1 #ifdef PETSC_RCS_HEADER 2 static char vcid[] = "$Id: baijfact.c,v 1.58 1998/03/26 22:11:31 balay Exp balay $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "src/mat/impls/baij/seq/baij.h" 9 #include "src/vec/vecimpl.h" 10 #include "src/inline/ilu.h" 11 12 13 /* 14 The symbolic factorization code is identical to that for AIJ format, 15 except for very small changes since this is now a SeqBAIJ datastructure. 16 NOT good code reuse. 17 */ 18 #undef __FUNC__ 19 #define __FUNC__ "MatLUFactorSymbolic_SeqBAIJ" 20 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 21 { 22 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 23 IS isicol; 24 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 25 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 26 int *idnew, idx, row,m,fm, nnz, nzi,realloc = 0,nzbd,*im; 27 28 PetscFunctionBegin; 29 PetscValidHeaderSpecific(isrow,IS_COOKIE); 30 PetscValidHeaderSpecific(iscol,IS_COOKIE); 31 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 32 PLogObjectParent(*B,isicol); 33 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 34 35 /* get new row pointers */ 36 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 37 ainew[0] = 0; 38 /* don't know how many column pointers are needed so estimate */ 39 jmax = (int) (f*ai[n] + 1); 40 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 41 /* fill is a linked list of nonzeros in active row */ 42 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 43 im = fill + n + 1; 44 /* idnew is location of diagonal in factor */ 45 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 46 idnew[0] = 0; 47 48 for ( i=0; i<n; i++ ) { 49 /* first copy previous fill into linked list */ 50 nnz = nz = ai[r[i]+1] - ai[r[i]]; 51 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 52 ajtmp = aj + ai[r[i]]; 53 fill[n] = n; 54 while (nz--) { 55 fm = n; 56 idx = ic[*ajtmp++]; 57 do { 58 m = fm; 59 fm = fill[m]; 60 } while (fm < idx); 61 fill[m] = idx; 62 fill[idx] = fm; 63 } 64 row = fill[n]; 65 while ( row < i ) { 66 ajtmp = ajnew + idnew[row] + 1; 67 nzbd = 1 + idnew[row] - ainew[row]; 68 nz = im[row] - nzbd; 69 fm = row; 70 while (nz-- > 0) { 71 idx = *ajtmp++; 72 nzbd++; 73 if (idx == i) im[row] = nzbd; 74 do { 75 m = fm; 76 fm = fill[m]; 77 } while (fm < idx); 78 if (fm != idx) { 79 fill[m] = idx; 80 fill[idx] = fm; 81 fm = idx; 82 nnz++; 83 } 84 } 85 row = fill[row]; 86 } 87 /* copy new filled row into permanent storage */ 88 ainew[i+1] = ainew[i] + nnz; 89 if (ainew[i+1] > jmax) { 90 91 /* estimate how much additional space we will need */ 92 /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */ 93 /* just double the memory each time */ 94 int maxadd = jmax; 95 /* maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); */ 96 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 97 jmax += maxadd; 98 99 /* allocate a longer ajnew */ 100 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 101 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 102 PetscFree(ajnew); 103 ajnew = ajtmp; 104 realloc++; /* count how many times we realloc */ 105 } 106 ajtmp = ajnew + ainew[i]; 107 fm = fill[n]; 108 nzi = 0; 109 im[i] = nnz; 110 while (nnz--) { 111 if (fm < i) nzi++; 112 *ajtmp++ = fm; 113 fm = fill[fm]; 114 } 115 idnew[i] = ainew[i] + nzi; 116 } 117 118 if (ai[n] != 0) { 119 double af = ((double)ainew[n])/((double)ai[n]); 120 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 121 realloc,f,af); 122 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:Run with -pc_lu_fill %g or use \n",af); 123 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:PCLUSetFill(pc,%g);\n",af); 124 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:for best performance.\n"); 125 } else { 126 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:Empty matrix.\n"); 127 } 128 129 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 130 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 131 132 PetscFree(fill); 133 134 /* put together the new matrix */ 135 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 136 PLogObjectParent(*B,isicol); 137 b = (Mat_SeqBAIJ *) (*B)->data; 138 PetscFree(b->imax); 139 b->singlemalloc = 0; 140 /* the next line frees the default space generated by the Create() */ 141 PetscFree(b->a); PetscFree(b->ilen); 142 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 143 b->j = ajnew; 144 b->i = ainew; 145 b->diag = idnew; 146 b->ilen = 0; 147 b->imax = 0; 148 b->row = isrow; 149 b->col = iscol; 150 b->icol = isicol; 151 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 152 /* In b structure: Free imax, ilen, old a, old j. 153 Allocate idnew, solve_work, new a, new j */ 154 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 155 b->maxnz = b->nz = ainew[n]; 156 157 (*B)->info.factor_mallocs = realloc; 158 (*B)->info.fill_ratio_given = f; 159 if (ai[i] != 0) { 160 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 161 } else { 162 (*B)->info.fill_ratio_needed = 0.0; 163 } 164 165 166 PetscFunctionReturn(0); 167 } 168 169 /* ----------------------------------------------------------- */ 170 #undef __FUNC__ 171 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 172 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 173 { 174 Mat C = *B; 175 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 