1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.18 1996/04/20 04:20:18 bsmith Exp balay $"; 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 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 220 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 221 } 222 223 PetscFree(rtmp); PetscFree(v_work); 224 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 225 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 226 ierr = ISDestroy(isicol); CHKERRQ(ierr); 227 C->factor = FACTOR_LU; 228 C->assembled = PETSC_TRUE; 229 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 230 return 0; 231 } 232 /* ------------------------------------------------------------*/ 233 /* 234 Version for when blocks are 5 by 5 235 */ 236 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 237 { 238 Mat C = *B; 239 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 240 IS iscol = b->col, isrow = b->row, isicol; 241 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 242 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 243 int *diag_offset = b->diag,bs = 5,idx; 244 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 245 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 246 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 247 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 248 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 249 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 250 register int *pj; 251 252 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 253 PLogObjectParent(*B,isicol); 254 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 255 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 256 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 257 258 for ( i=0; i<n; i++ ) { 259 nz = ai[i+1] - ai[i]; 260 ajtmp = aj + ai[i]; 261 for ( j=0; j<nz; j++ ) { 262 x = rtmp+25*ajtmp[j]; 263 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 264 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 265 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 266 } 267 /* load in initial (unfactored row) */ 268 idx = r[i]; 269 nz = a->i[idx+1] - a->i[idx]; 270 ajtmpold = a->j + a->i[idx]; 271 v = a->a + 25*a->i[idx]; 272 for ( j=0; j<nz; j++ ) { 273 x = rtmp+25*ic[ajtmpold[j]]; 274 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 275 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 276 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 277 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 278 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 279 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 280 v += 25; 281 } 282 row = *ajtmp++; 283 while (row < i) { 284 pc = rtmp + 25*row; 285 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 286 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 287 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 288 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 289 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 290 p25 = pc[24]; 291 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 292 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 293 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 294 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 295 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 296 p23 != 0.0 || p24 != 0.0) { 297 pv = b->a + 25*diag_offset[row]; 298 pj = b->j + diag_offset[row] + 1; 299 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 300 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 301 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 302 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 303 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 304 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 305 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 306 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 307 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 308 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 309 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 310 311 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 312 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 313 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 314 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 315 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 316 317 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 318 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 319 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 320 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 321 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 322 323 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 324 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 325 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 326 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 327 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 328 329 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 330 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 331 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 332 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 333 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 334 335 nz = ai[row+1] - diag_offset[row] - 1; 336 pv += 25; 337 for (j=0; j<nz; j++) { 338 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 339 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 340 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 341 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 342 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 343 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 344 x = rtmp + 25*pj[j]; 345 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 346 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 347 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 348 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 349 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 350 351 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 352 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 353 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 354 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 355 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 356 357 