1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.31 1996/12/08 20:09:19 bsmith Exp balay $"; 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 #undef __FUNCTION__ 19 #define __FUNCTION__ "MatLUFactorSymbolic_SeqBAIJ" 20 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 21 { 22 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 23 IS isicol; 24 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 25 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 26 int *idnew, idx, row,m,fm, nnz, nzi,realloc = 0,nzbd,*im; 27 28 PetscValidHeaderSpecific(isrow,IS_COOKIE); 29 PetscValidHeaderSpecific(iscol,IS_COOKIE); 30 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 31 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 32 33 /* get new row pointers */ 34 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 35 ainew[0] = 0; 36 /* don't know how many column pointers are needed so estimate */ 37 jmax = (int) (f*ai[n] + 1); 38 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 39 /* fill is a linked list of nonzeros in active row */ 40 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 41 im = fill + n + 1; 42 /* idnew is location of diagonal in factor */ 43 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 44 idnew[0] = 0; 45 46 for ( i=0; i<n; i++ ) { 47 /* first copy previous fill into linked list */ 48 nnz = nz = ai[r[i]+1] - ai[r[i]]; 49 ajtmp = aj + ai[r[i]]; 50 fill[n] = n; 51 while (nz--) { 52 fm = n; 53 idx = ic[*ajtmp++]; 54 do { 55 m = fm; 56 fm = fill[m]; 57 } while (fm < idx); 58 fill[m] = idx; 59 fill[idx] = fm; 60 } 61 row = fill[n]; 62 while ( row < i ) { 63 ajtmp = ajnew + idnew[row] + 1; 64 nzbd = 1 + idnew[row] - ainew[row]; 65 nz = im[row] - nzbd; 66 fm = row; 67 while (nz-- > 0) { 68 idx = *ajtmp++; 69 nzbd++; 70 if (idx == i) im[row] = nzbd; 71 do { 72 m = fm; 73 fm = fill[m]; 74 } while (fm < idx); 75 if (fm != idx) { 76 fill[m] = idx; 77 fill[idx] = fm; 78 fm = idx; 79 nnz++; 80 } 81 } 82 row = fill[row]; 83 } 84 /* copy new filled row into permanent storage */ 85 ainew[i+1] = ainew[i] + nnz; 86 if (ainew[i+1] > jmax) { 87 /* allocate a longer ajnew */ 88 int maxadd; 89 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 90 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 91 jmax += maxadd; 92 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 93 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 94 PetscFree(ajnew); 95 ajnew = ajtmp; 96 realloc++; /* count how many times we realloc */ 97 } 98 ajtmp = ajnew + ainew[i]; 99 fm = fill[n]; 100 nzi = 0; 101 im[i] = nnz; 102 while (nnz--) { 103 if (fm < i) nzi++; 104 *ajtmp++ = fm; 105 fm = fill[fm]; 106 } 107 idnew[i] = ainew[i] + nzi; 108 } 109 110 PLogInfo(A, 111 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 112 realloc,f,((double)ainew[n])/((double)ai[i])); 113 114 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 115 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 116 117 PetscFree(fill); 118 119 /* put together the new matrix */ 120 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 121 PLogObjectParent(*B,isicol); 122 ierr = ISDestroy(isicol); CHKERRQ(ierr); 123 b = (Mat_SeqBAIJ *) (*B)->data; 124 PetscFree(b->imax); 125 b->singlemalloc = 0; 126 /* the next line frees the default space generated by the Create() */ 127 PetscFree(b->a); PetscFree(b->ilen); 128 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 129 b->j = ajnew; 130 b->i = ainew; 131 b->diag = idnew; 132 b->ilen = 0; 133 b->imax = 0; 134 b->row = isrow; 135 b->col = iscol; 136 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 137 CHKPTRQ(b->solve_work); 138 /* In b structure: Free imax, ilen, old a, old j. 139 Allocate idnew, solve_work, new a, new j */ 140 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 141 b->maxnz = b->nz = ainew[n]; 142 143 (*B)->info.factor_mallocs = realloc; 144 (*B)->info.fill_ratio_given = f; 145 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 146 147 return 0; 148 } 149 150 /* ----------------------------------------------------------- */ 151 #undef __FUNCTION__ 152 #define __FUNCTION__ "MatLUFactorNumeric_SeqBAIJ_N" 153 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 154 { 155 Mat C = *B; 156 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 157 IS iscol = b->col, isrow = b->row, isicol; 158 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 159 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j; 160 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = bs*bs,*v_pivots; 161 Scalar *ba = b->a,*aa = a->a; 162 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 163 register int *pj; 164 165 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 166 PLogObjectParent(*B,isicol); 167 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 168 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 169 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 170 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 171 /* generate work space needed by dense LU factorization */ 172 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 173 CHKPTRQ(v_work); 174 multiplier = v_work + bs; 175 v_pivots = (int *) (multiplier + bs2); 176 177 /* flops in while loop */ 178 bslog = 2*bs*bs2; 179 180 for ( i=0; i<n; i++ ) { 181 nz = bi[i+1] - bi[i]; 182 ajtmp = bj + bi[i]; 183 for ( j=0; j<nz; j++ ) { 184 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 185 } 186 /* load in initial (unfactored row) */ 187 nz = ai[r[i]+1] - ai[r[i]]; 188 ajtmpold = aj + ai[r[i]]; 189 v = aa + bs2*ai[r[i]]; 190 for ( j=0; j<nz; j++ ) { 191 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 192 } 193 row = *ajtmp++; 194 while (row < i) { 195 pc = rtmp + bs2*row; 196 /* if (*pc) { */ 197 pv = ba + bs2*diag_offset[row]; 198 pj = bj + diag_offset[row] + 1; 199 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 200 nz = bi[row+1] - diag_offset[row] - 1; 201 pv += bs2; 202 for (j=0; j<nz; j++) { 203 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 204 } 205 PLogFlops(bslog*(nz+1)-bs); 206 /* } */ 207 row = *ajtmp++; 208 } 209 /* finished row so stick it into b->a */ 210 pv = ba + bs2*bi[i]; 211 pj = bj + bi[i]; 