1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.23 1996/05/03 19:27:07 bsmith Exp bsmith $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "baij.h" 9 #include "src/inline/ilu.h" 10 11 /* 12 The symbolic factorization code is identical to that for AIJ format, 13 except for very small changes since this is now a SeqBAIJ datastructure. 14 NOT good code reuse. 15 */ 16 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 17 { 18 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 19 IS isicol; 20 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 21 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 22 int *idnew, idx, row,m,fm, nnz, nzi,len, realloc = 0,nzbd,*im; 23 24 if (a->m != a->n) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must be square"); 25 if (!isrow) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must have row permutation"); 26 if (!iscol) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must have col. permutation"); 27 28 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 29 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 30 31 /* get new row pointers */ 32 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 33 ainew[0] = 0; 34 /* don't know how many column pointers are needed so estimate */ 35 jmax = (int) (f*ai[n] + 1); 36 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 37 /* fill is a linked list of nonzeros in active row */ 38 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 39 im = fill + n + 1; 40 /* idnew is location of diagonal in factor */ 41 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 42 idnew[0] = 0; 43 44 for ( i=0; i<n; i++ ) { 45 /* first copy previous fill into linked list */ 46 nnz = nz = ai[r[i]+1] - ai[r[i]]; 47 ajtmp = aj + ai[r[i]]; 48 fill[n] = n; 49 while (nz--) { 50 fm = n; 51 idx = ic[*ajtmp++]; 52 do { 53 m = fm; 54 fm = fill[m]; 55 } while (fm < idx); 56 fill[m] = idx; 57 fill[idx] = fm; 58 } 59 row = fill[n]; 60 while ( row < i ) { 61 ajtmp = ajnew + idnew[row] + 1; 62 nzbd = 1 + idnew[row] - ainew[row]; 63 nz = im[row] - nzbd; 64 fm = row; 65 while (nz-- > 0) { 66 idx = *ajtmp++; 67 nzbd++; 68 if (idx == i) im[row] = nzbd; 69 do { 70 m = fm; 71 fm = fill[m]; 72 } while (fm < idx); 73 if (fm != idx) { 74 fill[m] = idx; 75 fill[idx] = fm; 76 fm = idx; 77 nnz++; 78 } 79 } 80 row = fill[row]; 81 } 82 /* copy new filled row into permanent storage */ 83 ainew[i+1] = ainew[i] + nnz; 84 if (ainew[i+1] > jmax) { 85 /* allocate a longer ajnew */ 86 int maxadd; 87 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 88 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 89 jmax += maxadd; 90 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 91 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 92 PetscFree(ajnew); 93 ajnew = ajtmp; 94 realloc++; /* count how many times we realloc */ 95 } 96 ajtmp = ajnew + ainew[i]; 97 fm = fill[n]; 98 nzi = 0; 99 im[i] = nnz; 100 while (nnz--) { 101 if (fm < i) nzi++; 102 *ajtmp++ = fm; 103 fm = fill[fm]; 104 } 105 idnew[i] = ainew[i] + nzi; 106 } 107 108 PLogInfo(A, 109 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 110 realloc,f,((double)ainew[n])/((double)ai[i])); 111 112 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 113 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 114 115 PetscFree(fill); 116 117 /* put together the new matrix */ 118 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 119 PLogObjectParent(*B,isicol); 120 ierr = ISDestroy(isicol); CHKERRQ(ierr); 121 b = (Mat_SeqBAIJ *) (*B)->data; 122 PetscFree(b->imax); 123 b->singlemalloc = 0; 124 len = ainew[n]*sizeof(Scalar); 125 /* the next line frees the default space generated by the Create() */ 126 PetscFree(b->a); PetscFree(b->ilen); 127 b->a = (Scalar *) PetscMalloc( len*bs2 ); CHKPTRQ(b->a); 128 b->j = ajnew; 129 b->i = ainew; 130 b->diag = idnew; 131 b->ilen = 0; 132 b->imax = 0; 133 b->row = isrow; 134 b->col = iscol; 135 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 136 CHKPTRQ(b->solve_work); 137 /* In b structure: Free imax, ilen, old a, old j. 138 Allocate idnew, solve_work, new a, new j */ 139 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 140 b->maxnz = b->nz = ainew[n]; 141 142 return 0; 143 } 144 145 /* ----------------------------------------------------------- */ 146 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 147 { 148 Mat C = *B; 149 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 150 IS iscol = b->col, isrow = b->row, isicol; 151 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 152 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j; 153 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = bs*bs,*v_pivots; 154 Scalar *ba = b->a,*aa = a->a; 155 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 156 register int *pj; 157 158 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 159 PLogObjectParent(*B,isicol); 160 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 161 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 162 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 163 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 164 /* generate work space needed by dense LU factorization */ 165 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 166 CHKPTRQ(v_work); 167 multiplier = v_work + bs; 168 v_pivots = (int *) (multiplier + bs2); 169 170 /* flops in while loop */ 171 bslog = 2*bs*bs2; 172 173 for ( i=0; i<n; i++ ) { 174 nz = bi[i+1] - bi[i]; 175 ajtmp = bj + bi[i]; 176 for ( j=0; j<nz; j++ ) { 177 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 178 } 179 /* load in initial (unfactored row) */ 180 nz = ai[r[i]+1] - ai[r[i]]; 181 ajtmpold = aj + ai[r[i]]; 182 v = aa + bs2*ai[r[i]]; 183 for ( j=0; j<nz; j++ ) { 184 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 185 } 186 row = *ajtmp++; 187 while (row < i) { 188 pc = rtmp + bs2*row; 189 /* if (*pc) { */ 190 pv = ba + bs2*diag_offset[row]; 191 pj = bj + diag_offset[row] + 1; 192 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 193 nz = bi[row+1] - diag_offset[row] - 1; 194 pv += bs2; 195 