1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.14 1996/04/12 23:31:32 balay Exp bsmith $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "baij.h" 9 10 /* 11 The symbolic factorization code is identical to that for AIJ format, 12 except for very small changes since this is now a SeqBAIJ datastructure. 13 NOT good code reuse. 14 */ 15 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 16 { 17 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 18 IS isicol; 19 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 20 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 21 int *idnew, idx, row,m,fm, nnz, nzi,len, realloc = 0,nzbd,*im; 22 23 if (a->m != a->n) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must be square"); 24 if (!isrow) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must have row permutation"); 25 if (!iscol) SETERRQ(1,"MatLUFactorSymbolic_SeqBAIJ:Must have col. permutation"); 26 27 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 28 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 29 30 /* get new row pointers */ 31 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 32 ainew[0] = 0; 33 /* don't know how many column pointers are needed so estimate */ 34 jmax = (int) (f*ai[n] + 1); 35 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 36 /* fill is a linked list of nonzeros in active row */ 37 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 38 im = fill + n + 1; 39 /* idnew is location of diagonal in factor */ 40 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 41 idnew[0] = 0; 42 43 for ( i=0; i<n; i++ ) { 44 /* first copy previous fill into linked list */ 45 nnz = nz = ai[r[i]+1] - ai[r[i]]; 46 ajtmp = aj + ai[r[i]]; 47 fill[n] = n; 48 while (nz--) { 49 fm = n; 50 idx = ic[*ajtmp++]; 51 do { 52 m = fm; 53 fm = fill[m]; 54 } while (fm < idx); 55 fill[m] = idx; 56 fill[idx] = fm; 57 } 58 row = fill[n]; 59 while ( row < i ) { 60 ajtmp = ajnew + idnew[row] + 1; 61 nzbd = 1 + idnew[row] - ainew[row]; 62 nz = im[row] - nzbd; 63 fm = row; 64 while (nz-- > 0) { 65 idx = *ajtmp++; 66 nzbd++; 67 if (idx == i) im[row] = nzbd; 68 do { 69 m = fm; 70 fm = fill[m]; 71 } while (fm < idx); 72 if (fm != idx) { 73 fill[m] = idx; 74 fill[idx] = fm; 75 fm = idx; 76 nnz++; 77 } 78 } 79 row = fill[row]; 80 } 81 /* copy new filled row into permanent storage */ 82 ainew[i+1] = ainew[i] + nnz; 83 if (ainew[i+1] > jmax+1) { 84 /* allocate a longer ajnew */ 85 int maxadd; 86 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 87 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 88 jmax += maxadd; 89 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 90 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 91 PetscFree(ajnew); 92 ajnew = ajtmp; 93 realloc++; /* count how many times we realloc */ 94 } 95 ajtmp = ajnew + ainew[i]; 96 fm = fill[n]; 97 nzi = 0; 98 im[i] = nnz; 99 while (nnz--) { 100 if (fm < i) nzi++; 101 *ajtmp++ = fm; 102 fm = fill[fm]; 103 } 104 idnew[i] = ainew[i] + nzi; 105 } 106 107 PLogInfo(A, 108 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 109 realloc,f,((double)ainew[n])/((double)ai[i])); 110 111 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 112 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 113 114 PetscFree(fill); 115 116 /* put together the new matrix */ 117 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 118 PLogObjectParent(*B,isicol); 119 ierr = ISDestroy(isicol); CHKERRQ(ierr); 120 b = (Mat_SeqBAIJ *) (*B)->data; 121 PetscFree(b->imax); 122 b->singlemalloc = 0; 123 len = ainew[n]*sizeof(Scalar); 124 /* the next line frees the default space generated by the Create() */ 125 PetscFree(b->a); PetscFree(b->ilen); 126 b->a = (Scalar *) PetscMalloc( len*bs2 ); CHKPTRQ(b->a); 127 b->j = ajnew; 128 b->i = ainew; 129 b->diag = idnew; 130 b->ilen = 0; 131 b->imax = 0; 132 b->row = isrow; 133 b->col = iscol; 134 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 135 CHKPTRQ(b->solve_work); 136 /* In b structure: Free imax, ilen, old a, old j. 137 Allocate idnew, solve_work, new a, new j */ 138 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 139 b->maxnz = b->nz = ainew[n]; 140 141 return 0; 142 } 143 144 #include "pinclude/plapack.h" 145 int Linpack_DGEFA(Scalar *,int, int *); 146 int Linpack_DGEDI(Scalar *,int, int *,Scalar *); 147 148 149 /* ----------------------------------------------------------- */ 150 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 151 { 152 Mat C = *B; 153 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 154 IS iscol = b->col, isrow = b->row, isicol; 155 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 156 int *ajtmpold, *ajtmp, nz, row, bslog; 157 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs,*v_pivots; 158 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 159 Scalar one = 1.0, zero = 0.0, mone = -1.0; 160 register int *pj; 161 162 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 163 PLogObjectParent(*B,isicol); 164 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 165 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 166 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 167 168 /* generate work space needed by dense LU factorization */ 169 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 170 CHKPTRQ(v_work); 171 multiplier = v_work + bs; 172 v_pivots = (int *) (multiplier + bs2); 173 174 /* flops in while loop */ 175 bslog = 2*bs*bs2; 176 177 for ( i=0; i<n; i++ ) { 178 nz = ai[i+1] - ai[i]; 179 ajtmp = aj + ai[i]; 180 for ( j=0; j<nz; j++ ) { 181 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 182 } 183 /* load in initial (unfactored row) */ 184 nz = a->i[r[i]+1] - a->i[r[i]]; 185 ajtmpold = a->j + a->i[r[i]]; 186 v = a->a + bs2*a->i[r[i]]; 187 for ( j=0; j<nz; j++ ) { 188 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 189 } 190 row = *ajtmp++; 191 while (row < i) { 192 pc = rtmp + bs2*row; 193 /* if (*pc) { */ 194 pv = b->a + bs2*diag_offset[row]; 195 pj = b->j + diag_offset[row] + 1; 196 BLgemm_("N","N",&bs,&bs,&bs,&one,pc,&bs,pv,&bs,&zero, 197 multiplier,&bs); 198 PetscMemcpy(pc,multiplier,bs2*sizeof(Scalar)); 199 nz = ai[row+1] - diag_offset[row] - 1; 200 pv += bs2; 201 for (j=0; j<nz; j++) { 202 BLgemm_("N","N",&bs,&bs,&bs,&mone,multiplier,&bs,pv+bs2*j,&bs, 203 &one,rtmp+bs2*pj[j],&bs); 204 } 205 PLogFlops(bslog*(nz+1)-bs); 206 /* } */ 207 row = *ajtmp++; 208 } 209 /* finished row so stick it into b->a */ 210 pv = b->a + bs2*ai[i]; 211 pj = b->j + ai[i]; 212 nz = ai[i+1] - ai[i]; 213 for ( j=0; j<nz; j++ ) { 214 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 215 } 216 diag = diag_offset[i] - ai[i]; 217 /* invert diagonal block */ 218 w = pv + bs2*diag; 219 220 /* Old version that used LAPACK, make v_work bs2 length for this 221 LAgetrf_(&bs,&bs,w,&bs,v_pivots,&info); CHKERRQ(info); 222 PetscMemzero(v_work,bs2*sizeof(Scalar)); 223 for ( j=0; j<bs; j++ ) { v_work[j + bs*j] = 1.0; } 224 LAgetrs_("N",&bs,&bs,w,&bs,v_pivots,v_work,&bs, &info);CHKERRQ(info); 225 PetscMemcpy(w,v_work,bs2*sizeof(Scalar)); 226 */ 227 228 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 229 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 