1 2 #ifndef lint 3 static char vcid[] = "$Id: baijfact.c,v 1.41 1997/03/26 01:36:13 bsmith 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 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 154 155 return 0; 156 } 157 158 /* ----------------------------------------------------------- */ 159 #undef __FUNC__ 160 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 161 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 162 { 163 Mat C = *B; 164 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 165 IS iscol = b->col, isrow = b->row, isicol; 166 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 167 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 168 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 169 Scalar *ba = b->a,*aa = a->a; 170 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 171 register int *pj; 172 173 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 174 PLogObjectParent(*B,isicol); 175 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 176 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 177 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 178 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 179 /* generate work space needed by dense LU factorization */ 180 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 181 CHKPTRQ(v_work); 182 multiplier = v_work + bs; 183 v_pivots = (int *) (multiplier + bs2); 184 185 /* flops in while loop */ 186 bslog = 2*bs*bs2; 187 188 for ( i=0; i<n; i++ ) { 189 nz = bi[i+1] - bi[i]; 190 ajtmp = bj + bi[i]; 191 for ( j=0; j<nz; j++ ) { 192 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 193 } 194 /* load in initial (unfactored row) */ 195 nz = ai[r[i]+1] - ai[r[i]]; 196 ajtmpold = aj + ai[r[i]]; 197 v = aa + bs2*ai[r[i]]; 198 for ( j=0; j<nz; j++ ) { 199 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 200 } 201 row = *ajtmp++; 202 while (row < i) { 203 pc = rtmp + bs2*row; 204 /* if (*pc) { */ 205 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 206 if (flg) { 207 pv = ba + bs2*diag_offset[row]; 208 pj = bj + diag_offset[row] + 1; 209 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 210 nz = bi[row+1] - diag_offset[row] - 1; 211 pv += bs2; 212 for (j=0; j<nz; j++) { 213 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 214 } 215 PLogFlops(bslog*(nz+1)-bs); 216 } 217 row = *ajtmp++; 218 } 219 /* finished row so stick it into b->a */ 220 pv = ba + bs2*bi[i]; 221 pj = bj + bi[i]; 222 nz = bi[i+1] - bi[i]; 223 for ( j=0; j<nz; j++ ) { 224 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 225 } 226 diag = diag_offset[i] - bi[i]; 227 /* invert diagonal block */ 228 w = pv + bs2*diag; 229 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 230 } 231 232 PetscFree(rtmp); PetscFree(v_work); 233 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 234 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 235 ierr = ISDestroy(isicol); CHKERRQ(ierr); 236 C->factor = FACTOR_LU; 237 C->assembled = PETSC_TRUE; 238 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 239 return 0; 240 } 241 /* ------------------------------------------------------------*/ 242 /* 243 Version for when blocks are 5 by 5 244 */ 245 #undef __FUNC__ 246 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 247 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 248 { 249 Mat C = *B; 250 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 251 IS iscol = b->col, isrow = b->row, isicol; 252 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 253 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 254 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 255 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 256 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 257 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 258 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 259 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 260 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 261 Scalar *ba = b->a,*aa = a->a; 262 register int *pj; 263 264 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 265 PLogObjectParent(*B,isicol); 266 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 267 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 268 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 269 270 for ( i=0; i<n; i++ ) { 271 nz = bi[i+1] - bi[i]; 272 ajtmp = bj + bi[i]; 273 for ( j=0; j<nz; j++ ) { 274 x = rtmp+25*ajtmp[j]; 275 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 276 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 277 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 278 } 279 /* load in initial (unfactored row) */ 280 idx = r[i]; 281 nz = ai[idx+1] - ai[idx]; 282 ajtmpold = aj + ai[idx]; 283 v = aa + 25*ai[idx]; 284 for ( j=0; j<nz; j++ ) { 285 x = rtmp+25*ic[ajtmpold[j]]; 286 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 287 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 288 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 289 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 290 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 291 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 292 v += 25; 293 } 294 row = *ajtmp++; 295 while (row < i) { 296 pc = rtmp + 25*row; 297 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 298 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 299 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 300 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 301 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 302 p25 = pc[24]; 303 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 304 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 305 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 306 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 307 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 308 p24 != 0.0 || p25 != 0.0) { 309 pv = ba + 25*diag_offset[row]; 310 pj = bj + diag_offset[row] + 1; 311 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 312 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 313 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 314 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 315 