1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.30 1996/09/14 03:08:32 bsmith Exp bsmith $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "src/mat/impls/baij/seq/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,realloc = 0,nzbd,*im; 25 26 PetscValidHeaderSpecific(isrow,IS_COOKIE); 27 PetscValidHeaderSpecific(iscol,IS_COOKIE); 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 /* the next line frees the default space generated by the Create() */ 125 PetscFree(b->a); PetscFree(b->ilen); 126 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 127 b->j = ajnew; 128 b->i = ainew; 129 b->diag = idnew; 130 b->ilen = 0; 131 b->imax = 0; 132 b->row = isrow; 133 b->col = iscol; 134 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 135 CHKPTRQ(b->solve_work); 136 /* In b structure: Free imax, ilen, old a, old j. 137 Allocate idnew, solve_work, new a, new j */ 138 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 139 b->maxnz = b->nz = ainew[n]; 140 141 (*B)->info.factor_mallocs = realloc; 142 (*B)->info.fill_ratio_given = f; 143 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 144 145 return 0; 146 } 147 148 /* ----------------------------------------------------------- */ 149 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 150 { 151 Mat C = *B; 152 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 153 IS iscol = b->col, isrow = b->row, isicol; 154 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 155 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j; 156 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = bs*bs,*v_pivots; 157 Scalar *ba = b->a,*aa = a->a; 158 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 159 register int *pj; 160 161 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 162 PLogObjectParent(*B,isicol); 163 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 164 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 165 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 166 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 167 /* generate work space needed by dense LU factorization */ 168 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 169 CHKPTRQ(v_work); 170 multiplier = v_work + bs; 171 v_pivots = (int *) (multiplier + bs2); 172 173 /* flops in while loop */ 174 bslog = 2*bs*bs2; 175 176 for ( i=0; i<n; i++ ) { 177 nz = bi[i+1] - bi[i]; 178 ajtmp = bj + bi[i]; 179 for ( j=0; j<nz; j++ ) { 180 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 181 } 182 /* load in initial (unfactored row) */ 183 nz = ai[r[i]+1] - ai[r[i]]; 184 ajtmpold = aj + ai[r[i]]; 185 v = aa + bs2*ai[r[i]]; 186 for ( j=0; j<nz; j++ ) { 187 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 188 } 189 row = *ajtmp++; 190 while (row < i) { 191 pc = rtmp + bs2*row; 192 /* if (*pc) { */ 193 pv = ba + bs2*diag_offset[row]; 194 pj = bj + diag_offset[row] + 1; 195 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 196 nz = bi[row+1] - diag_offset[row] - 1; 197 pv += bs2; 198 for (j=0; j<nz; j++) { 199 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 200 } 201 PLogFlops(bslog*(nz+1)-bs); 202 /* } */ 203 row = *ajtmp++; 204 } 205 /* finished row so stick it into b->a */ 206 pv = ba + bs2*bi[i]; 207 pj = bj + bi[i]; 208 nz = bi[i+1] - bi[i]; 209 for ( j=0; j<nz; j++ ) { 210 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 211 } 212 diag = diag_offset[i] - bi[i]; 213 /* invert diagonal block */ 214 w = pv + bs2*diag; 215 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 216 } 217 218 PetscFree(rtmp); PetscFree(v_work); 219 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 220 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 221 ierr = ISDestroy(isicol); CHKERRQ(ierr); 222 C->factor = FACTOR_LU; 223 C->assembled = PETSC_TRUE; 224 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 225 return 0; 226 } 227 /* ------------------------------------------------------------*/ 228 /* 229 Version for when blocks are 5 by 5 230 */ 231 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 232 { 233 Mat C = *B; 234 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 235 IS iscol = b->col, isrow = b->row, isicol; 236 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 237 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 238 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 239 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 240 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 241 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 242 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 243 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 244 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 245 Scalar *ba = b->a,*aa = a->a; 246 register int *pj; 247 248 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 249 PLogObjectParent(*B,isicol); 250 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 251 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 252 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 253 254 for ( i=0; i<n; i++ ) { 255 nz = bi[i+1] - bi[i]; 256 ajtmp = bj + bi[i]; 257 for ( j=0; j<nz; j++ ) { 258 x = rtmp+25*ajtmp[j]; 259 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 260 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 261 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 262 } 263 /* load in initial (unfactored row) */ 264 idx = r[i]; 265 nz = ai[idx+1] - ai[idx]; 266 ajtmpold = aj + ai[idx]; 267 v = aa + 25*ai[idx]; 268 for ( j=0; j<nz; j++ ) { 269 x = rtmp+25*ic[ajtmpold[j]]; 270 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 271 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 272 