1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.37 1997/01/10 23:54:46 balay 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 __FUNC__ 19 #define __FUNC__ "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 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 50 ajtmp = aj + ai[r[i]]; 51 fill[n] = n; 52 while (nz--) { 53 fm = n; 54 idx = ic[*ajtmp++]; 55 do { 56 m = fm; 57 fm = fill[m]; 58 } while (fm < idx); 59 fill[m] = idx; 60 fill[idx] = fm; 61 } 62 row = fill[n]; 63 while ( row < i ) { 64 ajtmp = ajnew + idnew[row] + 1; 65 nzbd = 1 + idnew[row] - ainew[row]; 66 nz = im[row] - nzbd; 67 fm = row; 68 while (nz-- > 0) { 69 idx = *ajtmp++; 70 nzbd++; 71 if (idx == i) im[row] = nzbd; 72 do { 73 m = fm; 74 fm = fill[m]; 75 } while (fm < idx); 76 if (fm != idx) { 77 fill[m] = idx; 78 fill[idx] = fm; 79 fm = idx; 80 nnz++; 81 } 82 } 83 row = fill[row]; 84 } 85 /* copy new filled row into permanent storage */ 86 ainew[i+1] = ainew[i] + nnz; 87 if (ainew[i+1] > jmax) { 88 /* allocate a longer ajnew */ 89 int maxadd; 90 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 91 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 92 jmax += maxadd; 93 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 94 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 95 PetscFree(ajnew); 96 ajnew = ajtmp; 97 realloc++; /* count how many times we realloc */ 98 } 99 ajtmp = ajnew + ainew[i]; 100 fm = fill[n]; 101 nzi = 0; 102 im[i] = nnz; 103 while (nnz--) { 104 if (fm < i) nzi++; 105 *ajtmp++ = fm; 106 fm = fill[fm]; 107 } 108 idnew[i] = ainew[i] + nzi; 109 } 110 111 PLogInfo(A, 112 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 113 realloc,f,((double)ainew[n])/((double)ai[i])); 114 115 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 116 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 117 118 PetscFree(fill); 119 120 /* put together the new matrix */ 121 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 122 PLogObjectParent(*B,isicol); 123 ierr = ISDestroy(isicol); CHKERRQ(ierr); 124 b = (Mat_SeqBAIJ *) (*B)->data; 125 PetscFree(b->imax); 126 b->singlemalloc = 0; 127 /* the next line frees the default space generated by the Create() */ 128 PetscFree(b->a); PetscFree(b->ilen); 129 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 130 b->j = ajnew; 131 b->i = ainew; 132 b->diag = idnew; 133 b->ilen = 0; 134 b->imax = 0; 135 b->row = isrow; 136 b->col = iscol; 137 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 138 CHKPTRQ(b->solve_work); 139 /* In b structure: Free imax, ilen, old a, old j. 140 Allocate idnew, solve_work, new a, new j */ 141 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 142 b->maxnz = b->nz = ainew[n]; 143 144 (*B)->info.factor_mallocs = realloc; 145 (*B)->info.fill_ratio_given = f; 146 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 147 148 return 0; 149 } 150 151 /* ----------------------------------------------------------- */ 152 #undef __FUNC__ 153 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 154 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 155 { 156 Mat C = *B; 157 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 158 IS iscol = b->col, isrow = b->row, isicol; 159 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 160 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j; 161 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = bs*bs,*v_pivots; 162 Scalar *ba = b->a,*aa = a->a; 163 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 164 register int *pj; 165 166 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 167 PLogObjectParent(*B,isicol); 168 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 169 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 170 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 171 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 172 /* generate work space needed by dense LU factorization */ 173 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 174 CHKPTRQ(v_work); 175 multiplier = v_work + bs; 176 v_pivots = (int *) (multiplier + bs2); 177 178 /* flops in while loop */ 179 bslog = 2*bs*bs2; 180 181 for ( i=0; i<n; i++ ) { 182 nz = bi[i+1] - bi[i]; 183 ajtmp = bj + bi[i]; 184 for ( j=0; j<nz; j++ ) { 185 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 186 } 187 /* load in initial (unfactored row) */ 188 nz = ai[r[i]+1] - ai[r[i]]; 189 ajtmpold = aj + ai[r[i]]; 190 v = aa + bs2*ai[r[i]]; 191 for ( j=0; j<nz; j++ ) { 192 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 193 } 194 row = *ajtmp++; 195 while (row < i) { 196 pc = rtmp + bs2*row; 197 /* if (*pc) { */ 198 pv = ba + bs2*diag_offset[row]; 199 pj = bj + diag_offset[row] + 1; 200 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 201 nz = bi[row+1] - diag_offset[row] - 1; 202 pv += bs2; 203 for (j=0; j<nz; j++) { 204 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 205 } 206 PLogFlops(bslog*(nz+1)-bs); 207 /* } */ 208 row = *ajtmp++; 209 } 210 /* finished row so stick it into b->a */ 211 pv = ba + bs2*bi[i]; 212 pj = bj + bi[i]; 213 nz = bi[i+1] - bi[i]; 214 for ( j=0; j<nz; j++ ) { 215 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 216 } 217 diag = diag_offset[i] - bi[i]; 218 /* invert diagonal block */ 219 w = pv + bs2*diag; 220 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 221 } 222 223 PetscFree(rtmp); PetscFree(v_work); 224 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 225 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 226 ierr = ISDestroy(isicol); CHKERRQ(ierr); 227 C->factor = FACTOR_LU; 228 C->assembled = PETSC_TRUE; 229 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 230 return 0; 231 } 232 /* ------------------------------------------------------------*/ 233 /* 234 Version for when blocks are 5 by 5 235 */ 236 #undef __FUNC__ 237 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 238 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 239 { 240 Mat C = *B; 241 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 242 IS iscol = b->col, isrow = b->row, isicol; 