176 IS iscol = b->col, isrow = b->row, isicol = b->icol; 177 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 178 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 179 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 180 register int *pj; 181 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 182 Scalar *ba = b->a,*aa = a->a; 183 184 PetscFunctionBegin; 185 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 186 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 187 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 188 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 189 /* generate work space needed by dense LU factorization */ 190 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 191 CHKPTRQ(v_work); 192 multiplier = v_work + bs; 193 v_pivots = (int *) (multiplier + bs2); 194 195 /* flops in while loop */ 196 bslog = 2*bs*bs2; 197 198 for ( i=0; i<n; i++ ) { 199 nz = bi[i+1] - bi[i]; 200 ajtmp = bj + bi[i]; 201 for ( j=0; j<nz; j++ ) { 202 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 203 } 204 /* load in initial (unfactored row) */ 205 nz = ai[r[i]+1] - ai[r[i]]; 206 ajtmpold = aj + ai[r[i]]; 207 v = aa + bs2*ai[r[i]]; 208 for ( j=0; j<nz; j++ ) { 209 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 210 } 211 row = *ajtmp++; 212 while (row < i) { 213 pc = rtmp + bs2*row; 214 /* if (*pc) { */ 215 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 216 if (flg) { 217 pv = ba + bs2*diag_offset[row]; 218 pj = bj + diag_offset[row] + 1; 219 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 220 nz = bi[row+1] - diag_offset[row] - 1; 221 pv += bs2; 222 for (j=0; j<nz; j++) { 223 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 224 } 225 PLogFlops(bslog*(nz+1)-bs); 226 } 227 row = *ajtmp++; 228 } 229 /* finished row so stick it into b->a */ 230 pv = ba + bs2*bi[i]; 231 pj = bj + bi[i]; 232 nz = bi[i+1] - bi[i]; 233 for ( j=0; j<nz; j++ ) { 234 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 235 } 236 diag = diag_offset[i] - bi[i]; 237 /* invert diagonal block */ 238 w = pv + bs2*diag; 239 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 240 } 241 242 PetscFree(rtmp); PetscFree(v_work); 243 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 244 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 245 C->factor = FACTOR_LU; 246 C->assembled = PETSC_TRUE; 247 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 248 PetscFunctionReturn(0); 249 } 250 /* ------------------------------------------------------------*/ 251 /* 252 Version for when blocks are 5 by 5 253 */ 254 #undef __FUNC__ 255 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 256 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 257 { 258 Mat C = *B; 259 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 260 IS iscol = b->col, isrow = b->row, isicol; 261 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 262 int *ajtmpold, *ajtmp, nz, row; 263 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 264 register int *pj; 265 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 266 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 267 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 268 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 269 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 270 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 271 Scalar *ba = b->a,*aa = a->a; 272 273 PetscFunctionBegin; 274 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 275 PLogObjectParent(*B,isicol); 276 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 277 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 278 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 279 280 for ( i=0; i<n; i++ ) { 281 nz = bi[i+1] - bi[i]; 282 ajtmp = bj + bi[i]; 283 for ( j=0; j<nz; j++ ) { 284 x = rtmp+25*ajtmp[j]; 285 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 286 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 287 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 288 } 289 /* load in initial (unfactored row) */ 290 idx = r[i]; 291 nz = ai[idx+1] - ai[idx]; 292 ajtmpold = aj + ai[idx]; 293 v = aa + 25*ai[idx]; 294 for ( j=0; j<nz; j++ ) { 295 x = rtmp+25*ic[ajtmpold[j]]; 296 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 297 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 298 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 299 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 300 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 301 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 302 v += 25; 303 } 304 row = *ajtmp++; 305 while (row < i) { 306 pc = rtmp + 25*row; 307 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 308 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 309 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 310 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 311 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 312 p25 = pc[24]; 313 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 314 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 315 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 316 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 317 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 318 p24 != 0.0 || p25 != 0.0) { 319 pv = ba + 