for (j=0; j<nz; j++) { 196 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 197 } 198 PLogFlops(bslog*(nz+1)-bs); 199 /* } */ 200 row = *ajtmp++; 201 } 202 /* finished row so stick it into b->a */ 203 pv = ba + bs2*bi[i]; 204 pj = bj + bi[i]; 205 nz = bi[i+1] - bi[i]; 206 for ( j=0; j<nz; j++ ) { 207 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 208 } 209 diag = diag_offset[i] - bi[i]; 210 /* invert diagonal block */ 211 w = pv + bs2*diag; 212 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 213 } 214 215 PetscFree(rtmp); PetscFree(v_work); 216 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 217 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 218 ierr = ISDestroy(isicol); CHKERRQ(ierr); 219 C->factor = FACTOR_LU; 220 C->assembled = PETSC_TRUE; 221 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 222 return 0; 223 } 224 /* ------------------------------------------------------------*/ 225 /* 226 Version for when blocks are 5 by 5 227 */ 228 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 229 { 230 Mat C = *B; 231 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 232 IS iscol = b->col, isrow = b->row, isicol; 233 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 234 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 235 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 236 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 237 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 238 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 239 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 240 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 241 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 242 Scalar *ba = b->a,*aa = a->a; 243 register int *pj; 244 245 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 246 PLogObjectParent(*B,isicol); 247 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 248 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 249 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 250 251 for ( i=0; i<n; i++ ) { 252 nz = bi[i+1] - bi[i]; 253 ajtmp = bj + bi[i]; 254 for ( j=0; j<nz; j++ ) { 255 x = rtmp+25*ajtmp[j]; 256 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 257 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 258 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 259 } 260 /* load in initial (unfactored row) */ 261 idx = r[i]; 262 nz = ai[idx+1] - ai[idx]; 263 ajtmpold = aj + ai[idx]; 264 v = aa + 25*ai[idx]; 265 for ( j=0; j<nz; j++ ) { 266 x = rtmp+25*ic[ajtmpold[j]]; 267 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 268 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 269 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 270 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 271 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 272 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 273 v += 25; 274 } 275 row = *ajtmp++; 276 while (row < i) { 277 pc = rtmp + 25*row; 278 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 279 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 280 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 281 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 282 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 283 p25 = pc[24]; 284 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 285 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 286 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 287 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 288 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 289 p24 != 0.0 || p25 != 0.0) { 290 pv = ba + 25*diag_offset[row]; 291 pj = bj + diag_offset[row] + 1; 292 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 293 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 294 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 295 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 296 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 297 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 298 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 299 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 300 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 301 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 302 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 303 304 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 305 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 306 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 307 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 308 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 309 310 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 311 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 312 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 313 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 314 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 315 316 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 317 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 318 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 319 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 320 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 321 322 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 323 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 324 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 325 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 326 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 327 328 nz = bi[row+1] - diag_offset[row] - 1; 329 pv += 25; 330 for (j=0; j<nz; j++) { 331 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 332 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 333 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 334 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 335 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 336 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 337 x = rtmp + 25*pj[j]; 338 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 339 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 