1 #ifdef PETSC_RCS_HEADER 2 static char vcid[] = "$Id: baijfact.c,v 1.62 1998/03/30 19:48:59 balay Exp bsmith $"; 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 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 isrow = b->row, isicol = b->icol; 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 = ISGetIndices(isrow,&r); CHKERRQ(ierr); 274 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 275 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 276 277 for ( i=0; i<n; i++ ) { 278 nz = bi[i+1] - bi[i]; 279 ajtmp = bj + bi[i]; 280 for ( j=0; j<nz; j++ ) { 281 x = rtmp+25*ajtmp[j]; 282 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 283 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 284 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 285 } 286 /* load in initial (unfactored row) */ 287 idx = r[i]; 288 nz = ai[idx+1] - ai[idx]; 289 ajtmpold = aj + ai[idx]; 290 v = aa + 25*ai[idx]; 291 for ( j=0; j<nz; j++ ) { 292 x = rtmp+25*ic[ajtmpold[j]]; 293 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 294 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 295 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 296 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 297 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 298 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 299 v += 25; 300 } 301 row = *ajtmp++; 302 while (row < i) { 303 pc = rtmp + 25*row; 304 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 305 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 306 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 307 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 308 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 309 p25 = pc[24]; 310 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 311 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 312 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 313 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 314 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 315 p24 != 0.0 || p25 != 0.0) { 316 pv = ba + 25*diag_offset[row]; 317 pj = bj + diag_offset[row] + 1; 318 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 319 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 320 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 321 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 322 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 323 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 324 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 325 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 326 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 327 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 328 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 329 330 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 331 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 332 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 333 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 334 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 335 336 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 337 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 338 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 339 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 340 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 341 342 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 343 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 344 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 345 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 346 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 347 348 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 349 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 350 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 351 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 352 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 353 354 nz = bi[row+1] - diag_offset[row] - 1; 355 pv += 25; 356 for (j=0; j<nz; j++) { 357 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 358 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 359 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 360 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 361 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 362 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 363 x = rtmp + 25*pj[j]; 364 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 365 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 366 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 367 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 368 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 369 370 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 371 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 372 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 373 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 374 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 375 376 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 377 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 378 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 379 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 380 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 381 382 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 383 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 384 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 385 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 386 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 387 388 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 389 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 390 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 391 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 392 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 393 394 pv += 25; 395 } 396 PLogFlops(250*nz+225); 397 } 398 row = *ajtmp++; 399 } 400 /* finished row so stick it into b->a */ 401 pv = ba + 25*bi[i]; 402 pj = bj + bi[i]; 403 nz = bi[i+1] - bi[i]; 404 for ( j=0; j<nz; j++ ) { 405 x = rtmp+25*pj[j]; 406 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 407 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 408 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 409 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 410 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 411 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 412 pv += 25; 413 } 414 /* invert diagonal block */ 415 w = ba + 25*diag_offset[i]; 416 ierr = Kernel_A_gets_inverse_A_5(w); CHKERRQ(ierr); 417 } 418 419 PetscFree(rtmp); 420 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 421 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 422 C->factor = FACTOR_LU; 423 C->assembled = PETSC_TRUE; 424 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 425 PetscFunctionReturn(0); 426 } 427 428 /* ------------------------------------------------------------*/ 429 /* 