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 358 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 359 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 360 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 361 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 362 363 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 364 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 365 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 366 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 367 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 368 369 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 370 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 371 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 372 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 373 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 374 375 pv += 25; 376 } 377 PLogFlops(250*nz+225); 378 } 379 row = *ajtmp++; 380 } 381 /* finished row so stick it into b->a */ 382 pv = b->a + 25*ai[i]; 383 pj = b->j + ai[i]; 384 nz = ai[i+1] - ai[i]; 385 for ( j=0; j<nz; j++ ) { 386 x = rtmp+25*pj[j]; 387 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 388 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 389 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 390 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 391 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 392 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 393 pv += 25; 394 } 395 /* invert diagonal block */ 396 w = b->a + 25*diag_offset[i]; 397 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 398 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 399 } 400 401 PetscFree(rtmp); 402 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 403 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 404 ierr = ISDestroy(isicol); CHKERRQ(ierr); 405 C->factor = FACTOR_LU; 406 C->assembled = PETSC_TRUE; 407 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 408 return 0; 409 } 410 411 /* ------------------------------------------------------------*/ 412 /* 413 Version for when blocks are 4 by 4 414 */ 415 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 416 { 417 Mat C = *B; 418 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 419 IS iscol = b->col, isrow = b->row, isicol; 420 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 421 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 422 int *diag_offset = b->diag,bs = 4,idx; 423 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 424 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 425 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 426 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 427 Scalar m13,m14,m15,m16; 428 register int *pj; 429 430 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 431 PLogObjectParent(*B,isicol); 432 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 433 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 434 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 435 436 for ( i=0; i<n; i++ ) { 437 nz = ai[i+1] - ai[i]; 438 ajtmp = aj + ai[i]; 439 for ( j=0; j<nz; j++ ) { 440 x = rtmp+16*ajtmp[j]; 441 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 442 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 443 } 444 /* load in initial (unfactored row) */ 445 idx = r[i]; 446 nz = a->i[idx+1] - a->i[idx]; 447 ajtmpold = a->j + a->i[idx]; 448 v = a->a + 16*a->i[idx]; 449 for ( j=0; j<nz; j++ ) { 450 x = rtmp+16*ic[ajtmpold[j]]; 451 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 452 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 453 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 454 x[14] = v[14]; x[15] = v[15]; 455 v += 16; 456 } 457 row = *ajtmp++; 458 while (row < i) { 459 pc = rtmp + 16*row; 460 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 461 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 462 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 463 p15 = pc[14]; p16 = pc[15]; 464 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 465 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 466 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 467 || p16 != 0.0) { 468 pv = b->a + 16*diag_offset[row]; 469 pj = b->j + diag_offset[row] + 1; 470 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 471 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 472 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 473 x15 = pv[14]; x16 = pv[15]; 474 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 475 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 476 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 477 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 478 479 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 480 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 481 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 482 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 483 484 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 485 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 486 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 487 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 488 489 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 490 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 491 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 492 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 493 494 nz = ai[row+1] - diag_offset[row] - 1; 495 pv += 16; 496 for (j=0; j<nz; j++) { 497 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 498 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 499 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 500 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 501 x = rtmp + 16*pj[j]; 502 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 503 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 504 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 505 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 506 507 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 508 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 509 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 510 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 511 512 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 