212 nz = bi[i+1] - bi[i]; 213 for ( j=0; j<nz; j++ ) { 214 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 215 } 216 diag = diag_offset[i] - bi[i]; 217 /* invert diagonal block */ 218 w = pv + bs2*diag; 219 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 220 } 221 222 PetscFree(rtmp); PetscFree(v_work); 223 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 224 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 225 ierr = ISDestroy(isicol); CHKERRQ(ierr); 226 C->factor = FACTOR_LU; 227 C->assembled = PETSC_TRUE; 228 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 229 return 0; 230 } 231 /* ------------------------------------------------------------*/ 232 /* 233 Version for when blocks are 5 by 5 234 */ 235 #undef __FUNCTION__ 236 #define __FUNCTION__ "MatLUFactorNumeric_SeqBAIJ_5" 237 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 238 { 239 Mat C = *B; 240 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 241 IS iscol = b->col, isrow = b->row, isicol; 242 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 243 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 244 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 245 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 246 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 247 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 248 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 249 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 250 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 251 Scalar *ba = b->a,*aa = a->a; 252 register int *pj; 253 254 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 255 PLogObjectParent(*B,isicol); 256 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 257 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 258 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 259 260 for ( i=0; i<n; i++ ) { 261 nz = bi[i+1] - bi[i]; 262 ajtmp = bj + bi[i]; 263 for ( j=0; j<nz; j++ ) { 264 x = rtmp+25*ajtmp[j]; 265 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 266 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 267 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 268 } 269 /* load in initial (unfactored row) */ 270 idx = r[i]; 271 nz = ai[idx+1] - ai[idx]; 272 ajtmpold = aj + ai[idx]; 273 v = aa + 25*ai[idx]; 274 for ( j=0; j<nz; j++ ) { 275 x = rtmp+25*ic[ajtmpold[j]]; 276 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 277 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 278 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 279 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 280 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 281 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 282 v += 25; 283 } 284 row = *ajtmp++; 285 while (row < i) { 286 pc = rtmp + 25*row; 287 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 288 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 289 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 290 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 291 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 292 p25 = pc[24]; 293 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 294 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 295 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 296 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 297 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 298 p24 != 0.0 || p25 != 0.0) { 299 pv = ba + 25*diag_offset[row]; 300 pj = bj + diag_offset[row] + 1; 301 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 302 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 303 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 304 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 305 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 306 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 307 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 308 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 309 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 310 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 311 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 312 313 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 314 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 315 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 316 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 317 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 318 319 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 320 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 321 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 322 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 323 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 324 325 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 326 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 327 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 328 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 329 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 330 331 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 332 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 333 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 334 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 335 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 336 337 nz = bi[row+1] - diag_offset[row] - 1; 338 pv += 25; 339 for (j=0; j<nz; j++) { 340 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 341 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 342 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 343 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 344 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 345 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 346 x = rtmp + 25*pj[j]; 347 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 348 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 349 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 350 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 351 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 352 353 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 354 