230 } 231 232 PetscFree(rtmp); PetscFree(v_work); 233 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 234 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 235 ierr = ISDestroy(isicol); CHKERRQ(ierr); 236 C->factor = FACTOR_LU; 237 C->assembled = PETSC_TRUE; 238 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 239 return 0; 240 } 241 /* ------------------------------------------------------------*/ 242 /* 243 Version for when blocks are 5 by 5 244 */ 245 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 246 { 247 Mat C = *B; 248 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 249 IS iscol = b->col, isrow = b->row, isicol; 250 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 251 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 252 int *diag_offset = b->diag,bs = 5,idx; 253 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 254 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 255 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 256 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 257 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 258 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 259 register int *pj; 260 261 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 262 PLogObjectParent(*B,isicol); 263 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 264 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 265 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 266 267 for ( i=0; i<n; i++ ) { 268 nz = ai[i+1] - ai[i]; 269 ajtmp = aj + ai[i]; 270 for ( j=0; j<nz; j++ ) { 271 x = rtmp+25*ajtmp[j]; 272 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 273 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 274 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 275 } 276 /* load in initial (unfactored row) */ 277 idx = r[i]; 278 nz = a->i[idx+1] - a->i[idx]; 279 ajtmpold = a->j + a->i[idx]; 280 v = a->a + 25*a->i[idx]; 281 for ( j=0; j<nz; j++ ) { 282 x = rtmp+25*ic[ajtmpold[j]]; 283 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 284 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 285 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 286 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 287 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 288 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 289 v += 25; 290 } 291 row = *ajtmp++; 292 while (row < i) { 293 pc = rtmp + 25*row; 294 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 295 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 296 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 297 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 298 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 299 p25 = pc[24]; 300 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 301 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 302 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 303 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 304 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 305 p23 != 0.0 || p24 != 0.0) { 306 pv = b->a + 25*diag_offset[row]; 307 pj = b->j + diag_offset[row] + 1; 308 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 309 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 310 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 311 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 312 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 313 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 314 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 315 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 316 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 317 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 318 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 319 320 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 321 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 322 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 323 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 324 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 325 326 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 327 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 328 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 329 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 330 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 331 332 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 333 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 334 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 335 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 336 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 337 338 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 339 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 340 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 341 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 342 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 343 344 nz = ai[row+1] - diag_offset[row] - 1; 345 pv += 25; 346 for (j=0; j<nz; j++) { 347 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 348 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 349 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 350 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 351 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 352 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 353 x = rtmp + 25*pj[j]; 354 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 355 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 356 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 357 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 358 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 359 360 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 361 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 362 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 363 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 364 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 365 366 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 367 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 368 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 369 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 370 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 371 372 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 373 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 374 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 375 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 376 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 377 378 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 379 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 380 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 381 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 382 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 383 384 pv += 25; 385 } 386 PLogFlops(250*nz+225); 387 } 388 row = *ajtmp++; 389 } 390 /* finished row so stick it into b->a */ 391 pv = b->a + 25*ai[i]; 392 pj = b->j + ai[i]; 393 nz = ai[i+1] - ai[i]; 394 for ( j=0; j<nz; j++ ) { 395 x = rtmp+25*pj[j]; 396 