1 #ifdef PETSC_RCS_HEADER 2 static char vcid[] = "$Id: baijfact.c,v 1.51 1997/07/25 22:14:05 bsmith Exp bsmith $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "src/mat/impls/baij/seq/baij.h" 9 #include "src/vec/vecimpl.h" 10 #include "src/inline/ilu.h" 11 12 13 /* 14 The symbolic factorization code is identical to that for AIJ format, 15 except for very small changes since this is now a SeqBAIJ datastructure. 16 NOT good code reuse. 17 */ 18 #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 if (ai[n] != 0) { 112 double af = ((double)ainew[n])/((double)ai[n]); 113 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 114 realloc,f,af); 115 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Run with -pc_lu_fill %g or use \n",af); 116 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:PCLUSetFill(pc,%g);\n",af); 117 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:for best performance.\n"); 118 } else { 119 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Empty matrix.\n"); 120 } 121 122 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 123 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 124 125 PetscFree(fill); 126 127 /* put together the new matrix */ 128 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 129 PLogObjectParent(*B,isicol); 130 ierr = ISDestroy(isicol); CHKERRQ(ierr); 131 b = (Mat_SeqBAIJ *) (*B)->data; 132 PetscFree(b->imax); 133 b->singlemalloc = 0; 134 /* the next line frees the default space generated by the Create() */ 135 PetscFree(b->a); PetscFree(b->ilen); 136 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 137 b->j = ajnew; 138 b->i = ainew; 139 b->diag = idnew; 140 b->ilen = 0; 141 b->imax = 0; 142 b->row = isrow; 143 b->col = iscol; 144 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 145 CHKPTRQ(b->solve_work); 146 /* In b structure: Free imax, ilen, old a, old j. 147 Allocate idnew, solve_work, new a, new j */ 148 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 149 b->maxnz = b->nz = ainew[n]; 150 151 (*B)->info.factor_mallocs = realloc; 152 (*B)->info.fill_ratio_given = f; 153 if (ai[i] != 0) { 154 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 155 } else { 156 (*B)->info.fill_ratio_needed = 0.0; 157 } 158 159 160 return 0; 161 } 162 163 /* ----------------------------------------------------------- */ 164 #undef __FUNC__ 165 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 166 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 167 { 168 Mat C = *B; 169 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 170 IS iscol = b->col, isrow = b->row, isicol; 171 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 172 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 173 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 174 Scalar *ba = b->a,*aa = a->a; 175 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 176 register int *pj; 177 178 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 179 PLogObjectParent(*B,isicol); 180 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 181 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 182 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 183 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 184 /* generate work space needed by dense LU factorization */ 185 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 186 CHKPTRQ(v_work); 187 multiplier = v_work + bs; 188 v_pivots = (int *) (multiplier + bs2); 189 190 /* flops in while loop */ 191 bslog = 2*bs*bs2; 192 193 for ( i=0; i<n; i++ ) { 194 nz = bi[i+1] - bi[i]; 195 ajtmp = bj + bi[i]; 196 for ( j=0; j<nz; j++ ) { 197 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 198 } 199 /* load in initial (unfactored row) */ 200 nz = ai[r[i]+1] - ai[r[i]]; 201 ajtmpold = aj + ai[r[i]]; 202 v = aa + bs2*ai[r[i]]; 203 for ( j=0; j<nz; j++ ) { 204 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 205 } 206 row = *ajtmp++; 207 while (row < i) { 208 pc = rtmp + bs2*row; 209 /* if (*pc) { */ 210 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 211 if (flg) { 212 pv = ba + bs2*diag_offset[row]; 213 pj = bj + diag_offset[row] + 1; 214 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 215 nz = bi[row+1] - diag_offset[row] - 1; 216 pv += bs2; 217 for (j=0; j<nz; j++) { 218 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 219 } 220 PLogFlops(bslog*(nz+1)-bs); 221 } 222 row = *ajtmp++; 223 } 224 /* finished row so stick it into b->a */ 225 pv = ba + bs2*bi[i]; 226 pj = bj + bi[i]; 227 nz = bi[i+1] - bi[i]; 228 for ( j=0; j<nz; j++ ) { 229 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 230 } 231 diag = diag_offset[i] - bi[i]; 232 /* invert diagonal block */ 233 w = pv + bs2*diag; 234 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 235 } 236 237 PetscFree(rtmp); PetscFree(v_work); 238 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 239 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 240 ierr = ISDestroy(isicol); CHKERRQ(ierr); 241 C->factor = FACTOR_LU; 242 C->assembled = PETSC_TRUE; 243 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 244 return 0; 245 } 246 /* ------------------------------------------------------------*/ 247 /* 248 Version for when blocks are 5 by 5 249 */ 250 #undef __FUNC__ 251 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 252 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 253 { 254 Mat C = *B; 255 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 256 IS iscol = b->col, isrow = b->row, isicol; 257 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 258 int *ajtmpold, *ajtmp, nz, row; 259 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 260 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 261 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 262 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 263 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 264 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 265 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 266 Scalar *ba = b->a,*aa = a->a; 267 register int *pj; 268 269 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 270 PLogObjectParent(*B,isicol); 271 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 272 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 273 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 274 275 for ( i=0; i<n; i++ ) { 276 nz = bi[i+1] - bi[i]; 277 ajtmp = bj + bi[i]; 278 for ( j=0; j<nz; j++ ) { 279 x = rtmp+25*ajtmp[j]; 280 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 281 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 282 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 283 } 284 /* load in initial (unfactored row) */ 285 idx = r[i]; 286 nz = ai[idx+1] - ai[idx]; 287 ajtmpold = aj + ai[idx]; 288 v = aa + 25*ai[idx]; 289 for ( j=0; j<nz; j++ ) { 290 x = rtmp+25*ic[ajtmpold[j]]; 291 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 292 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 293 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 294 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 295 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 296 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 297 v += 25; 298 } 299 row = *ajtmp++; 300 while (row < i) { 301 pc = rtmp + 25*row; 302 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 303 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 304 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 305 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 306 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 