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 273 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 274 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 275 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 276 v += 25; 277 } 278 row = *ajtmp++; 279 while (row < i) { 280 pc = rtmp + 25*row; 281 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 282 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 283 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 284 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 285 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 286 p25 = pc[24]; 287 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 288 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 289 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 290 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 291 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 292 p24 != 0.0 || p25 != 0.0) { 293 pv = ba + 25*diag_offset[row]; 294 pj = bj + diag_offset[row] + 1; 295 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 296 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 297 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 298 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 299 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 300 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 301 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 302 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 303 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 304 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 305 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 306 307 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 308 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 309 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 310 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 311 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 312 313 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 314 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 315 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 316 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 317 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 318 319 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 320 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 321 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 322 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 323 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 324 325 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 326 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 327 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 328 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 329 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 330 331 nz = bi[row+1] - diag_offset[row] - 1; 332 pv += 25; 333 for (j=0; j<nz; j++) { 334 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 335 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 336 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 337 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 338 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 339 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 340 x = rtmp + 25*pj[j]; 341 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 342 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 343 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 344 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 345 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 346 347 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 348 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 349 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 350 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 351 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 352 353 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 354 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 355 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 356 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 357 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 358 359 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 360 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 361 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 362 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 363 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 364 365 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 366 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 367 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 368 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 369 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 370 371 pv += 25; 372 } 373 PLogFlops(250*nz+225); 374 } 375 row = *ajtmp++; 376 } 377 /* finished row so stick it into b->a */ 378 pv = ba + 25*bi[i]; 379 pj = bj + bi[i]; 380 nz = bi[i+1] - bi[i]; 381 for ( j=0; j<nz; j++ ) { 382 x = rtmp+25*pj[j]; 383 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 384 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 385 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 386 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 387 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 388 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 389 pv += 25; 390 } 391 /* invert diagonal block */ 392 w = ba + 25*diag_offset[i]; 393 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 394 } 395 396 PetscFree(rtmp); 397 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 398 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 399 ierr = ISDestroy(isicol); CHKERRQ(ierr); 