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 316 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 317 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 318 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 319 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 320 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 321 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 322 323 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 324 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 325 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 326 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 327 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 328 329 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 330 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 331 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 332 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 333 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 334 335 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 336 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 337 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 338 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 339 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 340 341 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 342 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 343 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 344 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 345 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 346 347 nz = bi[row+1] - diag_offset[row] - 1; 348 pv += 25; 349 for (j=0; j<nz; j++) { 350 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 351 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 352 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 353 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 354 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 355 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 356 x = rtmp + 25*pj[j]; 357 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 358 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 359 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 360 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 361 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 362 363 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 364 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 365 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 366 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 367 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 368 369 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 370 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 371 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 372 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 373 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 374 375 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 376 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 377 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 378 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 379 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 380 381 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 382 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 383 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 384 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 385 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 386 387 pv += 25; 388 } 389 PLogFlops(250*nz+225); 390 } 391 row = *ajtmp++; 392 } 393 /* finished row so stick it into b->a */ 394 pv = ba + 25*bi[i]; 395 pj = bj + bi[i]; 396 nz = bi[i+1] - bi[i]; 397 for ( j=0; j<nz; j++ ) { 398 x = rtmp+25*pj[j]; 399 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 400 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 401 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 402 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 403 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 404 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 405 pv += 25; 406 } 407 /* invert diagonal block */ 408 w = ba + 25*diag_offset[i]; 409 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 410 } 411 412 PetscFree(rtmp); 413 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 414 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 415 ierr = ISDestroy(isicol); CHKERRQ(ierr); 416 C->factor = FACTOR_LU; 417 C->assembled = PETSC_TRUE; 418 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 419 return 0; 420 } 421 422 /* ------------------------------------------------------------*/ 423 /* 424 Version for when blocks are 4 by 4 425 */ 426 #undef __FUNC__ 427 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 428 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 429 { 430 Mat C = *B; 431 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 432 IS iscol = b->col, isrow = b->row, isicol; 433 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 434 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 435 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 436 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 437 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 438 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 439 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 440 Scalar m13,m14,m15,m16; 441 Scalar *ba = b->a,*aa = a->a; 442 register int *pj; 443 444 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 445 PLogObjectParent(*B,isicol); 446 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 447 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 448 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 449 450 for ( i=0; i<n; i++ ) { 451 nz = bi[i+1] - bi[i]; 452 ajtmp = bj + bi[i]; 453 for ( j=0; j<nz; j++ ) { 454 x = rtmp+16*ajtmp[j]; 455 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 456 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 457 } 458 /* load in initial (unfactored row) */ 459 idx = r[i]; 460 nz = ai[idx+1] - ai[idx]; 461 ajtmpold = aj + ai[idx]; 462 v = aa + 16*ai[idx]; 463 for ( j=0; j<nz; j++ ) { 464 x = rtmp+16*ic[ajtmpold[j]]; 465 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 