307 p25 = pc[24]; 308 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 309 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 310 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 311 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 312 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 313 p24 != 0.0 || p25 != 0.0) { 314 pv = ba + 25*diag_offset[row]; 315 pj = bj + diag_offset[row] + 1; 316 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 317 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 318 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 319 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 320 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 321 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 322 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 323 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 324 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 325 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 326 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 327 328 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 329 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 330 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 331 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 332 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 333 334 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 335 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 336 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 337 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 338 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 339 340 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 341 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 342 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 343 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 344 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 345 346 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 347 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 348 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 349 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 350 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 351 352 nz = bi[row+1] - diag_offset[row] - 1; 353 pv += 25; 354 for (j=0; j<nz; j++) { 355 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 356 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 357 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 358 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 359 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 360 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 361 x = rtmp + 25*pj[j]; 362 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 363 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 364 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 365 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 366 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 367 368 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 369 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 370 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 371 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 372 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 373 374 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 375 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 376 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 377 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 378 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 379 380 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 381 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 382 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 383 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 384 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 385 386 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 387 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 388 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 389 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 390 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 391 392 pv += 25; 393 } 394 PLogFlops(250*nz+225); 395 } 396 row = *ajtmp++; 397 } 398 /* finished row so stick it into b->a */ 399 pv = ba + 25*bi[i]; 400 pj = bj + bi[i]; 401 nz = bi[i+1] - bi[i]; 402 for ( j=0; j<nz; j++ ) { 403 x = rtmp+25*pj[j]; 404 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 405 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 406 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 407 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 408 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 409 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 410 pv += 25; 411 } 412 /* invert diagonal block */ 413 w = ba + 25*diag_offset[i]; 414 ierr = Kernel_A_gets_inverse_A_5(w); CHKERRQ(ierr); 415 } 416 417 PetscFree(rtmp); 418 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 419 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 420 ierr = ISDestroy(isicol); CHKERRQ(ierr); 421 C->factor = FACTOR_LU; 422 C->assembled = PETSC_TRUE; 423 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 424 return 0; 425 } 426 427 /* ------------------------------------------------------------*/ 428 /* 429 Version for when blocks are 4 by 4 430 */ 431 #undef __FUNC__ 432 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 433 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 434 { 435 Mat C = *B; 436 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 437 IS iscol = b->col, isrow = b->row, isicol; 438 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 439 int *ajtmpold, *ajtmp, nz, row; 440 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 441 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 442 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 443 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 444 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 445 Scalar m13,m14,m15,m16; 446 Scalar *ba = b->a,*aa = a->a; 447 register int *pj; 448 449 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 450 PLogObjectParent(*B,isicol); 451 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 452 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 453 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 454 455 for ( i=0; i<n; i++ ) { 456 nz = bi[i+1] - bi[i]; 457 ajtmp = bj + bi[i]; 458 for ( j=0; j<nz; j++ ) { 459 x = rtmp+16*ajtmp[j]; 460 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 461 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 462 } 463 /* load in initial (unfactored row) */ 464 idx = r[i]; 465 nz = ai[idx+1] - ai[idx]; 466 ajtmpold = aj + ai[idx]; 467 v = aa + 16*ai[idx]; 468 for ( j=0; j<nz; j++ ) { 469 x = rtmp+16*ic[ajtmpold[j]]; 470 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 471 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 472 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 473 x[14] = v[14]; x[15] = v[15]; 474 v += 16; 475 } 476 row = *ajtmp++; 477 while (row < i) { 478 pc = rtmp + 16*row; 479 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 480 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 481 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 482 p15 = pc[14]; p16 = pc[15]; 483 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 484 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 485 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 486 || p16 != 0.0) { 487 pv = ba + 