1 #ifdef PETSC_RCS_HEADER 2 static char vcid[] = "$Id: baijfact.c,v 1.59 1998/03/26 22:31:56 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 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 90 /* estimate how much additional space we will need */ 91 /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */ 92 /* just double the memory each time */ 93 int maxadd = jmax; 94 /* maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); */ 95 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 96 jmax += maxadd; 97 98 /* allocate a longer ajnew */ 99 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 100 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 101 PetscFree(ajnew); 102 ajnew = ajtmp; 103 realloc++; /* count how many times we realloc */ 104 } 105 ajtmp = ajnew + ainew[i]; 106 fm = fill[n]; 107 nzi = 0; 108 im[i] = nnz; 109 while (nnz--) { 110 if (fm < i) nzi++; 111 *ajtmp++ = fm; 112 fm = fill[fm]; 113 } 114 idnew[i] = ainew[i] + nzi; 115 } 116 117 if (ai[n] != 0) { 118 double af = ((double)ainew[n])/((double)ai[n]); 119 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 120 realloc,f,af); 121 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:Run with -pc_lu_fill %g or use \n",af); 122 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:PCLUSetFill(pc,%g);\n",af); 123 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:for best performance.\n"); 124 } else { 125 PLogInfo(A,"MatLUFactorSymbolic_SeqBAIJ:Empty matrix.\n"); 126 } 127 128 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 129 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 130 131 PetscFree(fill); 132 133 /* put together the new matrix */ 134 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 135 PLogObjectParent(*B,isicol); 136 b = (Mat_SeqBAIJ *) (*B)->data; 137 PetscFree(b->imax); 138 b->singlemalloc = 0; 139 /* the next line frees the default space generated by the Create() */ 140 PetscFree(b->a); PetscFree(b->ilen); 141 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 142 b->j = ajnew; 143 b->i = ainew; 144 b->diag = idnew; 145 b->ilen = 0; 146 b->imax = 0; 147 b->row = isrow; 148 b->col = iscol; 149 b->icol = isicol; 150 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 151 /* In b structure: Free imax, ilen, old a, old j. 152 Allocate idnew, solve_work, new a, new j */ 153 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 154 b->maxnz = b->nz = ainew[n]; 155 156 (*B)->info.factor_mallocs = realloc; 157 (*B)->info.fill_ratio_given = f; 158 if (ai[i] != 0) { 159 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 160 } else { 161 (*B)->info.fill_ratio_needed = 0.0; 162 } 163 164 165 PetscFunctionReturn(0); 166 } 167 168 /* ----------------------------------------------------------- */ 169 #undef __FUNC__ 170 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 171 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 172 { 173 Mat C = *B; 174 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 175 IS iscol = b->col, isrow = b->row, isicol = b->icol; 176 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 177 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 178 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 179 register int *pj; 180 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 181 Scalar *ba = b->a,*aa = a->a; 182 183 PetscFunctionBegin; 184 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 185 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 186 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 187 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 188 /* generate work space needed by dense LU factorization */ 189 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 190 CHKPTRQ(v_work); 191 multiplier = v_work + bs; 192 v_pivots = (int *) (multiplier + bs2); 193 194 /* flops in while loop */ 195 bslog = 2*bs*bs2; 196 197 for ( i=0; i<n; i++ ) { 198 nz = bi[i+1] - bi[i]; 199 ajtmp = bj + bi[i]; 200 for ( j=0; j<nz; j++ ) { 201 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 202 } 203 /* load in initial (unfactored row) */ 204 nz = ai[r[i]+1] - ai[r[i]]; 205 ajtmpold = aj + ai[r[i]]; 206 v = aa + bs2*ai[r[i]]; 207 for ( j=0; j<nz; j++ ) { 208 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 209 } 210 row = *ajtmp++; 211 while (row < i) { 212 pc = rtmp + bs2*row; 213 /* if (*pc) { */ 214 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 215 if (flg) { 216 pv = ba + bs2*diag_offset[row]; 217 pj = bj + diag_offset[row] + 1; 218 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 219 nz = bi[row+1] - diag_offset[row] - 1; 220 pv += bs2; 221 for (j=0; j<nz; j++) { 222 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 223 } 224 PLogFlops(bslog*(nz+1)-bs); 225 } 226 row = *ajtmp++; 227 } 228 /* finished row so stick it into b->a */ 229 pv = ba + bs2*bi[i]; 230 pj = bj + bi[i]; 231 nz = bi[i+1] - bi[i]; 232 for ( j=0; j<nz; j++ ) { 233 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 234 } 235 diag = diag_offset[i] - bi[i]; 236 /* invert diagonal block */ 237 w = pv + bs2*diag; 238 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 239 } 240 241 PetscFree(rtmp); PetscFree(v_work); 242 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 243 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 244 C->factor = FACTOR_LU; 245 C->assembled = PETSC_TRUE; 246 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 247 PetscFunctionReturn(0); 248 } 249 /* ------------------------------------------------------------*/ 250 /* 251 Version for when blocks are 5 by 5 252 */ 253 #undef __FUNC__ 254 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 255 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 256 { 257 Mat C = *B; 258 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 259 IS iscol = b->col, isrow = b->row, isicol; 260 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 261 int *ajtmpold, *ajtmp, nz, row; 262 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 263 register int *pj; 264 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 265 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 266 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 267 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 268 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 269 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 270 Scalar *ba = b->a,*aa = a->a; 271 272 PetscFunctionBegin; 273 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 274 PLogObjectParent(*B,isicol); 275 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 276 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 277 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 278 279 for ( i=0; i<n; i++ ) { 280 nz = bi[i+1] - bi[i]; 281 ajtmp = bj + bi[i]; 282 for ( j=0; j<nz; j++ ) { 283 x = rtmp+25*ajtmp[j]; 284 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 285 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 286 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 287 } 288 /* load in initial (unfactored row) */ 289 idx = r[i]; 290 nz = ai[idx+1] - ai[idx]; 291 ajtmpold = aj + ai[idx]; 292 v = aa + 25*ai[idx]; 293 for ( j=0; j<nz; j++ ) { 294 x = rtmp+25*ic[ajtmpold[j]]; 295 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 296 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 297 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 298 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 299 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 300 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 301 v += 25; 302 } 303 row = *ajtmp++; 304 while (row < i) { 305 pc = rtmp + 25*row; 306 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 307 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 308 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 309 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 310 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 311 p25 = pc[24]; 312 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 313 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 314 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 315 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 316 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 317 p24 != 0.0 || p25 != 0.0) { 318 pv = ba + 25*diag_offset[row]; 319 pj = bj + diag_offset[row] + 1; 320 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 321 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 322 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 323 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 324 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 325 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 326 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 327 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 328 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 329 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 330 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 331 332 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 333 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 334 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 335 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 336 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 337 338 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 339 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 340 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 341 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 342 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 343 344 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 345 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 346 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 347 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 348 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 349 350 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 351 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 352 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 353 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 354 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 355 356 nz = bi[row+1] - diag_offset[row] - 1; 357 pv += 25; 358 for (j=0; j<nz; j++) { 359 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 360 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 361 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 362 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 363 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 364 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 365 x = rtmp + 25*pj[j]; 366 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 367 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 368 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 369 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 370 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 371 372 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 373 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 374 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 375 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 376 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 377 378 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 379 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 380 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 381 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 382 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 383 384 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 385 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 