1 2 3 #ifndef lint 4 static char vcid[] = "$Id: baijfact.c,v 1.48 1997/07/07 19:35:27 bsmith Exp bsmith $"; 5 #endif 6 /* 7 Factorization code for BAIJ format. 8 */ 9 10 #include "src/mat/impls/baij/seq/baij.h" 11 #include "src/vec/vecimpl.h" 12 #include "src/inline/ilu.h" 13 14 15 /* 16 The symbolic factorization code is identical to that for AIJ format, 17 except for very small changes since this is now a SeqBAIJ datastructure. 18 NOT good code reuse. 19 */ 20 #undef __FUNC__ 21 #define __FUNC__ "MatLUFactorSymbolic_SeqBAIJ" 22 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 23 { 24 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 25 IS isicol; 26 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 27 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 28 int *idnew, idx, row,m,fm, nnz, nzi,realloc = 0,nzbd,*im; 29 30 PetscValidHeaderSpecific(isrow,IS_COOKIE); 31 PetscValidHeaderSpecific(iscol,IS_COOKIE); 32 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 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 /* allocate a longer ajnew */ 91 int maxadd; 92 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 93 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 94 jmax += maxadd; 95 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 96 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 97 PetscFree(ajnew); 98 ajnew = ajtmp; 99 realloc++; /* count how many times we realloc */ 100 } 101 ajtmp = ajnew + ainew[i]; 102 fm = fill[n]; 103 nzi = 0; 104 im[i] = nnz; 105 while (nnz--) { 106 if (fm < i) nzi++; 107 *ajtmp++ = fm; 108 fm = fill[fm]; 109 } 110 idnew[i] = ainew[i] + nzi; 111 } 112 113 if (ai[n] != 0) { 114 double af = ((double)ainew[n])/((double)ai[n]); 115 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 116 realloc,f,af); 117 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Run with -pc_lu_fill %g or use \n",af); 118 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:PCLUSetFill(pc,%g);\n",af); 119 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:for best performance.\n"); 120 } else { 121 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Empty matrix.\n"); 122 } 123 124 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 125 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 126 127 PetscFree(fill); 128 129 /* put together the new matrix */ 130 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 131 PLogObjectParent(*B,isicol); 132 ierr = ISDestroy(isicol); CHKERRQ(ierr); 133 b = (Mat_SeqBAIJ *) (*B)->data; 134 PetscFree(b->imax); 135 b->singlemalloc = 0; 136 /* the next line frees the default space generated by the Create() */ 137 PetscFree(b->a); PetscFree(b->ilen); 138 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 139 b->j = ajnew; 140 b->i = ainew; 141 b->diag = idnew; 142 b->ilen = 0; 143 b->imax = 0; 144 b->row = isrow; 145 b->col = iscol; 146 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 147 CHKPTRQ(b->solve_work); 148 /* In b structure: Free imax, ilen, old a, old j. 149 Allocate idnew, solve_work, new a, new j */ 150 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 151 b->maxnz = b->nz = ainew[n]; 152 153 (*B)->info.factor_mallocs = realloc; 154 (*B)->info.fill_ratio_given = f; 155 if (ai[i] != 0) { 156 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 157 } else { 158 (*B)->info.fill_ratio_needed = 0.0; 159 } 160 161 162 return 0; 163 } 164 165 /* ----------------------------------------------------------- */ 166 #undef __FUNC__ 167 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 168 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 169 { 170 Mat C = *B; 171 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 172 IS iscol = b->col, isrow = b->row, isicol; 173 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 174 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 175 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 176 Scalar *ba = b->a,*aa = a->a; 177 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 178 register int *pj; 179 180 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 181 PLogObjectParent(*B,isicol); 182 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 183 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 184 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 185 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 186 /* generate work space needed by dense LU factorization */ 187 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 188 CHKPTRQ(v_work); 189 multiplier = v_work + bs; 190 v_pivots = (int *) (multiplier + bs2); 191 192 /* flops in while loop */ 193 bslog = 2*bs*bs2; 194 195 for ( i=0; i<n; i++ ) { 196 nz = bi[i+1] - bi[i]; 197 ajtmp = bj + bi[i]; 198 for ( j=0; j<nz; j++ ) { 199 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 200 } 201 /* load in initial (unfactored row) */ 202 nz = ai[r[i]+1] - ai[r[i]]; 203 ajtmpold = aj + ai[r[i]]; 204 v = aa + bs2*ai[r[i]]; 205 for ( j=0; j<nz; j++ ) { 206 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 207 } 208 row = *ajtmp++; 209 while (row < i) { 210 pc = rtmp + bs2*row; 211 /* if (*pc) { */ 212 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 213 if (flg) { 