16*diag_offset[row]; 488 pj = bj + diag_offset[row] + 1; 489 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 490 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 491 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 492 x15 = pv[14]; x16 = pv[15]; 493 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 494 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 495 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 496 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 497 498 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 499 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 500 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 501 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 502 503 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 504 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 505 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 506 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 507 508 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 509 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 510 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 511 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 512 513 nz = bi[row+1] - diag_offset[row] - 1; 514 pv += 16; 515 for (j=0; j<nz; j++) { 516 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 517 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 518 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 519 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 520 x = rtmp + 16*pj[j]; 521 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 522 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 523 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 524 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 525 526 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 527 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 528 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 529 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 530 531 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 532 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 533 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 534 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 535 536 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 537 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 538 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 539 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 540 541 pv += 16; 542 } 543 PLogFlops(128*nz+112); 544 } 545 row = *ajtmp++; 546 } 547 /* finished row so stick it into b->a */ 548 pv = ba + 16*bi[i]; 549 pj = bj + bi[i]; 550 nz = bi[i+1] - bi[i]; 551 for ( j=0; j<nz; j++ ) { 552 x = rtmp+16*pj[j]; 553 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 554 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 555 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 556 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 557 pv += 16; 558 } 559 /* invert diagonal block */ 560 w = ba + 16*diag_offset[i]; 561 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 562 } 563 564 PetscFree(rtmp); 565 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 566 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 567 ierr = ISDestroy(isicol); CHKERRQ(ierr); 568 C->factor = FACTOR_LU; 569 C->assembled = PETSC_TRUE; 570 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 571 return 0; 572 } 573 /* 574 Version for when blocks are 4 by 4 Using natural ordering 575 */ 576 #undef __FUNC__ 577 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering" 578 int MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering(Mat A,Mat *B) 579 { 580 Mat C = *B; 581 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 582 int ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 583 int *ajtmpold, *ajtmp, nz, row; 584 int *diag_offset = b->diag,*ai=a->i,*aj=a->j; 585 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 586 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 587 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 588 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 589 Scalar m13,m14,m15,m16; 590 Scalar *ba = b->a,*aa = a->a; 591 register int *pj; 592 593 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 594 595 for ( i=0; i<n; i++ ) { 596 nz = bi[i+1] - bi[i]; 597 ajtmp = bj + bi[i]; 598 for ( j=0; j<nz; j++ ) { 599 x = rtmp+16*ajtmp[j]; 600 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 601 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 602 } 603 /* load in initial (unfactored row) */ 604 nz = ai[i+1] - ai[i]; 605 ajtmpold = aj + ai[i]; 606 v = aa + 16*ai[i]; 607 for ( j=0; j<nz; j++ ) { 608 x = rtmp+16*ajtmpold[j]; 609 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 610 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 611 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 612 x[14] = v[14]; x[15] = v[15]; 613 v += 16; 614 } 615 row = *ajtmp++; 616 while (row < i) { 617 pc = rtmp + 16*row; 618 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 619 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 620 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 621 p15 = pc[14]; p16 = pc[15]; 622 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 623 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 624 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 625 || p16 != 0.0) { 626 pv = ba + 16*diag_offset[row]; 627 pj = bj + diag_offset[row] + 1; 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 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 631 x15 = pv[14]; x16 = pv[15]; 632 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 633 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 634 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 635 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 636 637 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 638 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 639 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 640 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 641 642 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 643 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 644 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 645 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 646 647 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 648 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 649 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 650 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 651 652 nz = bi[row+1] - diag_offset[row] - 1; 653 pv += 16; 654 for (j=0; j<nz; j++) { 655 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 656 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 657 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 658 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 659 x = rtmp + 16*pj[j]; 660 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 661 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 662 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 663 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 664 665 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 666 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 667 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 668 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 669 670 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 671 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 672 