386 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 387 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 388 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 389 390 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 391 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 392 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 393 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 394 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 395 396 pv += 25; 397 } 398 PLogFlops(250*nz+225); 399 } 400 row = *ajtmp++; 401 } 402 /* finished row so stick it into b->a */ 403 pv = ba + 25*bi[i]; 404 pj = bj + bi[i]; 405 nz = bi[i+1] - bi[i]; 406 for ( j=0; j<nz; j++ ) { 407 x = rtmp+25*pj[j]; 408 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 409 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 410 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 411 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 412 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 413 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 414 pv += 25; 415 } 416 /* invert diagonal block */ 417 w = ba + 25*diag_offset[i]; 418 ierr = Kernel_A_gets_inverse_A_5(w); CHKERRQ(ierr); 419 } 420 421 PetscFree(rtmp); 422 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 423 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 424 ierr = ISDestroy(isicol); CHKERRQ(ierr); 425 C->factor = FACTOR_LU; 426 C->assembled = PETSC_TRUE; 427 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 428 PetscFunctionReturn(0); 429 } 430 431 /* ------------------------------------------------------------*/ 432 /* 433 Version for when blocks are 4 by 4 434 */ 435 #undef __FUNC__ 436 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 437 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 438 { 439 Mat C = *B; 440 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 441 IS iscol = b->col, isrow = b->row, isicol; 442 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 443 int *ajtmpold, *ajtmp, nz, row; 444 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 445 register int *pj; 446 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 447 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 448 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 449 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 450 Scalar m13,m14,m15,m16; 451 Scalar *ba = b->a,*aa = a->a; 452 453 PetscFunctionBegin; 454 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 455 PLogObjectParent(*B,isicol); 456 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 457 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 458 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 459 460 for ( i=0; i<n; i++ ) { 461 nz = bi[i+1] - bi[i]; 462 ajtmp = bj + bi[i]; 463 for ( j=0; j<nz; j++ ) { 464 x = rtmp+16*ajtmp[j]; 465 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 466 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 467 } 468 /* load in initial (unfactored row) */ 469 idx = r[i]; 470 nz = ai[idx+1] - ai[idx]; 471 ajtmpold = aj + ai[idx]; 472 v = aa + 16*ai[idx]; 473 for ( j=0; j<nz; j++ ) { 474 x = rtmp+16*ic[ajtmpold[j]]; 475 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 476 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 477 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 478 x[14] = v[14]; x[15] = v[15]; 479 v += 16; 480 } 481 row = *ajtmp++; 482 while (row < i) { 483 pc = rtmp + 16*row; 484 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 485 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 486 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 487 p15 = pc[14]; p16 = pc[15]; 488 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 489 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 490 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 491 || p16 != 0.0) { 492 pv = ba + 16*diag_offset[row]; 493 pj = bj + diag_offset[row] + 1; 494 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 495 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 496 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 497 x15 = pv[14]; x16 = pv[15]; 498 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 499 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 500 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 501 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 502 503 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 504 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 505 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 506 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 507 508 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 509 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 510 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 511 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 512 513 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 514 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 515 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 516 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 517 518 nz = bi[row+1] - diag_offset[row] - 1; 519 pv += 16; 520 for (j=0; j<nz; j++) { 521 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 522 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 523 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 524 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 525 x = rtmp + 16*pj[j]; 526 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 527 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 528 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 529 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 530 531 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 532 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 