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 397 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 398 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 399 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 400 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 401 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 402 pv += 25; 403 } 404 /* invert diagonal block */ 405 w = b->a + 25*diag_offset[i]; 406 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 407 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 408 } 409 410 PetscFree(rtmp); 411 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 412 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 413 ierr = ISDestroy(isicol); CHKERRQ(ierr); 414 C->factor = FACTOR_LU; 415 C->assembled = PETSC_TRUE; 416 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 417 return 0; 418 } 419 420 /* ------------------------------------------------------------*/ 421 /* 422 Version for when blocks are 4 by 4 423 */ 424 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 425 { 426 Mat C = *B; 427 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 428 IS iscol = b->col, isrow = b->row, isicol; 429 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 430 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 431 int *diag_offset = b->diag,bs = 4,idx; 432 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 433 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 434 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 435 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 436 Scalar m13,m14,m15,m16; 437 register int *pj; 438 439 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 440 PLogObjectParent(*B,isicol); 441 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 442 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 443 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 444 445 for ( i=0; i<n; i++ ) { 446 nz = ai[i+1] - ai[i]; 447 ajtmp = aj + ai[i]; 448 for ( j=0; j<nz; j++ ) { 449 x = rtmp+16*ajtmp[j]; 450 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 451 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 452 } 453 /* load in initial (unfactored row) */ 454 idx = r[i]; 455 nz = a->i[idx+1] - a->i[idx]; 456 ajtmpold = a->j + a->i[idx]; 457 v = a->a + 16*a->i[idx]; 458 for ( j=0; j<nz; j++ ) { 459 x = rtmp+16*ic[ajtmpold[j]]; 460 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 461 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 462 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 463 x[14] = v[14]; x[15] = v[15]; 464 v += 16; 465 } 466 row = *ajtmp++; 467 while (row < i) { 468 pc = rtmp + 16*row; 469 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 470 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 471 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 472 p15 = pc[14]; p16 = pc[15]; 473 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 474 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 475 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 476 || p16 != 0.0) { 477 pv = b->a + 16*diag_offset[row]; 478 pj = b->j + diag_offset[row] + 1; 479 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 480 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 481 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 482 x15 = pv[14]; x16 = pv[15]; 483 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 484 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 485 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 486 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 487 488 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 489 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 490 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 491 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 492 493 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 494 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 495 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 496 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 497 498 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 499 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 500 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 501 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 502 503 nz = ai[row+1] - diag_offset[row] - 1; 504 pv += 16; 505 for (j=0; j<nz; j++) { 506 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 507 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 508 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 509 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 510 x = rtmp + 16*pj[j]; 511 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 512 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 513 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 514 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 515 516 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 517 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 518 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 519 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 520 521 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 522 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 523 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 524 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 525 526 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 527 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 528 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 529 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 530 531 pv += 16; 532 } 533 PLogFlops(128*nz+112); 534 } 535 row = *ajtmp++; 536 } 537 /* finished row so stick it into b->a */ 538 pv = b->a + 16*ai[i]; 539 pj = b->j + ai[i]; 540 nz = ai[i+1] - ai[i]; 541 for ( j=0; j<nz; j++ ) { 542 x = rtmp+16*pj[j]; 543 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 544 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 545 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 546 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 547 pv += 16; 548 } 549 /* invert diagonal block */ 550 w = b->a + 16*diag_offset[i]; 551 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 552 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 553 } 554 555 PetscFree(rtmp); 556 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 557 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 558 ierr = ISDestroy(isicol); CHKERRQ(ierr); 559 C->factor = FACTOR_LU; 560 C->assembled = PETSC_TRUE; 561 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 562 return 0; 563 } 564 /* ------------------------------------------------------------*/ 565 /* 566 Version for when blocks are 3 by 3 567 */ 568 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 569 { 570 Mat C = *B; 571 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 572 IS iscol = b->col, isrow = b->row, isicol; 573 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 574 int *ajtmpold, *ajtmp, nz, row, v_pivots[3]; 575 int *diag_offset = b->diag,bs = 3,idx; 576 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 577 Scalar p1,p2,p3,p4,v_work[3],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 