214 pv = ba + bs2*diag_offset[row]; 215 pj = bj + diag_offset[row] + 1; 216 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 217 nz = bi[row+1] - diag_offset[row] - 1; 218 pv += bs2; 219 for (j=0; j<nz; j++) { 220 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 221 } 222 PLogFlops(bslog*(nz+1)-bs); 223 } 224 row = *ajtmp++; 225 } 226 /* finished row so stick it into b->a */ 227 pv = ba + bs2*bi[i]; 228 pj = bj + bi[i]; 229 nz = bi[i+1] - bi[i]; 230 for ( j=0; j<nz; j++ ) { 231 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 232 } 233 diag = diag_offset[i] - bi[i]; 234 /* invert diagonal block */ 235 w = pv + bs2*diag; 236 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 237 } 238 239 PetscFree(rtmp); PetscFree(v_work); 240 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 241 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 242 ierr = ISDestroy(isicol); CHKERRQ(ierr); 243 C->factor = FACTOR_LU; 244 C->assembled = PETSC_TRUE; 245 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 246 return 0; 247 } 248 /* ------------------------------------------------------------*/ 249 /* 250 Version for when blocks are 5 by 5 251 */ 252 #undef __FUNC__ 253 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 254 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 255 { 256 Mat C = *B; 257 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 258 IS iscol = b->col, isrow = b->row, isicol; 259 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 260 int *ajtmpold, *ajtmp, nz, row; 261 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 262 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 263 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 264 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 265 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 266 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 267 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 268 Scalar *ba = b->a,*aa = a->a; 269 register int *pj; 270 271 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 272 PLogObjectParent(*B,isicol); 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 ierr = ISDestroy(isicol); CHKERRQ(ierr); 423 C->factor = FACTOR_LU; 424 C->assembled = PETSC_TRUE; 425 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 426 return 0; 427 } 428 429 /* ------------------------------------------------------------*/ 430 /* 431 Version for when blocks are 4 by 4 432 */ 433 #undef __FUNC__ 434 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 435 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 436 { 437 Mat C = *B; 438 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 439 IS iscol = b->col, isrow = b->row, isicol; 440 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 441 int *ajtmpold, *ajtmp, nz, row; 442 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 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 register int *pj; 450 451 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 452 PLogObjectParent(*B,isicol); 453 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 454 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 455 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 456 457 for ( i=0; i<n; i++ ) { 458 nz = bi[i+1] - bi[i]; 459 ajtmp = bj + bi[i]; 460 for ( j=0; j<nz; j++ ) { 461 x = rtmp+16*ajtmp[j]; 462 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 463 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 464 } 465 /* load in initial (unfactored row) */ 466 idx = r[i]; 467 nz = ai[idx+1] - ai[idx]; 468 ajtmpold = aj + ai[idx]; 469 v = aa + 16*ai[idx]; 470 for ( j=0; j<nz; j++ ) { 471 x = rtmp+16*ic[ajtmpold[j]]; 472 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 473 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 474 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 475 x[14] = v[14]; x[15] = v[15]; 476 v += 16; 477 } 478 row = *ajtmp++; 479 while (row < i) { 480 pc = rtmp + 16*row; 481 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 482 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 483 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 484 p15 = pc[14]; p16 = pc[15]; 485 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 486 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 487 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 488 || p16 != 0.0) { 489 pv = ba + 16*diag_offset[row]; 490 pj = bj + diag_offset[row] + 1; 491 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 492 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 493 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 494 x15 = pv[14]; x16 = pv[15]; 495 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 496 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 497 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 498 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 499 500 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 501 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 502 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 503 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 504 505 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 506 