25*diag_offset[row]; 320 pj = bj + diag_offset[row] + 1; 321 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 322 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 323 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 324 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 325 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 326 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 327 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 328 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 329 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 330 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 331 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 332 333 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 334 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 335 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 336 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 337 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 338 339 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 340 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 341 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 342 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 343 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 344 345 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 346 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 347 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 348 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 349 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 350 351 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 352 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 353 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 354 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 355 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 356 357 nz = bi[row+1] - diag_offset[row] - 1; 358 pv += 25; 359 for (j=0; j<nz; j++) { 360 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 361 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 362 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 363 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 364 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 365 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 366 x = rtmp + 25*pj[j]; 367 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 368 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 369 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 370 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 371 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 372 373 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 374 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 375 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 376 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 377 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 378 379 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 380 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 381 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 382 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 383 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 384 385 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 386 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 387 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 388 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 389 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 390 391 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 392 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 393 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 394 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 395 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 396 397 pv += 25; 398 } 399 PLogFlops(250*nz+225); 400 } 401 row = *ajtmp++; 402 } 403 /* finished row so stick it into b->a */ 404 pv = ba + 25*bi[i]; 405 pj = bj + bi[i]; 406 nz = bi[i+1] - bi[i]; 407 for ( j=0; j<nz; j++ ) { 408 x = rtmp+25*pj[j]; 409 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 410 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 411 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 412 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 413 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 414 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 415 pv += 25; 416 } 417 /* invert diagonal block */ 418 w = ba + 25*diag_offset[i]; 419 ierr = Kernel_A_gets_inverse_A_5(w); CHKERRQ(ierr); 420 } 421 422 PetscFree(rtmp); 423 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 424 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 425 ierr = ISDestroy(isicol); CHKERRQ(ierr); 426 C->factor = FACTOR_LU; 427 C->assembled = PETSC_TRUE; 428 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 429 PetscFunctionReturn(0); 430 } 431 432 /* ------------------------------------------------------------*/ 433 /* 434 Version for when blocks are 4 by 4 435 */ 436 #undef __FUNC__ 437 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 438 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 439 { 440 Mat C = *B; 441 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 442 IS iscol = b->col, isrow = b->row, isicol; 443 