1 #ifdef PETSC_RCS_HEADER 2 static char vcid[] = "$Id: baijfact.c,v 1.50 1997/07/09 20:55:07 balay 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 printf("greetings\n"); 594 595 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 596 597 for ( i=0; i<n; i++ ) { 598 nz = bi[i+1] - bi[i]; 599 ajtmp = bj + bi[i]; 600 for ( j=0; j<nz; j++ ) { 601 x = rtmp+16*ajtmp[j]; 602 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 603 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 604 } 605 /* load in initial (unfactored row) */ 606 nz = ai[i+1] - ai[i]; 607 ajtmpold = aj + ai[i]; 608 v = aa + 16*ai[i]; 609 for ( j=0; j<nz; j++ ) { 610 x = rtmp+16*ajtmpold[j]; 611 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 612 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 613 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 614 x[14] = v[14]; x[15] = v[15]; 615 v += 16; 616 } 617 row = *ajtmp++; 618 while (row < i) { 619 pc = rtmp + 16*row; 620 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 621 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 622 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 623 p15 = pc[14]; p16 = pc[15]; 624 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 625 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 626 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 627 || p16 != 0.0) { 628 pv = ba + 16*diag_offset[row]; 629 pj = bj + diag_offset[row] + 1; 630 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 631 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 632 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 633 x15 = pv[14]; x16 = pv[15]; 634 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 635 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 636 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 637 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 638 639 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 640 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 641 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 642 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 643 644 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 645 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 646 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 647 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 648 649 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 650 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 651 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 652 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 653 654 nz = bi[row+1] - diag_offset[row] - 1; 655 pv += 16; 656 for (j=0; j<nz; j++) { 657 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 658 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 659 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 660 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 661 x = rtmp + 16*pj[j]; 662 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 663 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 664 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 665 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 666 667 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 668 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 669 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 670 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 671 672 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 673 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 674 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 675 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 676 677 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 678 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 679 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 680 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 681 682 pv += 16; 683 } 684 PLogFlops(128*nz+112); 685 } 686 row = *ajtmp++; 687 } 688 /* finished row so stick it into b->a */ 689 pv = ba + 16*bi[i]; 690 pj = bj + bi[i]; 691 nz = bi[i+1] - bi[i]; 692 for ( j=0; j<nz; j++ ) { 693 x = rtmp+16*pj[j]; 694 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 695 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 696 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 697 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 698 pv += 16; 699 } 700 /* invert diagonal block */ 701 w = ba + 16*diag_offset[i]; 702 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 703 } 704 705 PetscFree(rtmp); 706 C->factor = FACTOR_LU; 707 C->assembled = PETSC_TRUE; 708 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 709 return 0; 710 } 711 712 /* ------------------------------------------------------------*/ 713 /* 714 Version for when blocks are 3 by 3 715 */ 716 #undef __FUNC__ 717 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 718 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 719 { 720 Mat C = *B; 721 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 722 IS iscol = b->col, isrow = b->row, isicol; 723 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 724 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 