243 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 244 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 245 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 246 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 247 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 248 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 249 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 250 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 251 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 252 Scalar *ba = b->a,*aa = a->a; 253 register int *pj; 254 255 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 256 PLogObjectParent(*B,isicol); 257 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 258 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 259 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 260 261 for ( i=0; i<n; i++ ) { 262 nz = bi[i+1] - bi[i]; 263 ajtmp = bj + bi[i]; 264 for ( j=0; j<nz; j++ ) { 265 x = rtmp+25*ajtmp[j]; 266 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 267 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 268 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 269 } 270 /* load in initial (unfactored row) */ 271 idx = r[i]; 272 nz = ai[idx+1] - ai[idx]; 273 ajtmpold = aj + ai[idx]; 274 v = aa + 25*ai[idx]; 275 for ( j=0; j<nz; j++ ) { 276 x = rtmp+25*ic[ajtmpold[j]]; 277 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 278 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 279 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 280 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 281 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 282 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 283 v += 25; 284 } 285 row = *ajtmp++; 286 while (row < i) { 287 pc = rtmp + 25*row; 288 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 289 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 290 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 291 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 292 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 293 p25 = pc[24]; 294 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 295 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 296 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 297 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 298 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 299 p24 != 0.0 || p25 != 0.0) { 300 pv = ba + 25*diag_offset[row]; 301 pj = bj + diag_offset[row] + 1; 302 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 303 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 304 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 305 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 306 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 307 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 308 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 309 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 310 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 311 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 312 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 313 314 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 315 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 316 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 317 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 318 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 319 320 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 321 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 322 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 323 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 324 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 325 326 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 327 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 328 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 329 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 330 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 331 332 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 333 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 334 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 335 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 336 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 337 338 nz = bi[row+1] - diag_offset[row] - 1; 339 pv += 25; 340 for (j=0; j<nz; j++) { 341 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 342 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 343 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 344 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 345 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 346 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 347 x = rtmp + 25*pj[j]; 348 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 349 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 350 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 351 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 352 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 353 354 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 355 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 356 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 357 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 358 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 359 360 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 361 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 362 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 363 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 364 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 365 366 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 367 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 368 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 369 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 370 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 371 372 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 373 