466 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 467 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 468 x[14] = v[14]; x[15] = v[15]; 469 v += 16; 470 } 471 row = *ajtmp++; 472 while (row < i) { 473 pc = rtmp + 16*row; 474 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 475 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 476 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 477 p15 = pc[14]; p16 = pc[15]; 478 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 479 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 480 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 481 || p16 != 0.0) { 482 pv = ba + 16*diag_offset[row]; 483 pj = bj + diag_offset[row] + 1; 484 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 485 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 486 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 487 x15 = pv[14]; x16 = pv[15]; 488 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 489 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 490 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 491 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 492 493 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 494 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 495 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 496 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 497 498 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 499 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 500 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 501 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 502 503 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 504 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 505 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 506 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 507 508 nz = bi[row+1] - diag_offset[row] - 1; 509 pv += 16; 510 for (j=0; j<nz; j++) { 511 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 512 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 513 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 514 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 515 x = rtmp + 16*pj[j]; 516 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 517 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 518 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 519 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 520 521 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 522 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 523 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 524 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 525 526 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 527 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 528 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 529 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 530 531 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 532 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 533 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 534 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 535 536 pv += 16; 537 } 538 PLogFlops(128*nz+112); 539 } 540 row = *ajtmp++; 541 } 542 /* finished row so stick it into b->a */ 543 pv = ba + 16*bi[i]; 544 pj = bj + bi[i]; 545 nz = bi[i+1] - bi[i]; 546 for ( j=0; j<nz; j++ ) { 547 x = rtmp+16*pj[j]; 548 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 549 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 550 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 551 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 552 pv += 16; 553 } 554 /* invert diagonal block */ 555 w = ba + 16*diag_offset[i]; 556 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 557 } 558 559 PetscFree(rtmp); 560 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 561 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 562 ierr = ISDestroy(isicol); CHKERRQ(ierr); 563 C->factor = FACTOR_LU; 564 C->assembled = PETSC_TRUE; 565 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 566 return 0; 567 } 568 /* ------------------------------------------------------------*/ 569 /* 570 Version for when blocks are 3 by 3 571 */ 572 #undef __FUNC__ 573 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 574 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 575 { 576 Mat C = *B; 577 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 578 IS iscol = b->col, isrow = b->row, isicol; 579 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 580 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 581 int *diag_offset = b->diag,idx; 582 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 583 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 584 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 585 Scalar *ba = b->a,*aa = a->a; 586 register int *pj; 587 588 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 589 PLogObjectParent(*B,isicol); 590 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 591 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 592 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 593 594 for ( i=0; i<n; i++ ) { 595 nz = bi[i+1] - bi[i]; 596 ajtmp = bj + bi[i]; 597 for ( j=0; j<nz; j++ ) { 598 x = rtmp + 9*ajtmp[j]; 599 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 600 } 601 /* load in initial (unfactored row) */ 602 idx = r[i]; 603 nz = ai[idx+1] - ai[idx]; 604 ajtmpold = aj + ai[idx]; 605 v = aa + 9*ai[idx]; 606 for ( j=0; j<nz; j++ ) { 607 x = rtmp + 9*ic[ajtmpold[j]]; 608 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 609 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 610 v += 9; 611 } 612 row = *ajtmp++; 613 while (row < i) { 614 pc = rtmp + 9*row; 615 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 616 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 617 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 618 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 619 pv = ba + 9*diag_offset[row]; 620 pj = bj + diag_offset[row] + 1; 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 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 624 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 