340 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 341 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 342 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 343 344 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 345 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 346 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 347 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 348 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 349 350 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 351 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 352 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 353 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 354 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 355 356 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 357 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 358 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 359 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 360 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 361 362 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 363 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 364 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 365 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 366 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 367 368 pv += 25; 369 } 370 PLogFlops(250*nz+225); 371 } 372 row = *ajtmp++; 373 } 374 /* finished row so stick it into b->a */ 375 pv = ba + 25*bi[i]; 376 pj = bj + bi[i]; 377 nz = bi[i+1] - bi[i]; 378 for ( j=0; j<nz; j++ ) { 379 x = rtmp+25*pj[j]; 380 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 381 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 382 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 383 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 384 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 385 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 386 pv += 25; 387 } 388 /* invert diagonal block */ 389 w = ba + 25*diag_offset[i]; 390 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 391 } 392 393 PetscFree(rtmp); 394 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 395 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 396 ierr = ISDestroy(isicol); CHKERRQ(ierr); 397 C->factor = FACTOR_LU; 398 C->assembled = PETSC_TRUE; 399 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 400 return 0; 401 } 402 403 /* ------------------------------------------------------------*/ 404 /* 405 Version for when blocks are 4 by 4 406 */ 407 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 408 { 409 Mat C = *B; 410 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 411 IS iscol = b->col, isrow = b->row, isicol; 412 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 413 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 414 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 415 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 416 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 417 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 418 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 419 Scalar m13,m14,m15,m16; 420 Scalar *ba = b->a,*aa = a->a; 421 register int *pj; 422 423 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 424 PLogObjectParent(*B,isicol); 425 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 426 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 427 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 428 429 for ( i=0; i<n; i++ ) { 430 nz = bi[i+1] - bi[i]; 431 ajtmp = bj + bi[i]; 432 for ( j=0; j<nz; j++ ) { 433 x = rtmp+16*ajtmp[j]; 434 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 435 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 436 } 437 /* load in initial (unfactored row) */ 438 idx = r[i]; 439 nz = ai[idx+1] - ai[idx]; 440 ajtmpold = aj + ai[idx]; 441 v = aa + 16*ai[idx]; 442 for ( j=0; j<nz; j++ ) { 443 x = rtmp+16*ic[ajtmpold[j]]; 444 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 445 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 446 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 447 x[14] = v[14]; x[15] = v[15]; 448 v += 16; 449 } 450 row = *ajtmp++; 451 while (row < i) { 452 pc = rtmp + 16*row; 453 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 454 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 455 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 456 p15 = pc[14]; p16 = pc[15]; 457 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 458 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 459 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 460 || p16 != 0.0) { 461 pv = ba + 16*diag_offset[row]; 462 pj = bj + diag_offset[row] + 1; 463 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 464 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 465 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 466 x15 = pv[14]; x16 = pv[15]; 467 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 468 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 469 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 470 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 471 472 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 473 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 474 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 475 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 476 477 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 478 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 479 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 480 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 481 482 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 483 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 484 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 485 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 486 487 nz = bi[row+1] - diag_offset[row] - 1; 488 pv += 16; 489 for (j=0; j<nz; j++) { 490 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 491 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 