1 2 #ifndef lint 3 static char vcid[] = "$Id: baijfact.c,v 1.42 1997/05/14 21:47:16 balay Exp balay $"; 4 #endif 5 /* 6 Factorization code for BAIJ format. 7 */ 8 9 #include "src/mat/impls/baij/seq/baij.h" 10 #include "src/vec/vecimpl.h" 11 #include "src/inline/ilu.h" 12 13 14 /* 15 The symbolic factorization code is identical to that for AIJ format, 16 except for very small changes since this is now a SeqBAIJ datastructure. 17 NOT good code reuse. 18 */ 19 #undef __FUNC__ 20 #define __FUNC__ "MatLUFactorSymbolic_SeqBAIJ" 21 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 22 { 23 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 24 IS isicol; 25 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 26 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 27 int *idnew, idx, row,m,fm, nnz, nzi,realloc = 0,nzbd,*im; 28 29 PetscValidHeaderSpecific(isrow,IS_COOKIE); 30 PetscValidHeaderSpecific(iscol,IS_COOKIE); 31 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 32 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 33 34 /* get new row pointers */ 35 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 36 ainew[0] = 0; 37 /* don't know how many column pointers are needed so estimate */ 38 jmax = (int) (f*ai[n] + 1); 39 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 40 /* fill is a linked list of nonzeros in active row */ 41 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 42 im = fill + n + 1; 43 /* idnew is location of diagonal in factor */ 44 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 45 idnew[0] = 0; 46 47 for ( i=0; i<n; i++ ) { 48 /* first copy previous fill into linked list */ 49 nnz = nz = ai[r[i]+1] - ai[r[i]]; 50 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 51 ajtmp = aj + ai[r[i]]; 52 fill[n] = n; 53 while (nz--) { 54 fm = n; 55 idx = ic[*ajtmp++]; 56 do { 57 m = fm; 58 fm = fill[m]; 59 } while (fm < idx); 60 fill[m] = idx; 61 fill[idx] = fm; 62 } 63 row = fill[n]; 64 while ( row < i ) { 65 ajtmp = ajnew + idnew[row] + 1; 66 nzbd = 1 + idnew[row] - ainew[row]; 67 nz = im[row] - nzbd; 68 fm = row; 69 while (nz-- > 0) { 70 idx = *ajtmp++; 71 nzbd++; 72 if (idx == i) im[row] = nzbd; 73 do { 74 m = fm; 75 fm = fill[m]; 76 } while (fm < idx); 77 if (fm != idx) { 78 fill[m] = idx; 79 fill[idx] = fm; 80 fm = idx; 81 nnz++; 82 } 83 } 84 row = fill[row]; 85 } 86 /* copy new filled row into permanent storage */ 87 ainew[i+1] = ainew[i] + nnz; 88 if (ainew[i+1] > jmax) { 89 /* allocate a longer ajnew */ 90 int maxadd; 91 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 92 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 93 jmax += maxadd; 94 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 95 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 96 PetscFree(ajnew); 97 ajnew = ajtmp; 98 realloc++; /* count how many times we realloc */ 99 } 100 ajtmp = ajnew + ainew[i]; 101 fm = fill[n]; 102 nzi = 0; 103 im[i] = nnz; 104 while (nnz--) { 105 if (fm < i) nzi++; 106 *ajtmp++ = fm; 107 fm = fill[fm]; 108 } 109 idnew[i] = ainew[i] + nzi; 110 } 111 112 if (ai[i] != 0) { 113 PLogInfo(A, 114 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 115 realloc,f,((double)ainew[n])/((double)ai[i])); 116 } else { 117 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Empty matrix.\n"); 118 } 119 120 121 122 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 123 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 124 125 PetscFree(fill); 126 127 /* put together the new matrix */ 128 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 129 PLogObjectParent(*B,isicol); 130 ierr = ISDestroy(isicol); CHKERRQ(ierr); 131 b = (Mat_SeqBAIJ *) (*B)->data; 132 PetscFree(b->imax); 133 b->singlemalloc = 0; 134 /* the next line frees the default space generated by the Create() */ 135 PetscFree(b->a); PetscFree(b->ilen); 136 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 137 b->j = ajnew; 138 b->i = ainew; 139 b->diag = idnew; 140 b->ilen = 0; 141 b->imax = 0; 142 b->row = isrow; 143 b->col = iscol; 144 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 145 CHKPTRQ(b->solve_work); 146 /* In b structure: Free imax, ilen, old a, old j. 147 Allocate idnew, solve_work, new a, new j */ 148 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 149 b->maxnz = b->nz = ainew[n]; 150 151 (*B)->info.factor_mallocs = realloc; 152 (*B)->info.fill_ratio_given = f; 153 if (ai[i] != 0) { 154 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 155 } else { 156 (*B)->info.fill_ratio_needed = 0.0; 157 } 158 159 160 return 0; 161 } 162 163 /* ----------------------------------------------------------- */ 164 #undef __FUNC__ 165 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 166 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 167 { 168 Mat C = *B; 169 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 170 IS iscol = b->col, isrow = b->row, isicol; 171 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 172 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 173 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 174 Scalar *ba = b->a,*aa = a->a; 175 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 176 register int *pj; 177 178 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 179 PLogObjectParent(*B,isicol); 180 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 181 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 182 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 183 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 184 /* generate work space needed by dense LU factorization */ 185 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 186 CHKPTRQ(v_work); 187 multiplier = v_work + bs; 188 v_pivots = (int *) (multiplier + bs2); 189 190 /* flops in while loop */ 191 bslog = 2*bs*bs2; 192 193 for ( i=0; i<n; i++ ) { 194 nz = bi[i+1] - bi[i]; 195 ajtmp = bj + bi[i]; 196 for ( j=0; j<nz; j++ ) { 197 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 198 } 199 /* load in initial (unfactored row) */ 200 nz = ai[r[i]+1] - ai[r[i]]; 201 ajtmpold = aj + ai[r[i]]; 202 v = aa + bs2*ai[r[i]]; 203 for ( j=0; j<nz; j++ ) { 204 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 205 } 206 row = *ajtmp++; 207 while (row < i) { 208 pc = rtmp + bs2*row; 209 /* if (*pc) { */ 210 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 211 if (flg) { 212 pv = ba + bs2*diag_offset[row]; 213 pj = bj + diag_offset[row] + 1; 214 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 215 nz = bi[row+1] - diag_offset[row] - 1; 216 pv += bs2; 217 for (j=0; j<nz; j++) { 218 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 219 } 220 PLogFlops(bslog*(nz+1)-bs); 221 } 222 row = *ajtmp++; 223 } 224 /* finished row so stick it into b->a */ 225 pv = ba + bs2*bi[i]; 226 pj = bj + bi[i]; 227 nz = bi[i+1] - bi[i]; 228 for ( j=0; j<nz; j++ ) { 229 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 230 } 231 diag = diag_offset[i] - bi[i]; 232 /* invert diagonal block */ 233 w = pv + bs2*diag; 234 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 235 } 236 237 PetscFree(rtmp); PetscFree(v_work); 238 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 239 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 240 ierr = ISDestroy(isicol); CHKERRQ(ierr); 241 C->factor = FACTOR_LU; 242 C->assembled = PETSC_TRUE; 243 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 244 return 0; 245 } 246 /* ------------------------------------------------------------*/ 247 /* 248 Version for when blocks are 5 by 5 249 */ 250 #undef __FUNC__ 251 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 252 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 253 { 254 Mat C = *B; 255 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 256 IS iscol = b->col, isrow = b->row, isicol; 257 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 258 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 259 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 260 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 261 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 262 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 263 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 264 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 265 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 266 Scalar *ba = b->a,*aa = a->a; 267 register int *pj; 268 269 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 270 PLogObjectParent(*B,isicol); 271 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 272 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 273 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 274 275 for ( i=0; i<n; i++ ) { 276 nz = bi[i+1] - bi[i]; 277 ajtmp = bj + bi[i]; 278 for ( j=0; j<nz; j++ ) { 279 x = rtmp+25*ajtmp[j]; 280 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 281 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 282 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 283 } 284 /* load in initial (unfactored row) */ 285 idx = r[i]; 286 nz = ai[idx+1] - ai[idx]; 287 ajtmpold = aj + ai[idx]; 288 v = aa + 25*ai[idx]; 289 for ( j=0; j<nz; j++ ) { 290 x = rtmp+25*ic[ajtmpold[j]]; 291 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 292 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 293 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 294 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 295 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 296 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 297 v += 25; 298 } 299 row = *ajtmp++; 300 while (row < i) { 301 pc = rtmp + 25*row; 302 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 303 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 304 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 305 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 306 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 307 p25 = pc[24]; 308 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 309 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 310 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 311 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 312 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 313 p24 != 0.0 || p25 != 0.0) { 314 pv = ba + 25*diag_offset[row]; 315 pj = bj + diag_offset[row] + 1; 316 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 317 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 318 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 319 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 320 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 321 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 322 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 323 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 324 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 325 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 326 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 327 328 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 329 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 330 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 331 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 332 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 333 334 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 335 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 336 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 337 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 338 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 339 340 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 341 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 342 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 343 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 344 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 345 346 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 347 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 348 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 349 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 350 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 351 352 nz = bi[row+1] - diag_offset[row] - 1; 353 pv += 25; 354 for (j=0; j<nz; j++) { 355 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 356 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 357 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 358 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 359 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 360 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 361 x = rtmp + 25*pj[j]; 362 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 363 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 364 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 365 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 366 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 367 368 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 369 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 370 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 371 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 372 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 373 374 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 375 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 376 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 377 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 378 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 379 380 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 381 