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 355 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 356 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 357 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 358 359 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 360 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 361 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 362 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 363 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 364 365 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 366 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 367 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 368 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 369 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 370 371 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 372 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 373 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 374 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 375 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 376 377 pv += 25; 378 } 379 PLogFlops(250*nz+225); 380 } 381 row = *ajtmp++; 382 } 383 /* finished row so stick it into b->a */ 384 pv = ba + 25*bi[i]; 385 pj = bj + bi[i]; 386 nz = bi[i+1] - bi[i]; 387 for ( j=0; j<nz; j++ ) { 388 x = rtmp+25*pj[j]; 389 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 390 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 391 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 392 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 393 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 394 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 395 pv += 25; 396 } 397 /* invert diagonal block */ 398 w = ba + 25*diag_offset[i]; 399 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 400 } 401 402 PetscFree(rtmp); 403 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 404 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 405 ierr = ISDestroy(isicol); CHKERRQ(ierr); 406 C->factor = FACTOR_LU; 407 C->assembled = PETSC_TRUE; 408 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 409 return 0; 410 } 411 412 /* ------------------------------------------------------------*/ 413 /* 414 Version for when blocks are 4 by 4 415 */ 416 #undef __FUNCTION__ 417 #define __FUNCTION__ "MatLUFactorNumeric_SeqBAIJ_4" 418 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 419 { 420 Mat C = *B; 421 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 422 IS iscol = b->col, isrow = b->row, isicol; 423 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 424 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 425 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 426 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 427 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 428 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 429 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 430 Scalar m13,m14,m15,m16; 431 Scalar *ba = b->a,*aa = a->a; 432 register int *pj; 433 434 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 435 PLogObjectParent(*B,isicol); 436 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 437 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 438 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 439 440 for ( i=0; i<n; i++ ) { 441 nz = bi[i+1] - bi[i]; 442 ajtmp = bj + bi[i]; 443 for ( j=0; j<nz; j++ ) { 444 x = rtmp+16*ajtmp[j]; 445 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 446 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 447 } 448 /* load in initial (unfactored row) */ 449 idx = r[i]; 450 nz = ai[idx+1] - ai[idx]; 451 ajtmpold = aj + ai[idx]; 452 v = aa + 16*ai[idx]; 453 for ( j=0; j<nz; j++ ) { 454 x = rtmp+16*ic[ajtmpold[j]]; 455 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 456 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 457 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 458 x[14] = v[14]; x[15] = v[15]; 459 v += 16; 460 } 461 row = *ajtmp++; 462 while (row < i) { 463 pc = rtmp + 16*row; 464 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 465 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 466 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 467 p15 = pc[14]; p16 = pc[15]; 468 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 469 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 470 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 471 || p16 != 0.0) { 472 pv = ba + 16*diag_offset[row]; 473 pj = bj + diag_offset[row] + 1; 474 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 475 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 476 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 477 x15 = pv[14]; x16 = pv[15]; 478 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 479 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 480 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 481 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 482 483 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 484 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 485 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 486 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 487 488 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 489 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 490 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 491 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 492 493 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 494 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 495 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 496 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 497 498 nz = bi[row+1] - diag_offset[row] - 1; 499 pv += 16; 500 for (j=0; j<nz; j++) { 501 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 502 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 503 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 504 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 505 x = rtmp + 16*pj[j]; 506 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 507 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 