578 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 579 register int *pj; 580 581 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 582 PLogObjectParent(*B,isicol); 583 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 584 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 585 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 586 587 for ( i=0; i<n; i++ ) { 588 nz = ai[i+1] - ai[i]; 589 ajtmp = aj + ai[i]; 590 for ( j=0; j<nz; j++ ) { 591 x = rtmp+9*ajtmp[j]; 592 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 593 } 594 /* load in initial (unfactored row) */ 595 idx = r[i]; 596 nz = a->i[idx+1] - a->i[idx]; 597 ajtmpold = a->j + a->i[idx]; 598 v = a->a + 9*a->i[idx]; 599 for ( j=0; j<nz; j++ ) { 600 x = rtmp+9*ic[ajtmpold[j]]; 601 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 602 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 603 v += 9; 604 } 605 row = *ajtmp++; 606 while (row < i) { 607 pc = rtmp + 9*row; 608 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 609 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 610 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 611 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 612 pv = b->a + 9*diag_offset[row]; 613 pj = b->j + diag_offset[row] + 1; 614 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 615 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 616 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 617 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 618 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 619 620 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 621 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 622 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 623 624 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 625 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 626 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 627 nz = ai[row+1] - diag_offset[row] - 1; 628 pv += 9; 629 for (j=0; j<nz; j++) { 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 x = rtmp + 9*pj[j]; 633 x[0] -= m1*x1 + m4*x2 + m7*x3; 634 x[1] -= m2*x1 + m5*x2 + m8*x3; 635 x[2] -= m3*x1 + m6*x2 + m9*x3; 636 637 x[3] -= m1*x4 + m4*x5 + m7*x6; 638 x[4] -= m2*x4 + m5*x5 + m8*x6; 639 x[5] -= m3*x4 + m6*x5 + m9*x6; 640 641 x[6] -= m1*x7 + m4*x8 + m7*x9; 642 x[7] -= m2*x7 + m5*x8 + m8*x9; 643 x[8] -= m3*x7 + m6*x8 + m9*x9; 644 pv += 9; 645 } 646 PLogFlops(54*nz+36); 647 } 648 row = *ajtmp++; 649 } 650 /* finished row so stick it into b->a */ 651 pv = b->a + 9*ai[i]; 652 pj = b->j + ai[i]; 653 nz = ai[i+1] - ai[i]; 654 for ( j=0; j<nz; j++ ) { 655 x = rtmp+9*pj[j]; 656 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 657 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 658 pv += 9; 659 } 660 /* invert diagonal block */ 661 w = b->a + 9*diag_offset[i]; 662 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 663 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 664 } 665 666 PetscFree(rtmp); 667 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 668 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 669 ierr = ISDestroy(isicol); CHKERRQ(ierr); 670 C->factor = FACTOR_LU; 671 C->assembled = PETSC_TRUE; 672 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 673 return 0; 674 } 675 676 /* ------------------------------------------------------------*/ 677 /* 678 Version for when blocks are 2 by 2 679 */ 680 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 681 { 682 Mat C = *B; 683 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 684 IS iscol = b->col, isrow = b->row, isicol; 685 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 686 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 687 int *diag_offset=b->diag,bs = 2,idx; 688 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 689 Scalar p1,p2,p3,p4,v_work[2]; 690 register int *pj; 691 692 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 693 PLogObjectParent(*B,isicol); 694 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 695 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 696 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 697 698 for ( i=0; i<n; i++ ) { 699 nz = ai[i+1] - ai[i]; 700 ajtmp = aj + ai[i]; 701 for ( j=0; j<nz; j++ ) { 702 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 703 } 704 /* load in initial (unfactored row) */ 705 idx = r[i]; 706 nz = a->i[idx+1] - a->i[idx]; 707 ajtmpold = a->j + a->i[idx]; 708 v = a->a + 4*a->i[idx]; 709 for ( j=0; j<nz; j++ ) { 710 x = rtmp+4*ic[ajtmpold[j]]; 711 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 712 v += 4; 713 } 714 row = *ajtmp++; 715 while (row < i) { 716 pc = rtmp + 4*row; 717 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 718 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 719 pv = b->a + 4*diag_offset[row]; 720 pj = b->j + diag_offset[row] + 1; 721 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 722 pc[0] = m1 = p1*x1 + p3*x2; 723 pc[1] = m2 = p2*x1 + p4*x2; 724 pc[2] = m3 = p1*x3 + p3*x4; 725 pc[3] = m4 = p2*x3 + p4*x4; 726 nz = ai[row+1] - diag_offset[row] - 1; 727 pv += 4; 728 for (j=0; j<nz; j++) { 729 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 730 x = rtmp + 4*pj[j]; 731 x[0] -= m1*x1 + m3*x2; 732 x[1] -= m2*x1 + m4*x2; 733 x[2] -= m1*x3 + m3*x4; 734 x[3] -= m2*x3 + m4*x4; 735 pv += 4; 736 } 737 PLogFlops(16*nz+12); 738 } 739 row = *ajtmp++; 740 } 741 /* finished row so stick it into b->a */ 742 pv = b->a + 4*ai[i]; 743 pj = b->j + ai[i]; 744 nz = ai[i+1] - ai[i]; 745 for ( j=0; j<nz; j++ ) { 746 x = rtmp+4*pj[j]; 747 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 748 pv += 4; 749 } 750 /* invert diagonal block */ 751 w = b->a + 4*diag_offset[i]; 752 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 753 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 754 } 755 756 PetscFree(rtmp); 757 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 758 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 759 ierr = ISDestroy(isicol); CHKERRQ(ierr); 760 C->factor = FACTOR_LU; 761 C->assembled = PETSC_TRUE; 762 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 763 return 0; 764 } 765 766 /* ----------------------------------------------------------- */ 767 /* 768 Version for when blocks are 1 by 1. 