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 382 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 383 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 384 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 385 386 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 387 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 388 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 389 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 390 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 391 392 pv += 25; 393 } 394 PLogFlops(250*nz+225); 395 } 396 row = *ajtmp++; 397 } 398 /* finished row so stick it into b->a */ 399 pv = ba + 25*bi[i]; 400 pj = bj + bi[i]; 401 nz = bi[i+1] - bi[i]; 402 for ( j=0; j<nz; j++ ) { 403 x = rtmp+25*pj[j]; 404 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 405 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 406 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 407 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 408 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 409 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 410 pv += 25; 411 } 412 /* invert diagonal block */ 413 w = ba + 25*diag_offset[i]; 414 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 415 } 416 417 PetscFree(rtmp); 418 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 419 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 420 ierr = ISDestroy(isicol); CHKERRQ(ierr); 421 C->factor = FACTOR_LU; 422 C->assembled = PETSC_TRUE; 423 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 424 return 0; 425 } 426 427 /* ------------------------------------------------------------*/ 428 /* 429 Version for when blocks are 4 by 4 430 */ 431 #undef __FUNC__ 432 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 433 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 434 { 435 Mat C = *B; 436 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 437 IS iscol = b->col, isrow = b->row, isicol; 438 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 439 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 440 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 441 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 442 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 443 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 444 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 445 Scalar m13,m14,m15,m16; 446 Scalar *ba = b->a,*aa = a->a; 447 register int *pj; 448 449 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 450 PLogObjectParent(*B,isicol); 451 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 452 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 453 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 454 455 for ( i=0; i<n; i++ ) { 456 nz = bi[i+1] - bi[i]; 457 ajtmp = bj + bi[i]; 458 for ( j=0; j<nz; j++ ) { 459 x = rtmp+16*ajtmp[j]; 460 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 461 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 462 } 463 /* load in initial (unfactored row) */ 464 idx = r[i]; 465 nz = ai[idx+1] - ai[idx]; 466 ajtmpold = aj + ai[idx]; 467 v = aa + 16*ai[idx]; 468 for ( j=0; j<nz; j++ ) { 469 x = rtmp+16*ic[ajtmpold[j]]; 470 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 471 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 472 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 473 x[14] = v[14]; x[15] = v[15]; 474 v += 16; 475 } 476 row = *ajtmp++; 477 while (row < i) { 478 pc = rtmp + 16*row; 479 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 480 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 481 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 482 p15 = pc[14]; p16 = pc[15]; 483 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 484 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 485 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 486 || p16 != 0.0) { 487 pv = ba + 16*diag_offset[row]; 488 pj = bj + diag_offset[row] + 1; 489 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 490 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 491 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 492 x15 = pv[14]; x16 = pv[15]; 493 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 494 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 495 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 496 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 497 498 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 499 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 500 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 501 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 502 503 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 504 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 505 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 506 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 507 508 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 509 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 510 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 511 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 512 513 nz = bi[row+1] - diag_offset[row] - 1; 514 pv += 16; 515 for (j=0; j<nz; j++) { 516 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 517 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 518 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 519 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 520 x = rtmp + 16*pj[j]; 521 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 522 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 523 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 524 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 525 526 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 527 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 528 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 529 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 530 531 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 532 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 533 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 534 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 535 536 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 537 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 538 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 539 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 540 541 pv += 16; 542 } 543 PLogFlops(128*nz+112); 544 } 545 row = *ajtmp++; 546 } 547 /* finished row so stick it into b->a */ 548 pv = ba + 16*bi[i]; 549 pj = bj + bi[i]; 550 nz = bi[i+1] - bi[i]; 551 for ( j=0; j<nz; j++ ) { 552 x = rtmp+16*pj[j]; 553 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 554 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 555 