508 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 509 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 510 511 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 512 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 513 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 514 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 515 516 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 517 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 518 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 519 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 520 521 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 522 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 523 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 524 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 525 526 pv += 16; 527 } 528 PLogFlops(128*nz+112); 529 } 530 row = *ajtmp++; 531 } 532 /* finished row so stick it into b->a */ 533 pv = ba + 16*bi[i]; 534 pj = bj + bi[i]; 535 nz = bi[i+1] - bi[i]; 536 for ( j=0; j<nz; j++ ) { 537 x = rtmp+16*pj[j]; 538 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 539 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 540 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 541 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 542 pv += 16; 543 } 544 /* invert diagonal block */ 545 w = ba + 16*diag_offset[i]; 546 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 547 } 548 549 PetscFree(rtmp); 550 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 551 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 552 ierr = ISDestroy(isicol); CHKERRQ(ierr); 553 C->factor = FACTOR_LU; 554 C->assembled = PETSC_TRUE; 555 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 556 return 0; 557 } 558 /* ------------------------------------------------------------*/ 559 /* 560 Version for when blocks are 3 by 3 561 */ 562 #undef __FUNCTION__ 563 #define __FUNCTION__ "MatLUFactorNumeric_SeqBAIJ_3" 564 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 565 { 566 Mat C = *B; 567 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 568 IS iscol = b->col, isrow = b->row, isicol; 569 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 570 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 571 int *diag_offset = b->diag,idx; 572 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 573 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 574 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 575 Scalar *ba = b->a,*aa = a->a; 576 register int *pj; 577 578 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 579 PLogObjectParent(*B,isicol); 580 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 581 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 582 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 583 584 for ( i=0; i<n; i++ ) { 585 nz = bi[i+1] - bi[i]; 586 ajtmp = bj + bi[i]; 587 for ( j=0; j<nz; j++ ) { 588 x = rtmp + 9*ajtmp[j]; 589 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 590 } 591 /* load in initial (unfactored row) */ 592 idx = r[i]; 593 nz = ai[idx+1] - ai[idx]; 594 ajtmpold = aj + ai[idx]; 595 v = aa + 9*ai[idx]; 596 for ( j=0; j<nz; j++ ) { 597 x = rtmp + 9*ic[ajtmpold[j]]; 598 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 599 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 600 v += 9; 601 } 602 row = *ajtmp++; 603 while (row < i) { 604 pc = rtmp + 9*row; 605 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 606 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 607 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 608 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 609 pv = ba + 9*diag_offset[row]; 610 pj = bj + diag_offset[row] + 1; 611 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 612 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 613 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 614 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 615 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 616 617 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 618 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 619 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 620 621 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 622 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 623 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 624 nz = bi[row+1] - diag_offset[row] - 1; 625 pv += 9; 626 for (j=0; j<nz; j++) { 627 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 628 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 629 x = rtmp + 9*pj[j]; 630 x[0] -= m1*x1 + m4*x2 + m7*x3; 631 x[1] -= m2*x1 + m5*x2 + m8*x3; 632 x[2] -= m3*x1 + m6*x2 + m9*x3; 633 634 x[3] -= m1*x4 + m4*x5 + m7*x6; 635 x[4] -= m2*x4 + m5*x5 + m8*x6; 636 x[5] -= m3*x4 + m6*x5 + m9*x6; 637 638 x[6] -= m1*x7 + m4*x8 + m7*x9; 639 x[7] -= m2*x7 + m5*x8 + m8*x9; 640 x[8] -= m3*x7 + m6*x8 + m9*x9; 641 pv += 9; 642 } 643 PLogFlops(54*nz+36); 644 } 645 row = *ajtmp++; 646 } 647 /* finished row so stick it into b->a */ 648 pv = ba + 9*bi[i]; 649 pj = bj + bi[i]; 650 nz = bi[i+1] - bi[i]; 651 for ( j=0; j<nz; j++ ) { 652 x = rtmp + 9*pj[j]; 653 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 654 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 655 pv += 9; 656 } 657 /* invert diagonal block */ 658 w = ba + 9*diag_offset[i]; 659 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 660 } 661 662 PetscFree(rtmp); 663 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 664 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 665 ierr = ISDestroy(isicol); CHKERRQ(ierr); 666 C->factor = FACTOR_LU; 667 C->assembled = PETSC_TRUE; 668 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 669 return 0; 670 } 671 672 /* ------------------------------------------------------------*/ 673 /* 674 Version for when blocks are 2 by 2 675 */ 676 #undef __FUNCTION__ 677 #define __FUNCTION__ "MatLUFactorNumeric_SeqBAIJ_2" 678 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 679 { 680 Mat C = *B; 681 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 682 IS iscol = b->col, isrow = b->row, isicol; 683 