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 507 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 508 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 509 510 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 511 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 512 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 513 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 514 515 nz = bi[row+1] - diag_offset[row] - 1; 516 pv += 16; 517 for (j=0; j<nz; j++) { 518 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 519 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 520 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 521 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 522 x = rtmp + 16*pj[j]; 523 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 524 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 525 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 526 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 527 528 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 529 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 530 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 531 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 532 533 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 534 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 535 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 536 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 537 538 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 539 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 540 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 541 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 542 543 pv += 16; 544 } 545 PLogFlops(128*nz+112); 546 } 547 row = *ajtmp++; 548 } 549 /* finished row so stick it into b->a */ 550 pv = ba + 16*bi[i]; 551 pj = bj + bi[i]; 552 nz = bi[i+1] - bi[i]; 553 for ( j=0; j<nz; j++ ) { 554 x = rtmp+16*pj[j]; 555 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 556 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 557 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 558 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 559 pv += 16; 560 } 561 /* invert diagonal block */ 562 w = ba + 16*diag_offset[i]; 563 ierr = Kernel_A_gets_inverse_A_4(w); CHKERRQ(ierr); 564 } 565 566 PetscFree(rtmp); 567 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 568 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 569 ierr = ISDestroy(isicol); CHKERRQ(ierr); 570 C->factor = FACTOR_LU; 571 C->assembled = PETSC_TRUE; 572 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 573 return 0; 574 } 575 /* ------------------------------------------------------------*/ 576 /* 577 Version for when blocks are 3 by 3 578 */ 579 #undef __FUNC__ 580 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 581 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 582 { 583 Mat C = *B; 584 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 585 IS iscol = b->col, isrow = b->row, isicol; 586 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 587 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 588 int *diag_offset = b->diag,idx; 589 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 590 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 591 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 592 Scalar *ba = b->a,*aa = a->a; 593 register int *pj; 594 595 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 596 PLogObjectParent(*B,isicol); 597 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 598 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 599 rtmp = (Scalar *) PetscMalloc(9*(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 + 9*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 } 608 /* load in initial (unfactored row) */ 609 idx = r[i]; 610 nz = ai[idx+1] - ai[idx]; 611 ajtmpold = aj + ai[idx]; 612 v = aa + 9*ai[idx]; 613 for ( j=0; j<nz; j++ ) { 614 x = rtmp + 9*ic[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 v += 9; 618 } 619 row = *ajtmp++; 620 while (row < i) { 621 pc = rtmp + 9*row; 622 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 623 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 624 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 625 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 626 pv = ba + 9*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 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 631 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 632 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 633 634 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 635 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 636 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 637 638 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 639 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 640 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 641 nz = bi[row+1] - diag_offset[row] - 1; 642 pv += 9; 643 for (j=0; j<nz; j++) { 644 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 645 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 646 x = rtmp + 9*pj[j]; 647 x[0] -= m1*x1 + m4*x2 + m7*x3; 648 x[1] -= m2*x1 + m5*x2 + m8*x3; 649 x[2] -= m3*x1 + m6*x2 + m9*x3; 650 651 x[3] -= m1*x4 + m4*x5 + m7*x6; 652 x[4] -= m2*x4 + m5*x5 + m8*x6; 653 x[5] -= m3*x4 + m6*x5 + m9*x6; 654 655 x[6] -= m1*x7 + m4*x8 + m7*x9; 656 x[7] -= m2*x7 + m5*x8 + m8*x9; 657 x[8] -= m3*x7 + m6*x8 + m9*x9; 658 pv += 9; 659 } 660 PLogFlops(54*nz+36); 661 } 662 row = *ajtmp++; 663 } 664 /* finished row so stick it into b->a */ 665 pv = ba + 9*bi[i]; 666 pj = bj + bi[i]; 667 nz = bi[i+1] - bi[i]; 668 for ( j=0; j<nz; j++ ) { 669 x = rtmp + 9*pj[j]; 670 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 671 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 672 pv += 9; 673 } 674 /* invert diagonal block */ 675 w = ba + 9*diag_offset[i]; 676 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 677 } 678 679 PetscFree(rtmp); 680 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 681 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 682 ierr = ISDestroy(isicol); CHKERRQ(ierr); 683 C->factor = FACTOR_LU; 684 C->assembled = PETSC_TRUE; 685 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 686 return 0; 687 } 688 689 /* ------------------------------------------------------------*/ 690 /* 691 Version for when blocks are 2 by 2 692 */ 693 #undef __FUNC__ 694 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 695 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 696 { 697 Mat C = *B; 698 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 699 IS iscol = b->col, isrow = b->row, isicol; 700 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 701 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 702 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 703 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 704 Scalar p1,p2,p3,p4,v_work[2]; 705 Scalar *ba = b->a,*aa = a->a; 706 register int *pj; 707 708 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 709 PLogObjectParent(*B,isicol); 710 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 711 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 712 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 713 714 for ( i=0; i<n; i++ ) { 715 nz = bi[i+1] - bi[i]; 716 ajtmp = bj + bi[i]; 717 for ( j=0; j<nz; j++ ) { 718 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 719 } 720 /* load in initial (unfactored row) */ 721 idx = r[i]; 722 nz = ai[idx+1] - ai[idx]; 723 ajtmpold = aj + ai[idx]; 724 v = aa + 4*ai[idx]; 725 for ( j=0; j<nz; j++ ) { 726 x = rtmp+4*ic[ajtmpold[j]]; 727 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 728 v += 4; 729 } 730 row = *ajtmp++; 731 while (row < i) { 732 pc = rtmp + 4*row; 733 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 734 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 735 pv = ba + 4*diag_offset[row]; 736 pj = bj + diag_offset[row] + 1; 737 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 738 pc[0] = m1 = p1*x1 + p3*x2; 739 pc[1] = m2 = p2*x1 + p4*x2; 740 pc[2] = m3 = p1*x3 + p3*x4; 741 pc[3] = m4 = p2*x3 + p4*x4; 742 nz = bi[row+1] - diag_offset[row] - 1; 743 pv += 4; 744 for (j=0; j<nz; j++) { 745 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 746 x = rtmp + 4*pj[j]; 747 x[0] -= m1*x1 + m3*x2; 748 x[1] -= m2*x1 + m4*x2; 749 x[2] -= m1*x3 + m3*x4; 750 x[3] -= m2*x3 + m4*x4; 751 pv += 4; 752 } 753 PLogFlops(16*nz+12); 754 } 755 row = *ajtmp++; 756 } 757 /* finished row so stick it into b->a */ 758 pv = ba + 4*bi[i]; 759 pj = bj + bi[i]; 760 nz = bi[i+1] - bi[i]; 761 for ( j=0; j<nz; j++ ) { 762 x = rtmp+4*pj[j]; 763 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 764 pv += 4; 765 } 766 /* invert diagonal block */ 767 w = ba + 4*diag_offset[i]; 768 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 769 } 770 771 PetscFree(rtmp); 772 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 773 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 774 ierr = ISDestroy(isicol); CHKERRQ(ierr); 775 C->factor = FACTOR_LU; 776 C->assembled = PETSC_TRUE; 777 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 778 return 0; 779 } 780 781 /* ----------------------------------------------------------- */ 782 /* 783 Version for when blocks are 1 by 1. 784 */ 785 #undef __FUNC__ 786 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 787 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 788 { 789 Mat C = *B; 790 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 791 IS iscol = b->col, isrow = b->row, isicol; 792 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 793 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 794 int *diag_offset = b->diag,diag; 795 register Scalar *pv,*v,*rtmp,multiplier,*pc; 796 Scalar *ba = b->a,*aa = a->a; 797 register int *pj; 798 799 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 800 PLogObjectParent(*B,isicol); 801 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 802 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 803 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 804 805 for ( i=0; i<n; i++ ) { 806 nz = bi[i+1] - bi[i]; 807 ajtmp = bj + bi[i]; 808 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 809 810 /* load in initial (unfactored row) */ 811 nz = ai[r[i]+1] - ai[r[i]]; 812 ajtmpold = aj + ai[r[i]]; 