492 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 493 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 494 x = rtmp + 16*pj[j]; 495 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 496 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 497 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 498 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 499 500 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 501 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 502 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 503 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 504 505 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 506 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 507 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 508 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 509 510 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 511 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 512 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 513 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 514 515 pv += 16; 516 } 517 PLogFlops(128*nz+112); 518 } 519 row = *ajtmp++; 520 } 521 /* finished row so stick it into b->a */ 522 pv = ba + 16*bi[i]; 523 pj = bj + bi[i]; 524 nz = bi[i+1] - bi[i]; 525 for ( j=0; j<nz; j++ ) { 526 x = rtmp+16*pj[j]; 527 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 528 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 529 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 530 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 531 pv += 16; 532 } 533 /* invert diagonal block */ 534 w = ba + 16*diag_offset[i]; 535 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 536 } 537 538 PetscFree(rtmp); 539 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 540 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 541 ierr = ISDestroy(isicol); CHKERRQ(ierr); 542 C->factor = FACTOR_LU; 543 C->assembled = PETSC_TRUE; 544 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 545 return 0; 546 } 547 /* ------------------------------------------------------------*/ 548 /* 549 Version for when blocks are 3 by 3 550 */ 551 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 552 { 553 Mat C = *B; 554 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 555 IS iscol = b->col, isrow = b->row, isicol; 556 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 557 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 558 int *diag_offset = b->diag,idx; 559 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 560 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 561 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 562 Scalar *ba = b->a,*aa = a->a; 563 register int *pj; 564 565 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 566 PLogObjectParent(*B,isicol); 567 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 568 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 569 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 570 571 for ( i=0; i<n; i++ ) { 572 nz = bi[i+1] - bi[i]; 573 ajtmp = bj + bi[i]; 574 for ( j=0; j<nz; j++ ) { 575 x = rtmp + 9*ajtmp[j]; 576 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 577 } 578 /* load in initial (unfactored row) */ 579 idx = r[i]; 580 nz = ai[idx+1] - ai[idx]; 581 ajtmpold = aj + ai[idx]; 582 v = aa + 9*ai[idx]; 583 for ( j=0; j<nz; j++ ) { 584 x = rtmp + 9*ic[ajtmpold[j]]; 585 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 586 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 587 v += 9; 588 } 589 row = *ajtmp++; 590 while (row < i) { 591 pc = rtmp + 9*row; 592 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 593 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 594 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 595 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 596 pv = ba + 9*diag_offset[row]; 597 pj = bj + diag_offset[row] + 1; 598 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 599 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 600 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 601 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 602 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 603 604 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 605 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 606 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 607 608 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 609 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 610 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 611 nz = bi[row+1] - diag_offset[row] - 1; 612 pv += 9; 613 for (j=0; j<nz; j++) { 614 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 615 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 616 x = rtmp + 9*pj[j]; 617 x[0] -= m1*x1 + m4*x2 + m7*x3; 618 x[1] -= m2*x1 + m5*x2 + m8*x3; 619 x[2] -= m3*x1 + m6*x2 + m9*x3; 620 621 x[3] -= m1*x4 + m4*x5 + m7*x6; 622 x[4] -= m2*x4 + m5*x5 + m8*x6; 623 x[5] -= m3*x4 + m6*x5 + m9*x6; 624 625 x[6] -= m1*x7 + m4*x8 + m7*x9; 626 x[7] -= m2*x7 + m5*x8 + m8*x9; 627 x[8] -= m3*x7 + m6*x8 + m9*x9; 628 pv += 9; 629 } 630 PLogFlops(54*nz+36); 631 } 632 row = *ajtmp++; 633 } 634 /* finished row so stick it into b->a */ 635 pv = ba + 9*bi[i]; 636 pj = bj + bi[i]; 637 nz = bi[i+1] - bi[i]; 638 for ( j=0; j<nz; j++ ) { 639 x = rtmp + 9*pj[j]; 640 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 641 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 642 pv += 9; 643 } 644 /* invert diagonal block */ 645 w = ba + 9*diag_offset[i]; 646 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 647 } 648 649 PetscFree(rtmp); 650 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 651 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 652 ierr = ISDestroy(isicol); CHKERRQ(ierr); 653 C->factor = FACTOR_LU; 654 C->assembled = PETSC_TRUE; 655 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 656 return 0; 657 } 658 659 /* ------------------------------------------------------------*/ 660 /* 661 Version for when blocks are 2 by 2 662 */ 663 