400 C->factor = FACTOR_LU; 401 C->assembled = PETSC_TRUE; 402 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 403 return 0; 404 } 405 406 /* ------------------------------------------------------------*/ 407 /* 408 Version for when blocks are 4 by 4 409 */ 410 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 411 { 412 Mat C = *B; 413 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 414 IS iscol = b->col, isrow = b->row, isicol; 415 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 416 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 417 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 418 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 419 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 420 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 421 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 422 Scalar m13,m14,m15,m16; 423 Scalar *ba = b->a,*aa = a->a; 424 register int *pj; 425 426 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 427 PLogObjectParent(*B,isicol); 428 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 429 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 430 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 431 432 for ( i=0; i<n; i++ ) { 433 nz = bi[i+1] - bi[i]; 434 ajtmp = bj + bi[i]; 435 for ( j=0; j<nz; j++ ) { 436 x = rtmp+16*ajtmp[j]; 437 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 438 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 439 } 440 /* load in initial (unfactored row) */ 441 idx = r[i]; 442 nz = ai[idx+1] - ai[idx]; 443 ajtmpold = aj + ai[idx]; 444 v = aa + 16*ai[idx]; 445 for ( j=0; j<nz; j++ ) { 446 x = rtmp+16*ic[ajtmpold[j]]; 447 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 448 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 449 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 450 x[14] = v[14]; x[15] = v[15]; 451 v += 16; 452 } 453 row = *ajtmp++; 454 while (row < i) { 455 pc = rtmp + 16*row; 456 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 457 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 458 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 459 p15 = pc[14]; p16 = pc[15]; 460 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 461 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 462 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 463 || p16 != 0.0) { 464 pv = ba + 16*diag_offset[row]; 465 pj = bj + diag_offset[row] + 1; 466 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 467 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 468 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 469 x15 = pv[14]; x16 = pv[15]; 470 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 471 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 472 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 473 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 474 475 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 476 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 477 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 478 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 479 480 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 481 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 482 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 483 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 484 485 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 486 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 487 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 488 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 489 490 nz = bi[row+1] - diag_offset[row] - 1; 491 pv += 16; 492 for (j=0; j<nz; j++) { 493 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 494 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 495 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 496 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 497 x = rtmp + 16*pj[j]; 498 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 499 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 500 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 501 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 502 503 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 504 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 505 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 506 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 507 508 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 509 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 510 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 511 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 512 513 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 514 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 515 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 516 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 517 518 pv += 16; 519 } 520 PLogFlops(128*nz+112); 521 } 522 row = *ajtmp++; 523 } 524 /* finished row so stick it into b->a */ 525 pv = ba + 16*bi[i]; 526 pj = bj + bi[i]; 527 nz = bi[i+1] - bi[i]; 528 for ( j=0; j<nz; j++ ) { 529 x = rtmp+16*pj[j]; 530 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 531 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 532 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 533 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 534 pv += 16; 535 } 536 /* invert diagonal block */ 537 w = ba + 16*diag_offset[i]; 538 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 539 } 540 541 PetscFree(rtmp); 542 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 543 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 544 