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 684 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 685 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 686 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 687 Scalar p1,p2,p3,p4,v_work[2]; 688 Scalar *ba = b->a,*aa = a->a; 689 register int *pj; 690 691 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 692 PLogObjectParent(*B,isicol); 693 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 694 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 695 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 696 697 for ( i=0; i<n; i++ ) { 698 nz = bi[i+1] - bi[i]; 699 ajtmp = bj + bi[i]; 700 for ( j=0; j<nz; j++ ) { 701 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 702 } 703 /* load in initial (unfactored row) */ 704 idx = r[i]; 705 nz = ai[idx+1] - ai[idx]; 706 ajtmpold = aj + ai[idx]; 707 v = aa + 4*ai[idx]; 708 for ( j=0; j<nz; j++ ) { 709 x = rtmp+4*ic[ajtmpold[j]]; 710 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 711 v += 4; 712 } 713 row = *ajtmp++; 714 while (row < i) { 715 pc = rtmp + 4*row; 716 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 717 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 718 pv = ba + 4*diag_offset[row]; 719 pj = bj + diag_offset[row] + 1; 720 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 721 pc[0] = m1 = p1*x1 + p3*x2; 722 pc[1] = m2 = p2*x1 + p4*x2; 723 pc[2] = m3 = p1*x3 + p3*x4; 724 pc[3] = m4 = p2*x3 + p4*x4; 725 nz = bi[row+1] - diag_offset[row] - 1; 726 pv += 4; 727 for (j=0; j<nz; j++) { 728 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 729 x = rtmp + 4*pj[j]; 730 x[0] -= m1*x1 + m3*x2; 731 x[1] -= m2*x1 + m4*x2; 732 x[2] -= m1*x3 + m3*x4; 733 x[3] -= m2*x3 + m4*x4; 734 pv += 4; 735 } 736 PLogFlops(16*nz+12); 737 } 738 row = *ajtmp++; 739 } 740 /* finished row so stick it into b->a */ 741 pv = ba + 4*bi[i]; 742 pj = bj + bi[i]; 743 nz = bi[i+1] - bi[i]; 744 for ( j=0; j<nz; j++ ) { 745 x = rtmp+4*pj[j]; 746 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 747 pv += 4; 748 } 749 /* invert diagonal block */ 750 w = ba + 4*diag_offset[i]; 751 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 752 } 753 754 PetscFree(rtmp); 755 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 756 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 757 ierr = ISDestroy(isicol); CHKERRQ(ierr); 758 C->factor = FACTOR_LU; 759 C->assembled = PETSC_TRUE; 760 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 761 return 0; 762 } 763 764 /* ----------------------------------------------------------- */ 765 /* 766 Version for when blocks are 1 by 1. 767 */ 768 #undef __FUNCTION__ 769 #define __FUNCTION__ "MatLUFactorNumeric_SeqBAIJ_1" 770 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 771 { 772 Mat C = *B; 773 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 774 IS iscol = b->col, isrow = b->row, isicol; 775 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 776 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 777 int *diag_offset = b->diag,diag; 778 register Scalar *pv,*v,*rtmp,multiplier,*pc; 779 Scalar *ba = b->a,*aa = a->a; 780 register int *pj; 781 782 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 783 PLogObjectParent(*B,isicol); 784 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 785 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 786 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 787 788 for ( i=0; i<n; i++ ) { 789 nz = bi[i+1] - bi[i]; 790 ajtmp = bj + bi[i]; 791 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 792 793 /* load in initial (unfactored row) */ 794 nz = ai[r[i]+1] - ai[r[i]]; 795 ajtmpold = aj + ai[r[i]]; 796 v = aa + ai[r[i]]; 797 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 798 799 row = *ajtmp++; 800 while (row < i) { 801 pc = rtmp + row; 802 if (*pc != 0.0) { 803 pv = ba + diag_offset[row]; 804 pj = bj + diag_offset[row] + 1; 805 multiplier = *pc * *pv++; 806 *pc = multiplier; 807 nz = bi[row+1] - diag_offset[row] - 1; 808 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 809 PLogFlops(1+2*nz); 810 } 811 row = *ajtmp++; 812 } 813 /* finished row so stick it into b->a */ 814 pv = ba + bi[i]; 815 pj = bj + bi[i]; 816 nz = bi[i+1] - bi[i]; 817 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 818 diag = diag_offset[i] - bi[i]; 819 /* check pivot entry for current row */ 820 if (pv[diag] == 0.0) { 821 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"MatLUFactorNumeric_SeqBAIJ_1:Zero pivot"); 822 } 823 pv[diag] = 1.0/pv[diag]; 824 } 825 826 PetscFree(rtmp); 827 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 828 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 829 ierr = ISDestroy(isicol); CHKERRQ(ierr); 830 C->factor = FACTOR_LU; 831 C->assembled = PETSC_TRUE; 832 PLogFlops(b->n); 833 return 0; 834 } 835 836 /* ----------------------------------------------------------- */ 837 #undef __FUNCTION__ 838 #define __FUNCTION__ "MatLUFactor_SeqBAIJ" 839 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 840 { 841 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 842 int ierr; 843 Mat C; 844 845 ierr = MatLUFactorSymbolic_SeqBAIJ(A,row,col,f,&C); CHKERRQ(ierr); 846 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 847 848 /* free all the data structures from mat */ 849 PetscFree(mat->a); 850 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 851 if (mat->diag) PetscFree(mat->diag); 852 if (mat->ilen) PetscFree(mat->ilen); 853 if (mat->imax) PetscFree(mat->imax); 854 if (mat->solve_work) PetscFree(mat->solve_work); 855 if (mat->mult_work) PetscFree(mat->mult_work); 856 PetscFree(mat); 857 858 PetscMemcpy(A,C,sizeof(struct _Mat)); 859 PetscHeaderDestroy(C); 860 return 0; 861 } 862 /* ----------------------------------------------------------- */ 863 #undef __FUNCTION__ 864 #define __FUNCTION__ "MatSolve_SeqBAIJ_N" 865 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 866 { 867 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 868 IS iscol=a->col,isrow=a->row; 869 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 870 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 871 