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 664 { 665 Mat C = *B; 666 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 667 IS iscol = b->col, isrow = b->row, isicol; 668 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 669 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 670 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 671 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 672 Scalar p1,p2,p3,p4,v_work[2]; 673 Scalar *ba = b->a,*aa = a->a; 674 register int *pj; 675 676 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 677 PLogObjectParent(*B,isicol); 678 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 679 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 680 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 681 682 for ( i=0; i<n; i++ ) { 683 nz = bi[i+1] - bi[i]; 684 ajtmp = bj + bi[i]; 685 for ( j=0; j<nz; j++ ) { 686 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 687 } 688 /* load in initial (unfactored row) */ 689 idx = r[i]; 690 nz = ai[idx+1] - ai[idx]; 691 ajtmpold = aj + ai[idx]; 692 v = aa + 4*ai[idx]; 693 for ( j=0; j<nz; j++ ) { 694 x = rtmp+4*ic[ajtmpold[j]]; 695 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 696 v += 4; 697 } 698 row = *ajtmp++; 699 while (row < i) { 700 pc = rtmp + 4*row; 701 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 702 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 703 pv = ba + 4*diag_offset[row]; 704 pj = bj + diag_offset[row] + 1; 705 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 706 pc[0] = m1 = p1*x1 + p3*x2; 707 pc[1] = m2 = p2*x1 + p4*x2; 708 pc[2] = m3 = p1*x3 + p3*x4; 709 pc[3] = m4 = p2*x3 + p4*x4; 710 nz = bi[row+1] - diag_offset[row] - 1; 711 pv += 4; 712 for (j=0; j<nz; j++) { 713 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 714 x = rtmp + 4*pj[j]; 715 x[0] -= m1*x1 + m3*x2; 716 x[1] -= m2*x1 + m4*x2; 717 x[2] -= m1*x3 + m3*x4; 718 x[3] -= m2*x3 + m4*x4; 719 pv += 4; 720 } 721 PLogFlops(16*nz+12); 722 } 723 row = *ajtmp++; 724 } 725 /* finished row so stick it into b->a */ 726 pv = ba + 4*bi[i]; 727 pj = bj + bi[i]; 728 nz = bi[i+1] - bi[i]; 729 for ( j=0; j<nz; j++ ) { 730 x = rtmp+4*pj[j]; 731 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 732 pv += 4; 733 } 734 /* invert diagonal block */ 735 w = ba + 4*diag_offset[i]; 736 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 737 } 738 739 PetscFree(rtmp); 740 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 741 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 742 ierr = ISDestroy(isicol); CHKERRQ(ierr); 743 C->factor = FACTOR_LU; 744 C->assembled = PETSC_TRUE; 745 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 746 return 0; 747 } 748 749 /* ----------------------------------------------------------- */ 750 /* 751 Version for when blocks are 1 by 1. 752 */ 753 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 754 { 755 Mat C = *B; 756 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 757 IS iscol = b->col, isrow = b->row, isicol; 758 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 759 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 760 int *diag_offset = b->diag,diag; 761 register Scalar *pv,*v,*rtmp,multiplier,*pc; 762 Scalar *ba = b->a,*aa = a->a; 763 register int *pj; 764 765 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 766 PLogObjectParent(*B,isicol); 767 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 768 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 769 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 770 771 for ( i=0; i<n; i++ ) { 772 nz = bi[i+1] - bi[i]; 773 ajtmp = bj + bi[i]; 774 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 775 776 /* load in initial (unfactored row) */ 777 nz = ai[r[i]+1] - ai[r[i]]; 778 ajtmpold = aj + ai[r[i]]; 779 v = aa + ai[r[i]]; 780 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 781 782 row = *ajtmp++; 783 while (row < i) { 784 pc = rtmp + row; 785 if (*pc != 0.0) { 786 pv = ba + diag_offset[row]; 787 pj = bj + diag_offset[row] + 1; 788 multiplier = *pc * *pv++; 789 *pc = multiplier; 790 nz = bi[row+1] - diag_offset[row] - 1; 791 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 792 PLogFlops(1+2*nz); 793 } 794 row = *ajtmp++; 795 } 796 /* finished row so stick it into b->a */ 797 pv = ba + bi[i]; 798 pj = bj + bi[i]; 799 nz = bi[i+1] - bi[i]; 800 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 801 diag = diag_offset[i] - bi[i]; 802 /* check pivot entry for current row */ 803 if (pv[diag] == 0.0) { 804 SETERRQ(1,"MatLUFactorNumeric_SeqAIJ:Zero pivot"); 805 } 806 pv[diag] = 1.0/pv[diag]; 807 } 808 809 PetscFree(rtmp); 810 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 811 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 812 ierr = ISDestroy(isicol); CHKERRQ(ierr); 813 C->factor = FACTOR_LU; 814 C->assembled = PETSC_TRUE; 815 PLogFlops(b->n); 816 return 0; 817 } 818 819 /* ----------------------------------------------------------- */ 820 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 821 { 822 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 823 int ierr; 824 Mat C; 825 826 ierr = MatLUFactorSymbolic_SeqBAIJ(A,row,col,f,&C); CHKERRQ(ierr); 827 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 828 829 /* free all the data structures from mat */ 830 PetscFree(mat->a); 831 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 832 if (mat->diag) PetscFree(mat->diag); 833 if (mat->ilen) PetscFree(mat->ilen); 834 if (mat->imax) PetscFree(mat->imax); 835 if (mat->solve_work) PetscFree(mat->solve_work); 836 if (mat->mult_work) PetscFree(mat->mult_work); 837 PetscFree(mat); 838 839 PetscMemcpy(A,C,sizeof(struct _Mat)); 840 PetscHeaderDestroy(C); 841 return 0; 842 } 843 /* ----------------------------------------------------------- */ 844 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 845 { 846 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 847 IS iscol=a->col,isrow=a->row; 848 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 849 int nz,bs=a->bs,bs2=a->bs2; 850 Scalar *xa,*ba,*aa=a->a,*sum; 851 