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 374 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 375 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 376 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 377 378 pv += 25; 379 } 380 PLogFlops(250*nz+225); 381 } 382 row = *ajtmp++; 383 } 384 /* finished row so stick it into b->a */ 385 pv = ba + 25*bi[i]; 386 pj = bj + bi[i]; 387 nz = bi[i+1] - bi[i]; 388 for ( j=0; j<nz; j++ ) { 389 x = rtmp+25*pj[j]; 390 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 391 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 392 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 393 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 394 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 395 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 396 pv += 25; 397 } 398 /* invert diagonal block */ 399 w = ba + 25*diag_offset[i]; 400 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 401 } 402 403 PetscFree(rtmp); 404 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 405 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 406 ierr = ISDestroy(isicol); CHKERRQ(ierr); 407 C->factor = FACTOR_LU; 408 C->assembled = PETSC_TRUE; 409 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 410 return 0; 411 } 412 413 /* ------------------------------------------------------------*/ 414 /* 415 Version for when blocks are 4 by 4 416 */ 417 #undef __FUNC__ 418 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 419 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 420 { 421 Mat C = *B; 422 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 423 IS iscol = b->col, isrow = b->row, isicol; 424 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 425 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 426 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 427 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 428 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 429 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 430 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 431 Scalar m13,m14,m15,m16; 432 Scalar *ba = b->a,*aa = a->a; 433 register int *pj; 434 435 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 436 PLogObjectParent(*B,isicol); 437 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 438 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 439 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 440 441 for ( i=0; i<n; i++ ) { 442 nz = bi[i+1] - bi[i]; 443 ajtmp = bj + bi[i]; 444 for ( j=0; j<nz; j++ ) { 445 x = rtmp+16*ajtmp[j]; 446 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 447 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 448 } 449 /* load in initial (unfactored row) */ 450 idx = r[i]; 451 nz = ai[idx+1] - ai[idx]; 452 ajtmpold = aj + ai[idx]; 453 v = aa + 16*ai[idx]; 454 for ( j=0; j<nz; j++ ) { 455 x = rtmp+16*ic[ajtmpold[j]]; 456 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 457 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 458 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 459 x[14] = v[14]; x[15] = v[15]; 460 v += 16; 461 } 462 row = *ajtmp++; 463 while (row < i) { 464 pc = rtmp + 16*row; 465 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 466 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 467 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 468 p15 = pc[14]; p16 = pc[15]; 469 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 470 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 471 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 472 || p16 != 0.0) { 473 pv = ba + 16*diag_offset[row]; 474 pj = bj + diag_offset[row] + 1; 475 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 476 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 477 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 478 x15 = pv[14]; x16 = pv[15]; 479 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 480 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 481 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 482 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 483 484 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 485 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 486 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 487 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 488 489 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 490 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 491 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 492 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 493 494 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 495 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 496 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 497 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 498 499 nz = bi[row+1] - diag_offset[row] - 1; 500 pv += 16; 501 for (j=0; j<nz; j++) { 502 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 503 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 504 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 505 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 506 x = rtmp + 16*pj[j]; 507 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 508 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 509 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 510 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 511 512 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 513 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 514 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 515 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 516 517 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 518 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 519 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 520 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 521 522 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 523 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 524 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 