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 556 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 557 pv += 16; 558 } 559 /* invert diagonal block */ 560 w = ba + 16*diag_offset[i]; 561 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 562 } 563 564 PetscFree(rtmp); 565 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 566 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 567 ierr = ISDestroy(isicol); CHKERRQ(ierr); 568 C->factor = FACTOR_LU; 569 C->assembled = PETSC_TRUE; 570 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 571 return 0; 572 } 573 /* ------------------------------------------------------------*/ 574 /* 575 Version for when blocks are 3 by 3 576 */ 577 #undef __FUNC__ 578 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 579 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 580 { 581 Mat C = *B; 582 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 583 IS iscol = b->col, isrow = b->row, isicol; 584 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 585 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 586 int *diag_offset = b->diag,idx; 587 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 588 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 589 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 590 Scalar *ba = b->a,*aa = a->a; 591 register int *pj; 592 593 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 594 PLogObjectParent(*B,isicol); 595 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 596 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 597 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 598 599 for ( i=0; i<n; i++ ) { 600 nz = bi[i+1] - bi[i]; 601 ajtmp = bj + bi[i]; 602 for ( j=0; j<nz; j++ ) { 603 x = rtmp + 9*ajtmp[j]; 604 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 605 } 606 /* load in initial (unfactored row) */ 607 idx = r[i]; 608 nz = ai[idx+1] - ai[idx]; 609 ajtmpold = aj + ai[idx]; 610 v = aa + 9*ai[idx]; 611 for ( j=0; j<nz; j++ ) { 612 x = rtmp + 9*ic[ajtmpold[j]]; 613 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 614 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 615 v += 9; 616 } 617 row = *ajtmp++; 618 while (row < i) { 619 pc = rtmp + 9*row; 620 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 621 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 622 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 623 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 624 pv = ba + 9*diag_offset[row]; 625 pj = bj + diag_offset[row] + 1; 626 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 627 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 628 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 629 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 630 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 631 632 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 633 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 634 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 635 636 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 637 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 638 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 639 nz = bi[row+1] - diag_offset[row] - 1; 640 pv += 9; 641 for (j=0; j<nz; j++) { 642 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 643 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 644 x = rtmp + 9*pj[j]; 645 x[0] -= m1*x1 + m4*x2 + m7*x3; 646 x[1] -= m2*x1 + m5*x2 + m8*x3; 647 x[2] -= m3*x1 + m6*x2 + m9*x3; 648 649 x[3] -= m1*x4 + m4*x5 + m7*x6; 650 x[4] -= m2*x4 + m5*x5 + m8*x6; 651 x[5] -= m3*x4 + m6*x5 + m9*x6; 652 653 x[6] -= m1*x7 + m4*x8 + m7*x9; 654 x[7] -= m2*x7 + m5*x8 + m8*x9; 655 x[8] -= m3*x7 + m6*x8 + m9*x9; 656 pv += 9; 657 } 658 PLogFlops(54*nz+36); 659 } 660 row = *ajtmp++; 661 } 662 /* finished row so stick it into b->a */ 663 pv = ba + 9*bi[i]; 664 pj = bj + bi[i]; 665 nz = bi[i+1] - bi[i]; 666 for ( j=0; j<nz; j++ ) { 667 x = rtmp + 9*pj[j]; 668 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 669 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 670 pv += 9; 671 } 672 /* invert diagonal block */ 673 w = ba + 9*diag_offset[i]; 674 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 675 } 676 677 PetscFree(rtmp); 678 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 679 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 680 ierr = ISDestroy(isicol); CHKERRQ(ierr); 681 C->factor = FACTOR_LU; 682 C->assembled = PETSC_TRUE; 683 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 684 return 0; 685 } 686 687 /* ------------------------------------------------------------*/ 688 /* 689 Version for when blocks are 2 by 2 690 */ 691 #undef __FUNC__ 692 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 693 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 694 { 695 Mat C = *B; 696 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 697 IS iscol = b->col, isrow = b->row, isicol; 698 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 699 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 700 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 701 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 702 Scalar p1,p2,p3,p4,v_work[2]; 703 Scalar *ba = b->a,*aa = a->a; 704 register int *pj; 705 706 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 707 PLogObjectParent(*B,isicol); 708 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 709 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 710 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 711 712 for ( i=0; i<n; i++ ) { 713 nz = bi[i+1] - bi[i]; 714 ajtmp = bj + bi[i]; 715 for ( j=0; j<nz; j++ ) { 716 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 717 } 718 /* load in initial (unfactored row) */ 719 idx = r[i]; 720 nz = ai[idx+1] - ai[idx]; 721 ajtmpold = aj + ai[idx]; 722 v = aa + 4*ai[idx]; 723 for ( j=0; j<nz; j++ ) { 724 x = rtmp+4*ic[ajtmpold[j]]; 725 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 726 v += 4; 727 } 728 row = *ajtmp++; 729 while (row < i) { 730 pc = rtmp + 4*row; 731 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 732 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 733 pv = ba + 4*diag_offset[row]; 734 pj = bj + diag_offset[row] + 1; 735 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 736 pc[0] = m1 = p1*x1 + p3*x2; 737 pc[1] = m2 = p2*x1 + p4*x2; 738 pc[2] = m3 = p1*x3 + p3*x4; 739 pc[3] = m4 = p2*x3 + p4*x4; 740 nz = bi[row+1] - diag_offset[row] - 1; 741 pv += 4; 742 for (j=0; j<nz; j++) { 743 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 744 x = rtmp + 4*pj[j]; 745 x[0] -= m1*x1 + m3*x2; 746 x[1] -= m2*x1 + m4*x2; 747 x[2] -= m1*x3 + m3*x4; 748 x[3] -= m2*x3 + m4*x4; 749 pv += 4; 750 } 751 PLogFlops(16*nz+12); 752 } 753 row = *ajtmp++; 754 } 755 /* finished row so stick it into b->a */ 756 pv = ba + 4*bi[i]; 757 pj = bj + bi[i]; 758 nz = bi[i+1] - bi[i]; 759 for ( j=0; j<nz; j++ ) { 760 x = rtmp+4*pj[j]; 761 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 762 pv += 4; 763 } 764 /* invert diagonal block */ 765 w = ba + 4*diag_offset[i]; 766 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 767 } 768 769 PetscFree(rtmp); 770 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 771 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 772 ierr = ISDestroy(isicol); CHKERRQ(ierr); 773 C->factor = FACTOR_LU; 774 C->assembled = PETSC_TRUE; 775 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 776 return 0; 777 } 778 779 /* ----------------------------------------------------------- */ 780 /* 781 Version for when blocks are 1 by 1. 