ierr = ISDestroy(isicol); CHKERRQ(ierr); 545 C->factor = FACTOR_LU; 546 C->assembled = PETSC_TRUE; 547 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 548 return 0; 549 } 550 /* ------------------------------------------------------------*/ 551 /* 552 Version for when blocks are 3 by 3 553 */ 554 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 555 { 556 Mat C = *B; 557 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 558 IS iscol = b->col, isrow = b->row, isicol; 559 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 560 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 561 int *diag_offset = b->diag,idx; 562 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 563 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 564 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 565 Scalar *ba = b->a,*aa = a->a; 566 register int *pj; 567 568 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 569 PLogObjectParent(*B,isicol); 570 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 571 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 572 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 573 574 for ( i=0; i<n; i++ ) { 575 nz = bi[i+1] - bi[i]; 576 ajtmp = bj + bi[i]; 577 for ( j=0; j<nz; j++ ) { 578 x = rtmp + 9*ajtmp[j]; 579 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 580 } 581 /* load in initial (unfactored row) */ 582 idx = r[i]; 583 nz = ai[idx+1] - ai[idx]; 584 ajtmpold = aj + ai[idx]; 585 v = aa + 9*ai[idx]; 586 for ( j=0; j<nz; j++ ) { 587 x = rtmp + 9*ic[ajtmpold[j]]; 588 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 589 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 590 v += 9; 591 } 592 row = *ajtmp++; 593 while (row < i) { 594 pc = rtmp + 9*row; 595 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 596 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 597 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 598 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 599 pv = ba + 9*diag_offset[row]; 600 pj = bj + diag_offset[row] + 1; 601 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 602 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 603 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 604 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 605 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 606 607 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 608 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 609 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 610 611 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 612 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 613 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 614 nz = bi[row+1] - diag_offset[row] - 1; 615 pv += 9; 616 for (j=0; j<nz; j++) { 617 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 618 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 619 x = rtmp + 9*pj[j]; 620 x[0] -= m1*x1 + m4*x2 + m7*x3; 621 x[1] -= m2*x1 + m5*x2 + m8*x3; 622 x[2] -= m3*x1 + m6*x2 + m9*x3; 623 624 x[3] -= m1*x4 + m4*x5 + m7*x6; 625 x[4] -= m2*x4 + m5*x5 + m8*x6; 626 x[5] -= m3*x4 + m6*x5 + m9*x6; 627 628 x[6] -= m1*x7 + m4*x8 + m7*x9; 629 x[7] -= m2*x7 + m5*x8 + m8*x9; 630 x[8] -= m3*x7 + m6*x8 + m9*x9; 631 pv += 9; 632 } 633 PLogFlops(54*nz+36); 634 } 635 row = *ajtmp++; 636 } 637 /* finished row so stick it into b->a */ 638 pv = ba + 9*bi[i]; 639 pj = bj + bi[i]; 640 nz = bi[i+1] - bi[i]; 641 for ( j=0; j<nz; j++ ) { 642 x = rtmp + 9*pj[j]; 643 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 644 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 645 pv += 9; 646 } 647 /* invert diagonal block */ 648 w = ba + 9*diag_offset[i]; 649 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 650 } 651 652 PetscFree(rtmp); 653 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 654 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 655 ierr = ISDestroy(isicol); CHKERRQ(ierr); 656 C->factor = FACTOR_LU; 657 C->assembled = PETSC_TRUE; 658 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 659 return 0; 660 } 661 662 /* ------------------------------------------------------------*/ 663 /* 664 Version for when blocks are 2 by 2 665 */ 666 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 667 { 668 Mat C = *B; 669 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 670 IS iscol = b->col, isrow = b->row, isicol; 671 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 672 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 673 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 674 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 675 Scalar p1,p2,p3,p4,v_work[2]; 676 Scalar *ba = b->a,*aa = a->a; 677 register int *pj; 678 679 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 680 PLogObjectParent(*B,isicol); 681 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 682 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 683 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 684 685 for ( i=0; i<n; i++ ) { 686 nz = bi[i+1] - bi[i]; 687 ajtmp = bj + bi[i]; 688 for ( j=0; j<nz; j++ ) { 689 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 690 } 691 /* load in initial (unfactored row) */ 692 idx = r[i]; 693 nz = ai[idx+1] - ai[idx]; 694 ajtmpold = aj + ai[idx]; 695 v = aa + 4*ai[idx]; 696 for ( j=0; j<nz; j++ ) { 697 x = rtmp+4*ic[ajtmpold[j]]; 698 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 699 v += 4; 700 } 701 row = *ajtmp++; 702 while (row < i) { 703 pc = rtmp + 4*row; 704 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 705 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 706 pv = ba + 4*diag_offset[row]; 707 pj = bj + diag_offset[row] + 1; 708 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 709 pc[0] = m1 = p1*x1 + p3*x2; 710 pc[1] = m2 = p2*x1 + p4*x2; 711 pc[2] = m3 = p1*x3 + p3*x4; 712 pc[3] = m4 = p2*x3 + p4*x4; 713 nz = bi[row+1] - diag_offset[row] - 1; 714 pv += 4; 715 for (j=0; j<nz; j++) { 716 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 717 x = rtmp + 4*pj[j]; 718 x[0] -= m1*x1 + m3*x2; 719 x[1] -= m2*x1 + m4*x2; 720 x[2] -= m1*x3 + m3*x4; 721 x[3] -= m2*x3 + m4*x4; 722 pv += 4; 723 } 724 PLogFlops(16*nz+12); 725 } 726 row = *ajtmp++; 727 } 728 /* finished row so stick it into b->a */ 729 pv = ba + 4*bi[i]; 730 pj = bj + bi[i]; 731 nz = bi[i+1] - bi[i]; 732 for ( j=0; j<nz; j++ ) { 733 x = rtmp+4*pj[j]; 734 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 735 pv += 4; 736 } 737 /* invert diagonal block */ 738 w = ba + 4*diag_offset[i]; 739 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 740 } 741 742 PetscFree(rtmp); 743 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 744 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 745 ierr = ISDestroy(isicol); CHKERRQ(ierr); 746 C->factor = FACTOR_LU; 747 C->assembled = PETSC_TRUE; 748 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 749 return 0; 750 } 751 752 /* ----------------------------------------------------------- */ 753 /* 754 Version for when blocks are 1 by 1. 755 */ 756 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 757 { 758 Mat C = *B; 759 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 760 IS iscol = b->col, isrow = b->row, isicol; 761 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 762 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 763 int *diag_offset = b->diag,diag; 764 register Scalar *pv,*v,*rtmp,multiplier,*pc; 765 Scalar *ba = b->a,*aa = a->a; 766 register int *pj; 767 768 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 769 PLogObjectParent(*B,isicol); 770 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 771 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 772 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 773 774 for ( i=0; i<n; i++ ) { 775 nz = bi[i+1] - bi[i]; 776 ajtmp = bj + bi[i]; 777 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 778 779 /* load in initial (unfactored row) */ 780 nz = ai[r[i]+1] - ai[r[i]]; 781 ajtmpold = aj + ai[r[i]]; 782 v = aa + ai[r[i]]; 783 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 784 785 row = *ajtmp++; 786 while (row < i) { 787 pc = rtmp + row; 788 if (*pc != 0.0) { 789 pv = ba + diag_offset[row]; 790 pj = bj + diag_offset[row] + 1; 791 multiplier = *pc * *pv++; 792 *pc = multiplier; 793 nz = bi[row+1] - diag_offset[row] - 1; 794 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 795 PLogFlops(1+2*nz); 796 } 797 row = *ajtmp++; 798 } 799 /* finished row so stick it into b->a */ 800 pv = ba + bi[i]; 801 pj = bj + bi[i]; 802 nz = bi[i+1] - bi[i]; 803 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 804 diag = diag_offset[i] - bi[i]; 805 /* check pivot entry for current row */ 806 if (pv[diag] == 0.0) { 807 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"MatLUFactorNumeric_SeqBAIJ_1:Zero pivot"); 808 } 809 pv[diag] = 1.0/pv[diag]; 810 } 811 812 PetscFree(rtmp); 813 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 814 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 815 ierr = ISDestroy(isicol); CHKERRQ(ierr); 816 C->factor = FACTOR_LU; 817 C->assembled = PETSC_TRUE; 818 PLogFlops(b->n); 819 return 0; 820 } 821 822 /* ----------------------------------------------------------- */ 823 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 824 { 825 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 826 int ierr; 827 Mat C; 828 829 ierr = MatLUFactorSymbolic_SeqBAIJ(A,row,col,f,&C); CHKERRQ(ierr); 830 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 831 832 /* free all the data structures from mat */ 833 PetscFree(mat->a); 834 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 835 if (mat->diag) PetscFree(mat->diag); 836 if (mat->ilen) PetscFree(mat->ilen); 837 if (mat->imax) PetscFree(mat->imax); 838 if (mat->solve_work) PetscFree(mat->solve_work); 839 if (mat->mult_work) PetscFree(mat->mult_work); 840 PetscFree(mat); 841 842 PetscMemcpy(A,C,sizeof(struct _Mat)); 843 PetscHeaderDestroy(C); 844 return 0; 845 } 846 /* ----------------------------------------------------------- */ 847 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 848 { 849 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 850 IS iscol=a->col,isrow=a->row; 851 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 852 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 853 Scalar *aa=a->a,*sum; 854 register Scalar *x,*b,*lsum,*tmp,*v; 855 856 VecGetArray_Fast(bb,b); 857 VecGetArray_Fast(xx,x); 858 tmp = a->solve_work; 859 860 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 861 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 862 863 /* forward solve the lower triangular */ 864 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 865 for ( i=1; i<n; i++ ) { 866 v = aa + bs2*ai[i]; 867 vi = aj + ai[i]; 868 nz = a->diag[i] - ai[i]; 869 sum = tmp + bs*i; 870 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 871 while (nz--) { 872 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 873 v += bs2; 874 } 875 } 876 /* backward solve the upper triangular */ 877 lsum = a->solve_work + a->n; 878 for ( i=n-1; i>=0; i-- ){ 879 v = aa + bs2*(a->diag[i] + 1); 880 vi = aj + a->diag[i] + 1; 881 nz = ai[i+1] - a->diag[i] - 1; 882 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 883 while (nz--) { 884 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 