525 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 526 527 pv += 16; 528 } 529 PLogFlops(128*nz+112); 530 } 531 row = *ajtmp++; 532 } 533 /* finished row so stick it into b->a */ 534 pv = ba + 16*bi[i]; 535 pj = bj + bi[i]; 536 nz = bi[i+1] - bi[i]; 537 for ( j=0; j<nz; j++ ) { 538 x = rtmp+16*pj[j]; 539 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 540 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 541 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 542 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 543 pv += 16; 544 } 545 /* invert diagonal block */ 546 w = ba + 16*diag_offset[i]; 547 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 548 } 549 550 PetscFree(rtmp); 551 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 552 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 553 ierr = ISDestroy(isicol); CHKERRQ(ierr); 554 C->factor = FACTOR_LU; 555 C->assembled = PETSC_TRUE; 556 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 557 return 0; 558 } 559 /* ------------------------------------------------------------*/ 560 /* 561 Version for when blocks are 3 by 3 562 */ 563 #undef __FUNC__ 564 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 565 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 566 { 567 Mat C = *B; 568 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 569 IS iscol = b->col, isrow = b->row, isicol; 570 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 571 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 572 int *diag_offset = b->diag,idx; 573 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 574 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 575 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 576 Scalar *ba = b->a,*aa = a->a; 577 register int *pj; 578 579 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 580 PLogObjectParent(*B,isicol); 581 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 582 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 583 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 584 585 for ( i=0; i<n; i++ ) { 586 nz = bi[i+1] - bi[i]; 587 ajtmp = bj + bi[i]; 588 for ( j=0; j<nz; j++ ) { 589 x = rtmp + 9*ajtmp[j]; 590 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 591 } 592 /* load in initial (unfactored row) */ 593 idx = r[i]; 594 nz = ai[idx+1] - ai[idx]; 595 ajtmpold = aj + ai[idx]; 596 v = aa + 9*ai[idx]; 597 for ( j=0; j<nz; j++ ) { 598 x = rtmp + 9*ic[ajtmpold[j]]; 599 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 600 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 601 v += 9; 602 } 603 row = *ajtmp++; 604 while (row < i) { 605 pc = rtmp + 9*row; 606 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 607 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 608 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 609 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 610 pv = ba + 9*diag_offset[row]; 611 pj = bj + diag_offset[row] + 1; 612 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 613 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 614 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 615 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 616 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 617 618 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 619 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 620 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 621 622 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 623 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 624 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 625 nz = bi[row+1] - diag_offset[row] - 1; 626 pv += 9; 627 for (j=0; j<nz; j++) { 628 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 629 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 630 x = rtmp + 9*pj[j]; 631 x[0] -= m1*x1 + m4*x2 + m7*x3; 632 x[1] -= m2*x1 + m5*x2 + m8*x3; 633 x[2] -= m3*x1 + m6*x2 + m9*x3; 634 635 x[3] -= m1*x4 + m4*x5 + m7*x6; 636 x[4] -= m2*x4 + m5*x5 + m8*x6; 637 x[5] -= m3*x4 + m6*x5 + m9*x6; 638 639 x[6] -= m1*x7 + m4*x8 + m7*x9; 640 x[7] -= m2*x7 + m5*x8 + m8*x9; 641 x[8] -= m3*x7 + m6*x8 + m9*x9; 642 pv += 9; 643 } 644 PLogFlops(54*nz+36); 645 } 646 row = *ajtmp++; 647 } 648 /* finished row so stick it into b->a */ 649 pv = ba + 9*bi[i]; 650 pj = bj + bi[i]; 651 nz = bi[i+1] - bi[i]; 652 for ( j=0; j<nz; j++ ) { 653 x = rtmp + 9*pj[j]; 654 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 655 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 656 pv += 9; 657 } 658 /* invert diagonal block */ 659 w = ba + 9*diag_offset[i]; 660 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 661 } 662 663 PetscFree(rtmp); 664 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 665 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 666 ierr = ISDestroy(isicol); CHKERRQ(ierr); 667 C->factor = FACTOR_LU; 668 C->assembled = PETSC_TRUE; 669 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 670 return 0; 671 } 672 673 /* ------------------------------------------------------------*/ 674 /* 675 Version for when blocks are 2 by 2 676 */ 677 #undef __FUNC__ 678 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 679 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 680 { 681 Mat C = *B; 682 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 683 IS iscol = b->col, isrow = b->row, isicol; 684 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 685 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 686 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 687 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 688 Scalar p1,p2,p3,p4,v_work[2]; 689 Scalar *ba = b->a,*aa = a->a; 690 register int *pj; 691 692 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 693 PLogObjectParent(*B,isicol); 694 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 695 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 696 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 697 698 for ( i=0; i<n; i++ ) { 699 nz = bi[i+1] - bi[i]; 700 ajtmp = bj + bi[i]; 701 for ( j=0; j<nz; j++ ) { 702 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 703 } 704 /* load in initial (unfactored row) */ 705 idx = r[i]; 706 nz = ai[idx+1] - ai[idx]; 707 ajtmpold = aj + ai[idx]; 708 v = aa + 4*ai[idx]; 709 for ( j=0; j<nz; j++ ) { 710 x = rtmp+4*ic[ajtmpold[j]]; 711 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 712 v += 4; 713 } 714 row = *ajtmp++; 715 while (row < i) { 716 pc = rtmp + 4*row; 717 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 718 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 719 pv = ba + 4*diag_offset[row]; 720 pj = bj + diag_offset[row] + 1; 721 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 722 pc[0] = m1 = p1*x1 + p3*x2; 723 pc[1] = m2 = p2*x1 + p4*x2; 724 pc[2] = m3 = p1*x3 + p3*x4; 725 pc[3] = m4 = p2*x3 + p4*x4; 726 nz = bi[row+1] - diag_offset[row] - 1; 727 pv += 4; 728 for (j=0; j<nz; j++) { 729 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 730 x = rtmp + 4*pj[j]; 731 x[0] -= m1*x1 + m3*x2; 732 x[1] -= m2*x1 + m4*x2; 733 x[2] -= m1*x3 + m3*x4; 734 x[3] -= m2*x3 + m4*x4; 735 pv += 4; 736 } 737 PLogFlops(16*nz+12); 738 } 739 row = *ajtmp++; 740 } 741 /* finished row so stick it into b->a */ 742 pv = ba + 4*bi[i]; 743 pj = bj + bi[i]; 744 nz = bi[i+1] - bi[i]; 745 for ( j=0; j<nz; j++ ) { 746 x = rtmp+4*pj[j]; 747 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 748 pv += 4; 749 } 750 /* invert diagonal block */ 751 w = ba + 4*diag_offset[i]; 752 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 753 } 754 755 PetscFree(rtmp); 756 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 757 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 758 ierr = ISDestroy(isicol); CHKERRQ(ierr); 759 C->factor = FACTOR_LU; 760 C->assembled = PETSC_TRUE; 761 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 762 return 0; 763 } 764 765 /* ----------------------------------------------------------- */ 766 /* 767 Version for when blocks are 1 by 1. 768 */ 769 #undef __FUNC__ 770 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 771 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 772 { 773 Mat C = *B; 774 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 775 IS iscol = b->col, isrow = b->row, isicol; 776 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 777 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 778 int *diag_offset = b->diag,diag; 779 register Scalar *pv,*v,*rtmp,multiplier,*pc; 780 Scalar *ba = b->a,*aa = a->a; 781 register int *pj; 782 783 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 784 PLogObjectParent(*B,isicol); 785 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 786 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 787 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 788 789 for ( i=0; i<n; i++ ) { 790 nz = bi[i+1] - bi[i]; 791 ajtmp = bj + bi[i]; 792 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 793 794 /* load in initial (unfactored row) */ 795 nz = ai[r[i]+1] - ai[r[i]]; 796 ajtmpold = aj + ai[r[i]]; 797 v = aa + ai[r[i]]; 798 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 799 800 row = *ajtmp++; 801 while (row < i) { 802 pc = rtmp + row; 803 if (*pc != 0.0) { 804 pv = ba + diag_offset[row]; 805 pj = bj + diag_offset[row] + 1; 806 multiplier = *pc * *pv++; 807 *pc = multiplier; 808 nz = bi[row+1] - diag_offset[row] - 1; 809 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 810 PLogFlops(1+2*nz); 811 } 812 row = *ajtmp++; 813 } 814 /* finished row so stick it into b->a */ 815 pv = ba + bi[i]; 816 pj = bj + bi[i]; 817 nz = bi[i+1] - bi[i]; 818 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 819 diag = diag_offset[i] - bi[i]; 820 /* check pivot entry for current row */ 821 if (pv[diag] == 0.0) { 822 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 823 } 824 pv[diag] = 1.0/pv[diag]; 825 } 826 827 PetscFree(rtmp); 828 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 829 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 830 ierr = ISDestroy(isicol); CHKERRQ(ierr); 831 C->factor = FACTOR_LU; 832 C->assembled = PETSC_TRUE; 833 PLogFlops(b->n); 834 return 0; 835 } 836 837 /* ----------------------------------------------------------- */ 838 #undef __FUNC__ 839 #define __FUNC__ "MatLUFactor_SeqBAIJ" 840 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 841 { 842 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 843 int ierr; 844 Mat C; 845 846 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 847 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 848 849 /* free all the data structures from mat */ 850 PetscFree(mat->a); 851 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 852 if (mat->diag) PetscFree(mat->diag); 853 if (mat->ilen) PetscFree(mat->ilen); 854 if (mat->imax) PetscFree(mat->imax); 855 if (mat->solve_work) PetscFree(mat->solve_work); 856 if (mat->mult_work) PetscFree(mat->mult_work); 857 PetscFree(mat); 858 859 PetscMemcpy(A,C,sizeof(struct _Mat)); 860 PetscHeaderDestroy(C); 861 return 0; 862 } 863 /* ----------------------------------------------------------- */ 864 #undef __FUNC__ 865 #define __FUNC__ "MatSolve_SeqBAIJ_N" 866 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 867 { 868 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 869 IS iscol=a->col,isrow=a->row; 870 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 871 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 872 Scalar *aa=a->a,*sum; 873 register Scalar *x,*b,*lsum,*tmp,*v; 874 875 VecGetArray_Fast(bb,b); 876 VecGetArray_Fast(xx,x); 877 tmp = a->solve_work; 878 879 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 880 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 881 882 /* forward solve the lower triangular */ 883 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 884 for ( i=1; i<n; i++ ) { 885 v = aa + bs2*ai[i]; 886 vi = aj + ai[i]; 887 nz = a->diag[i] - ai[i]; 888 sum = tmp + bs*i; 889 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 890 while (nz--) { 891 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 