782 */ 783 #undef __FUNC__ 784 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 785 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 786 { 787 Mat C = *B; 788 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 789 IS iscol = b->col, isrow = b->row, isicol; 790 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 791 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 792 int *diag_offset = b->diag,diag; 793 register Scalar *pv,*v,*rtmp,multiplier,*pc; 794 Scalar *ba = b->a,*aa = a->a; 795 register int *pj; 796 797 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 798 PLogObjectParent(*B,isicol); 799 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 800 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 801 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 802 803 for ( i=0; i<n; i++ ) { 804 nz = bi[i+1] - bi[i]; 805 ajtmp = bj + bi[i]; 806 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 807 808 /* load in initial (unfactored row) */ 809 nz = ai[r[i]+1] - ai[r[i]]; 810 ajtmpold = aj + ai[r[i]]; 811 v = aa + ai[r[i]]; 812 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 813 814 row = *ajtmp++; 815 while (row < i) { 816 pc = rtmp + row; 817 if (*pc != 0.0) { 818 pv = ba + diag_offset[row]; 819 pj = bj + diag_offset[row] + 1; 820 multiplier = *pc * *pv++; 821 *pc = multiplier; 822 nz = bi[row+1] - diag_offset[row] - 1; 823 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 824 PLogFlops(1+2*nz); 825 } 826 row = *ajtmp++; 827 } 828 /* finished row so stick it into b->a */ 829 pv = ba + bi[i]; 830 pj = bj + bi[i]; 831 nz = bi[i+1] - bi[i]; 832 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 833 diag = diag_offset[i] - bi[i]; 834 /* check pivot entry for current row */ 835 if (pv[diag] == 0.0) { 836 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 837 } 838 pv[diag] = 1.0/pv[diag]; 839 } 840 841 PetscFree(rtmp); 842 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 843 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 844 ierr = ISDestroy(isicol); CHKERRQ(ierr); 845 C->factor = FACTOR_LU; 846 C->assembled = PETSC_TRUE; 847 PLogFlops(b->n); 848 return 0; 849 } 850 851 /* ----------------------------------------------------------- */ 852 #undef __FUNC__ 853 #define __FUNC__ "MatLUFactor_SeqBAIJ" 854 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 855 { 856 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 857 int ierr; 858 Mat C; 859 860 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 861 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 862 863 /* free all the data structures from mat */ 864 PetscFree(mat->a); 865 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 866 if (mat->diag) PetscFree(mat->diag); 867 if (mat->ilen) PetscFree(mat->ilen); 868 if (mat->imax) PetscFree(mat->imax); 869 if (mat->solve_work) PetscFree(mat->solve_work); 870 if (mat->mult_work) PetscFree(mat->mult_work); 871 PetscFree(mat); 872 873 PetscMemcpy(A,C,sizeof(struct _Mat)); 874 PetscHeaderDestroy(C); 875 return 0; 876 } 877 /* ----------------------------------------------------------- */ 878 #undef __FUNC__ 879 #define __FUNC__ "MatSolve_SeqBAIJ_N" 880 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 881 { 882 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 883 IS iscol=a->col,isrow=a->row; 884 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 885 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 886 Scalar *aa=a->a,*sum; 887 register Scalar *x,*b,*lsum,*tmp,*v; 888 889 VecGetArray_Fast(bb,b); 890 VecGetArray_Fast(xx,x); 891 tmp = a->solve_work; 892 893 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 894 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 895 896 /* forward solve the lower triangular */ 897 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 898 for ( i=1; i<n; i++ ) { 899 v = aa + bs2*ai[i]; 900 vi = aj + ai[i]; 901 nz = a->diag[i] - ai[i]; 902 sum = tmp + bs*i; 903 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 904 while (nz--) { 905 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 906 v += bs2; 907 } 908 } 909 /* backward solve the upper triangular */ 910 lsum = a->solve_work + a->n; 911 for ( i=n-1; i>=0; i-- ){ 912 v = aa + bs2*(a->diag[i] + 1); 913 vi = aj + a->diag[i] + 1; 914 nz = ai[i+1] - a->diag[i] - 1; 915 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 916 while (nz--) { 917 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 918 v += bs2; 919 } 920 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 921 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 922 } 923 924 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 925 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 926 VecRestoreArray_Fast(bb,b); 927 VecRestoreArray_Fast(xx,x); 928 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 929 return 0; 930 } 931 932 #undef __FUNC__ 933 #define __FUNC__ "MatSolve_SeqBAIJ_7" 934 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 935 { 936 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 937 IS iscol=a->col,isrow=a->row; 938 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 939 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 940 register Scalar *x,*b,*tmp,*v; 941 942 VecGetArray_Fast(bb,b); 943 VecGetArray_Fast(xx,x); 944 tmp = a->solve_work; 945 946 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 947 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 948 949 /* forward solve the lower triangular */ 950 idx = 7*(*r++); 951 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 952 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 953 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 954 955 for ( i=1; i<n; i++ ) { 956 v = aa + 49*ai[i]; 957 vi = aj + ai[i]; 958 nz = a->diag[i] - ai[i]; 959 idx = 7*(*r++); 960 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 961 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 962 while (nz--) { 963 