Scalar *aa=a->a,*sum; 872 register Scalar *x,*b,*lsum,*tmp,*v; 873 874 VecGetArray_Fast(bb,b); 875 VecGetArray_Fast(xx,x); 876 tmp = a->solve_work; 877 878 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 879 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 880 881 /* forward solve the lower triangular */ 882 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 883 for ( i=1; i<n; i++ ) { 884 v = aa + bs2*ai[i]; 885 vi = aj + ai[i]; 886 nz = a->diag[i] - ai[i]; 887 sum = tmp + bs*i; 888 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 889 while (nz--) { 890 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 891 v += bs2; 892 } 893 } 894 /* backward solve the upper triangular */ 895 lsum = a->solve_work + a->n; 896 for ( i=n-1; i>=0; i-- ){ 897 v = aa + bs2*(a->diag[i] + 1); 898 vi = aj + a->diag[i] + 1; 899 nz = ai[i+1] - a->diag[i] - 1; 900 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 901 while (nz--) { 902 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 903 v += bs2; 904 } 905 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 906 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 907 } 908 909 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 910 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 911 VecRestoreArray_Fast(bb,b); 912 VecRestoreArray_Fast(xx,x); 913 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 914 return 0; 915 } 916 917 #undef __FUNCTION__ 918 #define __FUNCTION__ "MatSolve_SeqBAIJ_5" 919 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 920 { 921 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 922 IS iscol=a->col,isrow=a->row; 923 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 924 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 925 register Scalar *x,*b,*tmp,*v; 926 927 VecGetArray_Fast(bb,b); 928 VecGetArray_Fast(xx,x); 929 tmp = a->solve_work; 930 931 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 932 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 933 934 /* forward solve the lower triangular */ 935 idx = 5*(*r++); 936 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 937 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 938 for ( i=1; i<n; i++ ) { 939 v = aa + 25*ai[i]; 940 vi = aj + ai[i]; 941 nz = a->diag[i] - ai[i]; 942 idx = 5*(*r++); 943 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 944 sum5 = b[4+idx]; 945 while (nz--) { 946 idx = 5*(*vi++); 947 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 948 x4 = tmp[3+idx];x5 = tmp[4+idx]; 949 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 950 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 951 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 952 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 953 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 954 v += 25; 955 } 956 idx = 5*i; 957 tmp[idx] = sum1;tmp[1+idx] = sum2; 958 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 959 } 960 /* backward solve the upper triangular */ 961 for ( i=n-1; i>=0; i-- ){ 962 v = aa + 25*a->diag[i] + 25; 963 vi = aj + a->diag[i] + 1; 964 nz = ai[i+1] - a->diag[i] - 1; 965 idt = 5*i; 966 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 967 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 968 while (nz--) { 969 idx = 5*(*vi++); 970 x1 = tmp[idx]; x2 = tmp[1+idx]; 971 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 972 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 973 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 974 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 975 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 976 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 977 v += 25; 978 } 979 idc = 5*(*c--); 980 v = aa + 25*a->diag[i]; 981 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 982 v[15]*sum4+v[20]*sum5; 983 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 984 v[16]*sum4+v[21]*sum5; 985 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 986 v[17]*sum4+v[22]*sum5; 987 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 988 v[18]*sum4+v[23]*sum5; 989 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 990 v[19]*sum4+v[24]*sum5; 991 } 992 993 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 994 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 995 VecRestoreArray_Fast(bb,b); 996 VecRestoreArray_Fast(xx,x); 997 PLogFlops(2*25*(a->nz) - a->n); 998 return 0; 999 } 1000 1001 #undef __FUNCTION__ 1002 #define __FUNCTION__ "MatSolve_SeqBAIJ_4" 1003 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1004 { 1005 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1006 IS iscol=a->col,isrow=a->row; 1007 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1008 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1009 register Scalar *x,*b,*tmp,*v; 1010 1011 VecGetArray_Fast(bb,b); 1012 VecGetArray_Fast(xx,x); 1013 tmp = a->solve_work; 1014 1015 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1016 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1017 1018 /* forward solve the lower triangular */ 1019 idx = 4*(*r++); 1020 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1021 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1022 for ( i=1; i<n; i++ ) { 1023 v = aa + 16*ai[i]; 1024 vi = aj + ai[i]; 1025 nz = a->diag[i] - ai[i]; 1026 idx = 4*(*r++); 1027 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1028 while (nz--) { 1029 idx = 4*(*vi++); 1030 x1 = tmp[idx];x2 = tmp[1+idx];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 idx = 4*i; 1038 tmp[idx] = sum1;tmp[1+idx] = sum2; 1039 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1040 } 1041 /* backward solve the upper triangular */ 1042 for ( i=n-1; i>=0; i-- ){ 1043 v = aa + 16*a->diag[i] + 16; 1044 vi = aj + a->diag[i] + 1; 1045 nz = ai[i+1] - a->diag[i] - 1; 1046 idt = 4*i; 1047 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1048 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1049 while (nz--) { 1050 idx = 4*(*vi++); 1051 x1 = tmp[idx]; x2 = tmp[1+idx]; 1052 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1053 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1054 