892 v += bs2; 893 } 894 } 895 /* backward solve the upper triangular */ 896 lsum = a->solve_work + a->n; 897 for ( i=n-1; i>=0; i-- ){ 898 v = aa + bs2*(a->diag[i] + 1); 899 vi = aj + a->diag[i] + 1; 900 nz = ai[i+1] - a->diag[i] - 1; 901 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 902 while (nz--) { 903 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 904 v += bs2; 905 } 906 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 907 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 908 } 909 910 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 911 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 912 VecRestoreArray_Fast(bb,b); 913 VecRestoreArray_Fast(xx,x); 914 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 915 return 0; 916 } 917 918 #undef __FUNC__ 919 #define __FUNC__ "MatSolve_SeqBAIJ_5" 920 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 921 { 922 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 923 IS iscol=a->col,isrow=a->row; 924 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 925 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 926 register Scalar *x,*b,*tmp,*v; 927 928 VecGetArray_Fast(bb,b); 929 VecGetArray_Fast(xx,x); 930 tmp = a->solve_work; 931 932 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 933 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 934 935 /* forward solve the lower triangular */ 936 idx = 5*(*r++); 937 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 938 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 939 for ( i=1; i<n; i++ ) { 940 v = aa + 25*ai[i]; 941 vi = aj + ai[i]; 942 nz = a->diag[i] - ai[i]; 943 idx = 5*(*r++); 944 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 945 sum5 = b[4+idx]; 946 while (nz--) { 947 idx = 5*(*vi++); 948 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 949 x4 = tmp[3+idx];x5 = tmp[4+idx]; 950 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 951 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 952 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 953 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 954 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 955 v += 25; 956 } 957 idx = 5*i; 958 tmp[idx] = sum1;tmp[1+idx] = sum2; 959 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 960 } 961 /* backward solve the upper triangular */ 962 for ( i=n-1; i>=0; i-- ){ 963 v = aa + 25*a->diag[i] + 25; 964 vi = aj + a->diag[i] + 1; 965 nz = ai[i+1] - a->diag[i] - 1; 966 idt = 5*i; 967 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 968 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 969 while (nz--) { 970 idx = 5*(*vi++); 971 x1 = tmp[idx]; x2 = tmp[1+idx]; 972 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 973 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 974 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 975 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 976 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 977 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 978 v += 25; 979 } 980 idc = 5*(*c--); 981 v = aa + 25*a->diag[i]; 982 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 983 v[15]*sum4+v[20]*sum5; 984 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 985 v[16]*sum4+v[21]*sum5; 986 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 987 v[17]*sum4+v[22]*sum5; 988 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 989 v[18]*sum4+v[23]*sum5; 990 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 991 v[19]*sum4+v[24]*sum5; 992 } 993 994 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 995 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 996 VecRestoreArray_Fast(bb,b); 997 VecRestoreArray_Fast(xx,x); 998 PLogFlops(2*25*(a->nz) - a->n); 999 return 0; 1000 } 1001 1002 #undef __FUNC__ 1003 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1004 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1005 { 1006 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1007 IS iscol=a->col,isrow=a->row; 1008 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1009 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1010 register Scalar *x,*b,*tmp,*v; 1011 1012 VecGetArray_Fast(bb,b); 1013 VecGetArray_Fast(xx,x); 1014 tmp = a->solve_work; 1015 1016 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1017 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1018 1019 /* forward solve the lower triangular */ 1020 idx = 4*(*r++); 1021 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1022 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1023 for ( i=1; i<n; i++ ) { 1024 v = aa + 16*ai[i]; 1025 vi = aj + ai[i]; 1026 nz = a->diag[i] - ai[i]; 1027 idx = 4*(*r++); 1028 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1029 while (nz--) { 1030 idx = 4*(*vi++); 1031 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1032 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1033 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1034 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1035 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1036 v += 16; 1037 } 1038 idx = 4*i; 1039 tmp[idx] = sum1;tmp[1+idx] = sum2; 1040 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1041 } 1042 /* backward solve the upper triangular */ 1043 for ( i=n-1; i>=0; i-- ){ 1044 v = aa + 16*a->diag[i] + 16; 1045 vi = aj + a->diag[i] + 1; 1046 nz = ai[i+1] - a->diag[i] - 1; 1047 idt = 4*i; 1048 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1049 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1050 while (nz--) { 1051 idx = 4*(*vi++); 1052 x1 = tmp[idx]; x2 = tmp[1+idx]; 1053 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1054 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1055 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1056 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1057 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1058 v += 16; 1059 } 1060 idc = 4*(*c--); 1061 v = aa + 16*a->diag[i]; 1062 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1063 