->diag); 859 if (mat->ilen) PetscFree(mat->ilen); 860 if (mat->imax) PetscFree(mat->imax); 861 if (mat->solve_work) PetscFree(mat->solve_work); 862 if (mat->mult_work) PetscFree(mat->mult_work); 863 PetscFree(mat); 864 865 PetscMemcpy(A,C,sizeof(struct _Mat)); 866 PetscHeaderDestroy(C); 867 return 0; 868 } 869 /* ----------------------------------------------------------- */ 870 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 871 { 872 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 873 IS iscol=a->col,isrow=a->row; 874 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 875 int nz,bs=a->bs,bs2=a->bs2,_One=1; 876 Scalar *xa,*ba,*aa=a->a,*sum,_DOne=1.0,_DMOne=-1.0,_DZero=0.0; 877 register Scalar *x,*b,*lsum,*tmp,*v; 878 879 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 880 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 881 tmp = a->solve_work; 882 883 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 884 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 885 886 /* forward solve the lower triangular */ 887 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 888 for ( i=1; i<n; i++ ) { 889 v = aa + bs2*ai[i]; 890 vi = aj + ai[i]; 891 nz = a->diag[i] - ai[i]; 892 sum = tmp + bs*i; 893 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 894 while (nz--) { 895 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,sum,&_One); 896 v += bs2; 897 } 898 } 899 /* backward solve the upper triangular */ 900 lsum = a->solve_work + a->n; 901 for ( i=n-1; i>=0; i-- ){ 902 v = aa + bs2*(a->diag[i] + 1); 903 vi = aj + a->diag[i] + 1; 904 nz = ai[i+1] - a->diag[i] - 1; 905 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 906 while (nz--) { 907 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,lsum,&_One); 908 v += bs2; 909 } 910 LAgemv_("N",&bs,&bs,&_DOne,aa+bs2*a->diag[i],&bs,lsum,&_One,&_DZero, 911 tmp+i*bs,&_One); 912 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 913 } 914 915 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 916 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 917 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 918 return 0; 919 } 920 921 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 922 { 923 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 924 IS iscol=a->col,isrow=a->row; 925 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 926 Scalar *xa,*ba,*aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 927 register Scalar *x,*b,*tmp,*v; 928 929 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 930 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 931 tmp = a->solve_work; 932 933 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 934 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 935 936 /* forward solve the lower triangular */ 937 idx = 5*(*r++); 938 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 939 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 940 for ( i=1; i<n; i++ ) { 941 v = aa + 25*ai[i]; 942 vi = aj + ai[i]; 943 nz = a->diag[i] - ai[i]; 944 idx = 5*(*r++); 945 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 946 sum5 = b[4+idx]; 947 while (nz--) { 948 idx = 5*(*vi++); 949 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 950 x4 = tmp[3+idx];x5 = tmp[4+idx]; 951 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 952 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 953 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 954 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 955 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 956 v += 25; 957 } 958 idx = 5*i; 959 tmp[idx] = sum1;tmp[1+idx] = sum2; 960 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 961 } 962 /* backward solve the upper triangular */ 963 for ( i=n-1; i>=0; i-- ){ 964 v = aa + 25*a->diag[i] + 25; 965 vi = aj + a->diag[i] + 1; 966 nz = ai[i+1] - a->diag[i] - 1; 967 idt = 5*i; 968 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 969 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 970 while (nz--) { 971 idx = 5*(*vi++); 972 x1 = tmp[idx]; x2 = tmp[1+idx]; 973 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 974 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 975 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 976 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 977 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 978 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 979 v += 25; 980 } 981 idc = 5*(*c--); 982 v = aa + 25*a->diag[i]; 983 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 984 v[15]*sum4+v[20]*sum5; 985 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 986 v[16]*sum4+v[21]*sum5; 987 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 988 v[17]*sum4+v[22]*sum5; 989 