513 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 514 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 515 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 516 517 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 518 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 519 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 520 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 521 522 pv += 16; 523 } 524 PLogFlops(128*nz+112); 525 } 526 row = *ajtmp++; 527 } 528 /* finished row so stick it into b->a */ 529 pv = b->a + 16*ai[i]; 530 pj = b->j + ai[i]; 531 nz = ai[i+1] - ai[i]; 532 for ( j=0; j<nz; j++ ) { 533 x = rtmp+16*pj[j]; 534 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 535 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 536 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 537 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 538 pv += 16; 539 } 540 /* invert diagonal block */ 541 w = b->a + 16*diag_offset[i]; 542 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 543 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 544 } 545 546 PetscFree(rtmp); 547 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 548 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 549 ierr = ISDestroy(isicol); CHKERRQ(ierr); 550 C->factor = FACTOR_LU; 551 C->assembled = PETSC_TRUE; 552 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 553 return 0; 554 } 555 /* ------------------------------------------------------------*/ 556 /* 557 Version for when blocks are 3 by 3 558 */ 559 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 560 { 561 Mat C = *B; 562 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 563 IS iscol = b->col, isrow = b->row, isicol; 564 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 565 int *ajtmpold, *ajtmp, nz, row, v_pivots[3]; 566 int *diag_offset = b->diag,bs = 3,idx; 567 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 568 Scalar p1,p2,p3,p4,v_work[3],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 569 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 570 register int *pj; 571 572 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 573 PLogObjectParent(*B,isicol); 574 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 575 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 576 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 577 578 for ( i=0; i<n; i++ ) { 579 nz = ai[i+1] - ai[i]; 580 ajtmp = aj + ai[i]; 581 for ( j=0; j<nz; j++ ) { 582 x = rtmp+9*ajtmp[j]; 583 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 584 } 585 /* load in initial (unfactored row) */ 586 idx = r[i]; 587 nz = a->i[idx+1] - a->i[idx]; 588 ajtmpold = a->j + a->i[idx]; 589 v = a->a + 9*a->i[idx]; 590 for ( j=0; j<nz; j++ ) { 591 x = rtmp+9*ic[ajtmpold[j]]; 592 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 593 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 594 v += 9; 595 } 596 row = *ajtmp++; 597 while (row < i) { 598 pc = rtmp + 9*row; 599 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 600 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 601 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 602 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 603 pv = b->a + 9*diag_offset[row]; 604 pj = b->j + diag_offset[row] + 1; 605 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 606 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 607 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 608 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 609 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 610 611 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 612 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 613 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 614 615 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 616 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 617 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 618 nz = ai[row+1] - diag_offset[row] - 1; 619 pv += 9; 620 for (j=0; j<nz; j++) { 621 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 622 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 623 x = rtmp + 9*pj[j]; 624 x[0] -= m1*x1 + m4*x2 + m7*x3; 625 x[1] -= m2*x1 + m5*x2 + m8*x3; 626 x[2] -= m3*x1 + m6*x2 + m9*x3; 627 628 x[3] -= m1*x4 + m4*x5 + m7*x6; 629 x[4] -= m2*x4 + m5*x5 + m8*x6; 630 x[5] -= m3*x4 + m6*x5 + m9*x6; 631 632 x[6] -= m1*x7 + m4*x8 + m7*x9; 633 x[7] -= m2*x7 + m5*x8 + m8*x9; 634 x[8] -= m3*x7 + m6*x8 + m9*x9; 635 pv += 9; 636 } 637 PLogFlops(54*nz+36); 638 } 639 row = *ajtmp++; 640 } 641 /* finished row so stick it into b->a */ 642 pv = b->a + 9*ai[i]; 643 pj = b->j + ai[i]; 644 nz = ai[i+1] - ai[i]; 645 for ( j=0; j<nz; j++ ) { 646 x = rtmp+9*pj[j]; 647 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 648 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 649 pv += 9; 650 } 651 /* invert diagonal block */ 652 w = b->a + 9*diag_offset[i]; 653 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 654 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 655 } 656 657 PetscFree(rtmp); 658 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 659 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 660 ierr = ISDestroy(isicol); CHKERRQ(ierr); 661 C->factor = FACTOR_LU; 662 C->assembled = PETSC_TRUE; 663 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 664 return 0; 665 } 666 667 /* ------------------------------------------------------------*/ 668 /* 669 Version for when blocks are 2 by 2 670 */ 671 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 672 { 673 Mat C = *B; 674 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 675 IS iscol = b->col, isrow = b->row, isicol; 676 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 677 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 678 int *diag_offset=b->diag,bs = 2,idx; 679 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 680 Scalar p1,p2,p3,p4,v_work[2]; 681 register int *pj; 682 683 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 684 PLogObjectParent(*B,isicol); 685 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 686 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 687 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 688 689 for ( i=0; i<n; i++ ) { 690 nz = ai[i+1] - ai[i]; 691 ajtmp = aj + ai[i]; 692 for ( j=0; j<nz; j++ ) { 693 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 694 } 695 /* load in initial (unfactored row) */ 696 idx = r[i]; 697 nz = a->i[idx+1] - a->i[idx]; 698 ajtmpold = a->j + a->i[idx]; 699 v = a->a + 4*a->i[idx]; 700 for ( j=0; j<nz; j++ ) { 701 x = rtmp+4*ic[ajtmpold[j]]; 702 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 703 v += 4; 704 } 705 row = *ajtmp++; 706 while (row < i) { 707 pc = rtmp + 4*row; 708 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 709 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 710 pv = b->a + 4*diag_offset[row]; 711 pj = b->j + diag_offset[row] + 1; 712 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 713 pc[0] = m1 = p1*x1 + p3*x2; 714 pc[1] = m2 = p2*x1 + p4*x2; 715 pc[2] = m3 = p1*x3 + p3*x4; 716 pc[3] = m4 = p2*x3 + p4*x4; 717 nz = ai[row+1] - diag_offset[row] - 1; 718 pv += 4; 719 for (j=0; j<nz; j++) { 720 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 721 x = rtmp + 4*pj[j]; 722 x[0] -= m1*x1 + m3*x2; 723 x[1] -= m2*x1 + m4*x2; 724 x[2] -= m1*x3 + m3*x4; 725 x[3] -= m2*x3 + m4*x4; 726 pv += 4; 727 } 728 PLogFlops(16*nz+12); 729 } 730 row = *ajtmp++; 731 } 732 /* finished row so stick it into b->a */ 733 pv = b->a + 4*ai[i]; 734 pj = b->j + ai[i]; 735 nz = ai[i+1] - ai[i]; 736 for ( j=0; j<nz; j++ ) { 737 x = rtmp+4*pj[j]; 738 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 739 pv += 4; 740 } 741 /* invert diagonal block */ 742 w = b->a + 4*diag_offset[i]; 743 ierr = Linpack_DGEFA(w,bs,v_pivots); CHKERRQ(ierr); 744 ierr = Linpack_DGEDI(w,bs,v_pivots,v_work); CHKERRQ(ierr); 745 } 746 747 PetscFree(rtmp); 748 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 749 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 750 ierr = ISDestroy(isicol); CHKERRQ(ierr); 751 C->factor = FACTOR_LU; 752 C->assembled = PETSC_TRUE; 753 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 754 return 0; 755 } 756 757 /* ----------------------------------------------------------- */ 758 /* 759 Version for when blocks are 1 by 1. 760 */ 761 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 762 { 763 Mat C = *B; 764 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 765 IS iscol = b->col, isrow = b->row, isicol; 766 int *r,*ic, ierr, i, j, n = a->mbs, *ai = b->i, *aj = b->j; 767 int *ajtmpold, *ajtmp, nz, row; 768 int *diag_offset = b->diag,diag; 769 register Scalar *pv,*v,*rtmp,multiplier,*pc; 770 register int *pj; 771 772 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 773 PLogObjectParent(*B,isicol); 774 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 775 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 776 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 777 778 for ( i=0; i<n; i++ ) { 779 nz = ai[i+1] - ai[i]; 780 ajtmp = aj + ai[i]; 781 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 782 783 /* load in initial (unfactored row) */ 784 nz = a->i[r[i]+1] - a->i[r[i]]; 785 ajtmpold = a->j + a->i[r[i]]; 786 v = a->a + a->i[r[i]]; 787 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 788 789 row = *ajtmp++; 790 while (row < i) { 791 pc = rtmp + row; 792 if (*pc != 0.0) { 793 pv = b->a + diag_offset[row]; 794 pj = b->j + diag_offset[row] + 1; 795 multiplier = *pc * *pv++; 796 *pc = multiplier; 797 nz = ai[row+1] - diag_offset[row] - 1; 798 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 799 PLogFlops(1+2*nz); 800 } 801 row = *ajtmp++; 802 } 803 /* finished row so stick it into b->a */ 804 pv = b->a + ai[i]; 805 pj = b->j + ai[i]; 806 nz = ai[i+1] - ai[i]; 807 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 808 diag = diag_offset[i] - ai[i]; 809 /* check pivot entry for current row */ 810 if (pv[diag] == 0.0) { 811 SETERRQ(1,"MatLUFactorNumeric_SeqAIJ:Zero pivot"); 812 } 813 pv[diag] = 1.0/pv[diag]; 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(b->n); 823 return 0; 824 } 825 826 /* ----------------------------------------------------------- */ 827 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 828 { 829 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 830 int ierr; 831 Mat C; 832 833 ierr = MatLUFactorSymbolic_SeqBAIJ(A,row,col,f,&C); CHKERRQ(ierr); 834 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 835 836 /* free all the data structures from mat */ 837 PetscFree(mat->a); 838 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 839 if (mat->diag) PetscFree(mat->diag); 840 if (mat->ilen) PetscFree(mat->ilen); 841 if (mat->imax) PetscFree(mat->imax); 842 if (mat->solve_work) PetscFree(mat->solve_work); 843 PetscFree(mat); 844 845 PetscMemcpy(A,C,sizeof(struct _Mat)); 846 PetscHeaderDestroy(C); 847 return 0; 848 } 849 /* ----------------------------------------------------------- */ 850 static int MatSolveAdd_SeqBAIJ_Private(Mat A,Vec bb,Vec yy,Vec xx) 851 { 852 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 853 IS iscol=a->col,isrow=a->row; 854 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 855 int nz,bs=a->bs,bs2=a->bs2,idx,idt,idc,_One=1,m=a->m; 856 Scalar *xa,*ba,*aa=a->a,*sum,_DOne=1.0,_DMOne=-1.0; 857 Scalar _DZero=0.0,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5,*y; 858 register Scalar *x,*b,*lsum,*tmp,*v; 859 860 if (A->factor != FACTOR_LU) SETERRQ(1,"MatSolveAdd_SeqBAIJ_Private:Not for unfactored matrix"); 861 862 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 863 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 864 tmp = a->solve_work; 865 866 if (yy==PETSC_NULL) PetscMemzero(x,m*sizeof(Scalar)); /* MatSolve() */ 867 else if (xx!