769 */ 770 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 771 { 772 Mat C = *B; 773 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 774 IS iscol = b->col, isrow = b->row, isicol; 775 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 776 int *ajtmpold, *ajtmp, nz, row; 777 int *diag_offset = b->diag,diag; 778 register Scalar *pv,*v,*rtmp,multiplier,*pc; 779 register int *pj; 780 781 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 782 PLogObjectParent(*B,isicol); 783 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 784 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 785 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 786 787 for ( i=0; i<n; i++ ) { 788 nz = ai[i+1] - ai[i]; 789 ajtmp = aj + ai[i]; 790 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 791 792 /* load in initial (unfactored row) */ 793 nz = a->i[r[i]+1] - a->i[r[i]]; 794 ajtmpold = a->j + a->i[r[i]]; 795 v = a->a + a->i[r[i]]; 796 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 797 798 row = *ajtmp++; 799 while (row < i) { 800 pc = rtmp + row; 801 if (*pc != 0.0) { 802 pv = b->a + diag_offset[row]; 803 pj = b->j + diag_offset[row] + 1; 804 multiplier = *pc * *pv++; 805 *pc = multiplier; 806 nz = ai[row+1] - diag_offset[row] - 1; 807 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 808 PLogFlops(1+2*nz); 809 } 810 row = *ajtmp++; 811 } 812 /* finished row so stick it into b->a */ 813 pv = b->a + ai[i]; 814 pj = b->j + ai[i]; 815 nz = ai[i+1] - ai[i]; 816 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 817 diag = diag_offset[i] - ai[i]; 818 /* check pivot entry for current row */ 819 if (pv[diag] == 0.0) { 820 SETERRQ(1,"MatLUFactorNumeric_SeqAIJ:Zero pivot"); 821 } 822 pv[diag] = 1.0/pv[diag]; 823 } 824 825 PetscFree(rtmp); 826 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 827 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 828 ierr = ISDestroy(isicol); CHKERRQ(ierr); 829 C->factor = FACTOR_LU; 830 C->assembled = PETSC_TRUE; 831 PLogFlops(b->n); 832 return 0; 833 } 834 835 /* ----------------------------------------------------------- */ 836 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 837 { 838 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 839 int ierr; 840 Mat C; 841 842 ierr = MatLUFactorSymbolic_SeqBAIJ(A,row,col,f,&C); CHKERRQ(ierr); 843 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 844 845 /* free all the data structures from mat */ 846 PetscFree(mat->a); 847 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 848 if (mat->diag) PetscFree(mat->diag); 849 if (mat->ilen) PetscFree(mat->ilen); 850 if (mat->imax) PetscFree(mat->imax); 851 if (mat->solve_work) PetscFree(mat->solve_work); 852 PetscFree(mat); 853 854 PetscMemcpy(A,C,sizeof(struct _Mat)); 855 PetscHeaderDestroy(C); 856 return 0; 857 } 858 /* ----------------------------------------------------------- */ 859 static int MatSolveAdd_SeqBAIJ_Private(Mat A,Vec bb,Vec yy,Vec xx) 860 { 861 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 862 IS iscol=a->col,isrow=a->row; 863 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 864 int nz,bs=a->bs,bs2=a->bs2,idx,idt,idc,_One=1,m=a->m; 865 Scalar *xa,*ba,*aa=a->a,*sum,_DOne=1.0,_DMOne=-1.0; 866 Scalar _DZero=0.0,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5,*y; 867 register Scalar *x,*b,*lsum,*tmp,*v; 868 869 if (A->factor != FACTOR_LU) SETERRQ(1,"MatSolveAdd_SeqBAIJ_Private:Not for unfactored matrix"); 870 871 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 872 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 873 tmp = a->solve_work; 874 875 if (yy==PETSC_NULL) PetscMemzero(x,m*sizeof(Scalar)); /* MatSolve() */ 876 else if (xx!=yy){ 877 ierr = VecGetArray(yy,&y); CHKERRQ(ierr); 878 PetscMemcpy(x,y,m*sizeof(Scalar)); 879 ierr = VecRestoreArray(yy,&y); CHKERRQ(ierr); 880 } 881 882 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 883 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 884 885 switch (bs) { 886 case 1: 887 /* forward solve the lower triangular */ 888 tmp[0] = b[*r++]; 889 for ( i=1; i<n; i++ ) { 890 v = aa + ai[i]; 891 vi = aj + ai[i]; 892 nz = a->diag[i] - ai[i]; 893 sum1 = b[*r++]; 894 while (nz--) { 895 sum1 -= (*v++)*tmp[*vi++]; 896 } 897 tmp[i] = sum1; 898 } 899 /* backward solve the upper triangular */ 900 for ( i=n-1; i>=0; i-- ){ 901 v = aa + a->diag[i] + 1; 902 vi = aj + a->diag[i] + 1; 903 nz = ai[i+1] - a->diag[i] - 1; 904 sum1 = tmp[i]; 905 while (nz--) { 906 sum1 -= (*v++)*tmp[*vi++]; 907 } 908 x[*c--] += tmp[i] = aa[a->diag[i]]*sum1; 909 } 910 break; 911 case 2: 912 /* forward solve the lower triangular */ 913 idx = 2*(*r++); 914 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 915 for ( i=1; i<n; i++ ) { 916 v = aa + 4*ai[i]; 917 vi = aj + ai[i]; 918 nz = a->diag[i] - ai[i]; 919 idx = 2*(*r++); 920 sum1 = b[idx]; sum2 = b[1+idx]; 921 while (nz--) { 922 idx = 2*(*vi++); 923 x1 = tmp[idx]; x2 = tmp[1+idx]; 924 sum1 -= v[0]*x1 + v[2]*x2; 925 sum2 -= v[1]*x1 + v[3]*x2; 926 v += 4; 927 } 928 idx = 2*i; 929 tmp[idx] = sum1; tmp[1+idx] = sum2; 930 } 931 /* backward solve the upper triangular */ 932 for ( i=n-1; i>=0; i-- ){ 933 v = aa + 4*a->diag[i] + 4; 934 vi = aj + a->diag[i] + 1; 935 nz = ai[i+1] - a->diag[i] - 1; 936 idt = 2*i; 937 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 938 while (nz--) { 939 idx = 2*(*vi++); 940 x1 = tmp[idx]; x2 = tmp[1+idx]; 941 sum1 -= v[0]*x1 + v[2]*x2; 942 sum2 -= v[1]*x1 + v[3]*x2; 943 v += 4; 944 } 945 idc = 2*(*c--); 946 v = aa + 4*a->diag[i]; 947 x[idc] += tmp[idt] = v[0]*sum1 + v[2]*sum2; 948 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 949 } 950 break; 951 case 3: 952 /* forward solve the lower triangular */ 953 idx = 3*(*r++); 954 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 955 for ( i=1; i<n; i++ ) { 956 v = aa + 9*ai[i]; 957 vi = aj + ai[i]; 958 nz = a->diag[i] - ai[i]; 959 idx = 3*(*r++); 960 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 961 while (nz--) { 962 idx = 3*(*vi++); 963 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 964 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 965 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 966 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 967 v += 9; 968 } 969 idx = 3*i; 970 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 971 } 972 /* backward solve the upper triangular */ 973 for ( i=n-1; i>=0; i-- ){ 974 v = aa + 9*a->diag[i] + 9; 975 vi = aj + a->diag[i] + 1; 976 nz = ai[i+1] - a->diag[i] - 1; 977 idt = 3*i; 978 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 979 while (nz--) { 980 idx = 3*(*vi++); 981 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 982 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 983 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 984 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 985 v += 9; 986 } 987 idc = 3*(*c--); 988 v = aa + 9*a->diag[i]; 989 x[idc] += tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 990 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 991 x[2+idc] += tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 992 } 993 break; 994 case 4: 995 /* forward solve the lower triangular */ 996 idx = 4*(*r++); 997 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 998 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 999 for ( i=1; i<n; i++ ) { 1000 v = aa + 16*ai[i]; 1001 vi = aj + ai[i]; 1002 nz = a->diag[i] - ai[i]; 1003 idx = 4*(*r++); 1004 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1005 while (nz--) { 1006 idx = 4*(*vi++); 1007 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1008 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1009 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1010 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1011 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1012 v += 16; 1013 } 1014 idx = 4*i; 1015 tmp[idx] = sum1;tmp[1+idx] = sum2; 1016 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1017 } 1018 /* backward solve