725 int *diag_offset = b->diag,idx; 726 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 727 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 728 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 729 Scalar *ba = b->a,*aa = a->a; 730 register int *pj; 731 732 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 733 PLogObjectParent(*B,isicol); 734 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 735 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 736 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 737 738 for ( i=0; i<n; i++ ) { 739 nz = bi[i+1] - bi[i]; 740 ajtmp = bj + bi[i]; 741 for ( j=0; j<nz; j++ ) { 742 x = rtmp + 9*ajtmp[j]; 743 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 744 } 745 /* load in initial (unfactored row) */ 746 idx = r[i]; 747 nz = ai[idx+1] - ai[idx]; 748 ajtmpold = aj + ai[idx]; 749 v = aa + 9*ai[idx]; 750 for ( j=0; j<nz; j++ ) { 751 x = rtmp + 9*ic[ajtmpold[j]]; 752 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 753 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 754 v += 9; 755 } 756 row = *ajtmp++; 757 while (row < i) { 758 pc = rtmp + 9*row; 759 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 760 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 761 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 762 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 763 pv = ba + 9*diag_offset[row]; 764 pj = bj + diag_offset[row] + 1; 765 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 766 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 767 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 768 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 769 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 770 771 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 772 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 773 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 774 775 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 776 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 777 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 778 nz = bi[row+1] - diag_offset[row] - 1; 779 pv += 9; 780 for (j=0; j<nz; j++) { 781 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 782 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 783 x = rtmp + 9*pj[j]; 784 x[0] -= m1*x1 + m4*x2 + m7*x3; 785 x[1] -= m2*x1 + m5*x2 + m8*x3; 786 x[2] -= m3*x1 + m6*x2 + m9*x3; 787 788 x[3] -= m1*x4 + m4*x5 + m7*x6; 789 x[4] -= m2*x4 + m5*x5 + m8*x6; 790 x[5] -= m3*x4 + m6*x5 + m9*x6; 791 792 x[6] -= m1*x7 + m4*x8 + m7*x9; 793 x[7] -= m2*x7 + m5*x8 + m8*x9; 794 x[8] -= m3*x7 + m6*x8 + m9*x9; 795 pv += 9; 796 } 797 PLogFlops(54*nz+36); 798 } 799 row = *ajtmp++; 800 } 801 /* finished row so stick it into b->a */ 802 pv = ba + 9*bi[i]; 803 pj = bj + bi[i]; 804 nz = bi[i+1] - bi[i]; 805 for ( j=0; j<nz; j++ ) { 806 x = rtmp + 9*pj[j]; 807 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 808 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 809 pv += 9; 810 } 811 /* invert diagonal block */ 812 w = ba + 9*diag_offset[i]; 813 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 814 } 815 816 PetscFree(rtmp); 817 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 818 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 819 ierr = ISDestroy(isicol); CHKERRQ(ierr); 820 C->factor = FACTOR_LU; 821 C->assembled = PETSC_TRUE; 822 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 823 return 0; 824 } 825 826 /* ------------------------------------------------------------*/ 827 /* 828 Version for when blocks are 2 by 2 829 */ 830 #undef __FUNC__ 831 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 832 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 833 { 834 Mat C = *B; 835 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 836 IS iscol = b->col, isrow = b->row, isicol; 837 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 838 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 839 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 840 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 841 Scalar p1,p2,p3,p4,v_work[2]; 842 Scalar *ba = b->a,*aa = a->a; 843 register int *pj; 844 845 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 846 PLogObjectParent(*B,isicol); 847 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 848 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 849 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 850 851 for ( i=0; i<n; i++ ) { 852 nz = bi[i+1] - bi[i]; 853 ajtmp = bj + bi[i]; 854 for ( j=0; j<nz; j++ ) { 855 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 856 } 857 /* load in initial (unfactored row) */ 858 idx = r[i]; 859 nz = ai[idx+1] - ai[idx]; 860 ajtmpold = aj + ai[idx]; 861 v = aa + 4*ai[idx]; 862 for ( j=0; j<nz; j++ ) { 863 x = rtmp+4*ic[ajtmpold[j]]; 864 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 865 v += 4; 866 } 867 row = *ajtmp++; 868 while (row < i) { 869 pc = rtmp + 4*row; 870 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 871 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 872 pv = ba + 4*diag_offset[row]; 873 pj = bj + diag_offset[row] + 1; 874 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 875 pc[0] = m1 = p1*x1 + p3*x2; 876 pc[1] = m2 = p2*x1 + p4*x2; 877 pc[2] = m3 = p1*x3 + p3*x4; 878 pc[3] = m4 = p2*x3 + p4*x4; 879 nz = bi[row+1] - diag_offset[row] - 1; 880 pv += 4; 881 for (j=0; j<nz; j++) { 882 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 883 x = rtmp + 4*pj[j]; 884 x[0] -= m1*x1 + m3*x2; 885 x[1] -= m2*x1 + m4*x2; 886 x[2] -= m1*x3 + m3*x4; 887 x[3] -= m2*x3 + m4*x4; 888 pv += 4; 889 } 890 PLogFlops(16*nz+12); 891 } 892 row = *ajtmp++; 893 } 894 /* finished row so stick it into b->a */ 895 pv = ba + 4*bi[i]; 896 pj = bj + bi[i]; 897 nz = bi[i+1] - bi[i]; 898 for ( j=0; j<nz; j++ ) { 899 x = rtmp+4*pj[j]; 900 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 901 pv += 4; 902 } 903 /* invert diagonal block */ 904 w = ba + 4*diag_offset[i]; 905 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 906 } 907 908 PetscFree(rtmp); 909 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 910 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 911 ierr = ISDestroy(isicol); CHKERRQ(ierr); 912 C->factor = FACTOR_LU; 913 C->assembled = PETSC_TRUE; 914 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 915 return 0; 916 } 917 918 /* ----------------------------------------------------------- */ 919 /* 920 Version for when blocks are 1 by 1. 921 */ 922 #undef __FUNC__ 923 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 924 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 925 { 926 Mat C = *B; 927 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 928 IS iscol = b->col, isrow = b->row, isicol; 929 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 930 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 931 int *diag_offset = b->diag,diag; 932 register Scalar *pv,*v,*rtmp,multiplier,*pc; 933 Scalar *ba = b->a,*aa = a->a; 934 register int *pj; 935 936 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 937 PLogObjectParent(*B,isicol); 938 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 939 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 940 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 941 942 for ( i=0; i<n; i++ ) { 943 nz = bi[i+1] - bi[i]; 944 ajtmp = bj + bi[i]; 945 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 946 947 /* load in initial (unfactored row) */ 948 nz = ai[r[i]+1] - ai[r[i]]; 949 ajtmpold = aj + ai[r[i]]; 950 v = aa + ai[r[i]]; 951 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 952 953 row = *ajtmp++; 954 while (row < i) { 955 pc = rtmp + row; 956 if (*pc != 0.0) { 957 pv = ba + diag_offset[row]; 958 pj = bj + diag_offset[row] + 1; 959 multiplier = *pc * *pv++; 960 *pc = multiplier; 961 nz = bi[row+1] - diag_offset[row] - 1; 962 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 963 PLogFlops(1+2*nz); 964 } 965 row = *ajtmp++; 966 } 967 /* finished row so stick it into b->a */ 968 pv = ba + bi[i]; 969 pj = bj + bi[i]; 970 nz = bi[i+1] - bi[i]; 971 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 972 diag = diag_offset[i] - bi[i]; 973 /* check pivot entry for current row */ 974 if (pv[diag] == 0.0) { 975 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 976 } 977 pv[diag] = 1.0/pv[diag]; 978 } 979 980 PetscFree(rtmp); 981 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 982 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 983 ierr = ISDestroy(isicol); CHKERRQ(ierr); 984 C->factor = FACTOR_LU; 985 C->assembled = PETSC_TRUE; 986 PLogFlops(b->n); 987 return 0; 988 } 989 990 /* ----------------------------------------------------------- */ 991 #undef __FUNC__ 992 #define __FUNC__ "MatLUFactor_SeqBAIJ" 993 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 994 { 995 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 996 int ierr; 997 Mat C; 998 999 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 1000 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 1001 1002 /* free all the data structures from mat */ 1003 PetscFree(mat->a); 1004 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 1005 if (mat->diag) PetscFree(mat->diag); 1006 if (mat->ilen) PetscFree(mat->ilen); 1007 if (mat->imax) PetscFree(mat->imax); 1008 if (mat->solve_work) PetscFree(mat->solve_work); 1009 if (mat->mult_work) PetscFree(mat->mult_work); 1010 PetscFree(mat); 1011 1012 PetscMemcpy(A,C,sizeof(struct _p_Mat)); 1013 PetscHeaderDestroy(C); 1014 return 0; 1015 } 1016 /* ----------------------------------------------------------- */ 1017 #undef __FUNC__ 1018 #define __FUNC__ "MatSolve_SeqBAIJ_N" 1019 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 1020 { 1021 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1022 IS iscol=a->col,isrow=a->row; 1023 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 1024 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 1025 Scalar *aa=a->a,*sum; 1026 register Scalar *x,*b,*lsum,*tmp,*v; 1027 1028 VecGetArray_Fast(bb,b); 1029 VecGetArray_Fast(xx,x); 1030 tmp = a->solve_work; 1031 1032 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1033 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1034 1035 /* forward solve the lower triangular */ 1036 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 1037 for ( i=1; i<n; i++ ) { 1038 v = aa + bs2*ai[i]; 1039 vi = aj + ai[i]; 1040 nz = a->diag[i] - ai[i]; 1041 sum = tmp + bs*i; 1042 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 1043 while (nz--) { 1044 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 1045 v += bs2; 1046 } 1047 } 1048 /* backward solve the upper triangular */ 1049 lsum = a->solve_work + a->n; 1050 for ( i=n-1; i>=0; i-- ){ 1051 v = aa + bs2*(a->diag[i] + 1); 1052 vi = aj + a->diag[i] + 1; 1053 nz = ai[i+1] - a->diag[i] - 1; 1054 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 1055 while (nz--) { 1056 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 1057 v += bs2; 1058 } 1059 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 1060 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 1061 } 1062 1063 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1064 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1065 VecRestoreArray_Fast(bb,b); 1066 VecRestoreArray_Fast(xx,x); 1067 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 1068 return 0; 1069 } 1070 1071 #undef __FUNC__ 1072 #define __FUNC__ "MatSolve_SeqBAIJ_7" 1073 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 1074 { 1075 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1076 IS iscol=a->col,isrow=a->row; 1077 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1078 int *diag = a->diag; 1079 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 1080 register Scalar *x,*b,*tmp,*v; 1081 1082 VecGetArray_Fast(bb,b); 1083 VecGetArray_Fast(xx,x); 1084 tmp = a->solve_work; 1085 1086 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1087 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1088 1089 /* forward solve the lower triangular */ 1090 idx = 7*(*r++); 1091 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1092 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1093 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 1094 1095 for ( i=1; i<n; i++ ) { 1096 v = aa + 49*ai[i]; 1097 vi = aj + ai[i]; 1098 nz = diag[i] - ai[i]; 1099 idx = 7*(*r++); 1100 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1101 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 1102 while (nz--) { 1103 idx = 7*(*vi++); 1104 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1105 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1106 x6 = tmp[5+idx];x7 = tmp[6+idx]; 1107 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 1108 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 1109 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 1110 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 1111 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 1112 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 1113 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 1114 v += 49; 1115 } 1116 idx = 7*i; 1117 tmp[idx] = sum1;tmp[1+idx] = sum2; 1118 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1119 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 1120 } 1121 /* backward solve the upper triangular */ 1122 for ( i=n-1; i>=0; i-- ){ 1123 v = aa + 49*diag[i] + 49; 1124 vi = aj + diag[i] + 1; 1125 nz = ai[i+1] - diag[i] - 1; 1126 idt = 7*i; 1127 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1128 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1129 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 1130 while (nz--) { 1131 idx = 7*(*vi++); 1132 x1 = tmp[idx]; x2 = tmp[1+idx]; 1133 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1134 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 1135 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 1136 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 1137 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 1138 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 1139 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 1140 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 1141 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 1142 v += 49; 1143 } 1144 idc = 7*(*c--); 1145 v = aa + 49*diag[i]; 1146 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 1147 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 1148 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 1149 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 1150 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 1151 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1152 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1153 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1154 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1155 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1156 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1157 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1158 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1159 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1160 } 1161 1162 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1163 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1164 VecRestoreArray_Fast(bb,b); 1165 VecRestoreArray_Fast(xx,x); 1166 PLogFlops(2*49*(a->nz) - a->n); 1167 return 0; 1168 } 1169 1170 #undef __FUNC__ 1171 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1172 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1173 { 1174 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1175 IS iscol=a->col,isrow=a->row; 1176 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1177 int *diag = a->diag; 1178 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1179 register Scalar *x,*b,*tmp,*v; 1180 1181 VecGetArray_Fast(bb,b); 1182 VecGetArray_Fast(xx,x); 1183 tmp = a->solve_work; 1184 1185 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1186 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1187 1188 /* forward solve the lower triangular */ 1189 idx = 5*(*r++); 1190 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1191 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1192 for ( i=1; i<n; i++ ) { 1193 v = aa + 25*ai[i]; 1194 vi = aj + ai[i]; 1195 nz = diag[i] - ai[i]; 1196 idx = 5*(*r++); 1197 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1198 sum5 = b[4+idx]; 1199 while (nz--) { 1200 idx = 5*(*vi++); 1201 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1202 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1203 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1204 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1205 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1206 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1207 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1208 v += 25; 1209 } 1210 idx = 5*i; 1211 tmp[idx] = sum1;tmp[1+idx] = sum2; 1212 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1213 } 1214 /* backward solve the upper triangular */ 1215 for ( i=n-1; i>=0; i-- ){ 1216 v = aa + 25*diag[i] + 25; 1217 vi = aj + diag[i] + 1; 1218 nz = ai[i+1] - diag[i] - 1; 1219 idt = 5*i; 1220 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1221 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1222 while (nz--) { 1223 idx = 5*(*vi++); 1224 x1 = tmp[idx]; x2 = tmp[1+idx]; 1225 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1226 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1227 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1228 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1229 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1230 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1231 v += 25; 1232 } 1233 idc = 5*(*c--); 1234 v = aa + 25*diag[i]; 1235 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1236 v[15]*sum4+v[20]*sum5; 1237 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1238 v[16]*sum4+v[21]*sum5; 1239 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1240 v[17]*sum4+v[22]*sum5; 1241 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1242 v[18]*sum4+v[23]*sum5; 1243 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1244 v[19]*sum4+v[24]*sum5; 1245 } 1246 1247 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1248 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1249 VecRestoreArray_Fast(bb,b); 1250 VecRestoreArray_Fast(xx,x); 1251 PLogFlops(2*25*(a->nz) - a->n); 1252 return 0; 1253 } 1254 1255 #undef __FUNC__ 1256 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1257 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1258 { 1259 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1260 IS iscol=a->col,isrow=a->row; 1261 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1262 int *diag = a->diag; 1263 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1264 register Scalar *x,*b,*tmp,*v; 1265 1266 VecGetArray_Fast(bb,b); 1267 VecGetArray_Fast(xx,x); 1268 tmp = a->solve_work; 1269 1270 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1271 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1272 1273 /* forward solve the lower triangular */ 1274 idx = 4*(*r++); 1275 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1276 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1277 for ( i=1; i<n; i++ ) { 1278 v = aa + 16*ai[i]; 1279 vi = aj + ai[i]; 1280 nz = diag[i] - ai[i]; 1281 idx = 4*(*r++); 1282 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1283 while (nz--) { 1284 idx = 4*(*vi++); 1285 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1286 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1287 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1288 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1289 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1290 v += 16; 1291 } 1292 idx = 4*i; 1293 tmp[idx] = sum1;tmp[1+idx] = sum2; 1294 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1295 } 1296 /* backward solve the upper triangular */ 1297 for ( i=n-1; i>=0; i-- ){ 1298 v = aa + 16*diag[i] + 16; 1299 vi = aj + diag[i] + 1; 1300 nz = ai[i+1] - diag[i] - 1; 1301 idt = 4*i; 1302 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1303 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1304 while (nz--) { 1305 idx = 4*(*vi++); 1306 x1 = tmp[idx]; x2 = tmp[1+idx]; 1307 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1308 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1309 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1310 