register Scalar *x,*b,*lsum,*tmp,*v; 852 853 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 854 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 855 tmp = a->solve_work; 856 857 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 858 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 859 860 /* forward solve the lower triangular */ 861 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 862 for ( i=1; i<n; i++ ) { 863 v = aa + bs2*ai[i]; 864 vi = aj + ai[i]; 865 nz = a->diag[i] - ai[i]; 866 sum = tmp + bs*i; 867 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 868 while (nz--) { 869 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 870 v += bs2; 871 } 872 } 873 /* backward solve the upper triangular */ 874 lsum = a->solve_work + a->n; 875 for ( i=n-1; i>=0; i-- ){ 876 v = aa + bs2*(a->diag[i] + 1); 877 vi = aj + a->diag[i] + 1; 878 nz = ai[i+1] - a->diag[i] - 1; 879 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 880 while (nz--) { 881 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 882 v += bs2; 883 } 884 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 885 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 886 } 887 888 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 889 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 890 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 891 return 0; 892 } 893 894 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 895 { 896 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 897 IS iscol=a->col,isrow=a->row; 898 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 899 Scalar *xa,*ba,*aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 900 register Scalar *x,*b,*tmp,*v; 901 902 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 903 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 904 tmp = a->solve_work; 905 906 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 907 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 908 909 /* forward solve the lower triangular */ 910 idx = 5*(*r++); 911 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 912 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 913 for ( i=1; i<n; i++ ) { 914 v = aa + 25*ai[i]; 915 vi = aj + ai[i]; 916 nz = a->diag[i] - ai[i]; 917 idx = 5*(*r++); 918 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 919 sum5 = b[4+idx]; 920 while (nz--) { 921 idx = 5*(*vi++); 922 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 923 x4 = tmp[3+idx];x5 = tmp[4+idx]; 924 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 925 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 926 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 927 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 928 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 929 v += 25; 930 } 931 idx = 5*i; 932 tmp[idx] = sum1;tmp[1+idx] = sum2; 933 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 934 } 935 /* backward solve the upper triangular */ 936 for ( i=n-1; i>=0; i-- ){ 937 v = aa + 25*a->diag[i] + 25; 938 vi = aj + a->diag[i] + 1; 939 nz = ai[i+1] - a->diag[i] - 1; 940 idt = 5*i; 941 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 942 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 943 while (nz--) { 944 idx = 5*(*vi++); 945 x1 = tmp[idx]; x2 = tmp[1+idx]; 946 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 947 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 948 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 949 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 950 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 951 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 952 v += 25; 953 } 954 idc = 5*(*c--); 955 v = aa + 25*a->diag[i]; 956 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 957 v[15]*sum4+v[20]*sum5; 958 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 959 v[16]*sum4+v[21]*sum5; 960 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 961 v[17]*sum4+v[22]*sum5; 962 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 963 v[18]*sum4+v[23]*sum5; 964 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 965 v[19]*sum4+v[24]*sum5; 966 } 967 968 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 969 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 970 PLogFlops(2*25*(a->nz) - a->n); 971 return 0; 972 } 973 974 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 975 { 976 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 977 IS iscol=a->col,isrow=a->row; 978 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 979 Scalar *xa,*ba,*aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 980 register Scalar *x,*b,*tmp,*v; 981 982 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 983 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 984 tmp = a->solve_work; 985 986 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 987 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 988 989 /* forward solve the lower triangular */ 990 idx = 4*(*r++); 991 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 992 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 993 for ( i=1; i<n; i++ ) { 994 v = aa + 16*ai[i]; 995 vi = aj + ai[i]; 996 nz = a->diag[i] - ai[i]; 997 idx = 4*(*r++); 998 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 999 while (nz--) { 1000 idx = 4*(*vi++); 1001 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1002 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1003 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1004 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1005 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1006 v += 16; 1007 } 1008 idx = 4*i; 1009 tmp[idx] = sum1;tmp[1+idx] = sum2; 1010 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1011 } 1012 /* backward solve the upper triangular */ 1013 for ( i=n-1; i>=0; i-- ){ 1014 v = aa + 16*a->diag[i] + 16; 1015 vi = aj + a->diag[i] + 1; 1016 nz = ai[i+1] - a->diag[i] - 1; 1017 idt = 4*i; 1018 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1019 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1020 