=yy){ 868 ierr = VecGetArray(yy,&y); CHKERRQ(ierr); 869 PetscMemcpy(x,y,m*sizeof(Scalar)); 870 ierr = VecRestoreArray(yy,&y); CHKERRQ(ierr); 871 } 872 873 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 874 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 875 876 switch (bs) { 877 case 1: 878 /* forward solve the lower triangular */ 879 tmp[0] = b[*r++]; 880 for ( i=1; i<n; i++ ) { 881 v = aa + ai[i]; 882 vi = aj + ai[i]; 883 nz = a->diag[i] - ai[i]; 884 sum1 = b[*r++]; 885 while (nz--) { 886 sum1 -= (*v++)*tmp[*vi++]; 887 } 888 tmp[i] = sum1; 889 } 890 /* backward solve the upper triangular */ 891 for ( i=n-1; i>=0; i-- ){ 892 v = aa + a->diag[i] + 1; 893 vi = aj + a->diag[i] + 1; 894 nz = ai[i+1] - a->diag[i] - 1; 895 sum1 = tmp[i]; 896 while (nz--) { 897 sum1 -= (*v++)*tmp[*vi++]; 898 } 899 x[*c--] += tmp[i] = aa[a->diag[i]]*sum1; 900 } 901 break; 902 case 2: 903 /* forward solve the lower triangular */ 904 idx = 2*(*r++); 905 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 906 for ( i=1; i<n; i++ ) { 907 v = aa + 4*ai[i]; 908 vi = aj + ai[i]; 909 nz = a->diag[i] - ai[i]; 910 idx = 2*(*r++); 911 sum1 = b[idx]; sum2 = b[1+idx]; 912 while (nz--) { 913 idx = 2*(*vi++); 914 x1 = tmp[idx]; x2 = tmp[1+idx]; 915 sum1 -= v[0]*x1 + v[2]*x2; 916 sum2 -= v[1]*x1 + v[3]*x2; 917 v += 4; 918 } 919 idx = 2*i; 920 tmp[idx] = sum1; tmp[1+idx] = sum2; 921 } 922 /* backward solve the upper triangular */ 923 for ( i=n-1; i>=0; i-- ){ 924 v = aa + 4*a->diag[i] + 4; 925 vi = aj + a->diag[i] + 1; 926 nz = ai[i+1] - a->diag[i] - 1; 927 idt = 2*i; 928 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 929 while (nz--) { 930 idx = 2*(*vi++); 931 x1 = tmp[idx]; x2 = tmp[1+idx]; 932 sum1 -= v[0]*x1 + v[2]*x2; 933 sum2 -= v[1]*x1 + v[3]*x2; 934 v += 4; 935 } 936 idc = 2*(*c--); 937 v = aa + 4*a->diag[i]; 938 x[idc] += tmp[idt] = v[0]*sum1 + v[2]*sum2; 939 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 940 } 941 break; 942 case 3: 943 /* forward solve the lower triangular */ 944 idx = 3*(*r++); 945 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 946 for ( i=1; i<n; i++ ) { 947 v = aa + 9*ai[i]; 948 vi = aj + ai[i]; 949 nz = a->diag[i] - ai[i]; 950 idx = 3*(*r++); 951 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 952 while (nz--) { 953 idx = 3*(*vi++); 954 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 955 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 956 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 957 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 958 v += 9; 959 } 960 idx = 3*i; 961 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 962 } 963 /* backward solve the upper triangular */ 964 for ( i=n-1; i>=0; i-- ){ 965 v = aa + 9*a->diag[i] + 9; 966 vi = aj + a->diag[i] + 1; 967 nz = ai[i+1] - a->diag[i] - 1; 968 idt = 3*i; 969 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 970 while (nz--) { 971 idx = 3*(*vi++); 972 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 973 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 974 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 975 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 976 v += 9; 977 } 978 idc = 3*(*c--); 979 v = aa + 9*a->diag[i]; 980 x[idc] += tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 981 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 982 x[2+idc] += tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 983 } 984 break; 985 case 4: 986 /* forward solve the lower triangular */ 987 idx = 4*(*r++); 988 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 989 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 990 for ( i=1; i<n; i++ ) { 991 v = aa + 16*ai[i]; 992 vi = aj + ai[i]; 993 nz = a->diag[i] - ai[i]; 994 idx = 4*(*r++); 995 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 996 while (nz--) { 997 idx = 4*(*vi++); 998 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 999 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1000 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1001 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1002 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1003 v += 16; 1004 } 1005 idx = 4*i; 1006 tmp[idx] = sum1;tmp[1+idx] = sum2; 1007 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1008 } 1009 /* backward solve the upper triangular */ 1010 for ( i=n-1; i>=0; i-- ){ 1011 v = aa + 16*a->diag[i] + 16; 1012 vi = aj + a->diag[i] + 1; 1013 nz = ai[i+1] - a->diag[i] - 1; 1014 idt = 4*i; 1015 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1016 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1017 while (nz--) { 1018 idx = 4*(*vi++); 1019 x1 = tmp[idx]; x2 = tmp[1+idx]; 1020 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1021 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1022 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1023 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1024 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1025 v += 16; 1026 } 1027 idc = 4*(*c--); 1028 v = aa + 16*a->diag[i]; 1029 x[idc] += tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1030 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1031 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1032 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1033 } 1034 break; 1035 case 5: 1036 /* forward solve the lower triangular */ 1037 idx = 5*(*r++); 1038 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1039 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1040 for ( i=1; i<n; i++ ) { 1041 v = aa + 25*ai[i]; 1042 vi = aj + ai[i]; 1043 nz = a->diag[i] - ai[i]; 1044 idx = 5*(*r++); 1045 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1046 sum5 = b[4+idx]; 1047 while (nz--) { 1048 idx = 5*(*vi++); 1049 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1050 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1051 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1052 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1053 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1054 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1055 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1056 v += 25; 1057 } 1058 idx = 5*i; 1059 tmp[idx] = sum1;tmp[1+idx] = sum2; 1060 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1061 } 1062 /* backward solve the upper triangular */ 1063 for ( i=n-1; i>=0; i-- ){ 1064 v = aa + 25*a->diag[i] + 25; 1065 vi = aj + a->diag[i] + 1; 1066 nz = ai[i+1] - a->diag[i] - 1; 1067 idt = 5*i; 1068 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1069 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1070 while (nz--) { 1071 idx = 5*(*vi++); 1072 x1 = tmp[idx]; x2 = tmp[1+idx]; 1073 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1074 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1075 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1076 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1077 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1078 