idx = 7*(*vi++); 964 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 965 x4 = tmp[3+idx];x5 = tmp[4+idx]; 966 x6 = tmp[5+idx];x7 = tmp[6+idx]; 967 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 968 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 969 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 970 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 971 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 972 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 973 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 974 v += 49; 975 } 976 idx = 7*i; 977 tmp[idx] = sum1;tmp[1+idx] = sum2; 978 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 979 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 980 } 981 /* backward solve the upper triangular */ 982 for ( i=n-1; i>=0; i-- ){ 983 v = aa + 49*a->diag[i] + 49; 984 vi = aj + a->diag[i] + 1; 985 nz = ai[i+1] - a->diag[i] - 1; 986 idt = 7*i; 987 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 988 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 989 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 990 while (nz--) { 991 idx = 7*(*vi++); 992 x1 = tmp[idx]; x2 = tmp[1+idx]; 993 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 994 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 995 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 996 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 997 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 998 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 999 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 1000 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 1001 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 1002 v += 49; 1003 } 1004 idc = 7*(*c--); 1005 v = aa + 49*a->diag[i]; 1006 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 1007 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 1008 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 1009 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 1010 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 1011 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1012 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1013 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1014 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1015 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1016 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1017 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1018 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1019 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1020 } 1021 1022 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1023 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1024 VecRestoreArray_Fast(bb,b); 1025 VecRestoreArray_Fast(xx,x); 1026 PLogFlops(2*49*(a->nz) - a->n); 1027 return 0; 1028 } 1029 1030 #undef __FUNC__ 1031 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1032 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1033 { 1034 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1035 IS iscol=a->col,isrow=a->row; 1036 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1037 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1038 register Scalar *x,*b,*tmp,*v; 1039 1040 VecGetArray_Fast(bb,b); 1041 VecGetArray_Fast(xx,x); 1042 tmp = a->solve_work; 1043 1044 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1045 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1046 1047 /* forward solve the lower triangular */ 1048 idx = 5*(*r++); 1049 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1050 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1051 for ( i=1; i<n; i++ ) { 1052 v = aa + 25*ai[i]; 1053 vi = aj + ai[i]; 1054 nz = a->diag[i] - ai[i]; 1055 idx = 5*(*r++); 1056 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1057 sum5 = b[4+idx]; 1058 while (nz--) { 1059 idx = 5*(*vi++); 1060 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1061 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1062 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1063 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1064 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1065 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1066 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1067 v += 25; 1068 } 1069 idx = 5*i; 1070 tmp[idx] = sum1;tmp[1+idx] = sum2; 1071 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1072 } 1073 /* backward solve the upper triangular */ 1074 for ( i=n-1; i>=0; i-- ){ 1075 v = aa + 25*a->diag[i] + 25; 1076 vi = aj + a->diag[i] + 1; 1077 nz = ai[i+1] - a->diag[i] - 1; 1078 idt = 5*i; 1079 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1080 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1081 while (nz--) { 1082 idx = 5*(*vi++); 1083 x1 = tmp[idx]; x2 = tmp[1+idx]; 1084 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1085 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1086 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1087 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1088 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1089 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1090 v += 25; 1091 } 1092 idc = 5*(*c--); 1093 v = aa + 25*a->diag[i]; 1094 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1095 v[15]*sum4+v[20]*sum5; 1096 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1097 v[16]*sum4+v[21]*sum5; 1098 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1099 v[17]*sum4+v[22]*sum5; 1100 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1101 v[18]*sum4+v[23]*sum5; 1102 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1103 v[19]*sum4+v[24]*sum5; 1104 } 1105 1106 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1107 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1108 VecRestoreArray_Fast(bb,b); 1109 VecRestoreArray_Fast(xx,x); 1110 PLogFlops(2*25*(a->nz) - a->n); 1111 return 0; 1112 } 1113 1114 #undef __FUNC__ 1115 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1116 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1117 { 1118 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1119 IS iscol=a->col,isrow=a->row; 1120 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1121 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1122 register Scalar *x,*b,*tmp,*v; 1123 1124 VecGetArray_Fast(bb,b); 1125 VecGetArray_Fast(xx,x); 1126 tmp = a->solve_work; 1127 1128 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1129 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1130 1131 /* forward solve the lower triangular */ 1132 idx = 4*(*r++); 1133 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1134 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1135 for ( i=1; i<n; i++ ) { 1136 v = aa + 16*ai[i]; 1137 vi = aj + ai[i]; 1138 nz = a->diag[i] - ai[i]; 1139 idx = 4*(*r++); 1140 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1141 while (nz--) { 1142 idx = 4*(*vi++); 1143 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1144 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1145 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1146 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1147 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1148 v += 16; 1149 } 1150 idx = 4*i; 1151 tmp[idx] = sum1;tmp[1+idx] = sum2; 1152 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1153 } 1154 /* backward solve the upper triangular */ 1155 for ( i=n-1; i>=0; i-- ){ 1156 v = aa + 16*a->diag[i] + 16; 1157 vi = aj + a->diag[i] + 1; 1158 nz = ai[i+1] - a->diag[i] - 1; 1159 idt = 4*i; 1160 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1161 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1162 while (nz--) { 1163 idx = 4*(*vi++); 1164 x1 = tmp[idx]; x2 = tmp[1+idx]; 1165 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1166 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1167 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1168 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1169 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1170 v += 16; 1171 } 1172 idc = 4*(*c--); 1173 v = aa + 16*a->diag[i]; 1174 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1175 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1176 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1177 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1178 } 1179 1180 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1181 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1182 VecRestoreArray_Fast(bb,b); 1183 VecRestoreArray_Fast(xx,x); 1184 PLogFlops(2*16*(a->nz) - a->n); 1185 return 0; 1186 } 1187 1188 1189 #undef __FUNC__ 1190 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1191 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1192 { 1193 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1194 IS iscol=a->col,isrow=a->row; 1195 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1196 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1197 register Scalar *x,*b,*tmp,*v; 1198 1199 VecGetArray_Fast(bb,b); 1200 VecGetArray_Fast(xx,x); 1201 tmp = a->solve_work; 1202 1203 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1204 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1205 1206 /* forward solve the lower triangular */ 1207 idx = 3*(*r++); 1208 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1209 for ( i=1; i<n; i++ ) { 1210 v = aa + 9*ai[i]; 1211 vi = aj + ai[i]; 1212 nz = a->diag[i] - ai[i]; 1213 idx = 3*(*r++); 1214 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1215 while (nz--) { 1216 idx = 3*(*vi++); 1217 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1218 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1219 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1220 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1221 v += 9; 1222 } 1223 idx = 3*i; 1224 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1225 } 1226 /* backward solve the upper triangular */ 1227 for ( i=n-1; i>=0; i-- ){ 1228 v = aa + 9*a->diag[i] + 9; 1229 vi = aj + a->diag[i] + 1; 1230 nz = ai[i+1] - a->diag[i] - 1; 1231 idt = 3*i; 1232 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1233 while (nz--) { 1234 idx = 3*(*vi++); 1235 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1236 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1237 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1238 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1239 v += 9; 1240 } 1241 idc = 3*(*c--); 1242 v = aa + 9*a->diag[i]; 1243 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1244 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1245 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1246 } 1247 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1248 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1249 VecRestoreArray_Fast(bb,b); 1250 VecRestoreArray_Fast(xx,x); 1251 PLogFlops(2*9*(a->nz) - a->n); 1252 return 0; 1253 } 1254 1255 #undef __FUNC__ 1256 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1257 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1258 { 1259 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1260 IS iscol=a->col,isrow=a->row; 1261 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1262 Scalar *aa=a->a,sum1,sum2,x1,x2; 1263 register Scalar *x,*b,*tmp,*v; 1264 1265 VecGetArray_Fast(bb,b); 1266 VecGetArray_Fast(xx,x); 1267 tmp = a->solve_work; 1268 1269 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1270 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1271 1272 /* forward solve the lower triangular */ 1273 idx = 2*(*r++); 1274 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1275 for ( i=1; i<n; i++ ) { 1276 v = aa + 4*ai[i]; 1277 vi = aj + ai[i]; 1278 nz = a->diag[i] - ai[i]; 1279 idx = 2*(*r++); 1280 sum1 = b[idx]; sum2 = b[1+idx]; 1281 while (nz--) { 1282 idx = 2*(*vi++); 1283 x1 = tmp[idx]; x2 = tmp[1+idx]; 1284 sum1 -= v[0]*x1 + v[2]*x2; 1285 sum2 -= v[1]*x1 + v[3]*x2; 1286 v += 4; 1287 } 1288 idx = 2*i; 1289 tmp[idx] = sum1; tmp[1+idx] = sum2; 1290 } 1291 /* backward solve the upper triangular */ 1292 for ( i=n-1; i>=0; i-- ){ 1293 v = aa + 4*a->diag[i] + 4; 1294 vi = aj + a->diag[i] + 1; 1295 nz = ai[i+1] - a->diag[i] - 1; 1296 idt = 2*i; 1297 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1298 while (nz--) { 1299 idx = 2*(*vi++); 1300 x1 = tmp[idx]; x2 = tmp[1+idx]; 1301 sum1 -= v[0]*x1 + v[2]*x2; 1302 sum2 -= v[1]*x1 + v[3]*x2; 1303 v += 4; 1304 } 1305 idc = 2*(*c--); 1306 v = aa + 4*a->diag[i]; 1307 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1308 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1309 } 1310 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1311 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1312 VecRestoreArray_Fast(bb,b); 1313 VecRestoreArray_Fast(xx,x); 1314 PLogFlops(2*4*(a->nz) - a->n); 1315 return 0; 1316 } 1317 1318 1319 #undef __FUNC__ 1320 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1321 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1322 { 1323 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1324 IS iscol=a->col,isrow=a->row; 1325 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1326 Scalar *aa=a->a,sum1; 1327 register Scalar *x,*b,*tmp,*v; 1328 1329 if (!n) return 0; 1330 1331 VecGetArray_Fast(bb,b); 1332 VecGetArray_Fast(xx,x); 1333 tmp = a->solve_work; 1334 1335 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1336 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1337 1338 /* forward solve the lower