813 v = aa + ai[r[i]]; 814 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 815 816 row = *ajtmp++; 817 while (row < i) { 818 pc = rtmp + row; 819 if (*pc != 0.0) { 820 pv = ba + diag_offset[row]; 821 pj = bj + diag_offset[row] + 1; 822 multiplier = *pc * *pv++; 823 *pc = multiplier; 824 nz = bi[row+1] - diag_offset[row] - 1; 825 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 826 PLogFlops(1+2*nz); 827 } 828 row = *ajtmp++; 829 } 830 /* finished row so stick it into b->a */ 831 pv = ba + bi[i]; 832 pj = bj + bi[i]; 833 nz = bi[i+1] - bi[i]; 834 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 835 diag = diag_offset[i] - bi[i]; 836 /* check pivot entry for current row */ 837 if (pv[diag] == 0.0) { 838 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 839 } 840 pv[diag] = 1.0/pv[diag]; 841 } 842 843 PetscFree(rtmp); 844 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 845 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 846 ierr = ISDestroy(isicol); CHKERRQ(ierr); 847 C->factor = FACTOR_LU; 848 C->assembled = PETSC_TRUE; 849 PLogFlops(b->n); 850 return 0; 851 } 852 853 /* ----------------------------------------------------------- */ 854 #undef __FUNC__ 855 #define __FUNC__ "MatLUFactor_SeqBAIJ" 856 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 857 { 858 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 859 int ierr; 860 Mat C; 861 862 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 863 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 864 865 /* free all the data structures from mat */ 866 PetscFree(mat->a); 867 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 868 if (mat->diag) PetscFree(mat->diag); 869 if (mat->ilen) PetscFree(mat->ilen); 870 if (mat->imax) PetscFree(mat->imax); 871 if (mat->solve_work) PetscFree(mat->solve_work); 872 if (mat->mult_work) PetscFree(mat->mult_work); 873 PetscFree(mat); 874 875 PetscMemcpy(A,C,sizeof(struct _p_Mat)); 876 PetscHeaderDestroy(C); 877 return 0; 878 } 879 /* ----------------------------------------------------------- */ 880 #undef __FUNC__ 881 #define __FUNC__ "MatSolve_SeqBAIJ_N" 882 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 883 { 884 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 885 IS iscol=a->col,isrow=a->row; 886 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 887 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 888 Scalar *aa=a->a,*sum; 889 register Scalar *x,*b,*lsum,*tmp,*v; 890 891 VecGetArray_Fast(bb,b); 892 VecGetArray_Fast(xx,x); 893 tmp = a->solve_work; 894 895 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 896 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 897 898 /* forward solve the lower triangular */ 899 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 900 for ( i=1; i<n; i++ ) { 901 v = aa + bs2*ai[i]; 902 vi = aj + ai[i]; 903 nz = a->diag[i] - ai[i]; 904 sum = tmp + bs*i; 905 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 906 while (nz--) { 907 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 908 v += bs2; 909 } 910 } 911 /* backward solve the upper triangular */ 912 lsum = a->solve_work + a->n; 913 for ( i=n-1; i>=0; i-- ){ 914 v = aa + bs2*(a->diag[i] + 1); 915 vi = aj + a->diag[i] + 1; 916 nz = ai[i+1] - a->diag[i] - 1; 917 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 918 while (nz--) { 919 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 920 v += bs2; 921 } 922 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 923 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 924 } 925 926 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 927 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 928 VecRestoreArray_Fast(bb,b); 929 VecRestoreArray_Fast(xx,x); 930 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 931 return 0; 932 } 933 934 #undef __FUNC__ 935 #define __FUNC__ "MatSolve_SeqBAIJ_7" 936 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 937 { 938 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 939 IS iscol=a->col,isrow=a->row; 940 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 941 int *diag = a->diag; 942 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 943 register Scalar *x,*b,*tmp,*v; 944 945 VecGetArray_Fast(bb,b); 946 VecGetArray_Fast(xx,x); 947 tmp = a->solve_work; 948 949 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 950 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 951 952 /* forward solve the lower triangular */ 953 idx = 7*(*r++); 954 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 955 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 956 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 957 958 for ( i=1; i<n; i++ ) { 959 v = aa + 49*ai[i]; 960 vi = aj + ai[i]; 961 nz = diag[i] - ai[i]; 962 idx = 7*(*r++); 963 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 964 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 965 while (nz--) { 966 idx = 7*(*vi++); 967 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 968 x4 = tmp[3+idx];x5 = tmp[4+idx]; 969 x6 = tmp[5+idx];x7 = tmp[6+idx]; 970 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 971 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 972 