625 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 626 627 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 628 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 629 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 630 631 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 632 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 633 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 634 nz = bi[row+1] - diag_offset[row] - 1; 635 pv += 9; 636 for (j=0; j<nz; j++) { 637 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 638 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 639 x = rtmp + 9*pj[j]; 640 x[0] -= m1*x1 + m4*x2 + m7*x3; 641 x[1] -= m2*x1 + m5*x2 + m8*x3; 642 x[2] -= m3*x1 + m6*x2 + m9*x3; 643 644 x[3] -= m1*x4 + m4*x5 + m7*x6; 645 x[4] -= m2*x4 + m5*x5 + m8*x6; 646 x[5] -= m3*x4 + m6*x5 + m9*x6; 647 648 x[6] -= m1*x7 + m4*x8 + m7*x9; 649 x[7] -= m2*x7 + m5*x8 + m8*x9; 650 x[8] -= m3*x7 + m6*x8 + m9*x9; 651 pv += 9; 652 } 653 PLogFlops(54*nz+36); 654 } 655 row = *ajtmp++; 656 } 657 /* finished row so stick it into b->a */ 658 pv = ba + 9*bi[i]; 659 pj = bj + bi[i]; 660 nz = bi[i+1] - bi[i]; 661 for ( j=0; j<nz; j++ ) { 662 x = rtmp + 9*pj[j]; 663 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 664 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 665 pv += 9; 666 } 667 /* invert diagonal block */ 668 w = ba + 9*diag_offset[i]; 669 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 670 } 671 672 PetscFree(rtmp); 673 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 674 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 675 ierr = ISDestroy(isicol); CHKERRQ(ierr); 676 C->factor = FACTOR_LU; 677 C->assembled = PETSC_TRUE; 678 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 679 return 0; 680 } 681 682 /* ------------------------------------------------------------*/ 683 /* 684 Version for when blocks are 2 by 2 685 */ 686 #undef __FUNC__ 687 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 688 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 689 { 690 Mat C = *B; 691 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 692 IS iscol = b->col, isrow = b->row, isicol; 693 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 694 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 695 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 696 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 697 Scalar p1,p2,p3,p4,v_work[2]; 698 Scalar *ba = b->a,*aa = a->a; 699 register int *pj; 700 701 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 702 PLogObjectParent(*B,isicol); 703 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 704 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 705 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 706 707 for ( i=0; i<n; i++ ) { 708 nz = bi[i+1] - bi[i]; 709 ajtmp = bj + bi[i]; 710 for ( j=0; j<nz; j++ ) { 711 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 712 } 713 /* load in initial (unfactored row) */ 714 idx = r[i]; 715 nz = ai[idx+1] - ai[idx]; 716 ajtmpold = aj + ai[idx]; 717 v = aa + 4*ai[idx]; 718 for ( j=0; j<nz; j++ ) { 719 x = rtmp+4*ic[ajtmpold[j]]; 720 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 721 v += 4; 722 } 723 row = *ajtmp++; 724 while (row < i) { 725 pc = rtmp + 4*row; 726 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 727 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 728 pv = ba + 4*diag_offset[row]; 729 pj = bj + diag_offset[row] + 1; 730 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 731 pc[0] = m1 = p1*x1 + p3*x2; 732 pc[1] = m2 = p2*x1 + p4*x2; 733 pc[2] = m3 = p1*x3 + p3*x4; 734 pc[3] = m4 = p2*x3 + p4*x4; 735 nz = bi[row+1] - diag_offset[row] - 1; 736 pv += 4; 737 for (j=0; j<nz; j++) { 738 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 739 x = rtmp + 4*pj[j]; 740 x[0] -= m1*x1 + m3*x2; 741 x[1] -= m2*x1 + m4*x2; 742 x[2] -= m1*x3 + m3*x4; 743 x[3] -= m2*x3 + m4*x4; 744 pv += 4; 745 } 746 PLogFlops(16*nz+12); 747 } 748 row = *ajtmp++; 749 } 750 /* finished row so stick it into b->a */ 751 pv = ba + 4*bi[i]; 752 pj = bj + bi[i]; 753 nz = bi[i+1] - bi[i]; 754 for ( j=0; j<nz; j++ ) { 755 x = rtmp+4*pj[j]; 756 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 757 pv += 4; 758 } 759 /* invert diagonal block */ 760 w = ba + 4*diag_offset[i]; 761 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 762 } 763 764 PetscFree(rtmp); 765 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 766 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 767 ierr = ISDestroy(isicol); CHKERRQ(ierr); 768 C->factor = FACTOR_LU; 769 C->assembled = PETSC_TRUE; 770 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 771 return 0; 772 } 773 774 /* ----------------------------------------------------------- */ 775 /* 776 Version for when blocks are 1 by 1. 777 */ 778 #undef __FUNC__ 779 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 780 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 781 { 782 Mat C = *B; 783 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 784 IS iscol = b->col, isrow = b->row, isicol; 785 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 786 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 787 int *diag_offset = b->diag,diag; 788 register Scalar *pv,*v,*rtmp,multiplier,*pc; 789 Scalar *ba = b->a,*aa = a->a; 790 register int *pj; 791 792 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 793 PLogObjectParent(*B,isicol); 794 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 795 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 796 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 797 798 for ( i=0; i<n; i++ ) { 799 nz = bi[i+1] - bi[i]; 800 ajtmp = bj + bi[i]; 801 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 802 803 /* load in initial (unfactored row) */ 804 nz = ai[r[i]+1] - ai[r[i]]; 805 ajtmpold = aj + ai[r[i]]; 806 v = aa + ai[r[i]]; 807 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 808 809 row = *ajtmp++; 810 while (row < i) { 811 pc = rtmp + row; 812 if (*pc != 0.0) { 813 pv = ba + diag_offset[row]; 814 pj = bj + diag_offset[row] + 1; 815 multiplier = *pc * *pv++; 816 *pc = multiplier; 817 nz = bi[row+1] - diag_offset[row] - 