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 444 int *ajtmpold, *ajtmp, nz, row; 445 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 446 register int *pj; 447 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 448 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 449 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 450 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 451 Scalar m13,m14,m15,m16; 452 Scalar *ba = b->a,*aa = a->a; 453 454 PetscFunctionBegin; 455 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 456 PLogObjectParent(*B,isicol); 457 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 458 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 459 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 460 461 for ( i=0; i<n; i++ ) { 462 nz = bi[i+1] - bi[i]; 463 ajtmp = bj + bi[i]; 464 for ( j=0; j<nz; j++ ) { 465 x = rtmp+16*ajtmp[j]; 466 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 467 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 468 } 469 /* load in initial (unfactored row) */ 470 idx = r[i]; 471 nz = ai[idx+1] - ai[idx]; 472 ajtmpold = aj + ai[idx]; 473 v = aa + 16*ai[idx]; 474 for ( j=0; j<nz; j++ ) { 475 x = rtmp+16*ic[ajtmpold[j]]; 476 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 477 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 478 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 479 x[14] = v[14]; x[15] = v[15]; 480 v += 16; 481 } 482 row = *ajtmp++; 483 while (row < i) { 484 pc = rtmp + 16*row; 485 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 486 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 487 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 488 p15 = pc[14]; p16 = pc[15]; 489 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 490 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 491 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 492 || p16 != 0.0) { 493 pv = ba + 16*diag_offset[row]; 494 pj = bj + diag_offset[row] + 1; 495 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 496 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 497 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 498 x15 = pv[14]; x16 = pv[15]; 499 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 500 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 501 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 502 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 503 504 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 505 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 506 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 507 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 508 509 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 510 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 511 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 512 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 513 514 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 515 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 516 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 517 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 518 519 nz = bi[row+1] - diag_offset[row] - 1; 520 pv += 16; 521 for (j=0; j<nz; j++) { 522 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 523 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 524 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 525 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 526 x = rtmp + 16*pj[j]; 527 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 528 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 529 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 530 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 531 532 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 533 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 534 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 535 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 536 537 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 538 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 539 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 540 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 541 542 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 543 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 544 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 545 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 546 547 pv += 16; 548 } 549 PLogFlops(128*nz+112); 550 } 551 row = *ajtmp++; 552 } 553 /* finished row so stick it into b->a */ 554 pv = ba + 16*bi[i]; 555 pj = bj + bi[i]; 556 nz = bi[i+1] - bi[i]; 557 for ( j=0; j<nz; j++ ) { 558 x = rtmp+16*pj[j]; 559 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 560 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 561 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 562 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 563 pv += 16; 564 } 565 /* invert diagonal block */ 566 w = ba + 16*diag_offset[i]; 567 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 568 } 569 570 PetscFree(rtmp); 571 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 572 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 573 ierr = ISDestroy(isicol); CHKERRQ(ierr); 574 C->factor = FACTOR_LU; 575 C->assembled = PETSC_TRUE; 576 