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 990 v[18]*sum4+v[23]*sum5; 991 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 992 v[19]*sum4+v[24]*sum5; 993 } 994 995 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 996 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 997 PLogFlops(2*25*(a->nz) - a->n); 998 return 0; 999 } 1000 1001 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1002 { 1003 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1004 IS iscol=a->col,isrow=a->row; 1005 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 1006 Scalar *xa,*ba,*aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1007 register Scalar *x,*b,*tmp,*v; 1008 1009 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1010 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1011 tmp = a->solve_work; 1012 1013 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1014 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1015 1016 /* forward solve the lower triangular */ 1017 idx = 4*(*r++); 1018 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1019 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1020 for ( i=1; i<n; i++ ) { 1021 v = aa + 16*ai[i]; 1022 vi = aj + ai[i]; 1023 nz = a->diag[i] - ai[i]; 1024 idx = 4*(*r++); 1025 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1026 while (nz--) { 1027 idx = 4*(*vi++); 1028 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1029 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1030 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1031 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1032 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1033 v += 16; 1034 } 1035 idx = 4*i; 1036 tmp[idx] = sum1;tmp[1+idx] = sum2; 1037 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1038 } 1039 /* backward solve the upper triangular */ 1040 for ( i=n-1; i>=0; i-- ){ 1041 v = aa + 16*a->diag[i] + 16; 1042 vi = aj + a->diag[i] + 1; 1043 nz = ai[i+1] - a->diag[i] - 1; 1044 idt = 4*i; 1045 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1046 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1047 while (nz--) { 1048 idx = 4*(*vi++); 1049 x1 = tmp[idx]; x2 = tmp[1+idx]; 1050 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1051 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1052 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1053 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1054 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1055 v += 16; 1056 } 1057 idc = 4*(*c--); 1058 v = aa + 16*a->diag[i]; 1059 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1060 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1061 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1062 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1063 } 1064 1065 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1066 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1067 PLogFlops(2*16*(a->nz) - a->n); 1068 return 0; 1069 } 1070 1071 1072 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1073 { 1074 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1075 IS iscol=a->col,isrow=a->row; 1076 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 1077 Scalar *xa,*ba,*aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1078 register Scalar *x,*b,*tmp,*v; 1079 1080 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1081 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1082 tmp = a->solve_work; 1083 1084 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1085 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1086 1087 /* forward solve the lower triangular */ 1088 idx = 3*(*r++); 1089 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1090 for ( i=1; i<n; i++ ) { 1091 v = aa + 9*ai[i]; 1092 vi = aj + ai[i]; 1093 nz = a->diag[i] - ai[i]; 1094 idx = 3*(*r++); 1095 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1096 while (nz--) { 1097 idx = 3*(*vi++); 1098 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1099 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1100 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1101 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1102 v += 9; 1103 } 1104 idx = 3*i; 1105 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1106 } 1107 /* backward solve the upper triangular */ 1108 for ( i=n-1; i>=0; i-- ){ 1109 v = aa + 9*a->diag[i] + 9; 1110 vi = aj + a->diag[i] + 1; 1111 nz = ai[i+1] - a->diag[i] - 1; 1112 idt = 3*i; 1113 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1114 while (nz--) { 1115 idx = 3*(*vi++); 1116 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1117 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1118 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1119 