533 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 534 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 535 536 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 537 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 538 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 539 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 540 541 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 542 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 543 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 544 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 545 546 pv += 16; 547 } 548 PLogFlops(128*nz+112); 549 } 550 row = *ajtmp++; 551 } 552 /* finished row so stick it into b->a */ 553 pv = ba + 16*bi[i]; 554 pj = bj + bi[i]; 555 nz = bi[i+1] - bi[i]; 556 for ( j=0; j<nz; j++ ) { 557 x = rtmp+16*pj[j]; 558 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 559 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 560 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 561 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 562 pv += 16; 563 } 564 /* invert diagonal block */ 565 w = ba + 16*diag_offset[i]; 566 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 567 } 568 569 PetscFree(rtmp); 570 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 571 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 572 ierr = ISDestroy(isicol); CHKERRQ(ierr); 573 C->factor = FACTOR_LU; 574 C->assembled = PETSC_TRUE; 575 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 576 PetscFunctionReturn(0); 577 } 578 /* 579 Version for when blocks are 4 by 4 Using natural ordering 580 */ 581 #undef __FUNC__ 582 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering" 583 int MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering(Mat A,Mat *B) 584 { 585 Mat C = *B; 586 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 587 int ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 588 int *ajtmpold, *ajtmp, nz, row; 589 int *diag_offset = b->diag,*ai=a->i,*aj=a->j; 590 register int *pj; 591 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 592 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 593 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 594 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 595 Scalar m13,m14,m15,m16; 596 Scalar *ba = b->a,*aa = a->a; 597 598 PetscFunctionBegin; 599 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 600 601 for ( i=0; i<n; i++ ) { 602 nz = bi[i+1] - bi[i]; 603 ajtmp = bj + bi[i]; 604 for ( j=0; j<nz; j++ ) { 605 x = rtmp+16*ajtmp[j]; 606 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 607 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 608 } 609 /* load in initial (unfactored row) */ 610 nz = ai[i+1] - ai[i]; 611 ajtmpold = aj + ai[i]; 612 v = aa + 16*ai[i]; 613 for ( j=0; j<nz; j++ ) { 614 x = rtmp+16*ajtmpold[j]; 615 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 616 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 617 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 618 x[14] = v[14]; x[15] = v[15]; 619 v += 16; 620 } 621 row = *ajtmp++; 622 while (row < i) { 623 pc = rtmp + 16*row; 624 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 625 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 626 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 627 p15 = pc[14]; p16 = pc[15]; 628 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 629 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 630 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 631 || p16 != 0.0) { 632 pv = ba + 16*diag_offset[row]; 633 pj = bj + diag_offset[row] + 1; 634 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 635 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 636 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 637 x15 = pv[14]; x16 = pv[15]; 638 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 639 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 640 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 641 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 642 643 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 644 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 645 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 646 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 647 648 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 649 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 650 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 651 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 652 653 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 654 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 655 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 656 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 657 658 nz = bi[row+1] - diag_offset[row] - 1; 659 pv += 16; 660 for (j=0; j<nz; j++) { 661 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 662 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 663 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 664 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 665 x = rtmp + 16*pj[j]; 666 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 667 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 668 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 669 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 670 671 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 672 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 673 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 674 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 675 676 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 677 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 678 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 679 