885 v += bs2; 886 } 887 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 888 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 889 } 890 891 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 892 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 893 VecRestoreArray_Fast(bb,b); 894 VecRestoreArray_Fast(xx,x); 895 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 896 return 0; 897 } 898 899 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 900 { 901 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 902 IS iscol=a->col,isrow=a->row; 903 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 904 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 905 register Scalar *x,*b,*tmp,*v; 906 907 VecGetArray_Fast(bb,b); 908 VecGetArray_Fast(xx,x); 909 tmp = a->solve_work; 910 911 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 912 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 913 914 /* forward solve the lower triangular */ 915 idx = 5*(*r++); 916 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 917 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 918 for ( i=1; i<n; i++ ) { 919 v = aa + 25*ai[i]; 920 vi = aj + ai[i]; 921 nz = a->diag[i] - ai[i]; 922 idx = 5*(*r++); 923 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 924 sum5 = b[4+idx]; 925 while (nz--) { 926 idx = 5*(*vi++); 927 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 928 x4 = tmp[3+idx];x5 = tmp[4+idx]; 929 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 930 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 931 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 932 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 933 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 934 v += 25; 935 } 936 idx = 5*i; 937 tmp[idx] = sum1;tmp[1+idx] = sum2; 938 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 939 } 940 /* backward solve the upper triangular */ 941 for ( i=n-1; i>=0; i-- ){ 942 v = aa + 25*a->diag[i] + 25; 943 vi = aj + a->diag[i] + 1; 944 nz = ai[i+1] - a->diag[i] - 1; 945 idt = 5*i; 946 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 947 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 948 while (nz--) { 949 idx = 5*(*vi++); 950 x1 = tmp[idx]; x2 = tmp[1+idx]; 951 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 952 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 953 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 954 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 955 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 956 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 957 v += 25; 958 } 959 idc = 5*(*c--); 960 v = aa + 25*a->diag[i]; 961 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 962 v[15]*sum4+v[20]*sum5; 963 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 964 v[16]*sum4+v[21]*sum5; 965 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 966 v[17]*sum4+v[22]*sum5; 967 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 968 v[18]*sum4+v[23]*sum5; 969 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 970 v[19]*sum4+v[24]*sum5; 971 } 972 973 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 974 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 975 VecRestoreArray_Fast(bb,b); 976 VecRestoreArray_Fast(xx,x); 977 PLogFlops(2*25*(a->nz) - a->n); 978 return 0; 979 } 980 981 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 982 { 983 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 984 IS iscol=a->col,isrow=a->row; 985 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 986 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 987 register Scalar *x,*b,*tmp,*v; 988 989 VecGetArray_Fast(bb,b); 990 VecGetArray_Fast(xx,x); 991 tmp = a->solve_work; 992 993 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 994 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 995 996 /* forward solve the lower triangular */ 997 idx = 4*(*r++); 998 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 999 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1000 for ( i=1; i<n; i++ ) { 1001 v = aa + 16*ai[i]; 1002 vi = aj + ai[i]; 1003 nz = a->diag[i] - ai[i]; 1004 idx = 4*(*r++); 1005 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1006 while (nz--) { 1007 idx = 4*(*vi++); 1008 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1009 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1010 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1011 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1012 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1013 v += 16; 1014 } 1015 idx = 4*i; 1016 tmp[idx] = sum1;tmp[1+idx] = sum2; 1017 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1018 } 1019 /* backward solve the upper triangular */ 1020 for ( i=n-1; i>=0; i-- ){ 1021 v = aa + 16*a->diag[i] + 16; 1022 vi = aj + a->diag[i] + 1; 1023 nz = ai[i+1] - a->diag[i] - 1; 1024 idt = 4*i; 1025 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1026 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1027 while (nz--) { 1028 idx = 4*(*vi++); 1029 x1 = tmp[idx]; x2 = tmp[1+idx]; 1030 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1031 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1032 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1033 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1034 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1035 v += 16; 1036 } 1037 idc = 4*(*c--); 1038 v = aa + 16*a->diag[i]; 1039 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1040 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1041 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1042 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1043 } 1044 1045 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1046 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1047 VecRestoreArray_Fast(bb,b); 1048 VecRestoreArray_Fast(xx,x); 1049 PLogFlops(2*16*(a->nz) - a->n); 1050 return 0; 1051 } 1052 1053 1054 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1055 { 1056 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1057 IS iscol=a->col,isrow=a->row; 1058 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1059 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1060 register Scalar *x,*b,*tmp,*v; 1061 1062 VecGetArray_Fast(bb,b); 1063 VecGetArray_Fast(xx,x); 1064 tmp = a->solve_work; 1065 1066 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1067 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1068 1069 /* forward solve the lower triangular */ 1070 idx = 3*(*r++); 1071 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1072 for ( i=1; i<n; i++ ) { 1073 v = aa + 9*ai[i]; 1074 vi = aj + ai[i]; 1075 nz = a->diag[i] - ai[i]; 1076 idx = 3*(*r++); 1077 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1078 while (nz--) { 1079 idx = 3*(*vi++); 1080 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1081 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1082 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1083 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1084 v += 9; 1085 } 1086 idx = 3*i; 1087 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1088 } 1089 /* backward solve the upper triangular */ 1090 for ( i=n-1; i>=0; i-- ){ 1091 v = aa + 9*a->diag[i] + 9; 1092 vi = aj + a->diag[i] + 1; 1093 nz = ai[i+1] - a->diag[i] - 1; 1094 idt = 3*i; 1095 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 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 idc = 3*(*c--); 1105 v = aa + 9*a->diag[i]; 1106 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1107 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1108 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1109 } 1110 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1111 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1112 VecRestoreArray_Fast(bb,b); 1113 VecRestoreArray_Fast(xx,x); 1114 PLogFlops(2*9*(a->nz) - a->n); 1115 return 0; 1116 } 1117 1118 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1119 { 1120 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1121 IS iscol=a->col,isrow=a->row; 1122 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1123 Scalar *aa=a->a,sum1,sum2,x1,x2; 1124 register Scalar *x,*b,*tmp,*v; 1125 1126 VecGetArray_Fast(bb,b); 1127 VecGetArray_Fast(xx,x); 1128 tmp = a->solve_work; 1129 1130 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1131 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1132 1133 /* forward solve the lower triangular */ 1134 idx = 2*(*r++); 1135 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1136 for ( i=1; i<n; i++ ) { 1137 v = aa + 4*ai[i]; 1138 vi = aj + ai[i]; 1139 nz = a->diag[i] - ai[i]; 1140 idx = 2*(*r++); 1141 sum1 = b[idx]; sum2 = b[1+idx]; 1142 while (nz--) { 1143 idx = 2*(*vi++); 1144 x1 = tmp[idx]; x2 = tmp[1+idx]; 1145 sum1 -= v[0]*x1 + v[2]*x2; 1146 sum2 -= v[1]*x1 + v[3]*x2; 1147 v += 4; 1148 } 1149 idx = 2*i; 1150 tmp[idx] = sum1; tmp[1+idx] = sum2; 1151 } 1152 /* backward solve the upper triangular */ 1153 for ( i=n-1; i>=0; i-- ){ 1154 v = aa + 4*a->diag[i] + 4; 1155 vi = aj + a->diag[i] + 1; 1156 nz = ai[i+1] - a->diag[i] - 1; 1157 idt = 2*i; 1158 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1159 while (nz--) { 1160 idx = 2*(*vi++); 1161 x1 = tmp[idx]; x2 = tmp[1+idx]; 1162 sum1 -= v[0]*x1 + v[2]*x2; 1163 sum2 -= v[1]*x1 + v[3]*x2; 1164 v += 4; 1165 } 1166 idc = 2*(*c--); 1167 v = aa + 4*a->diag[i]; 1168 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1169 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1170 } 1171 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1172 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1173 VecRestoreArray_Fast(bb,b); 1174 VecRestoreArray_Fast(xx,x); 1175 PLogFlops(2*4*(a->nz) - a->n); 1176 return 0; 1177 } 1178 1179 1180 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1181 { 1182 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1183 IS iscol=a->col,isrow=a->row; 1184 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1185 Scalar *aa=a->a,sum1; 1186 register Scalar *x,*b,*tmp,*v; 1187 1188 if (!n) return 0; 1189 1190 VecGetArray_Fast(bb,b); 1191 VecGetArray_Fast(xx,x); 1192 tmp = a->solve_work; 1193 1194 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1195 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1196 1197 /* forward solve the lower triangular */ 1198 tmp[0] = b[*r++]; 1199 for ( i=1; i<n; i++ ) { 1200 v = aa + ai[i]; 1201 vi = aj + ai[i]; 1202 nz = a->diag[i] - ai[i]; 1203 sum1 = b[*r++]; 1204 while (nz--) { 1205 sum1 -= (*v++)*tmp[*vi++]; 1206 } 1207 tmp[i] = sum1; 1208 } 1209 /* backward solve the upper triangular */ 1210 for ( i=n-1; i>=0; i-- ){ 1211 v = aa + a->diag[i] + 1; 1212 vi = aj + a->diag[i] + 1; 1213 nz = ai[i+1] - a->diag[i] - 1; 1214 sum1 = tmp[i]; 1215 while (nz--) { 1216 sum1 -= (*v++)*tmp[*vi++]; 1217 } 1218 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1219 } 1220 1221 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1222 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1223 VecRestoreArray_Fast(bb,b); 1224 VecRestoreArray_Fast(xx,x); 1225 PLogFlops(2*1*(a->nz) - a->n); 1226 return 0; 1227 } 1228 1229 /* ----------------------------------------------------------------*/ 1230 /* 1231 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1232 except that the data structure of Mat_SeqAIJ is slightly different. 