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1079 v += 25; 1080 } 1081 idc = 5*(*c--); 1082 v = aa + 25*a->diag[i]; 1083 x[idc] += tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1084 v[15]*sum4+v[20]*sum5; 1085 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1086 v[16]*sum4+v[21]*sum5; 1087 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1088 v[17]*sum4+v[22]*sum5; 1089 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1090 v[18]*sum4+v[23]*sum5; 1091 x[4+idc] += tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1092 v[19]*sum4+v[24]*sum5; 1093 } 1094 break; 1095 default: { 1096 /* forward solve the lower triangular */ 1097 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 1098 for ( i=1; i<n; i++ ) { 1099 v = aa + bs2*ai[i]; 1100 vi = aj + ai[i]; 1101 nz = a->diag[i] - ai[i]; 1102 sum = tmp + bs*i; 1103 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 1104 while (nz--) { 1105 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,sum,&_One); 1106 v += bs2; 1107 } 1108 } 1109 /* backward solve the upper triangular */ 1110 lsum = a->solve_work + a->n; 1111 for ( i=n-1; i>=0; i-- ){ 1112 v = aa + bs2*(a->diag[i] + 1); 1113 vi = aj + a->diag[i] + 1; 1114 nz = ai[i+1] - a->diag[i] - 1; 1115 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 1116 while (nz--) { 1117 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,lsum,&_One); 1118 v += bs2; 1119 } 1120 LAgemv_("N",&bs,&bs,&_DOne,aa+bs2*a->diag[i],&bs,lsum,&_One,&_DZero, 1121 tmp+i*bs,&_One); 1122 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 1123 } 1124 } 1125 } 1126 1127 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1128 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1129 return 0; 1130 } 1131 1132 /* ----------------------------------------------------------------*/ 1133 1134 int MatSolve_SeqBAIJ(Mat A,Vec bb, Vec xx) 1135 { 1136 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1137 int ierr; 1138 1139 ierr = MatSolveAdd_SeqBAIJ_Private(A,bb,PETSC_NULL,xx); CHKERRQ(ierr); 1140 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 1141 return 0; 1142 } 1143 1144 /* ----------------------------------------------------------------*/ 1145 1146 int MatSolveAdd_SeqBAIJ(Mat A,Vec bb,Vec yy,Vec xx) 1147 { 1148 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1149 int ierr; 1150 1151 ierr = MatSolveAdd_SeqBAIJ_Private(A,bb,yy,xx); CHKERRQ(ierr); 1152 PLogFlops(2*(a->bs2)*(a->nz)); 1153 return 0; 1154 } 1155 /* ----------------------------------------------------------- */ 1156 static int MatSolveTransAdd_SeqBAIJ_Private(Mat A,Vec bb,Vec yy,Vec xx) 1157 { 1158 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1159 IS iscol=a->col,isrow=a->row; 1160 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 1161 int nz,bs=a->bs,bs2=a->bs2,idx,idt,idc,_One=1,m=a->m; 1162 Scalar *xa,*ba,*aa=a->a,*sum,_DOne=1.0,_DMOne=-1.0; 1163 Scalar _DZero=0.0,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5,*y; 1164 register Scalar *x,*b,*lsum,*tmp,*v; 1165 1166 if (A->factor != FACTOR_LU) SETERRQ(1,"MatSolveAdd_SeqBAIJ_Private:Not for unfactored matrix"); 1167 1168 ierr = VecGetArray(bb,&ba); CHKERRQ(ierr); b = ba; 1169 ierr = VecGetArray(xx,&xa); CHKERRQ(ierr); x = xa; 1170 tmp = a->solve_work; 1171 1172 if (yy==PETSC_NULL) PetscMemzero(x,m*sizeof(Scalar)); /* MatSolve() */ 1173 else if (xx!=yy){ 1174 ierr = VecGetArray(yy,&y); CHKERRQ(ierr); 1175 PetscMemcpy(x,y,m*sizeof(Scalar)); 1176 ierr = VecRestoreArray(yy,&y); CHKERRQ(ierr); 1177 } 1178 1179 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1180 ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 1181 1182 switch (bs) { 1183 case 1: 1184 /* forward solve the lower triangular */ 1185 tmp[0] = b[*r++]; 1186 for ( i=1; i<n; i++ ) { 1187 v = aa + ai[i]; 1188 vi = aj + ai[i]; 1189 nz = a->diag[i] - ai[i]; 1190 sum1 = b[*r++]; 1191 while (nz--) { 1192 sum1 -= (*v++)*tmp[*vi++]; 1193 } 1194 tmp[i] = sum1; 1195 } 1196 /* backward solve the upper triangular */ 1197 for ( i=n-1; i>=0; i-- ){ 1198 v = aa + a->diag[i] + 1; 1199 vi = aj + a->diag[i] + 1; 1200 nz = ai[i+1] - a->diag[i] - 1; 1201 sum1 = tmp[i]; 1202 while (nz--) { 1203 sum1 -= (*v++)*tmp[*vi++]; 1204 } 1205 x[*c--] += tmp[i] = aa[a->diag[i]]*sum1; 1206 } 1207 break; 1208 case 2: 1209 /* forward solve the lower triangular */ 1210 idx = 2*(*r++); 1211 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1212 for ( i=1; i<n; i++ ) { 1213 v = aa + 4*ai[i]; 1214 vi = aj + ai[i]; 1215 nz = a->diag[i] - ai[i]; 1216 idx = 2*(*r++); 1217 sum1 = b[idx]; sum2 = b[1+idx]; 1218 while (nz--) { 1219 idx = 2*(*vi++); 1220 x1 = tmp[idx]; x2 = tmp[1+idx]; 1221 sum1 -= v[0]*x1 + v[2]*x2; 1222 sum2 -= v[1]*x1 + v[3]*x2; 1223 v += 4; 1224 } 1225 idx = 2*i; 1226 tmp[idx] = sum1; tmp[1+idx] = sum2; 1227 } 1228 /* backward solve the upper triangular */ 1229 for ( i=n-1; i>=0; i-- ){ 1230 v = aa + 4*a->diag[i] + 4; 1231 vi = aj + a->diag[i] + 1; 1232 nz = ai[i+1] - a->diag[i] - 1; 1233 idt = 2*i; 1234 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1235 while (nz--) { 1236 idx = 2*(*vi++); 1237 x1 = tmp[idx]; x2 = tmp[1+idx]; 1238 sum1 -= v[0]*x1 + v[2]*x2; 1239 sum2 -= v[1]*x1 + v[3]*x2; 1240 v += 4; 1241 } 1242 idc = 2*(*c--); 1243 v = aa + 4*a->diag[i]; 1244 x[idc] += tmp[idt] = v[0]*sum1 + v[2]*sum2; 1245 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1246 } 1247 break; 1248 case 3: 1249 /* forward solve the lower triangular */ 1250 idx = 3*(*r++); 1251 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1252 for ( i=1; i<n; i++ ) { 1253 v = aa + 9*ai[i]; 1254 vi = aj + ai[i]; 1255 nz = a->diag[i] - ai[i]; 1256 idx = 3*(*r++); 1257 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1258 while (nz--) { 1259 idx = 3*(*vi++); 1260 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1261 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1262 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1263 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1264 v += 9; 1265 } 1266 idx = 3*i; 1267 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1268 } 1269 /* backward solve the upper triangular */ 1270 for ( i=n-1; i>=0; i-- ){ 1271 v = aa + 9*a->diag[i] + 9; 1272 vi = aj + a->diag[i] + 1; 1273 nz = ai[i+1] - a->diag[i] - 