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 577 PetscFunctionReturn(0); 578 } 579 /* 580 Version for when blocks are 4 by 4 Using natural ordering 581 */ 582 #undef __FUNC__ 583 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering" 584 int MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering(Mat A,Mat *B) 585 { 586 Mat C = *B; 587 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 588 int ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 589 int *ajtmpold, *ajtmp, nz, row; 590 int *diag_offset = b->diag,*ai=a->i,*aj=a->j; 591 register int *pj; 592 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 593 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 594 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 595 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 596 Scalar m13,m14,m15,m16; 597 Scalar *ba = b->a,*aa = a->a; 598 599 PetscFunctionBegin; 600 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 601 602 for ( i=0; i<n; i++ ) { 603 nz = bi[i+1] - bi[i]; 604 ajtmp = bj + bi[i]; 605 for ( j=0; j<nz; j++ ) { 606 x = rtmp+16*ajtmp[j]; 607 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 608 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 609 } 610 /* load in initial (unfactored row) */ 611 nz = ai[i+1] - ai[i]; 612 ajtmpold = aj + ai[i]; 613 v = aa + 16*ai[i]; 614 for ( j=0; j<nz; j++ ) { 615 x = rtmp+16*ajtmpold[j]; 616 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 617 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 618 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 619 x[14] = v[14]; x[15] = v[15]; 620 v += 16; 621 } 622 row = *ajtmp++; 623 while (row < i) { 624 pc = rtmp + 16*row; 625 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 626 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 627 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 628 p15 = pc[14]; p16 = pc[15]; 629 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 630 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 631 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 632 || p16 != 0.0) { 633 pv = ba + 16*diag_offset[row]; 634 pj = bj + diag_offset[row] + 1; 635 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 636 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 637 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 638 x15 = pv[14]; x16 = pv[15]; 639 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 640 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 641 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 642 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 643 644 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 645 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 646 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 647 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 648 649 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 650 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 651 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 652 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 653 654 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 655 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 656 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 657 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 658 659 nz = bi[row+1] - diag_offset[row] - 1; 660 pv += 16; 661 for (j=0; j<nz; j++) { 662 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 663 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 664 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 665 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 666 x = rtmp + 16*pj[j]; 667 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 668 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 669 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 670 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 671 672 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 673 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 674 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 675 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 676 677 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 678 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 679 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 680 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 681 682 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 683 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 684 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 685 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 686 687 pv += 16; 688 } 689 PLogFlops(128*nz+112); 690 } 691 row = *ajtmp++; 692 } 693 /* finished row so stick it into b->a */ 694 pv = ba + 16*bi[i]; 695 pj = bj + bi[i]; 696 nz = bi[i+1] - bi[i]; 697 for ( j=0; j<nz; j++ ) { 698 x = rtmp+16*pj[j]; 699 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 700 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 701 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 702 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 703 pv += 16; 704 } 705 /* invert diagonal block */ 706 w = ba + 16*diag_offset[i]; 707 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 708 } 709 710 PetscFree(rtmp); 711 C->factor = FACTOR_LU; 712 C->assembled = PETSC_TRUE; 713 