1233 Not a good example of code reuse. 1234 */ 1235 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1236 Mat *fact) 1237 { 1238 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1239 IS isicol; 1240 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1241 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1242 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1243 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1244 PetscTruth col_identity, row_identity; 1245 1246 /* special case that simply copies fill pattern */ 1247 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1248 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1249 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1250 if (levels == 0 && row_identity && col_identity) { 1251 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1252 (*fact)->factor = FACTOR_LU; 1253 b = (Mat_SeqBAIJ *) (*fact)->data; 1254 if (!b->diag) { 1255 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1256 } 1257 b->row = isrow; 1258 b->col = iscol; 1259 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1260 return 0; 1261 } 1262 1263 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1264 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1265 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1266 1267 /* get new row pointers */ 1268 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1269 ainew[0] = 0; 1270 /* don't know how many column pointers are needed so estimate */ 1271 jmax = (int) (f*ai[n] + 1); 1272 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1273 /* ajfill is level of fill for each fill entry */ 1274 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1275 /* fill is a linked list of nonzeros in active row */ 1276 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1277 /* im is level for each filled value */ 1278 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1279 /* dloc is location of diagonal in factor */ 1280 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1281 dloc[0] = 0; 1282 for ( prow=0; prow<n; prow++ ) { 1283 /* first copy previous fill into linked list */ 1284 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1285 xi = aj + ai[r[prow]]; 1286 fill[n] = n; 1287 while (nz--) { 1288 fm = n; 1289 idx = ic[*xi++]; 1290 do { 1291 m = fm; 1292 fm = fill[m]; 1293 } while (fm < idx); 1294 fill[m] = idx; 1295 fill[idx] = fm; 1296 im[idx] = 0; 1297 } 1298 nzi = 0; 1299 row = fill[n]; 1300 while ( row < prow ) { 1301 incrlev = im[row] + 1; 1302 nz = dloc[row]; 1303 xi = ajnew + ainew[row] + nz; 1304 flev = ajfill + ainew[row] + nz + 1; 1305 nnz = ainew[row+1] - ainew[row] - nz - 1; 1306 if (*xi++ != row) { 1307 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"MatILUFactorSymbolic_SeqBAIJ:zero pivot"); 1308 } 1309 fm = row; 1310 while (nnz-- > 0) { 1311 idx = *xi++; 1312 if (*flev + incrlev > levels) { 1313 flev++; 1314 continue; 1315 } 1316 do { 1317 m = fm; 1318 fm = fill[m]; 1319 } while (fm < idx); 1320 if (fm != idx) { 1321 im[idx] = *flev + incrlev; 1322 fill[m] = idx; 1323 fill[idx] = fm; 1324 fm = idx; 1325 nzf++; 1326 } 1327 else { 1328 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1329 } 1330 flev++; 1331 } 1332 row = fill[row]; 1333 nzi++; 1334 } 1335 /* copy new filled row into permanent storage */ 1336 ainew[prow+1] = ainew[prow] + nzf; 1337 if (ainew[prow+1] > jmax) { 1338 /* allocate a longer ajnew */ 1339 int maxadd; 1340 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1341 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1342 jmax += maxadd; 1343 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1344 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1345 PetscFree(ajnew); 1346 ajnew = xi; 1347 /* allocate a longer ajfill */ 1348 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1349 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1350 PetscFree(ajfill); 1351 ajfill = xi; 1352 realloc++; 1353 } 1354 xi = ajnew + ainew[prow]; 1355 flev = ajfill + ainew[prow]; 1356 dloc[prow] = nzi; 1357 fm = fill[n]; 1358 while (nzf--) { 1359 *xi++ = fm; 1360 *flev++ = im[fm]; 1361 fm = fill[fm]; 1362 } 1363 } 1364 PetscFree(ajfill); 1365 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1366 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1367 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1368 PetscFree(fill); PetscFree(im); 1369 1370 PLogInfo(A, 1371 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1372 realloc,f,((double)ainew[n])/((double)ai[prow])); 1373 1374 /* put together the new matrix */ 1375 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1376 b = (Mat_SeqBAIJ *) (*fact)->data; 1377 PetscFree(b->imax); 1378 b->singlemalloc = 0; 1379 len = bs2*ainew[n]*sizeof(Scalar); 1380 /* the next line frees the default space generated by the Create() */ 1381 PetscFree(b->a); PetscFree(b->ilen); 1382 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1383 b->j = ajnew; 1384 b->i = ainew; 1385 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1386 b->diag = dloc; 1387 b->ilen = 0; 1388 b->imax = 0; 1389 b->row = isrow; 1390 b->col = iscol; 1391 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1392 CHKPTRQ(b->solve_work); 1393 /* In b structure: Free imax, ilen, old a, old j. 1394 Allocate dloc, solve_work, new a, new j */ 1395 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1396 b->maxnz = b->nz = ainew[n]; 1397 (*fact)->factor = FACTOR_LU; 1398 1399 (*fact)->info.factor_mallocs = realloc; 1400 (*fact)->info.fill_ratio_given = f; 1401 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1402 1403 return 0; 1404 } 1405 1406 1407 1408 1409