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 673 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 674 675 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 676 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 677 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 678 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 679 680 pv += 16; 681 } 682 PLogFlops(128*nz+112); 683 } 684 row = *ajtmp++; 685 } 686 /* finished row so stick it into b->a */ 687 pv = ba + 16*bi[i]; 688 pj = bj + bi[i]; 689 nz = bi[i+1] - bi[i]; 690 for ( j=0; j<nz; j++ ) { 691 x = rtmp+16*pj[j]; 692 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 693 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 694 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 695 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 696 pv += 16; 697 } 698 /* invert diagonal block */ 699 w = ba + 16*diag_offset[i]; 700 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 701 } 702 703 PetscFree(rtmp); 704 C->factor = FACTOR_LU; 705 C->assembled = PETSC_TRUE; 706 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 707 return 0; 708 } 709 710 /* ------------------------------------------------------------*/ 711 /* 712 Version for when blocks are 3 by 3 713 */ 714 #undef __FUNC__ 715 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 716 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 717 { 718 Mat C = *B; 719 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 720 IS iscol = b->col, isrow = b->row, isicol; 721 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 722 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 723 int *diag_offset = b->diag,idx; 724 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 725 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 726 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 727 Scalar *ba = b->a,*aa = a->a; 728 register int *pj; 729 730 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 731 PLogObjectParent(*B,isicol); 732 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 733 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 734 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 735 736 for ( i=0; i<n; i++ ) { 737 nz = bi[i+1] - bi[i]; 738 ajtmp = bj + bi[i]; 739 for ( j=0; j<nz; j++ ) { 740 x = rtmp + 9*ajtmp[j]; 741 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 742 } 743 /* load in initial (unfactored row) */ 744 idx = r[i]; 745 nz = ai[idx+1] - ai[idx]; 746 ajtmpold = aj + ai[idx]; 747 v = aa + 9*ai[idx]; 748 for ( j=0; j<nz; j++ ) { 749 x = rtmp + 9*ic[ajtmpold[j]]; 750 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 751 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 752 v += 9; 753 } 754 row = *ajtmp++; 755 while (row < i) { 756 pc = rtmp + 9*row; 757 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 758 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 759 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 760 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 761 pv = ba + 9*diag_offset[row]; 762 pj = bj + diag_offset[row] + 1; 763 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 764 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 765 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 766 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 767 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 768 769 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 770 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 771 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 772 773 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 774 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 775 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 776 nz = bi[row+1] - diag_offset[row] - 1; 777 pv += 9; 778 for (j=0; j<nz; j++) { 779 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 780 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 781 x = rtmp + 9*pj[j]; 782 x[0] -= m1*x1 + m4*x2 + m7*x3; 783 x[1] -= m2*x1 + m5*x2 + m8*x3; 784 x[2] -= m3*x1 + m6*x2 + m9*x3; 785 786 x[3] -= m1*x4 + m4*x5 + m7*x6; 787 x[4] -= m2*x4 + m5*x5 + m8*x6; 788 x[5] -= m3*x4 + m6*x5 + m9*x6; 789 790 x[6] -= m1*x7 + m4*x8 + m7*x9; 791 x[7] -= m2*x7 + m5*x8 + m8*x9; 792 x[8] -= m3*x7 + m6*x8 + m9*x9; 793 pv += 9; 794 } 795 PLogFlops(54*nz+36); 796 } 797 row = *ajtmp++; 798 } 799 /* finished row so stick it into b->a */ 800 pv = ba + 9*bi[i]; 801 pj = bj + bi[i]; 802 nz = bi[i+1] - bi[i]; 803 for ( j=0; j<nz; j++ ) { 804 x = rtmp + 9*pj[j]; 805 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 806 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 807 pv += 9; 808 } 809 /* invert diagonal block */ 810 w = ba + 9*diag_offset[i]; 811 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 812 } 813 814 PetscFree(rtmp); 815 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 816 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 817 ierr = ISDestroy(isicol); CHKERRQ(ierr); 818 C->factor = FACTOR_LU; 819 C->assembled = PETSC_TRUE; 820 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 821 return 0; 822 } 823 824 /* ------------------------------------------------------------*/ 825 /* 826 Version for when blocks are 2 by 2 827 */ 828 #undef __FUNC__ 829 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 830 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 831 { 832 Mat C = *B; 833 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 834 IS iscol = b->col, isrow = b->row, isicol; 835 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 836 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 837 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 838 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 839 Scalar p1,p2,p3,p4,v_work[2]; 840 Scalar *ba = b->a,*aa = a->a; 841 register int *pj; 842 843 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 844 PLogObjectParent(*B,isicol); 845 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 846 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 847 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 848 849 for ( i=0; i<n; i++ ) { 850 nz = bi[i+1] - bi[i]; 851 ajtmp = bj + bi[i]; 852 for ( j=0; j<nz; j++ ) { 853 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 854 } 855 /* load in initial (unfactored row) */ 856 idx = r[i]; 857 nz = ai[idx+1] - ai[idx]; 858 ajtmpold = aj + ai[idx]; 859 v = aa + 4*ai[idx]; 860 for ( j=0; j<nz; j++ ) { 861 x = rtmp+4*ic[ajtmpold[j]]; 862 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 863 v += 4; 864 } 865 row = *ajtmp++; 866 while (row < i) { 867 pc = rtmp + 4*row; 868 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 869 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 870 pv = ba + 4*diag_offset[row]; 871 pj = bj + diag_offset[row] + 1; 872 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 873 pc[0] = m1 = p1*x1 + p3*x2; 874 pc[1] = m2 = p2*x1 + p4*x2; 875 pc[2] = m3 = p1*x3 + p3*x4; 876 pc[3] = m4 = p2*x3 + p4*x4; 877 nz = bi[row+1] - diag_offset[row] - 1; 878 pv += 4; 879 for (j=0; j<nz; j++) { 880 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 881 x = rtmp + 4*pj[j]; 882 x[0] -= m1*x1 + m3*x2; 883 x[1] -= m2*x1 + m4*x2; 884 x[2] -= m1*x3 + m3*x4; 885 x[3] -= m2*x3 + m4*x4; 886 pv += 4; 887 } 888 PLogFlops(16*nz+12); 889 } 890 row = *ajtmp++; 891 } 892 /* finished row so stick it into b->a */ 893 pv = ba + 4*bi[i]; 894 pj = bj + bi[i]; 895 nz = bi[i+1] - bi[i]; 896 for ( j=0; j<nz; j++ ) { 897 x = rtmp+4*pj[j]; 898 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 899 pv += 4; 900 } 901 /* invert diagonal block */ 902 w = ba + 4*diag_offset[i]; 903 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 904 } 905 906 PetscFree(rtmp); 907 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 908 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 909 ierr = ISDestroy(isicol); CHKERRQ(ierr); 910 C->factor = FACTOR_LU; 911 C->assembled = PETSC_TRUE; 912 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 913 return 0; 914 } 915 916 /* ----------------------------------------------------------- */ 917 /* 918 Version for when blocks are 1 by 1. 