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 714 PetscFunctionReturn(0); 715 } 716 717 /* ------------------------------------------------------------*/ 718 /* 719 Version for when blocks are 3 by 3 720 */ 721 #undef __FUNC__ 722 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 723 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 724 { 725 Mat C = *B; 726 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 727 IS iscol = b->col, isrow = b->row, isicol; 728 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 729 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 730 int *diag_offset = b->diag,idx; 731 register int *pj; 732 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 733 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 734 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 735 Scalar *ba = b->a,*aa = a->a; 736 737 PetscFunctionBegin; 738 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 739 PLogObjectParent(*B,isicol); 740 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 741 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 742 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 743 744 for ( i=0; i<n; i++ ) { 745 nz = bi[i+1] - bi[i]; 746 ajtmp = bj + bi[i]; 747 for ( j=0; j<nz; j++ ) { 748 x = rtmp + 9*ajtmp[j]; 749 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 750 } 751 /* load in initial (unfactored row) */ 752 idx = r[i]; 753 nz = ai[idx+1] - ai[idx]; 754 ajtmpold = aj + ai[idx]; 755 v = aa + 9*ai[idx]; 756 for ( j=0; j<nz; j++ ) { 757 x = rtmp + 9*ic[ajtmpold[j]]; 758 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 759 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 760 v += 9; 761 } 762 row = *ajtmp++; 763 while (row < i) { 764 pc = rtmp + 9*row; 765 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 766 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 767 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 768 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 769 pv = ba + 9*diag_offset[row]; 770 pj = bj + diag_offset[row] + 1; 771 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 772 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 773 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 774 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 775 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 776 777 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 778 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 779 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 780 781 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 782 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 783 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 784 nz = bi[row+1] - diag_offset[row] - 1; 785 pv += 9; 786 for (j=0; j<nz; j++) { 787 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 788 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 789 x = rtmp + 9*pj[j]; 790 x[0] -= m1*x1 + m4*x2 + m7*x3; 791 x[1] -= m2*x1 + m5*x2 + m8*x3; 792 x[2] -= m3*x1 + m6*x2 + m9*x3; 793 794 x[3] -= m1*x4 + m4*x5 + m7*x6; 795 x[4] -= m2*x4 + m5*x5 + m8*x6; 796 x[5] -= m3*x4 + m6*x5 + m9*x6; 797 798 x[6] -= m1*x7 + m4*x8 + m7*x9; 799 x[7] -= m2*x7 + m5*x8 + m8*x9; 800 x[8] -= m3*x7 + m6*x8 + m9*x9; 801 pv += 9; 802 } 803 PLogFlops(54*nz+36); 804 } 805 row = *ajtmp++; 806 } 807 /* finished row so stick it into b->a */ 808 pv = ba + 9*bi[i]; 809 pj = bj + bi[i]; 810 nz = bi[i+1] - bi[i]; 811 for ( j=0; j<nz; j++ ) { 812 x = rtmp + 9*pj[j]; 813 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 814 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 815 pv += 9; 816 } 817 /* invert diagonal block */ 818 w = ba + 9*diag_offset[i]; 819 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 820 } 821 822 PetscFree(rtmp); 823 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 824 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 825 ierr = ISDestroy(isicol); CHKERRQ(ierr); 826 C->factor = FACTOR_LU; 827 C->assembled = PETSC_TRUE; 828 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 829 PetscFunctionReturn(0); 830 } 831 832 /* ------------------------------------------------------------*/ 833 /* 834 Version for when blocks are 2 by 2 835 */ 836 #undef __FUNC__ 837 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 838 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 839 { 840 Mat C = *B; 841 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 842 IS iscol = b->col, isrow = b->row, isicol; 843 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 844 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 845 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 846 register int *pj; 847 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 848 Scalar p1,p2,p3,p4,v_work[2]; 849 Scalar *ba = b->a,*aa = a->a; 850 851 PetscFunctionBegin; 852 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 853 PLogObjectParent(*B,isicol); 854 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 855 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 856 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 857 858 for ( i=0; i<n; i++ ) { 859 nz = bi[i+1] - bi[i]; 860 ajtmp = bj + bi[i]; 861 for ( j=0; j<nz; j++ ) { 862 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 863 } 864 /* load in initial (unfactored row) */ 865 idx = r[i]; 866 nz = ai[idx+1] - ai[idx]; 867 ajtmpold = aj + ai[idx]; 868 v = aa + 4*ai[idx]; 869 for ( j=0; j<nz; j++ ) { 870 x = rtmp+4*ic[ajtmpold[j]]; 871 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 872 v += 4; 873 } 874 row = *ajtmp++; 875 while (row < i) { 876 pc = rtmp + 4*row; 877 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 878 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 879 pv = ba + 4*diag_offset[row]; 880 pj = bj + diag_offset[row] + 1; 881 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 882 pc[0] = m1 = p1*x1 + p3*x2; 883 pc[1] = m2 = p2*x1 + p4*x2; 884 pc[2] = m3 = p1*x3 + p3*x4; 885 pc[3] = m4 = p2*x3 + p4*x4; 886 nz = bi[row+1] - diag_offset[row] - 1; 887 pv += 4; 888 for (j=0; j<nz; j++) { 889 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 890 x = rtmp + 4*pj[j]; 891 x[0] -= m1*x1 + m3*x2; 892 x[1] -= m2*x1 + m4*x2; 893 x[2] -= m1*x3 + m3*x4; 894 x[3] -= m2*x3 + m4*x4; 895 pv += 4; 896 } 897 PLogFlops(16*nz+12); 898 } 899 row = *ajtmp++; 900 } 901 /* finished row so stick it into b->a */ 902 pv = ba + 4*bi[i]; 903 pj = bj + bi[i]; 904 nz = bi[i+1] - bi[i]; 905 for ( j=0; j<nz; j++ ) { 906 x = rtmp+4*pj[j]; 907 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 908 pv += 4; 909 } 910 /* invert diagonal block */ 911 w = ba + 4*diag_offset[i]; 912 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 913 } 914 915 PetscFree(rtmp); 916 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 917 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 918 ierr = ISDestroy(isicol); CHKERRQ(ierr); 919 C->factor = FACTOR_LU; 920 C->assembled = PETSC_TRUE; 921 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 922 PetscFunctionReturn(0); 923 } 924 925 /* ----------------------------------------------------------- */ 926 /* 927 Version for when blocks are 1 by 1. 928 */ 929 #undef __FUNC__ 930 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 931 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 932 { 933 Mat C = *B; 934 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 935 IS iscol = b->col, isrow = b->row, isicol; 936 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 937 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 938 int *diag_offset = b->diag,diag; 939 register int *pj; 940 register Scalar *pv,*v,*rtmp,multiplier,*pc; 941 Scalar *ba = b->a,*aa = a->a; 942 943 PetscFunctionBegin; 944 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 945 PLogObjectParent(*B,isicol); 946 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 947 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 948 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 949 950 for ( i=0; i<n; i++ ) { 951 nz = bi[i+1] - bi[i]; 952 ajtmp = bj + bi[i]; 953 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 954 955 /* load in initial (unfactored row) */ 956 nz = ai[r[i]+1] - ai[r[i]]; 957 ajtmpold = aj + ai[r[i]]; 958 v = aa + ai[r[i]]; 959 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 960 961 row = *ajtmp++; 962 while (row < i) { 963 pc = rtmp + row; 964 if (*pc != 0.0) { 965 pv = ba + diag_offset[row]; 966 pj = bj + diag_offset[row] + 1; 967 multiplier = *pc * *pv++; 968 *pc = multiplier; 969 nz = bi[row+1] - diag_offset[row] - 1; 970 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 971 PLogFlops(1+2*nz); 972 } 973 row = *ajtmp++; 974 } 975 /* finished row so stick it into b->a */ 976 pv = ba + bi[i]; 977 pj = bj + bi[i]; 978 nz = bi[i+1] - bi[i]; 979 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 980 diag = diag_offset[i] - bi[i]; 981 /* check pivot entry for current row */ 982 if (pv[diag] == 0.0) { 983 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 984 } 985 pv[diag] = 1.0/pv[diag]; 986 } 987 988 PetscFree(rtmp); 989 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 990 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 991 ierr = ISDestroy(isicol); CHKERRQ(ierr); 992 C->factor = FACTOR_LU; 993 C->assembled = PETSC_TRUE; 994 PLogFlops(b->n); 995 PetscFunctionReturn(0); 996 } 997 998 /* ----------------------------------------------------------- */ 999 #undef __FUNC__ 1000 #define __FUNC__ "MatLUFactor_SeqBAIJ" 1001 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 1002 { 1003 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 1004 int ierr; 1005 Mat C; 1006 PetscOps *Abops; 1007 struct _MatOps *Aops; 1008 1009 PetscFunctionBegin; 1010 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 1011 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 1012 1013 /* free all the data structures from mat */ 1014 PetscFree(mat->a); 1015 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 1016 if (mat->diag) PetscFree(mat->diag); 1017 if (mat->ilen) PetscFree(mat->ilen); 1018 if (mat->imax) PetscFree(mat->imax); 1019 if (mat->solve_work) PetscFree(mat->solve_work); 1020 if (mat->mult_work) PetscFree(mat->mult_work); 1021 PetscFree(mat); 1022 1023 /* 1024 This is horrible, horrible code. We need to keep the 1025 A pointers for the bops and ops but copy everything 1026 else from C. 1027 */ 1028 Abops = A->bops; 1029 Aops = A->ops; 1030 PetscMemcpy(A,C,sizeof(struct _p_Mat)); 1031 A->bops = Abops; 1032 A->ops = Aops; 1033 1034 PetscHeaderDestroy(C); 1035 PetscFunctionReturn(0); 1036 } 1037