430 Version for when blocks are 4 by 4 431 */ 432 #undef __FUNC__ 433 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 434 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 435 { 436 Mat C = *B; 437 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 438 IS isrow = b->row, isicol = b->icol; 439 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 440 int *ajtmpold, *ajtmp, nz, row; 441 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 442 register int *pj; 443 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 444 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 445 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 446 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 447 Scalar m13,m14,m15,m16; 448 Scalar *ba = b->a,*aa = a->a; 449 450 PetscFunctionBegin; 451 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 452 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 453 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 454 455 for ( i=0; i<n; i++ ) { 456 nz = bi[i+1] - bi[i]; 457 ajtmp = bj + bi[i]; 458 for ( j=0; j<nz; j++ ) { 459 x = rtmp+16*ajtmp[j]; 460 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 461 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 462 } 463 /* load in initial (unfactored row) */ 464 idx = r[i]; 465 nz = ai[idx+1] - ai[idx]; 466 ajtmpold = aj + ai[idx]; 467 v = aa + 16*ai[idx]; 468 for ( j=0; j<nz; j++ ) { 469 x = rtmp+16*ic[ajtmpold[j]]; 470 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 471 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 472 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 473 x[14] = v[14]; x[15] = v[15]; 474 v += 16; 475 } 476 row = *ajtmp++; 477 while (row < i) { 478 pc = rtmp + 16*row; 479 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 480 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 481 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 482 p15 = pc[14]; p16 = pc[15]; 483 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 484 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 485 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 486 || p16 != 0.0) { 487 pv = ba + 16*diag_offset[row]; 488 pj = bj + diag_offset[row] + 1; 489 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 490 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 491 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 492 x15 = pv[14]; x16 = pv[15]; 493 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 494 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 495 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 496 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 497 498 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 499 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 500 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 501 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 502 503 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 504 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 505 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 506 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 507 508 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 509 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 510 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 511 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 512 513 nz = bi[row+1] - diag_offset[row] - 1; 514 pv += 16; 515 for (j=0; j<nz; j++) { 516 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 517 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 518 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 519 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 520 x = rtmp + 16*pj[j]; 521 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 522 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 523 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 524 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 525 526 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 527 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 528 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 529 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 530 531 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 532 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 533 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 534 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 535 536 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 537 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 538 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 539 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 540 541 pv += 16; 542 } 543 PLogFlops(128*nz+112); 544 } 545 row = *ajtmp++; 546 } 547 /* finished row so stick it into b->a */ 548 pv = ba + 16*bi[i]; 549 pj = bj + bi[i]; 550 nz = bi[i+1] - bi[i]; 551 for ( j=0; j<nz; j++ ) { 552 x = rtmp+16*pj[j]; 553 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 554 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 555 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 556 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 557 pv += 16; 558 } 559 /* invert diagonal block */ 560 w = ba + 16*diag_offset[i]; 561 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 562 } 563 564 PetscFree(rtmp); 565 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 566 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 567 C->factor = FACTOR_LU; 568 C->assembled = PETSC_TRUE; 569 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 570 PetscFunctionReturn(0); 571 } 572 /* 573 Version for when blocks are 4 by 4 Using natural ordering 574 */ 575 #undef __FUNC__ 576 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering" 577 int MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering(Mat A,Mat *B) 578 { 579 Mat C = *B; 580 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 581 int ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 582 int *ajtmpold, *ajtmp, nz, row; 583 int *diag_offset = b->diag,*ai=a->i,*aj=a->j; 584 register int *pj; 585 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 586 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 587 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 588 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 589 Scalar m13,m14,m15,m16; 590 Scalar *ba = b->a,*aa = a->a; 591 592 PetscFunctionBegin; 593 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 594 595 for ( i=0; i<n; i++ ) { 596 nz = bi[i+1] - bi[i]; 597 ajtmp = bj + bi[i]; 598 for ( j=0; j<nz; j++ ) { 599 x = rtmp+16*ajtmp[j]; 600 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 601 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 602 } 603 /* load in initial (unfactored row) */ 604 nz = ai[i+1] - ai[i]; 605 ajtmpold = aj + ai[i]; 606 v = aa + 16*ai[i]; 607 for ( j=0; j<nz; j++ ) { 608 x = rtmp+16*ajtmpold[j]; 609 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 610 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 611 