919 */ 920 #undef __FUNC__ 921 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 922 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 923 { 924 Mat C = *B; 925 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 926 IS iscol = b->col, isrow = b->row, isicol; 927 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 928 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 929 int *diag_offset = b->diag,diag; 930 register Scalar *pv,*v,*rtmp,multiplier,*pc; 931 Scalar *ba = b->a,*aa = a->a; 932 register int *pj; 933 934 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 935 PLogObjectParent(*B,isicol); 936 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 937 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 938 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 939 940 for ( i=0; i<n; i++ ) { 941 nz = bi[i+1] - bi[i]; 942 ajtmp = bj + bi[i]; 943 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 944 945 /* load in initial (unfactored row) */ 946 nz = ai[r[i]+1] - ai[r[i]]; 947 ajtmpold = aj + ai[r[i]]; 948 v = aa + ai[r[i]]; 949 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 950 951 row = *ajtmp++; 952 while (row < i) { 953 pc = rtmp + row; 954 if (*pc != 0.0) { 955 pv = ba + diag_offset[row]; 956 pj = bj + diag_offset[row] + 1; 957 multiplier = *pc * *pv++; 958 *pc = multiplier; 959 nz = bi[row+1] - diag_offset[row] - 1; 960 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 961 PLogFlops(1+2*nz); 962 } 963 row = *ajtmp++; 964 } 965 /* finished row so stick it into b->a */ 966 pv = ba + bi[i]; 967 pj = bj + bi[i]; 968 nz = bi[i+1] - bi[i]; 969 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 970 diag = diag_offset[i] - bi[i]; 971 /* check pivot entry for current row */ 972 if (pv[diag] == 0.0) { 973 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 974 } 975 pv[diag] = 1.0/pv[diag]; 976 } 977 978 PetscFree(rtmp); 979 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 980 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 981 ierr = ISDestroy(isicol); CHKERRQ(ierr); 982 C->factor = FACTOR_LU; 983 C->assembled = PETSC_TRUE; 984 PLogFlops(b->n); 985 return 0; 986 } 987 988 /* ----------------------------------------------------------- */ 989 #undef __FUNC__ 990 #define __FUNC__ "MatLUFactor_SeqBAIJ" 991 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 992 { 993 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 994 int ierr; 995 Mat C; 996 997 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 998 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 999 1000 /* free all the data structures from mat */ 1001 PetscFree(mat->a); 1002 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 1003 if (mat->diag) PetscFree(mat->diag); 1004 if (mat->ilen) PetscFree(mat->ilen); 1005 if (mat->imax) PetscFree(mat->imax); 1006 if (mat->solve_work) PetscFree(mat->solve_work); 1007 if (mat->mult_work) PetscFree(mat->mult_work); 1008 PetscFree(mat); 1009 1010 PetscMemcpy(A,C,sizeof(struct _p_Mat)); 1011 PetscHeaderDestroy(C); 1012 return 0; 1013 } 1014 /* ----------------------------------------------------------- */ 1015 #undef __FUNC__ 1016 #define __FUNC__ "MatSolve_SeqBAIJ_N" 1017 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 1018 { 1019 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1020 IS iscol=a->col,isrow=a->row; 1021 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 1022 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 1023 Scalar *aa=a->a,*sum; 1024 register Scalar *x,*b,*lsum,*tmp,*v; 1025 1026 VecGetArray_Fast(bb,b); 1027 VecGetArray_Fast(xx,x); 1028 tmp = a->solve_work; 1029 1030 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1031 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1032 1033 /* forward solve the lower triangular */ 1034 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 1035 for ( i=1; i<n; i++ ) { 1036 v = aa + bs2*ai[i]; 1037 vi = aj + ai[i]; 1038 nz = a->diag[i] - ai[i]; 1039 sum = tmp + bs*i; 1040 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 1041 while (nz--) { 1042 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 1043 v += bs2; 1044 } 1045 } 1046 /* backward solve the upper triangular */ 1047 lsum = a->solve_work + a->n; 1048 for ( i=n-1; i>=0; i-- ){ 1049 v = aa + bs2*(a->diag[i] + 1); 1050 vi = aj + a->diag[i] + 1; 1051 nz = ai[i+1] - a->diag[i] - 1; 1052 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 1053 while (nz--) { 1054 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 1055 v += bs2; 1056 } 1057 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 1058 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 1059 } 1060 1061 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1062 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1063 VecRestoreArray_Fast(bb,b); 1064 VecRestoreArray_Fast(xx,x); 1065 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 1066 return 0; 1067 } 1068 1069 #undef __FUNC__ 1070 #define __FUNC__ "MatSolve_SeqBAIJ_7" 1071 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 1072 { 1073 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1074 IS iscol=a->col,isrow=a->row; 1075 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1076 int *diag = a->diag; 1077 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 1078 register Scalar *x,*b,*tmp,*v; 1079 1080 VecGetArray_Fast(bb,b); 1081 VecGetArray_Fast(xx,x); 1082 tmp = a->solve_work; 1083 1084 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1085 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1086 1087 /* forward solve the lower triangular */ 1088 idx = 7*(*r++); 1089 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1090 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1091 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 1092 1093 for ( i=1; i<n; i++ ) { 1094 v = aa + 49*ai[i]; 1095 vi = aj + ai[i]; 1096 nz = diag[i] - ai[i]; 1097 idx = 7*(*r++); 1098 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1099 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 1100 while (nz--) { 1101 idx = 7*(*vi++); 1102 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1103 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1104 x6 = tmp[5+idx];x7 = tmp[6+idx]; 1105 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 1106 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 1107 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 1108 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 1109 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 1110 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 1111 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 1112 v += 49; 1113 } 1114 idx = 7*i; 1115 tmp[idx] = sum1;tmp[1+idx] = sum2; 1116 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1117 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 1118 } 1119 /* backward solve the upper triangular */ 1120 for ( i=n-1; i>=0; i-- ){ 1121 v = aa + 49*diag[i] + 49; 1122 