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1311 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1312 v += 16; 1313 } 1314 idc = 4*(*c--); 1315 v = aa + 16*diag[i]; 1316 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1317 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1318 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1319 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1320 } 1321 1322 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1323 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1324 VecRestoreArray_Fast(bb,b); 1325 VecRestoreArray_Fast(xx,x); 1326 PLogFlops(2*16*(a->nz) - a->n); 1327 return 0; 1328 } 1329 1330 /* 1331 Special case where the matrix was ILU(0) factored in the natural 1332 ordering. This eliminates the need for the column and row permutation. 1333 */ 1334 #undef __FUNC__ 1335 #define __FUNC__ "MatSolve_SeqBAIJ_4_NaturalOrdering" 1336 int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat A,Vec bb,Vec xx) 1337 { 1338 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1339 int i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt; 1340 int *diag = a->diag,jdx; 1341 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1342 register Scalar *x,*b,*v; 1343 1344 VecGetArray_Fast(bb,b); 1345 VecGetArray_Fast(xx,x); 1346 1347 /* forward solve the lower triangular */ 1348 idx = 0; 1349 x[0] = b[idx]; x[1] = b[1+idx]; x[2] = b[2+idx]; x[3] = b[3+idx]; 1350 for ( i=1; i<n; i++ ) { 1351 v = aa + 16*ai[i]; 1352 vi = aj + ai[i]; 1353 nz = diag[i] - ai[i]; 1354 idx = 4*i; 1355 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1356 while (nz--) { 1357 jdx = 4*(*vi++); 1358 x1 = x[jdx];x2 = x[1+jdx];x3 = x[2+jdx];x4 = x[3+jdx]; 1359 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1360 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1361 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1362 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1363 v += 16; 1364 } 1365 x[idx] = sum1; 1366 x[1+idx] = sum2; 1367 x[2+idx] = sum3; 1368 x[3+idx] = sum4; 1369 } 1370 /* backward solve the upper triangular */ 1371 for ( i=n-1; i>=0; i-- ){ 1372 v = aa + 16*diag[i] + 16; 1373 vi = aj + diag[i] + 1; 1374 nz = ai[i+1] - diag[i] - 1; 1375 idt = 4*i; 1376 sum1 = x[idt]; sum2 = x[1+idt]; 1377 sum3 = x[2+idt];sum4 = x[3+idt]; 1378 while (nz--) { 1379 idx = 4*(*vi++); 1380 x1 = x[idx]; x2 = x[1+idx];x3 = x[2+idx]; x4 = x[3+idx]; 1381 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1382 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1383 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1384 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1385 v += 16; 1386 } 1387 v = aa + 16*diag[i]; 1388 x[idt] = v[0]*sum1 + v[4]*sum2 + v[8]*sum3 + v[12]*sum4; 1389 x[1+idt] = v[1]*sum1 + v[5]*sum2 + v[9]*sum3 + v[13]*sum4; 1390 x[2+idt] = v[2]*sum1 + v[6]*sum2 + v[10]*sum3 + v[14]*sum4; 1391 x[3+idt] = v[3]*sum1 + v[7]*sum2 + v[11]*sum3 + v[15]*sum4; 1392 } 1393 1394 VecRestoreArray_Fast(bb,b); 1395 VecRestoreArray_Fast(xx,x); 1396 PLogFlops(2*16*(a->nz) - a->n); 1397 return 0; 1398 } 1399 1400 1401 #undef __FUNC__ 1402 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1403 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1404 { 1405 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1406 IS iscol=a->col,isrow=a->row; 1407 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1408 int *diag = a->diag; 1409 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1410 register Scalar *x,*b,*tmp,*v; 1411 1412 VecGetArray_Fast(bb,b); 1413 VecGetArray_Fast(xx,x); 1414 tmp = a->solve_work; 1415 1416 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1417 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1418 1419 /* forward solve the lower triangular */ 1420 idx = 3*(*r++); 1421 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1422 for ( i=1; i<n; i++ ) { 1423 v = aa + 9*ai[i]; 1424 vi = aj + ai[i]; 1425 nz = diag[i] - ai[i]; 1426 idx = 3*(*r++); 1427 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1428 while (nz--) { 1429 idx = 3*(*vi++); 1430 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1431 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1432 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1433 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1434 v += 9; 1435 } 1436 idx = 3*i; 1437 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1438 } 1439 /* backward solve the upper triangular */ 1440 for ( i=n-1; i>=0; i-- ){ 1441 v = aa + 9*diag[i] + 9; 1442 vi = aj + diag[i] + 1; 1443 nz = ai[i+1] - diag[i] - 1; 1444 idt = 3*i; 1445 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1446 while (nz--) { 1447 idx = 3*(*vi++); 1448 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1449 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1450 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1451 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1452 v += 9; 1453 } 1454 idc = 3*(*c--); 1455 v = aa + 9*diag[i]; 1456 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1457 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1458 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1459 } 1460 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1461 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1462 VecRestoreArray_Fast(bb,b); 1463 VecRestoreArray_Fast(xx,x); 1464 PLogFlops(2*9*(a->nz) - a->n); 1465 return 0; 1466 } 1467 1468 #undef __FUNC__ 1469 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1470 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1471 { 1472 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1473 IS iscol=a->col,isrow=a->row; 1474 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1475 int *diag = a->diag; 1476 Scalar *aa=a->a,sum1,sum2,x1,x2; 1477 register Scalar *x,*b,*tmp,*v; 1478 1479 VecGetArray_Fast(bb,b); 1480 VecGetArray_Fast(xx,x); 1481 tmp = a->solve_work; 1482 1483 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1484 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1485 1486 /* forward solve the lower triangular */ 1487 idx = 2*(*r++); 1488 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1489 for ( i=1; i<n; i++ ) { 1490 v = aa + 4*ai[i]; 1491 vi = aj + ai[i]; 1492 nz = diag[i] - ai[i]; 1493 idx = 2*(*r++); 1494 sum1 = b[idx]; sum2 = b[1+idx]; 1495 while (nz--) { 1496 idx = 2*(*vi++); 1497 x1 = tmp[idx]; x2 = tmp[1+idx]; 1498 sum1 -= v[0]*x1 + v[2]*x2; 1499 sum2 -= v[1]*x1 + v[3]*x2; 1500 v += 4; 1501 } 1502 idx = 2*i; 1503 tmp[idx] = sum1; tmp[1+idx] = sum2; 1504 } 1505 /* backward solve the upper triangular */ 1506 for ( i=n-1; i>=0; i-- ){ 1507 v = aa + 4*diag[i] + 4; 1508 vi = aj + diag[i] + 1; 1509 nz = ai[i+1] - diag[i] - 1; 1510 idt = 2*i; 1511 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1512 while (nz--) { 1513 idx = 2*(*vi++); 1514 x1 = tmp[idx]; x2 = tmp[1+idx]; 1515 sum1 -= v[0]*x1 + v[2]*x2; 1516 sum2 -= v[1]*x1 + v[3]*x2; 1517 v += 4; 1518 } 1519 idc = 2*(*c--); 1520 v = aa + 4*diag[i]; 1521 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1522 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1523 } 1524 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1525 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1526 VecRestoreArray_Fast(bb,b); 1527 VecRestoreArray_Fast(xx,x); 1528 PLogFlops(2*4*(a->nz) - a->n); 1529 return 0; 1530 } 1531 1532 1533 #undef __FUNC__ 1534 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1535 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1536 { 1537 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1538 IS iscol=a->col,isrow=a->row; 1539 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1540 int *diag = a->diag; 1541 Scalar *aa=a->a,sum1; 1542 register Scalar *x,*b,*tmp,*v; 1543 1544 if (!n) return 0; 1545 1546 VecGetArray_Fast(bb,b); 1547 VecGetArray_Fast(xx,x); 1548 tmp = a->solve_work; 1549 1550 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1551 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1552 1553 /* forward solve the lower triangular */ 1554 tmp[0] = b[*r++]; 1555 for ( i=1; i<n; i++ ) { 1556 v = aa + ai[i]; 1557 vi = aj + ai[i]; 1558 nz = diag[i] - ai[i]; 1559 sum1 = b[*r++]; 1560 while (nz--) { 1561 sum1 -= (*v++)*tmp[*vi++]; 1562 } 1563 tmp[i] = sum1; 1564 } 1565 /* backward solve the upper triangular */ 1566 for ( i=n-1; i>=0; i-- ){ 1567 v = aa + diag[i] + 1; 1568 vi = aj + diag[i] + 1; 1569 nz = ai[i+1] - diag[i] - 1; 1570 sum1 = tmp[i]; 1571 while (nz--) { 1572 sum1 -= (*v++)*tmp[*vi++]; 1573 } 1574 x[*c--] = tmp[i] = aa[diag[i]]*sum1; 1575 } 1576 1577 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1578 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1579 VecRestoreArray_Fast(bb,b); 1580 VecRestoreArray_Fast(xx,x); 1581 PLogFlops(2*1*(a->nz) - a->n); 1582 return 0; 1583 } 1584 1585 extern int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat,Vec,Vec); 1586 /* ----------------------------------------------------------------*/ 1587 /* 1588 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1589 except that the data structure of Mat_SeqAIJ is slightly different. 1590 Not a good example of code reuse. 1591 */ 1592 #undef __FUNC__ 1593 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1594 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1595 Mat *fact) 1596 { 1597 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1598 IS isicol; 1599 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1600 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1601 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1602 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1603 PetscTruth col_identity, row_identity; 1604 1605 /* special case that simply copies fill pattern */ 1606 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1607 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1608 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1609 if (levels == 0 && row_identity && col_identity) { 1610 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1611 (*fact)->factor = FACTOR_LU; 1612 b = (Mat_SeqBAIJ *) (*fact)->data; 1613 if (!b->diag) { 1614 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1615 } 1616 b->row = isrow; 1617 b->col = iscol; 1618 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1619 /* 1620 Blocksize 4 has a special