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 680 681 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 682 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 683 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 684 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 685 686 pv += 16; 687 } 688 PLogFlops(128*nz+112); 689 } 690 row = *ajtmp++; 691 } 692 /* finished row so stick it into b->a */ 693 pv = ba + 16*bi[i]; 694 pj = bj + bi[i]; 695 nz = bi[i+1] - bi[i]; 696 for ( j=0; j<nz; j++ ) { 697 x = rtmp+16*pj[j]; 698 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 699 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 700 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 701 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 702 pv += 16; 703 } 704 /* invert diagonal block */ 705 w = ba + 16*diag_offset[i]; 706 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 707 } 708 709 PetscFree(rtmp); 710 C->factor = FACTOR_LU; 711 C->assembled = PETSC_TRUE; 712 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 713 PetscFunctionReturn(0); 714 } 715 716 /* ------------------------------------------------------------*/ 717 /* 718 Version for when blocks are 3 by 3 719 */ 720 #undef __FUNC__ 721 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 722 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 723 { 724 Mat C = *B; 725 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 726 IS iscol = b->col, isrow = b->row, isicol; 727 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 728 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 729 int *diag_offset = b->diag,idx; 730 register int *pj; 731 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 732 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 733 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 734 Scalar *ba = b->a,*aa = a->a; 735 736 PetscFunctionBegin; 737 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 738 PLogObjectParent(*B,isicol); 739 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 740 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 741 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 742 743 for ( i=0; i<n; i++ ) { 744 nz = bi[i+1] - bi[i]; 745 ajtmp = bj + bi[i]; 746 for ( j=0; j<nz; j++ ) { 747 x = rtmp + 9*ajtmp[j]; 748 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 749 } 750 /* load in initial (unfactored row) */ 751 idx = r[i]; 752 nz = ai[idx+1] - ai[idx]; 753 ajtmpold = aj + ai[idx]; 754 v = aa + 9*ai[idx]; 755 for ( j=0; j<nz; j++ ) { 756 x = rtmp + 9*ic[ajtmpold[j]]; 757 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 758 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 759 v += 9; 760 } 761 row = *ajtmp++; 762 while (row < i) { 763 pc = rtmp + 9*row; 764 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 765 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 766 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 767 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 768 pv = ba + 9*diag_offset[row]; 769 pj = bj + diag_offset[row] + 1; 770 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 771 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 772 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 773 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 774 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 775 776 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 777 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 778 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 779 780 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 781 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 782 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 783 nz = bi[row+1] - diag_offset[row] - 1; 784 pv += 9; 785 for (j=0; j<nz; j++) { 786 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 787 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 788 x = rtmp + 9*pj[j]; 789 x[0] -= m1*x1 + m4*x2 + m7*x3; 790 x[1] -= m2*x1 + m5*x2 + m8*x3; 791 x[2] -= m3*x1 + m6*x2 + m9*x3; 792 793 x[3] -= m1*x4 + m4*x5 + m7*x6; 794 x[4] -= m2*x4 + m5*x5 + m8*x6; 795 x[5] -= m3*x4 + m6*x5 + m9*x6; 796 797 x[6] -= m1*x7 + m4*x8 + m7*x9; 798 x[7] -= m2*x7 + m5*x8 + m8*x9; 799 x[8] -= m3*x7 + m6*x8 + m9*x9; 800 pv += 9; 801 } 802 PLogFlops(54*nz+36); 803 } 804 row = *ajtmp++; 805 } 806 /* finished row so stick it into b->a */ 807 pv = ba + 9*bi[i]; 808 pj = bj + bi[i]; 809 nz = bi[i+1] - bi[i]; 810 for ( j=0; j<nz; j++ ) { 811 x = rtmp + 9*pj[j]; 812 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 813 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 814 pv += 9; 815 } 816 /* invert diagonal block */ 817 w = ba + 9*diag_offset[i]; 818 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 819 } 820 821 PetscFree(rtmp); 822 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 823 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 824 ierr = ISDestroy(isicol); CHKERRQ(ierr); 825 C->factor = FACTOR_LU; 826 C->assembled = PETSC_TRUE; 827 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 828 PetscFunctionReturn(0); 829 } 830 831 /* ------------------------------------------------------------*/ 832 /* 833 Version for when blocks are 2 by 2 834 */ 835 #undef __FUNC__ 836 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 837 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 838 { 839 Mat C = *B; 840 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 841 IS iscol = b->col, isrow = b->row, isicol; 842 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 843 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 844 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 845 register int *pj; 846 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 847 Scalar p1,p2,p3,p4,v_work[2]; 848 Scalar *ba = b->a,*aa = a->a; 849 850 PetscFunctionBegin; 851 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 852 PLogObjectParent(*B,isicol); 853 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 854 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 855 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 856 857 for ( i=0; i<n; i++ ) { 858 nz = bi[i+1] - bi[i]; 859 ajtmp = bj + bi[i]; 860 for ( j=0; j<nz; j++ ) { 861 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 862 } 863 /* load in initial (unfactored row) */ 864 idx = r[i]; 865 nz = ai[idx+1] - ai[idx]; 866 ajtmpold = aj + ai[idx]; 867 v = aa + 4*ai[idx]; 868 for ( j=0; j<nz; j++ ) { 869 x = rtmp+4*ic[ajtmpold[j]]; 870 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 871 v += 4; 872 } 873 row = *ajtmp++; 874 while (row < i) { 875 pc = rtmp + 4*row; 876 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 877 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 878 pv = ba + 4*diag_offset[row]; 879 pj = bj + diag_offset[row] + 1; 880 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 881 pc[0] = m1 = p1*x1 + p3*x2; 882 pc[1] = m2 = p2*x1 + p4*x2; 883 pc[2] = m3 = p1*x3 + p3*x4; 884 pc[3] = m4 = p2*x3 + p4*x4; 885 nz = bi[row+1] - diag_offset[row] - 1; 886 pv += 4; 887 for (j=0; j<nz; j++) { 888 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 889 x = rtmp + 4*pj[j]; 890 x[0] -= m1*x1 + m3*x2; 891 x[1] -= m2*x1 + m4*x2; 892 x[2] -= m1*x3 + m3*x4; 893 x[3] -= m2*x3 + m4*x4; 894 pv += 4; 895 } 896 PLogFlops(16*nz+12); 897 } 898 row = *ajtmp++; 899 } 900 /* finished row so stick it into b->a */ 901 pv = ba + 4*bi[i]; 902 pj = bj + bi[i]; 903 nz = bi[i+1] - bi[i]; 904 for ( j=0; j<nz; j++ ) { 905 x = rtmp+4*pj[j]; 906 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 907 pv += 4; 908 } 909 /* invert diagonal block */ 910 w = ba + 4*diag_offset[i]; 911 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 912 } 913 914 PetscFree(rtmp); 915 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 916 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 917 ierr = ISDestroy(isicol); CHKERRQ(ierr); 918 C->factor = FACTOR_LU; 919 C->assembled = PETSC_TRUE; 920 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 921 PetscFunctionReturn(0); 922 } 923 924 /* ----------------------------------------------------------- */ 925 /* 926 Version for when blocks are 1 by 1. 927 */ 928 #undef __FUNC__ 929 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 930 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 931 { 932 Mat C = *B; 933 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 934 IS iscol = b->col, isrow = b->row, isicol; 935 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 936 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 937 int *diag_offset = b->diag,diag; 938 register int *pj; 939 register Scalar *pv,*v,*rtmp,multiplier,*pc; 940 Scalar *ba = b->a,*aa = a->a; 941 942 PetscFunctionBegin; 943 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 944 PLogObjectParent(*B,isicol); 945 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 946 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 947 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 948 949 for ( i=0; i<n; i++ ) { 950 nz = bi[i+1] - bi[i]; 951 ajtmp = bj + bi[i]; 952 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 953 954 /* load in initial (unfactored row) */ 955 nz = ai[r[i]+1] - ai[r[i]]; 956 ajtmpold = aj + ai[r[i]]; 957 v = aa + ai[r[i]]; 958 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 959 960 row = *ajtmp++; 961 while (row < i) { 962 pc = rtmp + row; 963 if (*pc != 0.0) { 964 pv = ba + diag_offset[row]; 965 pj = bj + diag_offset[row] + 1; 966 multiplier = *pc * *pv++; 967 *pc = multiplier; 968 nz = bi[row+1] - diag_offset[row] - 1; 969 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 970 PLogFlops(1+2*nz); 971 } 972 row = *ajtmp++; 973 } 974 /* finished row so stick it into b->a */ 975 pv = ba + bi[i]; 976 pj = bj + bi[i]; 977 nz = bi[i+1] - bi[i]; 978 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 979 diag = diag_offset[i] - bi[i]; 980 /* check pivot entry for current row */ 981 if (pv[diag] == 0.0) { 982 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 983 } 984 pv[diag] = 1.0/pv[diag]; 985 } 986 987 PetscFree(rtmp); 988 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 989 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 990 ierr = ISDestroy(isicol); CHKERRQ(ierr); 991 C->factor = FACTOR_LU; 992 C->assembled = PETSC_TRUE; 993 PLogFlops(b->n); 994 PetscFunctionReturn(0); 995 } 996 997 /* ----------------------------------------------------------- */ 998 #undef __FUNC__ 999 #define __FUNC__ "MatLUFactor_SeqBAIJ" 1000 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 1001 { 1002 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 1003 int ierr; 1004 Mat C; 1005 PetscOps *Abops; 1006 struct _MatOps *Aops; 1007 1008 PetscFunctionBegin; 1009 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 1010 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 1011 1012 /* free all the data structures from mat */ 1013 PetscFree(mat->a); 1014 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 1015 if (mat->diag) PetscFree(mat->diag); 1016 if (mat->ilen) PetscFree(mat->ilen); 1017 if (mat->imax) PetscFree(mat->imax); 1018 if (mat->solve_work) PetscFree(mat->solve_work); 1019 if (mat->mult_work) PetscFree(mat->mult_work); 1020 PetscFree(mat); 1021 1022 /* 1023 This is horrible, horrible code. We need to keep the 1024 A pointers for the bops and ops but copy everything 1025 else from C. 1026 */ 1027 Abops = A->bops; 1028 Aops = A->ops; 1029 PetscMemcpy(A,C,sizeof(struct _p_Mat)); 1030 A->bops = Abops; 1031 A->ops = Aops; 1032 1033 PetscHeaderDestroy(C); 1034 PetscFunctionReturn(0); 1035 } 1036