triangular */ 1339 tmp[0] = b[*r++]; 1340 for ( i=1; i<n; i++ ) { 1341 v = aa + ai[i]; 1342 vi = aj + ai[i]; 1343 nz = a->diag[i] - ai[i]; 1344 sum1 = b[*r++]; 1345 while (nz--) { 1346 sum1 -= (*v++)*tmp[*vi++]; 1347 } 1348 tmp[i] = sum1; 1349 } 1350 /* backward solve the upper triangular */ 1351 for ( i=n-1; i>=0; i-- ){ 1352 v = aa + a->diag[i] + 1; 1353 vi = aj + a->diag[i] + 1; 1354 nz = ai[i+1] - a->diag[i] - 1; 1355 sum1 = tmp[i]; 1356 while (nz--) { 1357 sum1 -= (*v++)*tmp[*vi++]; 1358 } 1359 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1360 } 1361 1362 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1363 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1364 VecRestoreArray_Fast(bb,b); 1365 VecRestoreArray_Fast(xx,x); 1366 PLogFlops(2*1*(a->nz) - a->n); 1367 return 0; 1368 } 1369 1370 /* ----------------------------------------------------------------*/ 1371 /* 1372 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1373 except that the data structure of Mat_SeqAIJ is slightly different. 1374 Not a good example of code reuse. 1375 */ 1376 #undef __FUNC__ 1377 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1378 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1379 Mat *fact) 1380 { 1381 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1382 IS isicol; 1383 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1384 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1385 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1386 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1387 PetscTruth col_identity, row_identity; 1388 1389 /* special case that simply copies fill pattern */ 1390 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1391 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1392 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1393 if (levels == 0 && row_identity && col_identity) { 1394 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1395 (*fact)->factor = FACTOR_LU; 1396 b = (Mat_SeqBAIJ *) (*fact)->data; 1397 if (!b->diag) { 1398 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1399 } 1400 b->row = isrow; 1401 b->col = iscol; 1402 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1403 return 0; 1404 } 1405 1406 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1407 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1408 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1409 1410 /* get new row pointers */ 1411 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1412 ainew[0] = 0; 1413 /* don't know how many column pointers are needed so estimate */ 1414 jmax = (int) (f*ai[n] + 1); 1415 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1416 /* ajfill is level of fill for each fill entry */ 1417 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1418 /* fill is a linked list of nonzeros in active row */ 1419 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1420 /* im is level for each filled value */ 1421 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1422 /* dloc is location of diagonal in factor */ 1423 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1424 dloc[0] = 0; 1425 for ( prow=0; prow<n; prow++ ) { 1426 /* first copy previous fill into linked list */ 1427 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1428 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1429 xi = aj + ai[r[prow]]; 1430 fill[n] = n; 1431 while (nz--) { 1432 fm = n; 1433 idx = ic[*xi++]; 1434 do { 1435 m = fm; 1436 fm = fill[m]; 1437 } while (fm < idx); 1438 fill[m] = idx; 1439 fill[idx] = fm; 1440 im[idx] = 0; 1441 } 1442 nzi = 0; 1443 row = fill[n]; 1444 while ( row < prow ) { 1445 incrlev = im[row] + 1; 1446 nz = dloc[row]; 1447 xi = ajnew + ainew[row] + nz; 1448 flev = ajfill + ainew[row] + nz + 1; 1449 nnz = ainew[row+1] - ainew[row] - nz - 1; 1450 if (*xi++ != row) { 1451 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"zero pivot"); 1452 } 1453 fm = row; 1454 while (nnz-- > 0) { 1455 idx = *xi++; 1456 if (*flev + incrlev > levels) { 1457 flev++; 1458 continue; 1459 } 1460 do { 1461 m = fm; 1462 fm = fill[m]; 1463 } while (fm < idx); 1464 if (fm != idx) { 1465 im[idx] = *flev + incrlev; 1466 fill[m] = idx; 1467 fill[idx] = fm; 1468 fm = idx; 1469 nzf++; 1470 } 1471 else { 1472 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1473 } 1474 flev++; 1475 } 1476 row = fill[row]; 1477 nzi++; 1478 } 1479 /* copy new filled row into permanent storage */ 1480 ainew[prow+1] = ainew[prow] + nzf; 1481 if (ainew[prow+1] > jmax) { 1482 /* allocate a longer ajnew */ 1483 int maxadd; 1484 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1485 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1486 jmax += maxadd; 1487 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1488 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1489 PetscFree(ajnew); 1490 ajnew = xi; 1491 /* allocate a longer ajfill */ 1492 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1493 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1494 PetscFree(ajfill); 1495 ajfill = xi; 1496 realloc++; 1497 } 1498 xi = ajnew + ainew[prow]; 1499 flev = ajfill + ainew[prow]; 1500 dloc[prow] = nzi; 1501 fm = fill[n]; 1502 while (nzf--) { 1503 *xi++ = fm; 1504 *flev++ = im[fm]; 1505 fm = fill[fm]; 1506 } 1507 } 1508 PetscFree(ajfill); 1509 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1510 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1511 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1512 PetscFree(fill); PetscFree(im); 1513 1514 PLogInfo(A, 1515 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1516 realloc,f,((double)ainew[n])/((double)ai[prow])); 1517 1518 /* put together the new matrix */ 1519 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1520 b = (Mat_SeqBAIJ *) (*fact)->data; 1521 PetscFree(b->imax); 1522 b->singlemalloc = 0; 1523 len = bs2*ainew[n]*sizeof(Scalar); 1524 /* the next line frees the default space generated by the Create() */ 1525 PetscFree(b->a); PetscFree(b->ilen); 1526 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1527 b->j = ajnew; 1528 b->i = ainew; 1529 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1530 b->diag = dloc; 1531 b->ilen = 0; 1532 b->imax = 0; 1533 b->row = isrow; 1534 b->col = iscol; 1535 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1536 CHKPTRQ(b->solve_work); 1537 /* In b structure: Free imax, ilen, old a, old j. 1538 Allocate dloc, solve_work, new a, new j */ 1539 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1540 b->maxnz = b->nz = ainew[n]; 1541 (*fact)->factor = FACTOR_LU; 1542 1543 (*fact)->info.factor_mallocs = realloc; 1544 (*fact)->info.fill_ratio_given = f; 1545 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1546 1547 return 0; 1548 } 1549 1550 1551 1552 1553