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1055 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1056 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1057 v += 16; 1058 } 1059 idc = 4*(*c--); 1060 v = aa + 16*a->diag[i]; 1061 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1062 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1063 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1064 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1065 } 1066 1067 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1068 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1069 VecRestoreArray_Fast(bb,b); 1070 VecRestoreArray_Fast(xx,x); 1071 PLogFlops(2*16*(a->nz) - a->n); 1072 return 0; 1073 } 1074 1075 1076 #undef __FUNCTION__ 1077 #define __FUNCTION__ "MatSolve_SeqBAIJ_3" 1078 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1079 { 1080 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1081 IS iscol=a->col,isrow=a->row; 1082 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1083 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1084 register Scalar *x,*b,*tmp,*v; 1085 1086 VecGetArray_Fast(bb,b); 1087 VecGetArray_Fast(xx,x); 1088 tmp = a->solve_work; 1089 1090 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1091 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1092 1093 /* forward solve the lower triangular */ 1094 idx = 3*(*r++); 1095 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1096 for ( i=1; i<n; i++ ) { 1097 v = aa + 9*ai[i]; 1098 vi = aj + ai[i]; 1099 nz = a->diag[i] - ai[i]; 1100 idx = 3*(*r++); 1101 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1102 while (nz--) { 1103 idx = 3*(*vi++); 1104 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1105 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1106 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1107 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1108 v += 9; 1109 } 1110 idx = 3*i; 1111 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1112 } 1113 /* backward solve the upper triangular */ 1114 for ( i=n-1; i>=0; i-- ){ 1115 v = aa + 9*a->diag[i] + 9; 1116 vi = aj + a->diag[i] + 1; 1117 nz = ai[i+1] - a->diag[i] - 1; 1118 idt = 3*i; 1119 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1120 while (nz--) { 1121 idx = 3*(*vi++); 1122 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1123 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1124 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1125 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1126 v += 9; 1127 } 1128 idc = 3*(*c--); 1129 v = aa + 9*a->diag[i]; 1130 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1131 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1132 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1133 } 1134 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1135 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1136 VecRestoreArray_Fast(bb,b); 1137 VecRestoreArray_Fast(xx,x); 1138 PLogFlops(2*9*(a->nz) - a->n); 1139 return 0; 1140 } 1141 1142 #undef __FUNCTION__ 1143 #define __FUNCTION__ "MatSolve_SeqBAIJ_2" 1144 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1145 { 1146 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1147 IS iscol=a->col,isrow=a->row; 1148 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1149 Scalar *aa=a->a,sum1,sum2,x1,x2; 1150 register Scalar *x,*b,*tmp,*v; 1151 1152 VecGetArray_Fast(bb,b); 1153 VecGetArray_Fast(xx,x); 1154 tmp = a->solve_work; 1155 1156 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1157 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1158 1159 /* forward solve the lower triangular */ 1160 idx = 2*(*r++); 1161 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1162 for ( i=1; i<n; i++ ) { 1163 v = aa + 4*ai[i]; 1164 vi = aj + ai[i]; 1165 nz = a->diag[i] - ai[i]; 1166 idx = 2*(*r++); 1167 sum1 = b[idx]; sum2 = b[1+idx]; 1168 while (nz--) { 1169 idx = 2*(*vi++); 1170 x1 = tmp[idx]; x2 = tmp[1+idx]; 1171 sum1 -= v[0]*x1 + v[2]*x2; 1172 sum2 -= v[1]*x1 + v[3]*x2; 1173 v += 4; 1174 } 1175 idx = 2*i; 1176 tmp[idx] = sum1; tmp[1+idx] = sum2; 1177 } 1178 /* backward solve the upper triangular */ 1179 for ( i=n-1; i>=0; i-- ){ 1180 v = aa + 4*a->diag[i] + 4; 1181 vi = aj + a->diag[i] + 1; 1182 nz = ai[i+1] - a->diag[i] - 1; 1183 idt = 2*i; 1184 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1185 while (nz--) { 1186 idx = 2*(*vi++); 1187 x1 = tmp[idx]; x2 = tmp[1+idx]; 1188 sum1 -= v[0]*x1 + v[2]*x2; 1189 sum2 -= v[1]*x1 + v[3]*x2; 1190 v += 4; 1191 } 1192 idc = 2*(*c--); 1193 v = aa + 4*a->diag[i]; 1194 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1195 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1196 } 1197 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1198 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1199 VecRestoreArray_Fast(bb,b); 1200 VecRestoreArray_Fast(xx,x); 1201 PLogFlops(2*4*(a->nz) - a->n); 1202 return 0; 1203 } 1204 1205 1206 #undef __FUNCTION__ 1207 #define __FUNCTION__ "MatSolve_SeqBAIJ_1" 1208 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1209 { 1210 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1211 IS iscol=a->col,isrow=a->row; 1212 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1213 Scalar *aa=a->a,sum1; 1214 register Scalar *x,*b,*tmp,*v; 1215 1216 if (!n) return 0; 1217 1218 VecGetArray_Fast(bb,b); 1219 VecGetArray_Fast(xx,x); 1220 tmp = a->solve_work; 1221 1222 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1223 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1224 1225 /* forward solve the lower triangular */ 1226 tmp[0] = b[*r++]; 1227 for ( i=1; i<n; i++ ) { 1228 v = aa + ai[i]; 1229 vi = aj + ai[i]; 1230 nz = a->diag[i] - ai[i]; 1231 sum1 = b[*r++]; 1232 while (nz--) { 1233 sum1 -= (*v++)*tmp[*vi++]; 1234 } 1235 tmp[i] = sum1; 1236 } 1237 /* backward solve the upper triangular */ 1238 for ( i=n-1; i>=0; i-- ){ 1239 v = aa + a->diag[i] + 1; 1240 vi = aj + a->diag[i] + 1; 1241 nz = ai[i+1] - a->diag[i] - 1; 1242 sum1 = tmp[i]; 1243 while (nz--) { 1244 sum1 -= (*v++)*tmp[*vi++]; 1245 } 1246 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1247 } 1248 1249 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1250 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1251 VecRestoreArray_Fast(bb,b); 1252 VecRestoreArray_Fast(xx,x); 1253 PLogFlops(2*1*(a->nz) - a->n); 1254 return 0; 1255 } 1256 1257 /* ----------------------------------------------------------------*/ 1258 /* 1259 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1260 except that the data structure of Mat_SeqAIJ is slightly different. 