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 973 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 974 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 975 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 976 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 977 v += 49; 978 } 979 idx = 7*i; 980 tmp[idx] = sum1;tmp[1+idx] = sum2; 981 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 982 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 983 } 984 /* backward solve the upper triangular */ 985 for ( i=n-1; i>=0; i-- ){ 986 v = aa + 49*diag[i] + 49; 987 vi = aj + diag[i] + 1; 988 nz = ai[i+1] - diag[i] - 1; 989 idt = 7*i; 990 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 991 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 992 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 993 while (nz--) { 994 idx = 7*(*vi++); 995 x1 = tmp[idx]; x2 = tmp[1+idx]; 996 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 997 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 998 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 999 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 1000 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 1001 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 1002 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 1003 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 1004 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 1005 v += 49; 1006 } 1007 idc = 7*(*c--); 1008 v = aa + 49*diag[i]; 1009 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 1010 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 1011 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 1012 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 1013 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 1014 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1015 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1016 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1017 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1018 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1019 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1020 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1021 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1022 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1023 } 1024 1025 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1026 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1027 VecRestoreArray_Fast(bb,b); 1028 VecRestoreArray_Fast(xx,x); 1029 PLogFlops(2*49*(a->nz) - a->n); 1030 return 0; 1031 } 1032 1033 #undef __FUNC__ 1034 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1035 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1036 { 1037 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1038 IS iscol=a->col,isrow=a->row; 1039 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1040 int *diag = a->diag; 1041 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1042 register Scalar *x,*b,*tmp,*v; 1043 1044 VecGetArray_Fast(bb,b); 1045 VecGetArray_Fast(xx,x); 1046 tmp = a->solve_work; 1047 1048 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1049 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1050 1051 /* forward solve the lower triangular */ 1052 idx = 5*(*r++); 1053 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1054 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1055 for ( i=1; i<n; i++ ) { 1056 v = aa + 25*ai[i]; 1057 vi = aj + ai[i]; 1058 nz = diag[i] - ai[i]; 1059 idx = 5*(*r++); 1060 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1061 sum5 = b[4+idx]; 1062 while (nz--) { 1063 idx = 5*(*vi++); 1064 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1065 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1066 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1067 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1068 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1069 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1070 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1071 v += 25; 1072 } 1073 idx = 5*i; 1074 tmp[idx] = sum1;tmp[1+idx] = sum2; 1075 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1076 } 1077 /* backward solve the upper triangular */ 1078 for ( i=n-1; i>=0; i-- ){ 1079 v = aa + 25*diag[i] + 25; 1080 vi = aj + diag[i] + 1; 1081 nz = ai[i+1] - diag[i] - 1; 1082 idt = 5*i; 1083 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1084 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1085 while (nz--) { 1086 idx = 5*(*vi++); 1087 x1 = tmp[idx]; x2 = tmp[1+idx]; 1088 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1089 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1090 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1091 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1092 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1093 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1094 v += 25; 1095 } 1096 idc = 5*(*c--); 1097 v = aa + 25*diag[i]; 1098 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1099 v[15]*sum4+v[20]*sum5; 1100 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1101 v[16]*sum4+v[21]*sum5; 1102 