the upper triangular */ 1019 for ( i=n-1; i>=0; i-- ){ 1020 v = aa + 16*a->diag[i] + 16; 1021 vi = aj + a->diag[i] + 1; 1022 nz = ai[i+1] - a->diag[i] - 1; 1023 idt = 4*i; 1024 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1025 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1026 while (nz--) { 1027 idx = 4*(*vi++); 1028 x1 = tmp[idx]; x2 = tmp[1+idx]; 1029 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1030 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1031 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1032 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1033 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1034 v += 16; 1035 } 1036 idc = 4*(*c--); 1037 v = aa + 16*a->diag[i]; 1038 x[idc] += tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1039 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1040 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1041 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1042 } 1043 break; 1044 case 5: 1045 /* forward solve the lower triangular */ 1046 idx = 5*(*r++); 1047 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1048 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1049 for ( i=1; i<n; i++ ) { 1050 v = aa + 25*ai[i]; 1051 vi = aj + ai[i]; 1052 nz = a->diag[i] - ai[i]; 1053 idx = 5*(*r++); 1054 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1055 sum5 = b[4+idx]; 1056 while (nz--) { 1057 idx = 5*(*vi++); 1058 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1059 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1060 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1061 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1062 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1063 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1064 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1065 v += 25; 1066 } 1067 idx = 5*i; 1068 tmp[idx] = sum1;tmp[1+idx] = sum2; 1069 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1070 } 1071 /* backward solve the upper triangular */ 1072 for ( i=n-1; i>=0; i-- ){ 1073 v = aa + 25*a->diag[i] + 25; 1074 vi = aj + a->diag[i] + 1; 1075 nz = ai[i+1] - a->diag[i] - 1; 1076 idt = 5*i; 1077 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1078 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1079 while (nz--) { 1080 idx = 5*(*vi++); 1081 x1 = tmp[idx]; x2 = tmp[1+idx]; 1082 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1083 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1084 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1085 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1086 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1087 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1088 v += 25; 1089 } 1090 idc = 5*(*c--); 1091 v = aa + 25*a->diag[i]; 1092 x[idc] += tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1093 v[15]*sum4+v[20]*sum5; 1094 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1095 v[16]*sum4+v[21]*sum5; 1096 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1097 v[17]*sum4+v[22]*sum5; 1098 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1099 v[18]*sum4+v[23]*sum5; 1100 x[4+idc] += tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1101 v[19]*sum4+v[24]*sum5; 1102 } 1103 break; 1104 default: { 1105 /* forward solve the lower triangular */ 1106 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 1107 for ( i=1; i<n; i++ ) { 1108 v = aa + bs2*ai[i]; 1109 vi = aj + ai[i]; 1110 nz = a->diag[i] - ai[i]; 1111 sum = tmp + bs*i; 1112 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 1113 while (nz--) { 1114 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,sum,&_One); 1115 v += bs2; 1116 } 1117 } 1118 /* backward solve the upper triangular */ 1119 lsum = a->solve_work + a->n; 1120 for ( i=n-1; i>=0; i-- ){ 1121 v = aa + bs2*(a->diag[i] + 1); 1122 vi = aj + a->diag[i] + 1; 1123 nz = ai[i+1] - a->diag[i] - 1; 1124 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 1125 while (nz--) { 1126 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,lsum,&_One); 1127 v += bs2; 1128 } 1129 LAgemv_("N",&bs,&bs,&_DOne,aa+bs2*a->diag[i],&bs,lsum,&_One,&_DZero, 1130 tmp+i*bs,&_One); 1131 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 1132 } 1133 } 1134 } 1135 1136 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1137 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1138 return 0; 1139 } 1140 1141 /* ----------------------------------------------------------------*/ 1142 1143 int MatSolve_SeqBAIJ(Mat A,Vec bb, Vec xx) 1144 { 1145 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1146 int ierr,bs2=a->bs2; 1147 1148 ierr = MatSolveAdd_SeqBAIJ_Private(A,bb,PETSC_NULL,xx); CHKERRQ(ierr); 1149 PLogFlops(2*bs2*(a->nz) - a->n); 1150 return 0; 1151 } 1152 1153 /* ----------------------------------------------------------------*/ 1154 1155 int MatSolveAdd_SeqBAIJ(Mat A,Vec bb,Vec yy,Vec xx) 1156 { 1157 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1158 int ierr,bs2=a->bs2; 1159 1160 ierr = MatSolveAdd_SeqBAIJ_Private(A,bb,yy,xx); CHKERRQ(ierr); 1161 PLogFlops(2*bs2*(a->nz)); 1162 return 0; 1163 } 1164 /* ----------------------------------------------------------- */ 1165 static int MatSolveTransAdd_SeqBAIJ_Private(Mat A,Vec bb,Vec yy,Vec xx) 1166 { 1167 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1168 IS iscol=a->col,isrow=a->row; 1169 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 1170 int nz,bs=a->bs,bs2=a->bs2,idx,idt,idc,_One=1,N=a->n; 1171 Scalar *xa,*ba,*aa=a->a,*sum,_DOne=1.0,_DMOne=-1.0; 1172 Scalar _DZero=0.0,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5,*y; 1173 register Scalar *x,*b,*lsum,*tmp,*v; 1174 1175 if (A->factor != FACTOR_LU) SETERRQ(1,"MatSolveTransAdd_SeqBAIJ_Private:Not for unfactored matrix"); 1176 1177 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1178 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1179 tmp = a->solve_work; 1180 1181 if (yy==PETSC_NULL) PetscMemzero(x,N*sizeof(Scalar)); /* VecZero()??? MatSolve() */ 1182 else if (xx!=yy){ /* VecCopy() ??? */ 1183 ierr = VecGetArray(yy,&y); CHKERRQ(ierr); 1184 PetscMemcpy(x,y,N*sizeof(Scalar)); 1185 ierr = VecRestoreArray(yy,&y); CHKERRQ(ierr); 1186 } 1187 1188 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1189 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1190 1191 switch (bs) { 1192 case 1: 1193 /* forward solve the lower triangular */ 1194 tmp[0] = b[*r++]; 1195 for ( i=1; i<n; i++ ) { 1196 v = aa + ai[i]; 1197 vi = aj + ai[i]; 1198 nz = a->diag[i] - ai[i]; 1199 sum1 = b[*r++]; 1200 while (nz--) { 1201 sum1 -= (*v++)*tmp[*vi++]; 1202 } 1203 tmp[i] = sum1; 1204 } 1205 /* backward solve the upper triangular */ 1206 for ( i=n-1; i>=0; i-- ){ 1207 v = aa + a->diag[i] + 1; 1208 vi = aj + a->diag[i] + 1; 1209 nz = ai[i+1] - a->diag[i] - 1; 1210 sum1 = tmp[i]; 1211 while (nz--) { 1212 sum1 -= (*v++)*tmp[*vi++]; 1213 } 1214 x[*c--] += tmp[i] = aa[a->diag[i]]*sum1; 1215 } 1216 break; 1217 case 2: 1218 /* forward solve the lower triangular */ 1219 idx = 2*(*r++); 1220 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1221 for ( i=1; i<n; i++ ) { 1222 v = aa + 4*ai[i]; 1223 vi = aj + ai[i]; 1224 nz = a->diag[i] - ai[i]; 1225 idx = 2*(*r++); 1226 sum1 = b[idx]; sum2 = b[1+idx]; 1227 while (nz--) { 1228 idx = 2*(*vi++); 1229 x1 = tmp[idx]; x2 = tmp[1+idx]; 1230 sum1 -= v[0]*x1 + v[2]*x2; 1231 sum2 -= v[1]*x1 + v[3]*x2; 1232 v += 4; 1233 } 1234 idx = 2*i; 1235 tmp[idx] = sum1; tmp[1+idx] = sum2; 1236 } 1237 /* backward solve the upper triangular */ 1238 for ( i=n-1; i>=0; i-- ){ 1239 v = aa + 4*a->diag[i] + 4; 1240 vi = aj + a->diag[i] + 1; 1241 nz = ai[i+1] - a->diag[i] - 1; 1242 idt = 2*i; 1243 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1244 while (nz--) { 1245 idx = 2*(*vi++); 1246 x1 = tmp[idx]; x2 = tmp[1+idx]; 1247 sum1 -= v[0]*x1 + v[2]*x2; 1248 sum2 -= v[1]*x1 + v[3]*x2; 1249 v += 4; 1250 } 1251 idc = 2*(*c--); 1252 v = aa + 4*a->diag[i]; 1253 x[idc] += tmp[idt] = v[0]*sum1 + v[2]*sum2; 1254 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1255 } 1256 break; 1257 case 3: 1258 /* forward solve the lower triangular */ 1259 idx = 3*(*r++); 1260 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1261 for ( i=1; i<n; i++ ) { 1262 v = aa + 9*ai[i]; 1263 vi = aj + ai[i]; 1264 nz = a->diag[i] - ai[i]; 1265 idx = 3*(*r++); 1266 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1267 while (nz--) { 1268 idx = 3*(*vi++); 1269 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1270 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1271 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1272 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1273 v += 9; 1274 } 1275 idx = 3*i; 1276 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1277 } 1278 /* backward solve the upper triangular */ 1279 for ( i=n-1; i>=0; i-- ){ 1280 v = aa + 9*a->diag[i] + 9; 1281 vi = aj + a->diag[i] + 1; 1282 nz = ai[i+1] - a->diag[i] - 1; 1283 idt = 3*i; 1284 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1285 while (nz--) { 1286 idx = 3*(*vi++); 1287 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1288 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1289 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1290 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1291 v += 9; 1292 } 1293 idc = 3*(*c--); 1294 v = aa + 9*a->diag[i]; 1295 x[idc] += tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1296 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1297 x[2+idc] += tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1298 } 1299 break; 1300 case 4: 1301 /* forward solve the lower triangular */ 1302 idx = 4*(*r++); 1303 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1304 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1305 for ( i=1; i<n; i++ ) { 1306 v = aa + 16*ai[i]; 1307 vi = aj + ai[i]; 1308 nz = a->diag[i] - ai[i]; 1309 idx = 4*(*r++); 1310 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1311 while (nz--) { 1312 idx = 4*(*vi++); 1313 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1314 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1315 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1316 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1317 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1318 v += 16; 1319 } 1320 idx = 4*i; 1321 tmp[idx] = sum1;tmp[1+idx] = sum2; 1322 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1323 } 1324 /* backward solve the upper triangular */ 1325 for ( i=n-1; i>=0; i-- ){ 1326 v = aa + 16*a->diag[i] + 16; 1327 vi = aj + a->diag[i] + 1; 1328 nz = ai[i+1] - a->diag[i] - 1; 1329 idt = 4*i; 1330 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1331 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1332 while (nz--) { 1333 idx = 4*(*vi++); 1334 x1 = tmp[idx]; x2 = tmp[1+idx]; 1335 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1336 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1337 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1338 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1339 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1340 v += 16; 1341 } 1342 idc = 4*(*c--); 1343 v = aa + 16*a->diag[i]; 1344 x[idc] += tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1345 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1346 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1347 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1348 } 1349 break; 1350 case 5: 1351 /* forward solve the lower triangular */ 1352 idx = 5*(*r++); 1353 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1354 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1355 for ( i=1; i<n; i++ ) { 1356 v = aa + 25*ai[i]; 1357 vi = aj + ai[i]; 1358 nz = a->diag[i] - ai[i]; 1359 idx = 5*(*r++); 1360 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1361 sum5 = b[4+idx]; 1362 while (nz--) { 1363 idx = 5*(*vi++); 1364 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1365 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1366 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1367 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1368 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1369 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1370 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1371 v += 25; 1372 } 1373 idx = 5*i; 1374 tmp[idx] = sum1;tmp[1+idx] = sum2; 1375 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1376 } 1377 /* backward solve the upper triangular */ 1378 for ( i=n-1; i>=0; i-- ){ 1379 v = aa + 25*a->diag[i] + 25; 1380 vi = aj + a->diag[i] + 1; 1381 nz = ai[i+1] - a->diag[i] - 1; 1382 idt = 5*i; 1383 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1384 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1385 while (nz--) { 1386 idx = 5*(*vi++); 1387 x1 = tmp[idx]; x2 = tmp[1+idx]; 1388 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1389 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1390 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1391 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1392 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1393 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1394 v += 25; 1395 } 1396 idc = 5*(*c--); 1397 v = aa + 25*a->diag[i]; 1398 x[idc] += tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1399 v[15]*sum4+v[20]*sum5; 1400 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1401 v[16]*sum4+v[21]*sum5; 1402 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1403 v[17]*sum4+v[22]*sum5; 1404 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1405 v[18]*sum4+v[23]*sum5; 1406 x[4+idc] += tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1407 v[19]*sum4+v[24]*sum5; 1408 } 1409 break; 1410 default: { 1411 /* forward solve the lower triangular */ 1412 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 1413 for ( i=1; i<n; i++ ) { 1414 v = aa + bs2*ai[i]; 1415 vi = aj + ai[i]; 1416 nz = a->diag[i] - ai[i]; 1417 sum = tmp + bs*i; 1418 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 1419 while (nz--) { 1420 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,sum,&_One); 1421 v += bs2; 1422 } 1423 } 1424 /* backward solve the upper triangular */ 1425 lsum = a->solve_work + a->n; 1426 for ( i=n-1; i>=0; i-- ){ 1427 v = aa + bs2*(a->diag[i] + 