1261 Not a good example of code reuse. 1262 */ 1263 #undef __FUNCTION__ 1264 #define __FUNCTION__ "MatILUFactorSymbolic_SeqBAIJ" 1265 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1266 Mat *fact) 1267 { 1268 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1269 IS isicol; 1270 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1271 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1272 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1273 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1274 PetscTruth col_identity, row_identity; 1275 1276 /* special case that simply copies fill pattern */ 1277 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1278 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1279 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1280 if (levels == 0 && row_identity && col_identity) { 1281 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1282 (*fact)->factor = FACTOR_LU; 1283 b = (Mat_SeqBAIJ *) (*fact)->data; 1284 if (!b->diag) { 1285 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1286 } 1287 b->row = isrow; 1288 b->col = iscol; 1289 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1290 return 0; 1291 } 1292 1293 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1294 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1295 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1296 1297 /* get new row pointers */ 1298 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1299 ainew[0] = 0; 1300 /* don't know how many column pointers are needed so estimate */ 1301 jmax = (int) (f*ai[n] + 1); 1302 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1303 /* ajfill is level of fill for each fill entry */ 1304 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1305 /* fill is a linked list of nonzeros in active row */ 1306 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1307 /* im is level for each filled value */ 1308 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1309 /* dloc is location of diagonal in factor */ 1310 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1311 dloc[0] = 0; 1312 for ( prow=0; prow<n; prow++ ) { 1313 /* first copy previous fill into linked list */ 1314 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1315 xi = aj + ai[r[prow]]; 1316 fill[n] = n; 1317 while (nz--) { 1318 fm = n; 1319 idx = ic[*xi++]; 1320 do { 1321 m = fm; 1322 fm = fill[m]; 1323 } while (fm < idx); 1324 fill[m] = idx; 1325 fill[idx] = fm; 1326 im[idx] = 0; 1327 } 1328 nzi = 0; 1329 row = fill[n]; 1330 while ( row < prow ) { 1331 incrlev = im[row] + 1; 1332 nz = dloc[row]; 1333 xi = ajnew + ainew[row] + nz; 1334 flev = ajfill + ainew[row] + nz + 1; 1335 nnz = ainew[row+1] - ainew[row] - nz - 1; 1336 if (*xi++ != row) { 1337 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"MatILUFactorSymbolic_SeqBAIJ:zero pivot"); 1338 } 1339 fm = row; 1340 while (nnz-- > 0) { 1341 idx = *xi++; 1342 if (*flev + incrlev > levels) { 1343 flev++; 1344 continue; 1345 } 1346 do { 1347 m = fm; 1348 fm = fill[m]; 1349 } while (fm < idx); 1350 if (fm != idx) { 1351 im[idx] = *flev + incrlev; 1352 fill[m] = idx; 1353 fill[idx] = fm; 1354 fm = idx; 1355 nzf++; 1356 } 1357 else { 1358 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1359 } 1360 flev++; 1361 } 1362 row = fill[row]; 1363 nzi++; 1364 } 1365 /* copy new filled row into permanent storage */ 1366 ainew[prow+1] = ainew[prow] + nzf; 1367 if (ainew[prow+1] > jmax) { 1368 /* allocate a longer ajnew */ 1369 int maxadd; 1370 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1371 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1372 jmax += maxadd; 1373 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1374 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1375 PetscFree(ajnew); 1376 ajnew = xi; 1377 /* allocate a longer ajfill */ 1378 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1379 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1380 PetscFree(ajfill); 1381 ajfill = xi; 1382 realloc++; 1383 } 1384 xi = ajnew + ainew[prow]; 1385 flev = ajfill + ainew[prow]; 1386 dloc[prow] = nzi; 1387 fm = fill[n]; 1388 while (nzf--) { 1389 *xi++ = fm; 1390 *flev++ = im[fm]; 1391 fm = fill[fm]; 1392 } 1393 } 1394 PetscFree(ajfill); 1395 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1396 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1397 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1398 PetscFree(fill); PetscFree(im); 1399 1400 PLogInfo(A, 1401 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1402 realloc,f,((double)ainew[n])/((double)ai[prow])); 1403 1404 /* put together the new matrix */ 1405 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1406 b = (Mat_SeqBAIJ *) (*fact)->data; 1407 PetscFree(b->imax); 1408 b->singlemalloc = 0; 1409 len = bs2*ainew[n]*sizeof(Scalar); 1410 /* the next line frees the default space generated by the Create() */ 1411 PetscFree(b->a); PetscFree(b->ilen); 1412 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1413 b->j = ajnew; 1414 b->i = ainew; 1415 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1416 b->diag = dloc; 1417 b->ilen = 0; 1418 b->imax = 0; 1419 b->row = isrow; 1420 b->col = iscol; 1421 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1422 CHKPTRQ(b->solve_work); 1423 /* In b structure: Free imax, ilen, old a, old j. 1424 Allocate dloc, solve_work, new a, new j */ 1425 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1426 b->maxnz = b->nz = ainew[n]; 1427 (*fact)->factor = FACTOR_LU; 1428 1429 (*fact)->info.factor_mallocs = realloc; 1430 (*fact)->info.fill_ratio_given = f; 1431 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1432 1433 return 0; 1434 } 1435 1436 1437 1438 1439