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1120 v += 9; 1121 } 1122 idc = 3*(*c--); 1123 v = aa + 9*a->diag[i]; 1124 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1125 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1126 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1127 } 1128 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1129 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1130 PLogFlops(2*9*(a->nz) - a->n); 1131 return 0; 1132 } 1133 1134 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1135 { 1136 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1137 IS iscol=a->col,isrow=a->row; 1138 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 1139 Scalar *xa,*ba,*aa=a->a,sum1,sum2,x1,x2; 1140 register Scalar *x,*b,*tmp,*v; 1141 1142 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1143 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1144 tmp = a->solve_work; 1145 1146 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1147 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1148 1149 /* forward solve the lower triangular */ 1150 idx = 2*(*r++); 1151 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1152 for ( i=1; i<n; i++ ) { 1153 v = aa + 4*ai[i]; 1154 vi = aj + ai[i]; 1155 nz = a->diag[i] - ai[i]; 1156 idx = 2*(*r++); 1157 sum1 = b[idx]; sum2 = b[1+idx]; 1158 while (nz--) { 1159 idx = 2*(*vi++); 1160 x1 = tmp[idx]; x2 = tmp[1+idx]; 1161 sum1 -= v[0]*x1 + v[2]*x2; 1162 sum2 -= v[1]*x1 + v[3]*x2; 1163 v += 4; 1164 } 1165 idx = 2*i; 1166 tmp[idx] = sum1; tmp[1+idx] = sum2; 1167 } 1168 /* backward solve the upper triangular */ 1169 for ( i=n-1; i>=0; i-- ){ 1170 v = aa + 4*a->diag[i] + 4; 1171 vi = aj + a->diag[i] + 1; 1172 nz = ai[i+1] - a->diag[i] - 1; 1173 idt = 2*i; 1174 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1175 while (nz--) { 1176 idx = 2*(*vi++); 1177 x1 = tmp[idx]; x2 = tmp[1+idx]; 1178 sum1 -= v[0]*x1 + v[2]*x2; 1179 sum2 -= v[1]*x1 + v[3]*x2; 1180 v += 4; 1181 } 1182 idc = 2*(*c--); 1183 v = aa + 4*a->diag[i]; 1184 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1185 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1186 } 1187 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1188 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1189 PLogFlops(2*4*(a->nz) - a->n); 1190 return 0; 1191 } 1192 1193 1194 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1195 { 1196 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1197 IS iscol=a->col,isrow=a->row; 1198 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz; 1199 Scalar *xa,*ba,*aa=a->a,sum1; 1200 register Scalar *x,*b,*tmp,*v; 1201 1202 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1203 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1204 tmp = a->solve_work; 1205 1206 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1207 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1208 1209 /* forward solve the lower triangular */ 1210 tmp[0] = b[*r++]; 1211 for ( i=1; i<n; i++ ) { 1212 v = aa + ai[i]; 1213 vi = aj + ai[i]; 1214 nz = a->diag[i] - ai[i]; 1215 sum1 = b[*r++]; 1216 while (nz--) { 1217 sum1 -= (*v++)*tmp[*vi++]; 1218 } 1219 tmp[i] = sum1; 1220 } 1221 /* backward solve the upper triangular */ 1222 for ( i=n-1; i>=0; i-- ){ 1223 v = aa + a->diag[i] + 1; 1224 vi = aj + a->diag[i] + 1; 1225 nz = ai[i+1] - a->diag[i] - 1; 1226 sum1 = tmp[i]; 1227 while (nz--) { 1228 sum1 -= (*v++)*tmp[*vi++]; 1229 } 1230 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1231 } 1232 1233 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1234 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1235 PLogFlops(2*1*(a->nz) - a->n); 1236 return 0; 1237 } 1238 1239 /* ----------------------------------------------------------------*/ 1240 /* 1241 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1242 except that the data structure of Mat_SeqAIJ is slightly different. 1243 Not a good example of code reuse. 1244 */ 1245 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1246 Mat *fact) 1247 { 1248 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1249 IS isicol; 1250 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1251 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1252 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1253 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1254 PetscTruth col_identity, row_identity; 1255 1256 if (a->m != a->n) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Matrix must be square"); 1257 if (!isrow) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have row permutation"); 1258 if (!iscol) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have column permutation"); 1259 1260 /* special case that simply copies fill pattern */ 1261 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1262 