1; 818 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 819 PLogFlops(1+2*nz); 820 } 821 row = *ajtmp++; 822 } 823 /* finished row so stick it into b->a */ 824 pv = ba + bi[i]; 825 pj = bj + bi[i]; 826 nz = bi[i+1] - bi[i]; 827 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 828 diag = diag_offset[i] - bi[i]; 829 /* check pivot entry for current row */ 830 if (pv[diag] == 0.0) { 831 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 832 } 833 pv[diag] = 1.0/pv[diag]; 834 } 835 836 PetscFree(rtmp); 837 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 838 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 839 ierr = ISDestroy(isicol); CHKERRQ(ierr); 840 C->factor = FACTOR_LU; 841 C->assembled = PETSC_TRUE; 842 PLogFlops(b->n); 843 return 0; 844 } 845 846 /* ----------------------------------------------------------- */ 847 #undef __FUNC__ 848 #define __FUNC__ "MatLUFactor_SeqBAIJ" 849 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 850 { 851 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 852 int ierr; 853 Mat C; 854 855 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 856 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 857 858 /* free all the data structures from mat */ 859 PetscFree(mat->a); 860 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 861 if (mat->diag) PetscFree(mat->diag); 862 if (mat->ilen) PetscFree(mat->ilen); 863 if (mat->imax) PetscFree(mat->imax); 864 if (mat->solve_work) PetscFree(mat->solve_work); 865 if (mat->mult_work) PetscFree(mat->mult_work); 866 PetscFree(mat); 867 868 PetscMemcpy(A,C,sizeof(struct _Mat)); 869 PetscHeaderDestroy(C); 870 return 0; 871 } 872 /* ----------------------------------------------------------- */ 873 #undef __FUNC__ 874 #define __FUNC__ "MatSolve_SeqBAIJ_N" 875 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 876 { 877 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 878 IS iscol=a->col,isrow=a->row; 879 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 880 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 881 Scalar *aa=a->a,*sum; 882 register Scalar *x,*b,*lsum,*tmp,*v; 883 884 VecGetArray_Fast(bb,b); 885 VecGetArray_Fast(xx,x); 886 tmp = a->solve_work; 887 888 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 889 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 890 891 /* forward solve the lower triangular */ 892 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 893 for ( i=1; i<n; i++ ) { 894 v = aa + bs2*ai[i]; 895 vi = aj + ai[i]; 896 nz = a->diag[i] - ai[i]; 897 sum = tmp + bs*i; 898 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 899 while (nz--) { 900 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 901 v += bs2; 902 } 903 } 904 /* backward solve the upper triangular */ 905 lsum = a->solve_work + a->n; 906 for ( i=n-1; i>=0; i-- ){ 907 v = aa + bs2*(a->diag[i] + 1); 908 vi = aj + a->diag[i] + 1; 909 nz = ai[i+1] - a->diag[i] - 1; 910 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 911 while (nz--) { 912 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 913 v += bs2; 914 } 915 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 916 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 917 } 918 919 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 920 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 921 VecRestoreArray_Fast(bb,b); 922 VecRestoreArray_Fast(xx,x); 923 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 924 return 0; 925 } 926 927 #undef __FUNC__ 928 #define __FUNC__ "MatSolve_SeqBAIJ_7" 929 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 930 { 931 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 932 IS iscol=a->col,isrow=a->row; 933 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 934 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 935 register Scalar *x,*b,*tmp,*v; 936 937 VecGetArray_Fast(bb,b); 938 VecGetArray_Fast(xx,x); 939 tmp = a->solve_work; 940 941 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 942 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 943 944 /* forward solve the lower triangular */ 945 idx = 7*(*r++); 946 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 947 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 948 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 949 950 for ( i=1; i<n; i++ ) { 951 v = aa + 49*ai[i]; 952 vi = aj + ai[i]; 953 nz = a->diag[i] - ai[i]; 954 idx = 7*(*r++); 955 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 956 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 957 while (nz--) { 958 idx = 7*(*vi++); 959 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 960 x4 = tmp[3+idx];x5 = tmp[4+idx]; 961 x6 = tmp[5+idx];x7 = tmp[6+idx]; 962 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 963 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 964 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 965 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 966 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 967 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 968 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 969 v += 49; 970 } 971 idx = 7*i; 972 tmp[idx] = sum1;tmp[1+idx] = sum2; 973 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 974 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 975 } 976 /* backward solve the upper triangular */ 977 for ( i=n-1; i>=0; i-- ){ 978 v = aa + 49*a->diag[i] + 49; 979 vi = aj + a->diag[i] + 1; 980 nz = ai[i+1] - a->diag[i] - 1; 981 idt = 7*i; 982 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 983 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 984 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 985 while (nz--) { 986 idx = 7*(*vi++); 987 x1 = tmp[idx]; x2 = tmp[1+idx]; 988 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 989 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 990 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 991 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 992 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 993 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 994 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 995 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 