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 612 x[14] = v[14]; x[15] = v[15]; 613 v += 16; 614 } 615 row = *ajtmp++; 616 while (row < i) { 617 pc = rtmp + 16*row; 618 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 619 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 620 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 621 p15 = pc[14]; p16 = pc[15]; 622 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 623 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 624 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 625 || p16 != 0.0) { 626 pv = ba + 16*diag_offset[row]; 627 pj = bj + diag_offset[row] + 1; 628 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 629 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 630 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 631 x15 = pv[14]; x16 = pv[15]; 632 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 633 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 634 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 635 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 636 637 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 638 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 639 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 640 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 641 642 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 643 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 644 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 645 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 646 647 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 648 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 649 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 650 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 651 652 nz = bi[row+1] - diag_offset[row] - 1; 653 pv += 16; 654 for (j=0; j<nz; j++) { 655 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 656 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 657 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 658 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 659 x = rtmp + 16*pj[j]; 660 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 661 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 662 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 663 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 664 665 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 666 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 667 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 668 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 669 670 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 671 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 672 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 673 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 674 675 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 676 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 677 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 678 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 679 680 pv += 16; 681 } 682 PLogFlops(128*nz+112); 683 } 684 row = *ajtmp++; 685 } 686 /* finished row so stick it into b->a */ 687 pv = ba + 16*bi[i]; 688 pj = bj + bi[i]; 689 nz = bi[i+1] - bi[i]; 690 for ( j=0; j<nz; j++ ) { 691 x = rtmp+16*pj[j]; 692 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 693 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 694 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 695 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 696 pv += 16; 697 } 698 /* invert diagonal block */ 699 w = ba + 16*diag_offset[i]; 700 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 701 } 702 703 PetscFree(rtmp); 704 C->factor = FACTOR_LU; 705 C->assembled = PETSC_TRUE; 706 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 707 PetscFunctionReturn(0); 708 } 709 710 /* ------------------------------------------------------------*/ 711 /* 712 Version for when blocks are 3 by 3 713 */ 714 #undef __FUNC__ 715 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 716 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 717 { 718 Mat C = *B; 719 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 720 IS isrow = b->row, isicol = b->icol; 721 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 722 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 723 int *diag_offset = b->diag,idx; 724 register int *pj; 725 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 726 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 727 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 728 Scalar *ba = b->a,*aa = a->a; 729 730 PetscFunctionBegin; 731 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 732 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 733 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 734 735 for ( i=0; i<n; i++ ) { 736 nz = bi[i+1] - bi[i]; 737 ajtmp = bj + bi[i]; 738 for ( j=0; j<nz; j++ ) { 739 x = rtmp + 9*ajtmp[j]; 740 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 741 } 742 /* load in initial (unfactored row) */ 743 idx = r[i]; 744 nz = ai[idx+1] - ai[idx]; 745 ajtmpold = aj + ai[idx]; 746 v = aa + 9*ai[idx]; 747 for ( j=0; j<nz; j++ ) { 748 x = rtmp + 9*ic[ajtmpold[j]]; 749 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 750 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 751 v += 9; 752 } 753 row = *ajtmp++; 754 while (row < i) { 755 pc = rtmp + 9*row; 756 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 757 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 758 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 759 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 760 pv = ba + 9*diag_offset[row]; 761 pj = bj + diag_offset[row] + 1; 762 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 763 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 764 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 765 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 766 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 767 768 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 769 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 770 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 771 772 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 773 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 774 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 775 nz = bi[row+1] - diag_offset[row] - 1; 776 pv += 9; 777 for (j=0; j<nz; j++) { 778 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 779 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 780 x = rtmp + 9*pj[j]; 781 x[0] -= m1*x1 + m4*x2 + m7*x3; 782 x[1] -= m2*x1 + m5*x2 + m8*x3; 783 x[2] -= m3*x1 + m6*x2 + m9*x3; 784 785 x[3] -= m1*x4 + m4*x5 + m7*x6; 786 x[4] -= m2*x4 + m5*x5 + m8*x6; 787 x[5] -= m3*x4 + m6*x5 + m9*x6; 788 789 x[6] -= m1*x7 + m4*x8 + m7*x9; 790 x[7] -= m2*x7 + m5*x8 + m8*x9; 791 x[8] -= m3*x7 + m6*x8 + m9*x9; 792 pv += 9; 793 } 794 