vi = aj + diag[i] + 1; 1123 nz = ai[i+1] - diag[i] - 1; 1124 idt = 7*i; 1125 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1126 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1127 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 1128 while (nz--) { 1129 idx = 7*(*vi++); 1130 x1 = tmp[idx]; x2 = tmp[1+idx]; 1131 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1132 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 1133 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 1134 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 1135 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 1136 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 1137 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 1138 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 1139 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 1140 v += 49; 1141 } 1142 idc = 7*(*c--); 1143 v = aa + 49*diag[i]; 1144 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 1145 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 1146 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 1147 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 1148 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 1149 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1150 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1151 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1152 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1153 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1154 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1155 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1156 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1157 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1158 } 1159 1160 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1161 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1162 VecRestoreArray_Fast(bb,b); 1163 VecRestoreArray_Fast(xx,x); 1164 PLogFlops(2*49*(a->nz) - a->n); 1165 return 0; 1166 } 1167 1168 #undef __FUNC__ 1169 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1170 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1171 { 1172 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1173 IS iscol=a->col,isrow=a->row; 1174 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1175 int *diag = a->diag; 1176 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1177 register Scalar *x,*b,*tmp,*v; 1178 1179 VecGetArray_Fast(bb,b); 1180 VecGetArray_Fast(xx,x); 1181 tmp = a->solve_work; 1182 1183 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1184 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1185 1186 /* forward solve the lower triangular */ 1187 idx = 5*(*r++); 1188 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1189 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1190 for ( i=1; i<n; i++ ) { 1191 v = aa + 25*ai[i]; 1192 vi = aj + ai[i]; 1193 nz = diag[i] - ai[i]; 1194 idx = 5*(*r++); 1195 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1196 sum5 = b[4+idx]; 1197 while (nz--) { 1198 idx = 5*(*vi++); 1199 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1200 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1201 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1202 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1203 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1204 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1205 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1206 v += 25; 1207 } 1208 idx = 5*i; 1209 tmp[idx] = sum1;tmp[1+idx] = sum2; 1210 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1211 } 1212 /* backward solve the upper triangular */ 1213 for ( i=n-1; i>=0; i-- ){ 1214 v = aa + 25*diag[i] + 25; 1215 vi = aj + diag[i] + 1; 1216 nz = ai[i+1] - diag[i] - 1; 1217 idt = 5*i; 1218 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1219 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1220 while (nz--) { 1221 idx = 5*(*vi++); 1222 x1 = tmp[idx]; x2 = tmp[1+idx]; 1223 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1224 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1225 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1226 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1227 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1228 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1229 v += 25; 1230 } 1231 idc = 5*(*c--); 1232 v = aa + 25*diag[i]; 1233 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1234 v[15]*sum4+v[20]*sum5; 1235 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1236 v[16]*sum4+v[21]*sum5; 1237 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1238 v[17]*sum4+v[22]*sum5; 1239 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1240 v[18]*sum4+v[23]*sum5; 1241 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1242 v[19]*sum4+v[24]*sum5; 1243 } 1244 1245 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1246 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1247 VecRestoreArray_Fast(bb,b); 1248 VecRestoreArray_Fast(xx,x); 1249 PLogFlops(2*25*(a->nz) - a->n); 1250 return 0; 1251 } 1252 1253 #undef __FUNC__ 1254 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1255 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1256 { 1257 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1258 IS iscol=a->col,isrow=a->row; 1259 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1260 int *diag = a->diag; 1261 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1262 register Scalar *x,*b,*tmp,*v; 1263 1264 VecGetArray_Fast(bb,b); 1265 VecGetArray_Fast(xx,x); 1266 tmp = a->solve_work; 1267 1268 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1269 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1270 1271 /* forward solve the lower triangular */ 1272 idx = 4*(*r++); 1273 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1274 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1275 for ( i=1; i<n; i++ ) { 1276 v = aa + 16*ai[i]; 1277 vi = aj + ai[i]; 1278 nz = diag[i] - ai[i]; 1279 idx = 4*(*r++); 1280 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1281 while (nz--) { 1282 idx = 4*(*vi++); 1283 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1284 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1285 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1286 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1287 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1288 v += 16; 1289 } 1290 idx = 4*i; 1291 tmp[idx] = sum1;tmp[1+idx] = sum2; 1292 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1293 } 1294 /* backward solve the upper triangular */ 1295 for ( i=n-1; i>=0; i-- ){ 1296 v = aa + 16*diag[i] + 16; 1297 vi = aj + diag[i] + 1; 1298 nz = ai[i+1] - diag[i] - 1; 1299 idt = 4*i; 1300 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1301 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1302 while (nz--) { 1303 idx = 4*(*vi++); 1304 x1 = tmp[idx]; x2 = tmp[1+idx]; 1305 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1306 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1307 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1308 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1309 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1310 v += 16; 1311 } 1312 idc = 4*(*c--); 1313 v = aa + 16*diag[i]; 1314 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1315 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1316 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1317 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1318 } 1319 1320 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1321 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1322 VecRestoreArray_Fast(bb,b); 1323 VecRestoreArray_Fast(xx,x); 1324 PLogFlops(2*16*(a->nz) - a->n); 1325 return 0; 1326 } 1327 1328 /* 1329 Special case where the matrix was ILU(0) factored in the natural 1330 ordering. This eliminates the need for the column and row permutation. 