996 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 997 v += 49; 998 } 999 idc = 7*(*c--); 1000 v = aa + 49*a->diag[i]; 1001 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 1002 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 1003 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 1004 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 1005 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 1006 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1007 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1008 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1009 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1010 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1011 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1012 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1013 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1014 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1015 } 1016 1017 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1018 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1019 VecRestoreArray_Fast(bb,b); 1020 VecRestoreArray_Fast(xx,x); 1021 PLogFlops(2*49*(a->nz) - a->n); 1022 return 0; 1023 } 1024 1025 #undef __FUNC__ 1026 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1027 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1028 { 1029 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1030 IS iscol=a->col,isrow=a->row; 1031 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1032 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1033 register Scalar *x,*b,*tmp,*v; 1034 1035 VecGetArray_Fast(bb,b); 1036 VecGetArray_Fast(xx,x); 1037 tmp = a->solve_work; 1038 1039 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1040 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1041 1042 /* forward solve the lower triangular */ 1043 idx = 5*(*r++); 1044 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1045 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1046 for ( i=1; i<n; i++ ) { 1047 v = aa + 25*ai[i]; 1048 vi = aj + ai[i]; 1049 nz = a->diag[i] - ai[i]; 1050 idx = 5*(*r++); 1051 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1052 sum5 = b[4+idx]; 1053 while (nz--) { 1054 idx = 5*(*vi++); 1055 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1056 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1057 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1058 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1059 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1060 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1061 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1062 v += 25; 1063 } 1064 idx = 5*i; 1065 tmp[idx] = sum1;tmp[1+idx] = sum2; 1066 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1067 } 1068 /* backward solve the upper triangular */ 1069 for ( i=n-1; i>=0; i-- ){ 1070 v = aa + 25*a->diag[i] + 25; 1071 vi = aj + a->diag[i] + 1; 1072 nz = ai[i+1] - a->diag[i] - 1; 1073 idt = 5*i; 1074 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1075 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1076 while (nz--) { 1077 idx = 5*(*vi++); 1078 x1 = tmp[idx]; x2 = tmp[1+idx]; 1079 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1080 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1081 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1082 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1083 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1084 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1085 v += 25; 1086 } 1087 idc = 5*(*c--); 1088 v = aa + 25*a->diag[i]; 1089 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1090 v[15]*sum4+v[20]*sum5; 1091 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1092 v[16]*sum4+v[21]*sum5; 1093 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1094 v[17]*sum4+v[22]*sum5; 1095 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1096 v[18]*sum4+v[23]*sum5; 1097 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1098 v[19]*sum4+v[24]*sum5; 1099 } 1100 1101 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1102 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1103 VecRestoreArray_Fast(bb,b); 1104 VecRestoreArray_Fast(xx,x); 1105 PLogFlops(2*25*(a->nz) - a->n); 1106 return 0; 1107 } 1108 1109 #undef __FUNC__ 1110 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1111 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1112 { 1113 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1114 IS iscol=a->col,isrow=a->row; 1115 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1116 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1117 register Scalar *x,*b,*tmp,*v; 1118 1119 VecGetArray_Fast(bb,b); 1120 VecGetArray_Fast(xx,x); 1121 tmp = a->solve_work; 1122 1123 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1124 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1125 1126 /* forward solve the lower triangular */ 1127 idx = 4*(*r++); 1128 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1129 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1130 for ( i=1; i<n; i++ ) { 1131 v = aa + 16*ai[i]; 1132 vi = aj + ai[i]; 1133 nz = a->diag[i] - ai[i]; 1134 idx = 4*(*r++); 1135 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1136 while (nz--) { 1137 idx = 4*(*vi++); 1138 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1139 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1140 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1141 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1142 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1143 v += 16; 1144 } 1145 idx = 4*i; 1146 tmp[idx] = sum1;tmp[1+idx] = sum2; 1147 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1148 } 1149 /* backward solve the upper triangular */ 1150 for ( i=n-1; i>=0; i-- ){ 1151 v = aa + 16*a->diag[i] + 16; 1152 vi = aj + a->diag[i] + 1; 1153 nz = ai[i+1] - a->diag[i] - 1; 1154 idt = 4*i; 1155 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1156 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1157 while (nz--) { 1158 idx = 4*(*vi++); 1159 x1 = tmp[idx]; x2 = tmp[1+idx]; 1160 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1161 