if (levels == 0 && row_identity && col_identity) { 1263 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1264 (*fact)->factor = FACTOR_LU; 1265 b = (Mat_SeqBAIJ *) (*fact)->data; 1266 if (!b->diag) { 1267 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1268 } 1269 b->row = isrow; 1270 b->col = iscol; 1271 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1272 return 0; 1273 } 1274 1275 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1276 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1277 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1278 1279 /* get new row pointers */ 1280 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1281 ainew[0] = 0; 1282 /* don't know how many column pointers are needed so estimate */ 1283 jmax = (int) (f*ai[n] + 1); 1284 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1285 /* ajfill is level of fill for each fill entry */ 1286 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1287 /* fill is a linked list of nonzeros in active row */ 1288 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1289 /* im is level for each filled value */ 1290 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1291 /* dloc is location of diagonal in factor */ 1292 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1293 dloc[0] = 0; 1294 for ( prow=0; prow<n; prow++ ) { 1295 /* first copy previous fill into linked list */ 1296 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1297 xi = aj + ai[r[prow]]; 1298 fill[n] = n; 1299 while (nz--) { 1300 fm = n; 1301 idx = ic[*xi++]; 1302 do { 1303 m = fm; 1304 fm = fill[m]; 1305 } while (fm < idx); 1306 fill[m] = idx; 1307 fill[idx] = fm; 1308 im[idx] = 0; 1309 } 1310 nzi = 0; 1311 row = fill[n]; 1312 while ( row < prow ) { 1313 incrlev = im[row] + 1; 1314 nz = dloc[row]; 1315 xi = ajnew + ainew[row] + nz; 1316 flev = ajfill + ainew[row] + nz + 1; 1317 nnz = ainew[row+1] - ainew[row] - nz - 1; 1318 if (*xi++ != row) { 1319 SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:zero pivot"); 1320 } 1321 fm = row; 1322 while (nnz-- > 0) { 1323 idx = *xi++; 1324 if (*flev + incrlev > levels) { 1325 flev++; 1326 continue; 1327 } 1328 do { 1329 m = fm; 1330 fm = fill[m]; 1331 } while (fm < idx); 1332 if (fm != idx) { 1333 im[idx] = *flev + incrlev; 1334 fill[m] = idx; 1335 fill[idx] = fm; 1336 fm = idx; 1337 nzf++; 1338 } 1339 else { 1340 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1341 } 1342 flev++; 1343 } 1344 row = fill[row]; 1345 nzi++; 1346 } 1347 /* copy new filled row into permanent storage */ 1348 ainew[prow+1] = ainew[prow] + nzf; 1349 if (ainew[prow+1] > jmax) { 1350 /* allocate a longer ajnew */ 1351 int maxadd; 1352 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1353 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1354 jmax += maxadd; 1355 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1356 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1357 PetscFree(ajnew); 1358 ajnew = xi; 1359 /* allocate a longer ajfill */ 1360 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1361 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1362 PetscFree(ajfill); 1363 ajfill = xi; 1364 realloc++; 1365 } 1366 xi = ajnew + ainew[prow]; 1367 flev = ajfill + ainew[prow]; 1368 dloc[prow] = nzi; 1369 fm = fill[n]; 1370 while (nzf--) { 1371 *xi++ = fm; 1372 *flev++ = im[fm]; 1373 fm = fill[fm]; 1374 } 1375 } 1376 PetscFree(ajfill); 1377 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1378 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1379 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1380 PetscFree(fill); PetscFree(im); 1381 1382 PLogInfo(A, 1383 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1384 realloc,f,((double)ainew[n])/((double)ai[prow])); 1385 1386 /* put together the new matrix */ 1387 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1388 b = (Mat_SeqBAIJ *) (*fact)->data; 1389 PetscFree(b->imax); 1390 b->singlemalloc = 0; 1391 len = bs2*ainew[n]*sizeof(Scalar); 1392 /* the next line frees the default space generated by the Create() */ 1393 PetscFree(b->a); PetscFree(b->ilen); 1394 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1395 b->j = ajnew; 1396 b->i = ainew; 1397 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1398 b->diag = dloc; 1399 b->ilen = 0; 1400 b->imax = 0; 1401 b->row = isrow; 1402 b->col = iscol; 1403 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1404 CHKPTRQ(b->solve_work); 1405 /* In b structure: Free imax, ilen, old a, old j. 1406 Allocate dloc, solve_work, new a, new j */ 1407 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1408 b->maxnz = b->nz = ainew[n]; 1409 (*fact)->factor = FACTOR_LU; 1410 return 0; 1411 } 1412 1413 1414 1415 1416