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1064 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1065 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1066 } 1067 1068 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1069 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1070 VecRestoreArray_Fast(bb,b); 1071 VecRestoreArray_Fast(xx,x); 1072 PLogFlops(2*16*(a->nz) - a->n); 1073 return 0; 1074 } 1075 1076 1077 #undef __FUNC__ 1078 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1079 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1080 { 1081 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1082 IS iscol=a->col,isrow=a->row; 1083 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1084 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1085 register Scalar *x,*b,*tmp,*v; 1086 1087 VecGetArray_Fast(bb,b); 1088 VecGetArray_Fast(xx,x); 1089 tmp = a->solve_work; 1090 1091 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1092 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1093 1094 /* forward solve the lower triangular */ 1095 idx = 3*(*r++); 1096 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1097 for ( i=1; i<n; i++ ) { 1098 v = aa + 9*ai[i]; 1099 vi = aj + ai[i]; 1100 nz = a->diag[i] - ai[i]; 1101 idx = 3*(*r++); 1102 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1103 while (nz--) { 1104 idx = 3*(*vi++); 1105 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1106 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1107 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1108 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1109 v += 9; 1110 } 1111 idx = 3*i; 1112 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1113 } 1114 /* backward solve the upper triangular */ 1115 for ( i=n-1; i>=0; i-- ){ 1116 v = aa + 9*a->diag[i] + 9; 1117 vi = aj + a->diag[i] + 1; 1118 nz = ai[i+1] - a->diag[i] - 1; 1119 idt = 3*i; 1120 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1121 while (nz--) { 1122 idx = 3*(*vi++); 1123 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1124 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1125 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1126 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1127 v += 9; 1128 } 1129 idc = 3*(*c--); 1130 v = aa + 9*a->diag[i]; 1131 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1132 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1133 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1134 } 1135 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1136 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1137 VecRestoreArray_Fast(bb,b); 1138 VecRestoreArray_Fast(xx,x); 1139 PLogFlops(2*9*(a->nz) - a->n); 1140 return 0; 1141 } 1142 1143 #undef __FUNC__ 1144 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1145 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1146 { 1147 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1148 IS iscol=a->col,isrow=a->row; 1149 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1150 Scalar *aa=a->a,sum1,sum2,x1,x2; 1151 register Scalar *x,*b,*tmp,*v; 1152 1153 VecGetArray_Fast(bb,b); 1154 VecGetArray_Fast(xx,x); 1155 tmp = a->solve_work; 1156 1157 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1158 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1159 1160 /* forward solve the lower triangular */ 1161 idx = 2*(*r++); 1162 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1163 for ( i=1; i<n; i++ ) { 1164 v = aa + 4*ai[i]; 1165 vi = aj + ai[i]; 1166 nz = a->diag[i] - ai[i]; 1167 idx = 2*(*r++); 1168 sum1 = b[idx]; sum2 = b[1+idx]; 1169 while (nz--) { 1170 idx = 2*(*vi++); 1171 x1 = tmp[idx]; x2 = tmp[1+idx]; 1172 sum1 -= v[0]*x1 + v[2]*x2; 1173 sum2 -= v[1]*x1 + v[3]*x2; 1174 v += 4; 1175 } 1176 idx = 2*i; 1177 tmp[idx] = sum1; tmp[1+idx] = sum2; 1178 } 1179 /* backward solve the upper triangular */ 1180 for ( i=n-1; i>=0; i-- ){ 1181 v = aa + 4*a->diag[i] + 4; 1182 vi = aj + a->diag[i] + 1; 1183 nz = ai[i+1] - a->diag[i] - 1; 1184 idt = 2*i; 1185 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1186 while (nz--) { 1187 idx = 2*(*vi++); 1188 x1 = tmp[idx]; x2 = tmp[1+idx]; 1189 sum1 -= v[0]*x1 + v[2]*x2; 1190 sum2 -= v[1]*x1 + v[3]*x2; 1191 v += 4; 1192 } 1193 idc = 2*(*c--); 1194 v = aa + 4*a->diag[i]; 1195 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1196 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1197 } 1198 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1199 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1200 VecRestoreArray_Fast(bb,b); 1201 VecRestoreArray_Fast(xx,x); 1202 PLogFlops(2*4*(a->nz) - a->n); 1203 return 0; 1204 } 1205 1206 1207 #undef __FUNC__ 1208 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1209 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1210 { 1211 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1212 IS iscol=a->col,isrow=a->row; 1213 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1214 Scalar *aa=a->a,sum1; 1215 register Scalar *x,*b,*tmp,*v; 1216 1217 if (!n) return 0; 1218 1219 VecGetArray_Fast(bb,b); 1220 VecGetArray_Fast(xx,x); 1221 tmp = a->solve_work; 1222 1223 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1224 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1225 1226 /* forward solve the lower triangular */ 1227 tmp[0] = b[*r++]; 1228 for ( i=1; i<n; i++ ) { 1229 v = aa + ai[i]; 1230 vi = aj + ai[i]; 1231 nz = a->diag[i] - ai[i]; 1232 sum1 = b[*r++]; 1233 while (nz--) { 1234 sum1 -= (*v++)*tmp[*vi++]; 1235 } 1236 tmp[i] = sum1; 1237 } 1238 /* backward solve the upper triangular */ 1239 for ( i=n-1; i>=0; i-- ){ 1240 v = aa + a->diag[i] + 1; 1241 vi = aj + a->diag[i] + 1; 1242 nz = ai[i+1] - a->diag[i] - 1; 1243 sum1 = tmp[i]; 1244 while (nz--) { 1245 sum1 -= (*v++)*tmp[*vi++]; 1246 } 1247 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1248 } 1249 1250 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1251 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1252 VecRestoreArray_Fast(bb,b); 1253 VecRestoreArray_Fast(xx,x); 1254 PLogFlops(2*1*(a->nz) - a->n); 1255 return 0; 1256 } 1257 1258 /* ----------------------------------------------------------------*/ 1259 /* 1260 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1261 except that the data structure of Mat_SeqAIJ is slightly different. 