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1162 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1163 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1164 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1165 v += 16; 1166 } 1167 idc = 4*(*c--); 1168 v = aa + 16*a->diag[i]; 1169 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1170 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1171 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1172 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1173 } 1174 1175 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1176 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1177 VecRestoreArray_Fast(bb,b); 1178 VecRestoreArray_Fast(xx,x); 1179 PLogFlops(2*16*(a->nz) - a->n); 1180 return 0; 1181 } 1182 1183 1184 #undef __FUNC__ 1185 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1186 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1187 { 1188 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1189 IS iscol=a->col,isrow=a->row; 1190 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1191 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1192 register Scalar *x,*b,*tmp,*v; 1193 1194 VecGetArray_Fast(bb,b); 1195 VecGetArray_Fast(xx,x); 1196 tmp = a->solve_work; 1197 1198 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1199 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1200 1201 /* forward solve the lower triangular */ 1202 idx = 3*(*r++); 1203 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1204 for ( i=1; i<n; i++ ) { 1205 v = aa + 9*ai[i]; 1206 vi = aj + ai[i]; 1207 nz = a->diag[i] - ai[i]; 1208 idx = 3*(*r++); 1209 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1210 while (nz--) { 1211 idx = 3*(*vi++); 1212 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1213 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1214 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1215 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1216 v += 9; 1217 } 1218 idx = 3*i; 1219 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1220 } 1221 /* backward solve the upper triangular */ 1222 for ( i=n-1; i>=0; i-- ){ 1223 v = aa + 9*a->diag[i] + 9; 1224 vi = aj + a->diag[i] + 1; 1225 nz = ai[i+1] - a->diag[i] - 1; 1226 idt = 3*i; 1227 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1228 while (nz--) { 1229 idx = 3*(*vi++); 1230 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1231 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1232 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1233 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1234 v += 9; 1235 } 1236 idc = 3*(*c--); 1237 v = aa + 9*a->diag[i]; 1238 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1239 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1240 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1241 } 1242 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1243 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1244 VecRestoreArray_Fast(bb,b); 1245 VecRestoreArray_Fast(xx,x); 1246 PLogFlops(2*9*(a->nz) - a->n); 1247 return 0; 1248 } 1249 1250 #undef __FUNC__ 1251 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1252 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1253 { 1254 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1255 IS iscol=a->col,isrow=a->row; 1256 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1257 Scalar *aa=a->a,sum1,sum2,x1,x2; 1258 register Scalar *x,*b,*tmp,*v; 1259 1260 VecGetArray_Fast(bb,b); 1261 VecGetArray_Fast(xx,x); 1262 tmp = a->solve_work; 1263 1264 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1265 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1266 1267 /* forward solve the lower triangular */ 1268 idx = 2*(*r++); 1269 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1270 for ( i=1; i<n; i++ ) { 1271 v = aa + 4*ai[i]; 1272 vi = aj + ai[i]; 1273 nz = a->diag[i] - ai[i]; 1274 idx = 2*(*r++); 1275 sum1 = b[idx]; sum2 = b[1+idx]; 1276 while (nz--) { 1277 idx = 2*(*vi++); 1278 x1 = tmp[idx]; x2 = tmp[1+idx]; 1279 sum1 -= v[0]*x1 + v[2]*x2; 1280 sum2 -= v[1]*x1 + v[3]*x2; 1281 v += 4; 1282 } 1283 idx = 2*i; 1284 tmp[idx] = sum1; tmp[1+idx] = sum2; 1285 } 1286 /* backward solve the upper triangular */ 1287 for ( i=n-1; i>=0; i-- ){ 1288 v = aa + 4*a->diag[i] + 4; 1289 vi = aj + a->diag[i] + 1; 1290 nz = ai[i+1] - a->diag[i] - 1; 1291 idt = 2*i; 1292 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1293 while (nz--) { 1294 idx = 2*(*vi++); 1295 x1 = tmp[idx]; x2 = tmp[1+idx]; 1296 sum1 -= v[0]*x1 + v[2]*x2; 1297 sum2 -= v[1]*x1 + v[3]*x2; 1298 v += 4; 1299 } 1300 idc = 2*(*c--); 1301 v = aa + 4*a->diag[i]; 1302 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1303 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1304 } 1305 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1306 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1307 VecRestoreArray_Fast(bb,b); 1308 VecRestoreArray_Fast(xx,x); 1309 PLogFlops(2*4*(a->nz) - a->n); 1310 return 0; 1311 } 1312 1313 1314 #undef __FUNC__ 1315 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1316 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1317 { 1318 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1319 IS iscol=a->col,isrow=a->row; 1320 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1321 Scalar *aa=a->a,sum1; 1322 register Scalar *x,*b,*tmp,*v; 1323 1324 if (!n) return 0; 1325 1326 VecGetArray_Fast(bb,b); 1327 VecGetArray_Fast(xx,x); 1328 tmp = a->solve_work; 1329 1330 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1331 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1332 1333 /* forward solve the lower triangular */ 1334 tmp[0] = b[*r++]; 1335 for ( i=1; i<n; i++ ) { 1336 v = aa + ai[i]; 1337 vi = aj + ai[i]; 1338 nz = a->diag[i] - ai[i]; 1339 sum1 = b[*r++]; 1340 while (nz--) { 1341 sum1 -= (*v++)*tmp[*vi++]; 1342 } 1343 tmp[i] = sum1; 1344 } 1345 /* backward solve the upper triangular */ 1346 for ( i=n-1; i>=0; i-- ){ 1347 v = aa + a->diag[i] + 1; 1348 vi = aj + a->diag[i] + 1; 1349 nz = ai[i+1] - a->diag[i] - 1; 1350 sum1 = tmp[i]; 1351 while (nz--) { 1352 sum1 -= (*v++)*tmp[*vi++]; 1353 } 1354 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1355 } 1356 1357 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1358 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1359 VecRestoreArray_Fast(bb,b); 1360 VecRestoreArray_Fast(xx,x); 1361 PLogFlops(2*1*(a->nz) - a->n); 1362 return 0; 1363 } 1364 1365 /* ----------------------------------------------------------------*/ 1366 /* 1367 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1368 except that the data structure of Mat_SeqAIJ is slightly different. 