while (nz--) { 1021 idx = 4*(*vi++); 1022 x1 = tmp[idx]; x2 = tmp[1+idx]; 1023 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1024 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1025 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1026 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1027 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1028 v += 16; 1029 } 1030 idc = 4*(*c--); 1031 v = aa + 16*a->diag[i]; 1032 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1033 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1034 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1035 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1036 } 1037 1038 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1039 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1040 PLogFlops(2*16*(a->nz) - a->n); 1041 return 0; 1042 } 1043 1044 1045 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1046 { 1047 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1048 IS iscol=a->col,isrow=a->row; 1049 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 1050 Scalar *xa,*ba,*aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1051 register Scalar *x,*b,*tmp,*v; 1052 1053 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1054 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1055 tmp = a->solve_work; 1056 1057 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1058 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1059 1060 /* forward solve the lower triangular */ 1061 idx = 3*(*r++); 1062 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1063 for ( i=1; i<n; i++ ) { 1064 v = aa + 9*ai[i]; 1065 vi = aj + ai[i]; 1066 nz = a->diag[i] - ai[i]; 1067 idx = 3*(*r++); 1068 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1069 while (nz--) { 1070 idx = 3*(*vi++); 1071 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1072 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1073 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1074 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1075 v += 9; 1076 } 1077 idx = 3*i; 1078 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1079 } 1080 /* backward solve the upper triangular */ 1081 for ( i=n-1; i>=0; i-- ){ 1082 v = aa + 9*a->diag[i] + 9; 1083 vi = aj + a->diag[i] + 1; 1084 nz = ai[i+1] - a->diag[i] - 1; 1085 idt = 3*i; 1086 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1087 while (nz--) { 1088 idx = 3*(*vi++); 1089 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1090 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1091 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1092 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1093 v += 9; 1094 } 1095 idc = 3*(*c--); 1096 v = aa + 9*a->diag[i]; 1097 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1098 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1099 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1100 } 1101 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1102 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1103 PLogFlops(2*9*(a->nz) - a->n); 1104 return 0; 1105 } 1106 1107 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1108 { 1109 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1110 IS iscol=a->col,isrow=a->row; 1111 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc; 1112 Scalar *xa,*ba,*aa=a->a,sum1,sum2,x1,x2; 1113 register Scalar *x,*b,*tmp,*v; 1114 1115 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1116 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1117 tmp = a->solve_work; 1118 1119 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1120 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1121 1122 /* forward solve the lower triangular */ 1123 idx = 2*(*r++); 1124 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1125 for ( i=1; i<n; i++ ) { 1126 v = aa + 4*ai[i]; 1127 vi = aj + ai[i]; 1128 nz = a->diag[i] - ai[i]; 1129 idx = 2*(*r++); 1130 sum1 = b[idx]; sum2 = b[1+idx]; 1131 while (nz--) { 1132 idx = 2*(*vi++); 1133 x1 = tmp[idx]; x2 = tmp[1+idx]; 1134 sum1 -= v[0]*x1 + v[2]*x2; 1135 sum2 -= v[1]*x1 + v[3]*x2; 1136 v += 4; 1137 } 1138 idx = 2*i; 1139 tmp[idx] = sum1; tmp[1+idx] = sum2; 1140 } 1141 /* backward solve the upper triangular */ 1142 for ( i=n-1; i>=0; i-- ){ 1143 v = aa + 4*a->diag[i] + 4; 1144 vi = aj + a->diag[i] + 1; 1145 nz = ai[i+1] - a->diag[i] - 1; 1146 idt = 2*i; 1147 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1148 while (nz--) { 1149 idx = 2*(*vi++); 1150 x1 = tmp[idx]; x2 = tmp[1+idx]; 1151 sum1 -= v[0]*x1 + v[2]*x2; 1152 sum2 -= v[1]*x1 + v[3]*x2; 1153 v += 4; 1154 } 1155 idc = 2*(*c--); 1156 v = aa + 4*a->diag[i]; 1157 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1158 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1159 } 1160 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1161 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1162 PLogFlops(2*4*(a->nz) - a->n); 1163 return 0; 1164 } 1165 1166 1167 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1168 { 1169 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1170 IS iscol=a->col,isrow=a->row; 1171 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz; 1172 Scalar *xa,*ba,*aa=a->a,sum1; 1173 register Scalar *x,*b,*tmp,*v; 1174 1175 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1176 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1177 tmp = a->solve_work; 1178 1179 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1180 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1181 1182 /* forward solve the lower triangular */ 1183 tmp[0] = b[*r++]; 1184 for ( i=1; i<n; i++ ) { 1185 v = aa + ai[i]; 1186 vi = aj + ai[i]; 1187 nz = a->diag[i] - ai[i]; 1188 sum1 = b[*r++]; 1189 while (nz--) { 1190 sum1 -= (*v++)*tmp[*vi++]; 1191 } 1192 tmp[i] = sum1; 1193 } 1194 /* backward solve the upper triangular */ 1195 for ( i=n-1; i>=0; i-- ){ 1196 v = aa + a->diag[i] + 1; 1197 vi = aj + a->diag[i] + 1; 1198 nz = ai[i+1] - a->diag[i] - 1; 1199 sum1 = tmp[i]; 1200 while (nz--) { 1201 sum1 -= (*v++)*tmp[*vi++]; 1202 } 1203 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1204 } 1205 1206 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1207 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1208 PLogFlops(2*1*(a->nz) - a->n); 1209 return 0; 1210 } 1211 1212 /* ----------------------------------------------------------------*/ 1213 /* 1214 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1215 except that the data structure of Mat_SeqAIJ is slightly different. 