1331 */ 1332 #undef __FUNC__ 1333 #define __FUNC__ "MatSolve_SeqBAIJ_4_NaturalOrdering" 1334 int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat A,Vec bb,Vec xx) 1335 { 1336 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1337 int i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt; 1338 int *diag = a->diag,jdx; 1339 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1340 register Scalar *x,*b,*v; 1341 1342 VecGetArray_Fast(bb,b); 1343 VecGetArray_Fast(xx,x); 1344 1345 /* forward solve the lower triangular */ 1346 idx = 0; 1347 x[0] = b[idx]; x[1] = b[1+idx]; x[2] = b[2+idx]; x[3] = b[3+idx]; 1348 for ( i=1; i<n; i++ ) { 1349 v = aa + 16*ai[i]; 1350 vi = aj + ai[i]; 1351 nz = diag[i] - ai[i]; 1352 idx = 4*i; 1353 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1354 while (nz--) { 1355 jdx = 4*(*vi++); 1356 x1 = x[jdx];x2 = x[1+jdx];x3 = x[2+jdx];x4 = x[3+jdx]; 1357 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1358 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1359 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1360 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1361 v += 16; 1362 } 1363 x[idx] = sum1; 1364 x[1+idx] = sum2; 1365 x[2+idx] = sum3; 1366 x[3+idx] = sum4; 1367 } 1368 /* backward solve the upper triangular */ 1369 for ( i=n-1; i>=0; i-- ){ 1370 v = aa + 16*diag[i] + 16; 1371 vi = aj + diag[i] + 1; 1372 nz = ai[i+1] - diag[i] - 1; 1373 idt = 4*i; 1374 sum1 = x[idt]; sum2 = x[1+idt]; 1375 sum3 = x[2+idt];sum4 = x[3+idt]; 1376 while (nz--) { 1377 idx = 4*(*vi++); 1378 x1 = x[idx]; x2 = x[1+idx];x3 = x[2+idx]; x4 = x[3+idx]; 1379 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1380 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1381 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1382 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1383 v += 16; 1384 } 1385 v = aa + 16*diag[i]; 1386 x[idt] = v[0]*sum1 + v[4]*sum2 + v[8]*sum3 + v[12]*sum4; 1387 x[1+idt] = v[1]*sum1 + v[5]*sum2 + v[9]*sum3 + v[13]*sum4; 1388 x[2+idt] = v[2]*sum1 + v[6]*sum2 + v[10]*sum3 + v[14]*sum4; 1389 x[3+idt] = v[3]*sum1 + v[7]*sum2 + v[11]*sum3 + v[15]*sum4; 1390 } 1391 1392 VecRestoreArray_Fast(bb,b); 1393 VecRestoreArray_Fast(xx,x); 1394 PLogFlops(2*16*(a->nz) - a->n); 1395 return 0; 1396 } 1397 1398 1399 #undef __FUNC__ 1400 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1401 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1402 { 1403 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1404 IS iscol=a->col,isrow=a->row; 1405 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1406 int *diag = a->diag; 1407 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1408 register Scalar *x,*b,*tmp,*v; 1409 1410 VecGetArray_Fast(bb,b); 1411 VecGetArray_Fast(xx,x); 1412 tmp = a->solve_work; 1413 1414 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1415 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1416 1417 /* forward solve the lower triangular */ 1418 idx = 3*(*r++); 1419 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1420 for ( i=1; i<n; i++ ) { 1421 v = aa + 9*ai[i]; 1422 vi = aj + ai[i]; 1423 nz = diag[i] - ai[i]; 1424 idx = 3*(*r++); 1425 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1426 while (nz--) { 1427 idx = 3*(*vi++); 1428 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1429 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1430 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1431 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1432 v += 9; 1433 } 1434 idx = 3*i; 1435 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1436 } 1437 /* backward solve the upper triangular */ 1438 for ( i=n-1; i>=0; i-- ){ 1439 v = aa + 9*diag[i] + 9; 1440 vi = aj + diag[i] + 1; 1441 nz = ai[i+1] - diag[i] - 1; 1442 idt = 3*i; 1443 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1444 while (nz--) { 1445 idx = 3*(*vi++); 1446 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1447 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1448 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1449 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1450 v += 9; 1451 } 1452 idc = 3*(*c--); 1453 v = aa + 9*diag[i]; 1454 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1455 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1456 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1457 } 1458 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1459 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1460 VecRestoreArray_Fast(bb,b); 1461 VecRestoreArray_Fast(xx,x); 1462 PLogFlops(2*9*(a->nz) - a->n); 1463 return 0; 1464 } 1465 1466 #undef __FUNC__ 1467 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1468 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1469 { 1470 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1471 IS iscol=a->col,isrow=a->row; 1472 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1473 int *diag = a->diag; 1474 Scalar *aa=a->a,sum1,sum2,x1,x2; 1475 register Scalar *x,*b,*tmp,*v; 1476 1477 VecGetArray_Fast(bb,b); 1478 VecGetArray_Fast(xx,x); 1479 tmp = a->solve_work; 1480 1481 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1482 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1483 1484 /* forward solve the lower triangular */ 1485 idx = 2*(*r++); 1486 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1487 for ( i=1; i<n; i++ ) { 1488 v = aa + 4*ai[i]; 1489 vi = aj + ai[i]; 1490 nz = diag[i] - ai[i]; 1491 idx = 2*(*r++); 1492 sum1 = b[idx]; sum2 = b[1+idx]; 1493 while (nz--) { 1494 idx = 2*(*vi++); 1495 x1 = tmp[idx]; x2 = tmp[1+idx]; 1496 sum1 -= v[0]*x1 + v[2]*x2; 1497 sum2 -= v[1]*x1 + v[3]*x2; 1498 v += 4; 1499 } 1500 idx = 2*i; 1501 tmp[idx] = sum1; tmp[1+idx] = sum2; 1502 } 1503 /* backward solve the upper triangular */ 1504 for ( i=n-1; i>=0; i-- ){ 1505 v = aa + 4*diag[i] + 4; 1506 vi = aj + diag[i] + 1; 1507 nz = ai[i+1] - diag[i] - 1; 1508 idt = 2*i; 1509 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1510 while (nz--) { 1511 idx = 2*(*vi++); 1512 x1 = tmp[idx]; x2 = tmp[1+idx]; 1513 sum1 -= v[0]*x1 + v[2]*x2; 1514 sum2 -= v[1]*x1 + v[3]*x2; 1515 v += 4; 1516 } 1517 idc = 2*(*c--); 1518 v = aa + 4*diag[i]; 1519 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1520 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1521 } 1522 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1523 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1524 VecRestoreArray_Fast(bb,b); 1525 VecRestoreArray_Fast(xx,x); 1526 PLogFlops(2*4*(a->nz) - a->n); 1527 return 0; 1528 } 1529 1530 1531 #undef __FUNC__ 1532 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1533 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1534 { 1535 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1536 IS iscol=a->col,isrow=a->row; 1537 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1538 int *diag = a->diag; 1539 Scalar *aa=a->a,sum1; 1540 register Scalar *x,*b,*tmp,*v; 1541 1542 if (!n) return 0; 1543 1544 VecGetArray_Fast(bb,b); 1545 VecGetArray_Fast(xx,x); 1546 tmp = a->solve_work; 1547 1548 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1549 