1); 1428 vi = aj + a->diag[i] + 1; 1429 nz = ai[i+1] - a->diag[i] - 1; 1430 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 1431 while (nz--) { 1432 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,lsum,&_One); 1433 v += bs2; 1434 } 1435 LAgemv_("N",&bs,&bs,&_DOne,aa+bs2*a->diag[i],&bs,lsum,&_One,&_DZero, 1436 tmp+i*bs,&_One); 1437 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 1438 } 1439 } 1440 } 1441 1442 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1443 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1444 return 0; 1445 } 1446 1447 /* ----------------------------------------------------------------*/ 1448 1449 int MatSolveTrans_SeqBAIJ(Mat A,Vec bb, Vec xx) 1450 { 1451 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1452 int ierr,bs2=a->bs2; 1453 1454 ierr = MatSolveTransAdd_SeqBAIJ_Private(A,bb,PETSC_NULL,xx); CHKERRQ(ierr); 1455 PLogFlops(2*bs2*(a->nz) - a->n); 1456 return 0; 1457 } 1458 1459 /* ----------------------------------------------------------------*/ 1460 1461 int MatSolveTransAdd_SeqBAIJ(Mat A,Vec bb,Vec yy,Vec xx) 1462 { 1463 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1464 int ierr,bs2=a->bs2; 1465 1466 ierr = MatSolveTransAdd_SeqBAIJ_Private(A,bb,yy,xx); CHKERRQ(ierr); 1467 PLogFlops(2*bs2*(a->nz)); 1468 return 0; 1469 } 1470 1471 /* ----------------------------------------------------------------*/ 1472 /* 1473 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1474 except that the data structure of Mat_SeqAIJ is slightly different. 1475 Not a good example of code reuse. 1476 */ 1477 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1478 Mat *fact) 1479 { 1480 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1481 IS isicol; 1482 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1483 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1484 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1485 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1486 PetscTruth col_identity, row_identity; 1487 1488 if (a->m != a->n) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Matrix must be square"); 1489 if (!isrow) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have row permutation"); 1490 if (!iscol) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have column permutation"); 1491 1492 /* special case that simply copies fill pattern */ 1493 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1494 if (levels == 0 && row_identity && col_identity) { 1495 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1496 (*fact)->factor = FACTOR_LU; 1497 b = (Mat_SeqBAIJ *) (*fact)->data; 1498 if (!b->diag) { 1499 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1500 } 1501 b->row = isrow; 1502 b->col = iscol; 1503 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1504 return 0; 1505 } 1506 1507 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1508 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1509 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1510 1511 /* get new row pointers */ 1512 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1513 ainew[0] = 0; 1514 /* don't know how many column pointers are needed so estimate */ 1515 jmax = (int) (f*ai[n] + 1); 1516 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1517 /* ajfill is level of fill for each fill entry */ 1518 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1519 /* fill is a linked list of nonzeros in active row */ 1520 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1521 /* im is level for each filled value */ 1522 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1523 /* dloc is location of diagonal in factor */ 1524 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1525 dloc[0] = 0; 1526 for ( prow=0; prow<n; prow++ ) { 1527 /* first copy previous fill into linked list */ 1528 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1529 xi = aj + ai[r[prow]]; 1530 fill[n] = n; 1531 while (nz--) { 1532 fm = n; 1533 idx = ic[*xi++]; 1534 do { 1535 m = fm; 1536 fm = fill[m]; 1537 } while (fm < idx); 1538 fill[m] = idx; 1539 fill[idx] = fm; 1540 im[idx] = 0; 1541 } 1542 nzi = 0; 1543 row = fill[n]; 1544 while ( row < prow ) { 1545 incrlev = im[row] + 1; 1546 nz = dloc[row]; 1547 xi = ajnew + ainew[row] + nz; 1548 flev = ajfill + ainew[row] + nz + 1; 1549 nnz = ainew[row+1] - ainew[row] - nz - 1; 1550 if (*xi++ != row) { 1551 SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:zero pivot"); 1552 } 1553 fm = row; 1554 while (nnz-- > 0) { 1555 idx = *xi++; 1556 if (*flev + incrlev > levels) { 1557 flev++; 1558 continue; 1559 } 1560 do { 1561 m = fm; 1562 fm = fill[m]; 1563 } while (fm < idx); 1564 if (fm != idx) { 1565 im[idx] = *flev + incrlev; 1566 fill[m] = idx; 1567 fill[idx] = fm; 1568 fm = idx; 1569 nzf++; 1570 } 1571 else { 1572 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1573 } 1574 flev++; 1575 } 1576 row = fill[row]; 1577 nzi++; 1578 } 1579 /* copy new filled row into permanent storage */ 1580 ainew[prow+1] = ainew[prow] + nzf; 1581 if (ainew[prow+1] > jmax) { 1582 /* allocate a longer ajnew */ 1583 int maxadd; 1584 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1585 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1586 jmax += maxadd; 1587 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1588 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1589 PetscFree(ajnew); 1590 ajnew = xi; 1591 /* allocate a longer ajfill */ 1592 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1593 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1594 PetscFree(ajfill); 1595 ajfill = xi; 1596 realloc++; 1597 } 1598 xi = ajnew + ainew[prow]; 1599 flev = ajfill + ainew[prow]; 1600 dloc[prow] = nzi; 1601 fm = fill[n]; 1602 while (nzf--) { 1603 *xi++ = fm; 1604 *flev++ = im[fm]; 1605 fm = fill[fm]; 1606 } 1607 } 1608 PetscFree(ajfill); 1609 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1610 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1611 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1612 PetscFree(fill); PetscFree(im); 1613 1614 PLogInfo(A, 1615 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1616 realloc,f,((double)ainew[n])/((double)ai[prow])); 1617 1618 /* put together the new matrix */ 1619 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1620 b = (Mat_SeqBAIJ *) (*fact)->data; 1621 PetscFree(b->imax); 1622 b->singlemalloc = 0; 1623 len = bs2*ainew[n]*sizeof(Scalar); 1624 /* the next line frees the default space generated by the Create() */ 1625 PetscFree(b->a); PetscFree(b->ilen); 1626 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1627 b->j = ajnew; 1628 b->i = ainew; 1629 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1630 b->diag = dloc; 1631 b->ilen = 0; 1632 b->imax = 0; 1633 b->row = isrow; 1634 b->col = iscol; 1635 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1636 CHKPTRQ(b->solve_work); 1637 /* In b structure: Free imax, ilen, old a, old j. 1638 Allocate dloc, solve_work, new a, new j */ 1639 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1640 b->maxnz = b->nz = ainew[n]; 1641 (*fact)->factor = FACTOR_LU; 1642 return 0; 1643 } 1644 1645 1646 1647 1648