1216 Not a good example of code reuse. 1217 */ 1218 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1219 Mat *fact) 1220 { 1221 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1222 IS isicol; 1223 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1224 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1225 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1226 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1227 PetscTruth col_identity, row_identity; 1228 1229 if (a->m != a->n) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Matrix must be square"); 1230 if (!isrow) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have row permutation"); 1231 if (!iscol) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have column permutation"); 1232 1233 /* special case that simply copies fill pattern */ 1234 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1235 if (levels == 0 && row_identity && col_identity) { 1236 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1237 (*fact)->factor = FACTOR_LU; 1238 b = (Mat_SeqBAIJ *) (*fact)->data; 1239 if (!b->diag) { 1240 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1241 } 1242 b->row = isrow; 1243 b->col = iscol; 1244 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1245 return 0; 1246 } 1247 1248 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1249 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1250 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1251 1252 /* get new row pointers */ 1253 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1254 ainew[0] = 0; 1255 /* don't know how many column pointers are needed so estimate */ 1256 jmax = (int) (f*ai[n] + 1); 1257 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1258 /* ajfill is level of fill for each fill entry */ 1259 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1260 /* fill is a linked list of nonzeros in active row */ 1261 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1262 /* im is level for each filled value */ 1263 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1264 /* dloc is location of diagonal in factor */ 1265 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1266 dloc[0] = 0; 1267 for ( prow=0; prow<n; prow++ ) { 1268 /* first copy previous fill into linked list */ 1269 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1270 xi = aj + ai[r[prow]]; 1271 fill[n] = n; 1272 while (nz--) { 1273 fm = n; 1274 idx = ic[*xi++]; 1275 do { 1276 m = fm; 1277 fm = fill[m]; 1278 } while (fm < idx); 1279 fill[m] = idx; 1280 fill[idx] = fm; 1281 im[idx] = 0; 1282 } 1283 nzi = 0; 1284 row = fill[n]; 1285 while ( row < prow ) { 1286 incrlev = im[row] + 1; 1287 nz = dloc[row]; 1288 xi = ajnew + ainew[row] + nz; 1289 flev = ajfill + ainew[row] + nz + 1; 1290 nnz = ainew[row+1] - ainew[row] - nz - 1; 1291 if (*xi++ != row) { 1292 SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:zero pivot"); 1293 } 1294 fm = row; 1295 while (nnz-- > 0) { 1296 idx = *xi++; 1297 if (*flev + incrlev > levels) { 1298 flev++; 1299 continue; 1300 } 1301 do { 1302 m = fm; 1303 fm = fill[m]; 1304 } while (fm < idx); 1305 if (fm != idx) { 1306 im[idx] = *flev + incrlev; 1307 fill[m] = idx; 1308 fill[idx] = fm; 1309 fm = idx; 1310 nzf++; 1311 } 1312 else { 1313 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1314 } 1315 flev++; 1316 } 1317 row = fill[row]; 1318 nzi++; 1319 } 1320 /* copy new filled row into permanent storage */ 1321 ainew[prow+1] = ainew[prow] + nzf; 1322 if (ainew[prow+1] > jmax) { 1323 /* allocate a longer ajnew */ 1324 int maxadd; 1325 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1326 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1327 jmax += maxadd; 1328 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1329 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1330 PetscFree(ajnew); 1331 ajnew = xi; 1332 /* allocate a longer ajfill */ 1333 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1334 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1335 PetscFree(ajfill); 1336 ajfill = xi; 1337 realloc++; 1338 } 1339 xi = ajnew + ainew[prow]; 1340 flev = ajfill + ainew[prow]; 1341 dloc[prow] = nzi; 1342 fm = fill[n]; 1343 while (nzf--) { 1344 *xi++ = fm; 1345 *flev++ = im[fm]; 1346 fm = fill[fm]; 1347 } 1348 } 1349 PetscFree(ajfill); 1350 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1351 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1352 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1353 PetscFree(fill); PetscFree(im); 1354 1355 PLogInfo(A, 1356 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1357 realloc,f,((double)ainew[n])/((double)ai[prow])); 1358 1359 /* put together the new matrix */ 1360 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1361 b = (Mat_SeqBAIJ *) (*fact)->data; 1362 PetscFree(b->imax); 1363 b->singlemalloc = 0; 1364 len = bs2*ainew[n]*sizeof(Scalar); 1365 /* the next line frees the default space generated by the Create() */ 1366 PetscFree(b->a); PetscFree(b->ilen); 1367 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1368 b->j = ajnew; 1369 b->i = ainew; 1370 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1371 b->diag = dloc; 1372 b->ilen = 0; 1373 b->imax = 0; 1374 b->row = isrow; 1375 b->col = iscol; 1376 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1377 CHKPTRQ(b->solve_work); 1378 /* In b structure: Free imax, ilen, old a, old j. 1379 Allocate dloc, solve_work, new a, new j */ 1380 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1381 b->maxnz = b->nz = ainew[n]; 1382 (*fact)->factor = FACTOR_LU; 1383 return 0; 1384 } 1385 1386 1387 1388 1389