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1550 1551 /* forward solve the lower triangular */ 1552 tmp[0] = b[*r++]; 1553 for ( i=1; i<n; i++ ) { 1554 v = aa + ai[i]; 1555 vi = aj + ai[i]; 1556 nz = diag[i] - ai[i]; 1557 sum1 = b[*r++]; 1558 while (nz--) { 1559 sum1 -= (*v++)*tmp[*vi++]; 1560 } 1561 tmp[i] = sum1; 1562 } 1563 /* backward solve the upper triangular */ 1564 for ( i=n-1; i>=0; i-- ){ 1565 v = aa + diag[i] + 1; 1566 vi = aj + diag[i] + 1; 1567 nz = ai[i+1] - diag[i] - 1; 1568 sum1 = tmp[i]; 1569 while (nz--) { 1570 sum1 -= (*v++)*tmp[*vi++]; 1571 } 1572 x[*c--] = tmp[i] = aa[diag[i]]*sum1; 1573 } 1574 1575 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1576 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1577 VecRestoreArray_Fast(bb,b); 1578 VecRestoreArray_Fast(xx,x); 1579 PLogFlops(2*1*(a->nz) - a->n); 1580 return 0; 1581 } 1582 1583 extern int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat,Vec,Vec); 1584 /* ----------------------------------------------------------------*/ 1585 /* 1586 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1587 except that the data structure of Mat_SeqAIJ is slightly different. 1588 Not a good example of code reuse. 1589 */ 1590 #undef __FUNC__ 1591 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1592 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1593 Mat *fact) 1594 { 1595 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1596 IS isicol; 1597 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1598 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1599 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1600 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1601 PetscTruth col_identity, row_identity; 1602 1603 /* special case that simply copies fill pattern */ 1604 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1605 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1606 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1607 if (levels == 0 && row_identity && col_identity) { 1608 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1609 (*fact)->factor = FACTOR_LU; 1610 b = (Mat_SeqBAIJ *) (*fact)->data; 1611 if (!b->diag) { 1612 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1613 } 1614 b->row = isrow; 1615 b->col = iscol; 1616 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1617 /* 1618 Blocksize 4 has a special faster solver for ILU(0) factorization 1619 with natural ordering 1620 */ 1621 if (b->bs == 4) { 1622 (*fact)->ops.lufactornumeric = MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering; 1623 (*fact)->ops.solve = MatSolve_SeqBAIJ_4_NaturalOrdering; 1624 } 1625 return 0; 1626 } 1627 1628 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1629 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1630 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1631 1632 /* get new row pointers */ 1633 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1634 ainew[0] = 0; 1635 /* don't know how many column pointers are needed so estimate */ 1636 jmax = (int) (f*ai[n] + 1); 1637 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1638 /* ajfill is level of fill for each fill entry */ 1639 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1640 /* fill is a linked list of nonzeros in active row */ 1641 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1642 /* im is level for each filled value */ 1643 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1644 /* dloc is location of diagonal in factor */ 1645 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1646 dloc[0] = 0; 1647 for ( prow=0; prow<n; prow++ ) { 1648 /* first copy previous fill into linked list */ 1649 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1650 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1651 xi = aj + ai[r[prow]]; 1652 fill[n] = n; 1653 while (nz--) { 1654 fm = n; 1655 idx = ic[*xi++]; 1656 do { 1657 m = fm; 1658 fm = fill[m]; 1659 } while (fm < idx); 1660 fill[m] = idx; 1661 fill[idx] = fm; 1662 im[idx] = 0; 1663 } 1664 nzi = 0; 1665 row = fill[n]; 1666 while ( row < prow ) { 1667 incrlev = im[row] + 1; 1668 nz = dloc[row]; 1669 xi = ajnew + ainew[row] + nz; 1670 flev = ajfill + ainew[row] + nz + 1; 1671 nnz = ainew[row+1] - ainew[row] - nz - 1; 1672 if (*xi++ != row) { 1673 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot: try running with -pc_ilu_nonzeros_along_diagonal"); 1674 } 1675 fm = row; 1676 while (nnz-- > 0) { 1677 idx = *xi++; 1678 if (*flev + incrlev > levels) { 1679 flev++; 1680 continue; 1681 } 1682 do { 1683 m = fm; 1684 fm = fill[m]; 1685 } while (fm < idx); 1686 if (fm != idx) { 1687 im[idx] = *flev + incrlev; 1688 fill[m] = idx; 1689 fill[idx] = fm; 1690 fm = idx; 1691 nzf++; 1692 } 1693 else { 1694 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1695 } 1696 flev++; 1697 } 1698 row = fill[row]; 1699 nzi++; 1700 } 1701 /* copy new filled row into permanent storage */ 1702 ainew[prow+1] = ainew[prow] + nzf; 1703 if (ainew[prow+1] > jmax) { 1704 /* allocate a longer ajnew */ 1705 int maxadd; 1706 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1707 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1708 jmax += maxadd; 1709 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1710 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1711 PetscFree(ajnew); 1712 ajnew = xi; 1713 /* allocate a longer ajfill */ 1714 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1715 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1716 PetscFree(ajfill); 1717 ajfill = xi; 1718 realloc++; 1719 } 1720 xi = ajnew + ainew[prow]; 1721 flev = ajfill + ainew[prow]; 1722 dloc[prow] = nzi; 1723 fm = fill[n]; 1724 while (nzf--) { 1725 *xi++ = fm; 1726 *flev++ = im[fm]; 1727 fm = fill[fm]; 1728 } 1729 } 1730 PetscFree(ajfill); 1731 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1732 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1733 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1734 PetscFree(fill); PetscFree(im); 1735 1736 { 1737 double af = ((double)ainew[n])/((double)ai[n]); 1738 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 1739 realloc,f,af); 1740 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Run with -pc_ilu_fill %g or use \n",af); 1741 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:PCILUSetFill(pc,%g);\n",af); 1742 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:for best performance.\n"); 1743 } 1744 1745 /* put together the new matrix */ 1746 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1747 b = (Mat_SeqBAIJ *) (*fact)->data; 1748 PetscFree(b->imax); 1749 b->singlemalloc = 0; 1750 len = bs2*ainew[n]*sizeof(Scalar); 1751 /* the next line frees the default space generated by the Create() */ 1752 PetscFree(b->a); PetscFree(b->ilen); 1753 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1754 b->j = ajnew; 1755 b->i = ainew; 1756 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1757 b->diag = dloc; 1758 b->ilen = 0; 1759 b->imax = 0; 1760 b->row = isrow; 1761 b->col = iscol; 1762 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1763 CHKPTRQ(b->solve_work); 1764 /* In b structure: Free imax, ilen, old a, old j. 1765 Allocate dloc, solve_work, new a, new j */ 1766 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1767 b->maxnz = b->nz = ainew[n]; 1768 (*fact)->factor = FACTOR_LU; 1769 1770 (*fact)->info.factor_mallocs = realloc; 1771 (*fact)->info.fill_ratio_given = f; 1772 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1773 1774 return 0; 1775 } 1776 1777 1778 1779 1780