1262 Not a good example of code reuse. 1263 */ 1264 #undef __FUNC__ 1265 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1266 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1267 Mat *fact) 1268 { 1269 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1270 IS isicol; 1271 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1272 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1273 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1274 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1275 PetscTruth col_identity, row_identity; 1276 1277 /* special case that simply copies fill pattern */ 1278 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1279 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1280 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1281 if (levels == 0 && row_identity && col_identity) { 1282 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1283 (*fact)->factor = FACTOR_LU; 1284 b = (Mat_SeqBAIJ *) (*fact)->data; 1285 if (!b->diag) { 1286 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1287 } 1288 b->row = isrow; 1289 b->col = iscol; 1290 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1291 return 0; 1292 } 1293 1294 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1295 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1296 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1297 1298 /* get new row pointers */ 1299 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1300 ainew[0] = 0; 1301 /* don't know how many column pointers are needed so estimate */ 1302 jmax = (int) (f*ai[n] + 1); 1303 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1304 /* ajfill is level of fill for each fill entry */ 1305 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1306 /* fill is a linked list of nonzeros in active row */ 1307 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1308 /* im is level for each filled value */ 1309 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1310 /* dloc is location of diagonal in factor */ 1311 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1312 dloc[0] = 0; 1313 for ( prow=0; prow<n; prow++ ) { 1314 /* first copy previous fill into linked list */ 1315 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1316 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1317 xi = aj + ai[r[prow]]; 1318 fill[n] = n; 1319 while (nz--) { 1320 fm = n; 1321 idx = ic[*xi++]; 1322 do { 1323 m = fm; 1324 fm = fill[m]; 1325 } while (fm < idx); 1326 fill[m] = idx; 1327 fill[idx] = fm; 1328 im[idx] = 0; 1329 } 1330 nzi = 0; 1331 row = fill[n]; 1332 while ( row < prow ) { 1333 incrlev = im[row] + 1; 1334 nz = dloc[row]; 1335 xi = ajnew + ainew[row] + nz; 1336 flev = ajfill + ainew[row] + nz + 1; 1337 nnz = ainew[row+1] - ainew[row] - nz - 1; 1338 if (*xi++ != row) { 1339 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"zero pivot"); 1340 } 1341 fm = row; 1342 while (nnz-- > 0) { 1343 idx = *xi++; 1344 if (*flev + incrlev > levels) { 1345 flev++; 1346 continue; 1347 } 1348 do { 1349 m = fm; 1350 fm = fill[m]; 1351 } while (fm < idx); 1352 if (fm != idx) { 1353 im[idx] = *flev + incrlev; 1354 fill[m] = idx; 1355 fill[idx] = fm; 1356 fm = idx; 1357 nzf++; 1358 } 1359 else { 1360 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1361 } 1362 flev++; 1363 } 1364 row = fill[row]; 1365 nzi++; 1366 } 1367 /* copy new filled row into permanent storage */ 1368 ainew[prow+1] = ainew[prow] + nzf; 1369 if (ainew[prow+1] > jmax) { 1370 /* allocate a longer ajnew */ 1371 int maxadd; 1372 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1373 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1374 jmax += maxadd; 1375 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1376 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1377 PetscFree(ajnew); 1378 ajnew = xi; 1379 /* allocate a longer ajfill */ 1380 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1381 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1382 PetscFree(ajfill); 1383 ajfill = xi; 1384 realloc++; 1385 } 1386 xi = ajnew + ainew[prow]; 1387 flev = ajfill + ainew[prow]; 1388 dloc[prow] = nzi; 1389 fm = fill[n]; 1390 while (nzf--) { 1391 *xi++ = fm; 1392 *flev++ = im[fm]; 1393 fm = fill[fm]; 1394 } 1395 } 1396 PetscFree(ajfill); 1397 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1398 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1399 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1400 PetscFree(fill); PetscFree(im); 1401 1402 PLogInfo(A, 1403 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1404 realloc,f,((double)ainew[n])/((double)ai[prow])); 1405 1406 /* put together the new matrix */ 1407 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1408 b = (Mat_SeqBAIJ *) (*fact)->data; 1409 PetscFree(b->imax); 1410 b->singlemalloc = 0; 1411 len = bs2*ainew[n]*sizeof(Scalar); 1412 /* the next line frees the default space generated by the Create() */ 1413 PetscFree(b->a); PetscFree(b->ilen); 1414 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1415 b->j = ajnew; 1416 b->i = ainew; 1417 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1418 b->diag = dloc; 1419 b->ilen = 0; 1420 b->imax = 0; 1421 b->row = isrow; 1422 b->col = iscol; 1423 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1424 CHKPTRQ(b->solve_work); 1425 /* In b structure: Free imax, ilen, old a, old j. 1426 Allocate dloc, solve_work, new a, new j */ 1427 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1428 b->maxnz = b->nz = ainew[n]; 1429 (*fact)->factor = FACTOR_LU; 1430 1431 (*fact)->info.factor_mallocs = realloc; 1432 (*fact)->info.fill_ratio_given = f; 1433 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1434 1435 return 0; 1436 } 1437 1438 1439 1440 1441