1; 1274 idt = 3*i; 1275 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1276 while (nz--) { 1277 idx = 3*(*vi++); 1278 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1279 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1280 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1281 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1282 v += 9; 1283 } 1284 idc = 3*(*c--); 1285 v = aa + 9*a->diag[i]; 1286 x[idc] += tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1287 x[1+idc] += tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1288 x[2+idc] += tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1289 } 1290 break; 1291 case 4: 1292 /* forward solve the lower triangular */ 1293 idx = 4*(*r++); 1294 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1295 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1296 for ( i=1; i<n; i++ ) { 1297 v = aa + 16*ai[i]; 1298 vi = aj + ai[i]; 1299 nz = a->diag[i] - ai[i]; 1300 idx = 4*(*r++); 1301 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1302 while (nz--) { 1303 idx = 4*(*vi++); 1304 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1305 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1306 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1307 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1308 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1309 v += 16; 1310 } 1311 idx = 4*i; 1312 tmp[idx] = sum1;tmp[1+idx] = sum2; 1313 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1314 } 1315 /* backward solve the upper triangular */ 1316 for ( i=n-1; i>=0; i-- ){ 1317 v = aa + 16*a->diag[i] + 16; 1318 vi = aj + a->diag[i] + 1; 1319 nz = ai[i+1] - a->diag[i] - 1; 1320 idt = 4*i; 1321 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1322 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1323 while (nz--) { 1324 idx = 4*(*vi++); 1325 x1 = tmp[idx]; x2 = tmp[1+idx]; 1326 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1327 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1328 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1329 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1330 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1331 v += 16; 1332 } 1333 idc = 4*(*c--); 1334 v = aa + 16*a->diag[i]; 1335 x[idc] += tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1336 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1337 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1338 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1339 } 1340 break; 1341 case 5: 1342 /* forward solve the lower triangular */ 1343 idx = 5*(*r++); 1344 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1345 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1346 for ( i=1; i<n; i++ ) { 1347 v = aa + 25*ai[i]; 1348 vi = aj + ai[i]; 1349 nz = a->diag[i] - ai[i]; 1350 idx = 5*(*r++); 1351 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1352 sum5 = b[4+idx]; 1353 while (nz--) { 1354 idx = 5*(*vi++); 1355 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1356 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1357 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1358 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1359 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1360 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1361 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1362 v += 25; 1363 } 1364 idx = 5*i; 1365 tmp[idx] = sum1;tmp[1+idx] = sum2; 1366 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1367 } 1368 /* backward solve the upper triangular */ 1369 for ( i=n-1; i>=0; i-- ){ 1370 v = aa + 25*a->diag[i] + 25; 1371 vi = aj + a->diag[i] + 1; 1372 nz = ai[i+1] - a->diag[i] - 1; 1373 idt = 5*i; 1374 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1375 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1376 while (nz--) { 1377 idx = 5*(*vi++); 1378 x1 = tmp[idx]; x2 = tmp[1+idx]; 1379 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1380 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1381 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1382 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1383 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1384 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1385 v += 25; 1386 } 1387 idc = 5*(*c--); 1388 v = aa + 25*a->diag[i]; 1389 x[idc] += tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1390 v[15]*sum4+v[20]*sum5; 1391 x[1+idc] += tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1392 v[16]*sum4+v[21]*sum5; 1393 x[2+idc] += tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1394 v[17]*sum4+v[22]*sum5; 1395 x[3+idc] += tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1396 v[18]*sum4+v[23]*sum5; 1397 x[4+idc] += tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1398 v[19]*sum4+v[24]*sum5; 1399 } 1400 break; 1401 default: { 1402 /* forward solve the lower triangular */ 1403 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 1404 for ( i=1; i<n; i++ ) { 1405 v = aa + bs2*ai[i]; 1406 vi = aj + ai[i]; 1407 nz = a->diag[i] - ai[i]; 1408 sum = tmp + bs*i; 1409 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 1410 while (nz--) { 1411 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,sum,&_One); 1412 v += bs2; 1413 } 1414 } 1415 /* backward solve the upper triangular */ 1416 lsum = a->solve_work + a->n; 1417 for ( i=n-1; i>=0; i-- ){ 1418 v = aa + bs2*(a->diag[i] + 1); 1419 vi = aj + a->diag[i] + 1; 1420 nz = ai[i+1] - a->diag[i] - 1; 1421 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 1422 while (nz--) { 1423 LAgemv_("N",&bs,&bs,&_DMOne,v,&bs,tmp+bs*(*vi++),&_One,&_DOne,lsum,&_One); 1424 v += bs2; 1425 } 1426 LAgemv_("N",&bs,&bs,&_DOne,aa+bs2*a->diag[i],&bs,lsum,&_One,&_DZero, 1427 tmp+i*bs,&_One); 1428 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 1429 } 1430 } 1431 } 1432 1433 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1434 ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 1435 return 0; 1436 } 1437 1438 /* ----------------------------------------------------------------*/ 1439 1440 int MatSolveTrans_SeqBAIJ(Mat A,Vec bb, Vec xx) 1441 { 1442 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1443 int ierr; 1444 1445 ierr = MatSolveAdd_SeqBAIJ_Private(A,bb,PETSC_NULL,xx); CHKERRQ(ierr); 1446 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 1447 return 0; 1448 } 1449 1450 /* ----------------------------------------------------------------*/ 1451 1452 int MatSolveTransAdd_SeqBAIJ(Mat A,Vec bb,Vec yy,Vec xx) 1453 { 1454 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 1455 int ierr; 1456 1457 ierr = MatSolveAdd_SeqBAIJ_Private(A,bb,yy,xx); CHKERRQ(ierr); 1458 PLogFlops(2*(a->bs2)*(a->nz)); 1459 return 0; 1460 } 1461 1462 /* ----------------------------------------------------------------*/ 1463 /* 1464 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1465 except that the data structure of Mat_SeqAIJ is slightly different. 