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1103 v[17]*sum4+v[22]*sum5; 1104 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1105 v[18]*sum4+v[23]*sum5; 1106 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1107 v[19]*sum4+v[24]*sum5; 1108 } 1109 1110 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1111 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1112 VecRestoreArray_Fast(bb,b); 1113 VecRestoreArray_Fast(xx,x); 1114 PLogFlops(2*25*(a->nz) - a->n); 1115 return 0; 1116 } 1117 1118 #undef __FUNC__ 1119 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1120 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1121 { 1122 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1123 IS iscol=a->col,isrow=a->row; 1124 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1125 int *diag = a->diag; 1126 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1127 register Scalar *x,*b,*tmp,*v; 1128 1129 VecGetArray_Fast(bb,b); 1130 VecGetArray_Fast(xx,x); 1131 tmp = a->solve_work; 1132 1133 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1134 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1135 1136 /* forward solve the lower triangular */ 1137 idx = 4*(*r++); 1138 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1139 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1140 for ( i=1; i<n; i++ ) { 1141 v = aa + 16*ai[i]; 1142 vi = aj + ai[i]; 1143 nz = diag[i] - ai[i]; 1144 idx = 4*(*r++); 1145 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1146 while (nz--) { 1147 idx = 4*(*vi++); 1148 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1149 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1150 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1151 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1152 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1153 v += 16; 1154 } 1155 idx = 4*i; 1156 tmp[idx] = sum1;tmp[1+idx] = sum2; 1157 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1158 } 1159 /* backward solve the upper triangular */ 1160 for ( i=n-1; i>=0; i-- ){ 1161 v = aa + 16*diag[i] + 16; 1162 vi = aj + diag[i] + 1; 1163 nz = ai[i+1] - diag[i] - 1; 1164 idt = 4*i; 1165 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1166 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1167 while (nz--) { 1168 idx = 4*(*vi++); 1169 x1 = tmp[idx]; x2 = tmp[1+idx]; 1170 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1171 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1172 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1173 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1174 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1175 v += 16; 1176 } 1177 idc = 4*(*c--); 1178 v = aa + 16*diag[i]; 1179 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1180 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1181 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1182 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1183 } 1184 1185 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1186 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1187 VecRestoreArray_Fast(bb,b); 1188 VecRestoreArray_Fast(xx,x); 1189 PLogFlops(2*16*(a->nz) - a->n); 1190 return 0; 1191 } 1192 1193 /* 1194 Special case where the matrix was ILU(0) factored in the natural 1195 ordering. This eliminates the need for the column and row permutation. 1196 */ 1197 #undef __FUNC__ 1198 #define __FUNC__ "MatSolve_SeqBAIJ_4_NaturalOrdering" 1199 int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat A,Vec bb,Vec xx) 1200 { 1201 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1202 int i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt; 1203 int *diag = a->diag,jdx; 1204 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1205 register Scalar *x,*b,*v; 1206 1207 VecGetArray_Fast(bb,b); 1208 VecGetArray_Fast(xx,x); 1209 1210 /* forward solve the lower triangular */ 1211 idx = 0; 1212 x[0] = b[idx]; x[1] = b[1+idx]; x[2] = b[2+idx]; x[3] = b[3+idx]; 1213 for ( i=1; i<n; i++ ) { 1214 v = aa + 16*ai[i]; 1215 vi = aj + ai[i]; 1216 nz = diag[i] - ai[i]; 1217 idx = 4*i; 1218 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1219 while (nz--) { 1220 jdx = 4*(*vi++); 1221 x1 = x[jdx];x2 = x[1+jdx];x3 = x[2+jdx];x4 = x[3+jdx]; 1222 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1223 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1224 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1225 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1226 v += 16; 1227 } 1228 x[idx] = sum1; 1229 x[1+idx] = sum2; 1230 x[2+idx] = sum3; 1231 x[3+idx] = sum4; 1232 } 1233 /* backward solve the upper triangular */ 1234 for ( i=n-1; i>=0; i-- ){ 1235 v = aa + 16*diag[i] + 16; 1236 vi = aj + diag[i] + 1; 1237 nz = ai[i+1] - diag[i] - 1; 1238 idt = 4*i; 1239 sum1 = x[idt]; sum2 = x[1+idt]; 1240 sum3 = x[2+idt];sum4 = x[3+idt]; 1241 while (nz--) { 1242 idx = 4*(*vi++); 1243 x1 = x[idx]; x2 = x[1+idx];x3 = x[2+idx]; x4 = x[3+idx]; 1244 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1245 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1246 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1247 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1248 