PLogFlops(54*nz+36); 795 } 796 row = *ajtmp++; 797 } 798 /* finished row so stick it into b->a */ 799 pv = ba + 9*bi[i]; 800 pj = bj + bi[i]; 801 nz = bi[i+1] - bi[i]; 802 for ( j=0; j<nz; j++ ) { 803 x = rtmp + 9*pj[j]; 804 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 805 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 806 pv += 9; 807 } 808 /* invert diagonal block */ 809 w = ba + 9*diag_offset[i]; 810 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 811 } 812 813 PetscFree(rtmp); 814 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 815 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 816 C->factor = FACTOR_LU; 817 C->assembled = PETSC_TRUE; 818 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 819 PetscFunctionReturn(0); 820 } 821 822 /* ------------------------------------------------------------*/ 823 /* 824 Version for when blocks are 2 by 2 825 */ 826 #undef __FUNC__ 827 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 828 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 829 { 830 Mat C = *B; 831 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 832 IS isrow = b->row, isicol = b->icol; 833 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 834 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 835 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 836 register int *pj; 837 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 838 Scalar p1,p2,p3,p4,v_work[2]; 839 Scalar *ba = b->a,*aa = a->a; 840 841 PetscFunctionBegin; 842 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 843 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 844 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 845 846 for ( i=0; i<n; i++ ) { 847 nz = bi[i+1] - bi[i]; 848 ajtmp = bj + bi[i]; 849 for ( j=0; j<nz; j++ ) { 850 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 851 } 852 /* load in initial (unfactored row) */ 853 idx = r[i]; 854 nz = ai[idx+1] - ai[idx]; 855 ajtmpold = aj + ai[idx]; 856 v = aa + 4*ai[idx]; 857 for ( j=0; j<nz; j++ ) { 858 x = rtmp+4*ic[ajtmpold[j]]; 859 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 860 v += 4; 861 } 862 row = *ajtmp++; 863 while (row < i) { 864 pc = rtmp + 4*row; 865 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 866 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 867 pv = ba + 4*diag_offset[row]; 868 pj = bj + diag_offset[row] + 1; 869 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 870 pc[0] = m1 = p1*x1 + p3*x2; 871 pc[1] = m2 = p2*x1 + p4*x2; 872 pc[2] = m3 = p1*x3 + p3*x4; 873 pc[3] = m4 = p2*x3 + p4*x4; 874 nz = bi[row+1] - diag_offset[row] - 1; 875 pv += 4; 876 for (j=0; j<nz; j++) { 877 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 878 x = rtmp + 4*pj[j]; 879 x[0] -= m1*x1 + m3*x2; 880 x[1] -= m2*x1 + m4*x2; 881 x[2] -= m1*x3 + m3*x4; 882 x[3] -= m2*x3 + m4*x4; 883 pv += 4; 884 } 885 PLogFlops(16*nz+12); 886 } 887 row = *ajtmp++; 888 } 889 /* finished row so stick it into b->a */ 890 pv = ba + 4*bi[i]; 891 pj = bj + bi[i]; 892 nz = bi[i+1] - bi[i]; 893 for ( j=0; j<nz; j++ ) { 894 x = rtmp+4*pj[j]; 895 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 896 pv += 4; 897 } 898 /* invert diagonal block */ 899 w = ba + 4*diag_offset[i]; 900 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 901 } 902 903 PetscFree(rtmp); 904 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 905 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 906 C->factor = FACTOR_LU; 907 C->assembled = PETSC_TRUE; 908 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 909 PetscFunctionReturn(0); 910 } 911 912 /* ----------------------------------------------------------- */ 913 /* 914 Version for when blocks are 1 by 1. 915 */ 916 #undef __FUNC__ 917 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 918 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 919 { 920 Mat C = *B; 921 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 922 IS isrow = b->row, isicol = b->icol; 923 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 924 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 925 int *diag_offset = b->diag,diag; 926 register int *pj; 927 register Scalar *pv,*v,*rtmp,multiplier,*pc; 928 Scalar *ba = b->a,*aa = a->a; 929 930 PetscFunctionBegin; 931 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 932 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 933 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 934 935 for ( i=0; i<n; i++ ) { 936 nz = bi[i+1] - bi[i]; 937 ajtmp = bj + bi[i]; 938 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 939 940 /* load in initial (unfactored row) */ 941 nz = ai[r[i]+1] - ai[r[i]]; 942 ajtmpold = aj + ai[r[i]]; 943 v = aa + ai[r[i]]; 944 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 945 946 row = *ajtmp++; 947 while (row < i) { 948 pc = rtmp + row; 949 if (*pc != 0.0) { 950 pv = ba + diag_offset[row]; 951 pj = bj + diag_offset[row] + 1; 952 multiplier = *pc * *pv++; 953 *pc = multiplier; 954 nz = bi[row+1] - diag_offset[row] - 1; 955 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 956 PLogFlops(1+2*nz); 957 } 958 row = *ajtmp++; 959 } 960 /* finished row so stick it into b->a */ 961 pv = ba + bi[i]; 962 pj = bj + bi[i]; 963 nz = bi[i+1] - bi[i]; 964 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 965 diag = diag_offset[i] - bi[i]; 966 /* check pivot entry for current row */ 967 if (pv[diag] == 0.0) { 968 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 969 } 970 pv[diag] = 1.0/pv[diag]; 971 } 972 973 PetscFree(rtmp); 974 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 975 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 976 C->factor = FACTOR_LU; 977 C->assembled = PETSC_TRUE; 978 PLogFlops(b->n); 979 PetscFunctionReturn(0); 980 } 981 982 /* ----------------------------------------------------------- */ 983 #undef __FUNC__ 984 #define __FUNC__ "MatLUFactor_SeqBAIJ" 985 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 986 { 987 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 988 int ierr; 989 Mat C; 990 PetscOps *Abops; 991 struct _MatOps *Aops; 992 993 PetscFunctionBegin; 994 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 995 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 996 997 /* free all the data structures from mat */ 998 PetscFree(mat->a); 999 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 1000 if (mat->diag) PetscFree(mat->diag); 1001 if (mat->ilen) PetscFree(mat->ilen); 1002 if (mat->imax) PetscFree(mat->imax); 1003 if (mat->solve_work) PetscFree(mat->solve_work); 1004 if (mat->mult_work) PetscFree(mat->mult_work); 1005 PetscFree(mat); 1006 1007 /* 1008 This is horrible, horrible code. We need to keep the 1009 A pointers for the bops and ops but copy everything 1010 else from C. 1011 */ 1012 Abops = A->bops; 1013 Aops = A->ops; 1014 PetscMemcpy(A,C,sizeof(struct _p_Mat)); 1015 A->bops = Abops; 1016 A->ops = Aops; 1017 A->qlist = 0; 1018 1019 PetscHeaderDestroy(C); 1020 PetscFunctionReturn(0); 1021 } 1022