1466 Not a good example of code reuse. 1467 */ 1468 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1469 Mat *fact) 1470 { 1471 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1472 IS isicol; 1473 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1474 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1475 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1476 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1477 PetscTruth col_identity, row_identity; 1478 1479 if (a->m != a->n) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Matrix must be square"); 1480 if (!isrow) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have row permutation"); 1481 if (!iscol) SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:Must have column permutation"); 1482 1483 /* special case that simply copies fill pattern */ 1484 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1485 if (levels == 0 && row_identity && col_identity) { 1486 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1487 (*fact)->factor = FACTOR_LU; 1488 b = (Mat_SeqBAIJ *) (*fact)->data; 1489 if (!b->diag) { 1490 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1491 } 1492 b->row = isrow; 1493 b->col = iscol; 1494 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1495 return 0; 1496 } 1497 1498 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1499 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1500 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1501 1502 /* get new row pointers */ 1503 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1504 ainew[0] = 0; 1505 /* don't know how many column pointers are needed so estimate */ 1506 jmax = (int) (f*ai[n] + 1); 1507 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1508 /* ajfill is level of fill for each fill entry */ 1509 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1510 /* fill is a linked list of nonzeros in active row */ 1511 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1512 /* im is level for each filled value */ 1513 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1514 /* dloc is location of diagonal in factor */ 1515 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1516 dloc[0] = 0; 1517 for ( prow=0; prow<n; prow++ ) { 1518 /* first copy previous fill into linked list */ 1519 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1520 xi = aj + ai[r[prow]]; 1521 fill[n] = n; 1522 while (nz--) { 1523 fm = n; 1524 idx = ic[*xi++]; 1525 do { 1526 m = fm; 1527 fm = fill[m]; 1528 } while (fm < idx); 1529 fill[m] = idx; 1530 fill[idx] = fm; 1531 im[idx] = 0; 1532 } 1533 nzi = 0; 1534 row = fill[n]; 1535 while ( row < prow ) { 1536 incrlev = im[row] + 1; 1537 nz = dloc[row]; 1538 xi = ajnew + ainew[row] + nz; 1539 flev = ajfill + ainew[row] + nz + 1; 1540 nnz = ainew[row+1] - ainew[row] - nz - 1; 1541 if (*xi++ != row) { 1542 SETERRQ(1,"MatILUFactorSymbolic_SeqBAIJ:zero pivot"); 1543 } 1544 fm = row; 1545 while (nnz-- > 0) { 1546 idx = *xi++; 1547 if (*flev + incrlev > levels) { 1548 flev++; 1549 continue; 1550 } 1551 do { 1552 m = fm; 1553 fm = fill[m]; 1554 } while (fm < idx); 1555 if (fm != idx) { 1556 im[idx] = *flev + incrlev; 1557 fill[m] = idx; 1558 fill[idx] = fm; 1559 fm = idx; 1560 nzf++; 1561 } 1562 else { 1563 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1564 } 1565 flev++; 1566 } 1567 row = fill[row]; 1568 nzi++; 1569 } 1570 /* copy new filled row into permanent storage */ 1571 ainew[prow+1] = ainew[prow] + nzf; 1572 if (ainew[prow+1] > jmax) { 1573 /* allocate a longer ajnew */ 1574 int maxadd; 1575 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1576 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1577 jmax += maxadd; 1578 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1579 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1580 PetscFree(ajnew); 1581 ajnew = xi; 1582 /* allocate a longer ajfill */ 1583 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1584 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1585 PetscFree(ajfill); 1586 ajfill = xi; 1587 realloc++; 1588 } 1589 xi = ajnew + ainew[prow]; 1590 flev = ajfill + ainew[prow]; 1591 dloc[prow] = nzi; 1592 fm = fill[n]; 1593 while (nzf--) { 1594 *xi++ = fm; 1595 *flev++ = im[fm]; 1596 fm = fill[fm]; 1597 } 1598 } 1599 PetscFree(ajfill); 1600 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1601 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1602 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1603 PetscFree(fill); PetscFree(im); 1604 1605 PLogInfo(A, 1606 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1607 realloc,f,((double)ainew[n])/((double)ai[prow])); 1608 1609 /* put together the new matrix */ 1610 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1611 b = (Mat_SeqBAIJ *) (*fact)->data; 1612 PetscFree(b->imax); 1613 b->singlemalloc = 0; 1614 len = bs2*ainew[n]*sizeof(Scalar); 1615 /* the next line frees the default space generated by the Create() */ 1616 PetscFree(b->a); PetscFree(b->ilen); 1617 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1618 b->j = ajnew; 1619 b->i = ainew; 1620 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1621 b->diag = dloc; 1622 b->ilen = 0; 1623 b->imax = 0; 1624 b->row = isrow; 1625 b->col = iscol; 1626 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1627 CHKPTRQ(b->solve_work); 1628 /* In b structure: Free imax, ilen, old a, old j. 1629 Allocate dloc, solve_work, new a, new j */ 1630 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1631 b->maxnz = b->nz = ainew[n]; 1632 (*fact)->factor = FACTOR_LU; 1633 return 0; 1634 } 1635 1636 1637 1638 1639