faster solver for ILU(0) factorization 1621 with natural ordering 1622 */ 1623 if (b->bs == 4) { 1624 (*fact)->ops.lufactornumeric = MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering; 1625 (*fact)->ops.solve = MatSolve_SeqBAIJ_4_NaturalOrdering; 1626 } 1627 return 0; 1628 } 1629 1630 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1631 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1632 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1633 1634 /* get new row pointers */ 1635 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1636 ainew[0] = 0; 1637 /* don't know how many column pointers are needed so estimate */ 1638 jmax = (int) (f*ai[n] + 1); 1639 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1640 /* ajfill is level of fill for each fill entry */ 1641 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1642 /* fill is a linked list of nonzeros in active row */ 1643 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1644 /* im is level for each filled value */ 1645 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1646 /* dloc is location of diagonal in factor */ 1647 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1648 dloc[0] = 0; 1649 for ( prow=0; prow<n; prow++ ) { 1650 /* first copy previous fill into linked list */ 1651 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1652 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1653 xi = aj + ai[r[prow]]; 1654 fill[n] = n; 1655 while (nz--) { 1656 fm = n; 1657 idx = ic[*xi++]; 1658 do { 1659 m = fm; 1660 fm = fill[m]; 1661 } while (fm < idx); 1662 fill[m] = idx; 1663 fill[idx] = fm; 1664 im[idx] = 0; 1665 } 1666 nzi = 0; 1667 row = fill[n]; 1668 while ( row < prow ) { 1669 incrlev = im[row] + 1; 1670 nz = dloc[row]; 1671 xi = ajnew + ainew[row] + nz; 1672 flev = ajfill + ainew[row] + nz + 1; 1673 nnz = ainew[row+1] - ainew[row] - nz - 1; 1674 if (*xi++ != row) { 1675 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot: try running with -pc_ilu_nonzeros_along_diagonal"); 1676 } 1677 fm = row; 1678 while (nnz-- > 0) { 1679 idx = *xi++; 1680 if (*flev + incrlev > levels) { 1681 flev++; 1682 continue; 1683 } 1684 do { 1685 m = fm; 1686 fm = fill[m]; 1687 } while (fm < idx); 1688 if (fm != idx) { 1689 im[idx] = *flev + incrlev; 1690 fill[m] = idx; 1691 fill[idx] = fm; 1692 fm = idx; 1693 nzf++; 1694 } 1695 else { 1696 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1697 } 1698 flev++; 1699 } 1700 row = fill[row]; 1701 nzi++; 1702 } 1703 /* copy new filled row into permanent storage */ 1704 ainew[prow+1] = ainew[prow] + nzf; 1705 if (ainew[prow+1] > jmax) { 1706 /* allocate a longer ajnew */ 1707 int maxadd; 1708 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1709 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1710 jmax += maxadd; 1711 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1712 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1713 PetscFree(ajnew); 1714 ajnew = xi; 1715 /* allocate a longer ajfill */ 1716 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1717 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1718 PetscFree(ajfill); 1719 ajfill = xi; 1720 realloc++; 1721 } 1722 xi = ajnew + ainew[prow]; 1723 flev = ajfill + ainew[prow]; 1724 dloc[prow] = nzi; 1725 fm = fill[n]; 1726 while (nzf--) { 1727 *xi++ = fm; 1728 *flev++ = im[fm]; 1729 fm = fill[fm]; 1730 } 1731 } 1732 PetscFree(ajfill); 1733 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1734 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1735 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1736 PetscFree(fill); PetscFree(im); 1737 1738 { 1739 double af = ((double)ainew[n])/((double)ai[n]); 1740 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 1741 realloc,f,af); 1742 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Run with -pc_ilu_fill %g or use \n",af); 1743 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:PCILUSetFill(pc,%g);\n",af); 1744 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:for best performance.\n"); 1745 } 1746 1747 /* put together the new matrix */ 1748 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1749 b = (Mat_SeqBAIJ *) (*fact)->data; 1750 PetscFree(b->imax); 1751 b->singlemalloc = 0; 1752 len = bs2*ainew[n]*sizeof(Scalar); 1753 /* the next line frees the default space generated by the Create() */ 1754 PetscFree(b->a); PetscFree(b->ilen); 1755 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1756 b->j = ajnew; 1757 b->i = ainew; 1758 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1759 b->diag = dloc; 1760 b->ilen = 0; 1761 b->imax = 0; 1762 b->row = isrow; 1763 b->col = iscol; 1764 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1765 CHKPTRQ(b->solve_work); 1766 /* In b structure: Free imax, ilen, old a, old j. 1767 Allocate dloc, solve_work, new a, new j */ 1768 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1769 b->maxnz = b->nz = ainew[n]; 1770 (*fact)->factor = FACTOR_LU; 1771 1772 (*fact)->info.factor_mallocs = realloc; 1773 (*fact)->info.fill_ratio_given = f; 1774 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1775 1776 return 0; 1777 } 1778 1779 1780 1781 1782