1369 Not a good example of code reuse. 1370 */ 1371 #undef __FUNC__ 1372 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1373 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1374 Mat *fact) 1375 { 1376 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1377 IS isicol; 1378 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1379 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1380 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1381 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1382 PetscTruth col_identity, row_identity; 1383 1384 /* special case that simply copies fill pattern */ 1385 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1386 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1387 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1388 if (levels == 0 && row_identity && col_identity) { 1389 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1390 (*fact)->factor = FACTOR_LU; 1391 b = (Mat_SeqBAIJ *) (*fact)->data; 1392 if (!b->diag) { 1393 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1394 } 1395 b->row = isrow; 1396 b->col = iscol; 1397 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1398 return 0; 1399 } 1400 1401 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1402 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1403 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1404 1405 /* get new row pointers */ 1406 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1407 ainew[0] = 0; 1408 /* don't know how many column pointers are needed so estimate */ 1409 jmax = (int) (f*ai[n] + 1); 1410 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1411 /* ajfill is level of fill for each fill entry */ 1412 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1413 /* fill is a linked list of nonzeros in active row */ 1414 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1415 /* im is level for each filled value */ 1416 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1417 /* dloc is location of diagonal in factor */ 1418 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1419 dloc[0] = 0; 1420 for ( prow=0; prow<n; prow++ ) { 1421 /* first copy previous fill into linked list */ 1422 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1423 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1424 xi = aj + ai[r[prow]]; 1425 fill[n] = n; 1426 while (nz--) { 1427 fm = n; 1428 idx = ic[*xi++]; 1429 do { 1430 m = fm; 1431 fm = fill[m]; 1432 } while (fm < idx); 1433 fill[m] = idx; 1434 fill[idx] = fm; 1435 im[idx] = 0; 1436 } 1437 nzi = 0; 1438 row = fill[n]; 1439 while ( row < prow ) { 1440 incrlev = im[row] + 1; 1441 nz = dloc[row]; 1442 xi = ajnew + ainew[row] + nz; 1443 flev = ajfill + ainew[row] + nz + 1; 1444 nnz = ainew[row+1] - ainew[row] - nz - 1; 1445 if (*xi++ != row) { 1446 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"zero pivot"); 1447 } 1448 fm = row; 1449 while (nnz-- > 0) { 1450 idx = *xi++; 1451 if (*flev + incrlev > levels) { 1452 flev++; 1453 continue; 1454 } 1455 do { 1456 m = fm; 1457 fm = fill[m]; 1458 } while (fm < idx); 1459 if (fm != idx) { 1460 im[idx] = *flev + incrlev; 1461 fill[m] = idx; 1462 fill[idx] = fm; 1463 fm = idx; 1464 nzf++; 1465 } 1466 else { 1467 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1468 } 1469 flev++; 1470 } 1471 row = fill[row]; 1472 nzi++; 1473 } 1474 /* copy new filled row into permanent storage */ 1475 ainew[prow+1] = ainew[prow] + nzf; 1476 if (ainew[prow+1] > jmax) { 1477 /* allocate a longer ajnew */ 1478 int maxadd; 1479 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1480 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1481 jmax += maxadd; 1482 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1483 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1484 PetscFree(ajnew); 1485 ajnew = xi; 1486 /* allocate a longer ajfill */ 1487 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1488 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1489 PetscFree(ajfill); 1490 ajfill = xi; 1491 realloc++; 1492 } 1493 xi = ajnew + ainew[prow]; 1494 flev = ajfill + ainew[prow]; 1495 dloc[prow] = nzi; 1496 fm = fill[n]; 1497 while (nzf--) { 1498 *xi++ = fm; 1499 *flev++ = im[fm]; 1500 fm = fill[fm]; 1501 } 1502 } 1503 PetscFree(ajfill); 1504 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1505 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1506 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1507 PetscFree(fill); PetscFree(im); 1508 1509 PLogInfo(A, 1510 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1511 realloc,f,((double)ainew[n])/((double)ai[prow])); 1512 1513 /* put together the new matrix */ 1514 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1515 b = (Mat_SeqBAIJ *) (*fact)->data; 1516 PetscFree(b->imax); 1517 b->singlemalloc = 0; 1518 len = bs2*ainew[n]*sizeof(Scalar); 1519 /* the next line frees the default space generated by the Create() */ 1520 PetscFree(b->a); PetscFree(b->ilen); 1521 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1522 b->j = ajnew; 1523 b->i = ainew; 1524 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1525 b->diag = dloc; 1526 b->ilen = 0; 1527 b->imax = 0; 1528 b->row = isrow; 1529 b->col = iscol; 1530 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1531 CHKPTRQ(b->solve_work); 1532 /* In b structure: Free imax, ilen, old a, old j. 1533 Allocate dloc, solve_work, new a, new j */ 1534 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1535 b->maxnz = b->nz = ainew[n]; 1536 (*fact)->factor = FACTOR_LU; 1537 1538 (*fact)->info.factor_mallocs = realloc; 1539 (*fact)->info.fill_ratio_given = f; 1540 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1541 1542 return 0; 1543 } 1544 1545 1546 1547 1548