v += 16; 1249 } 1250 v = aa + 16*diag[i]; 1251 x[idt] = v[0]*sum1 + v[4]*sum2 + v[8]*sum3 + v[12]*sum4; 1252 x[1+idt] = v[1]*sum1 + v[5]*sum2 + v[9]*sum3 + v[13]*sum4; 1253 x[2+idt] = v[2]*sum1 + v[6]*sum2 + v[10]*sum3 + v[14]*sum4; 1254 x[3+idt] = v[3]*sum1 + v[7]*sum2 + v[11]*sum3 + v[15]*sum4; 1255 } 1256 1257 VecRestoreArray_Fast(bb,b); 1258 VecRestoreArray_Fast(xx,x); 1259 PLogFlops(2*16*(a->nz) - a->n); 1260 return 0; 1261 } 1262 1263 1264 #undef __FUNC__ 1265 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1266 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1267 { 1268 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1269 IS iscol=a->col,isrow=a->row; 1270 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1271 int *diag = a->diag; 1272 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1273 register Scalar *x,*b,*tmp,*v; 1274 1275 VecGetArray_Fast(bb,b); 1276 VecGetArray_Fast(xx,x); 1277 tmp = a->solve_work; 1278 1279 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1280 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1281 1282 /* forward solve the lower triangular */ 1283 idx = 3*(*r++); 1284 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1285 for ( i=1; i<n; i++ ) { 1286 v = aa + 9*ai[i]; 1287 vi = aj + ai[i]; 1288 nz = diag[i] - ai[i]; 1289 idx = 3*(*r++); 1290 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1291 while (nz--) { 1292 idx = 3*(*vi++); 1293 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1294 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1295 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1296 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1297 v += 9; 1298 } 1299 idx = 3*i; 1300 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1301 } 1302 /* backward solve the upper triangular */ 1303 for ( i=n-1; i>=0; i-- ){ 1304 v = aa + 9*diag[i] + 9; 1305 vi = aj + diag[i] + 1; 1306 nz = ai[i+1] - diag[i] - 1; 1307 idt = 3*i; 1308 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1309 while (nz--) { 1310 idx = 3*(*vi++); 1311 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1312 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1313 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1314 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1315 v += 9; 1316 } 1317 idc = 3*(*c--); 1318 v = aa + 9*diag[i]; 1319 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1320 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1321 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1322 } 1323 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1324 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1325 VecRestoreArray_Fast(bb,b); 1326 VecRestoreArray_Fast(xx,x); 1327 PLogFlops(2*9*(a->nz) - a->n); 1328 return 0; 1329 } 1330 1331 #undef __FUNC__ 1332 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1333 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1334 { 1335 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1336 IS iscol=a->col,isrow=a->row; 1337 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1338 int *diag = a->diag; 1339 Scalar *aa=a->a,sum1,sum2,x1,x2; 1340 register Scalar *x,*b,*tmp,*v; 1341 1342 VecGetArray_Fast(bb,b); 1343 VecGetArray_Fast(xx,x); 1344 tmp = a->solve_work; 1345 1346 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1347 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1348 1349 /* forward solve the lower triangular */ 1350 idx = 2*(*r++); 1351 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1352 for ( i=1; i<n; i++ ) { 1353 v = aa + 4*ai[i]; 1354 vi = aj + ai[i]; 1355 nz = diag[i] - ai[i]; 1356 idx = 2*(*r++); 1357 sum1 = b[idx]; sum2 = b[1+idx]; 1358 while (nz--) { 1359 idx = 2*(*vi++); 1360 x1 = tmp[idx]; x2 = tmp[1+idx]; 1361 sum1 -= v[0]*x1 + v[2]*x2; 1362 sum2 -= v[1]*x1 + v[3]*x2; 1363 v += 4; 1364 } 1365 idx = 2*i; 1366 tmp[idx] = sum1; tmp[1+idx] = sum2; 1367 } 1368 /* backward solve the upper triangular */ 1369 for ( i=n-1; i>=0; i-- ){ 1370 v = aa + 4*diag[i] + 4; 1371 vi = aj + diag[i] + 1; 1372 nz = ai[i+1] - diag[i] - 1; 1373 idt = 2*i; 1374 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1375 while (nz--) { 1376 idx = 2*(*vi++); 1377 x1 = tmp[idx]; x2 = tmp[1+idx]; 1378 sum1 -= v[0]*x1 + v[2]*x2; 1379 sum2 -= v[1]*x1 + v[3]*x2; 1380 v += 4; 1381 } 1382 idc = 2*(*c--); 1383 v = aa + 4*diag[i]; 1384 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1385 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1386 } 1387 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1388 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1389 VecRestoreArray_Fast(bb,b); 1390 VecRestoreArray_Fast(xx,x); 1391 PLogFlops(2*4*(a->nz) - a->n); 1392 return 0; 1393 } 1394 1395 1396 #undef __FUNC__ 1397 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1398 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1399 { 1400 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1401 IS iscol=a->col,isrow=a->row; 1402 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1403 int *diag = a->diag; 1404 Scalar *aa=a->a,sum1; 1405 register Scalar *x,*b,*tmp,*v; 1406 1407 if (!n) return 0; 1408 1409 VecGetArray_Fast(bb,b); 1410 VecGetArray_Fast(xx,x); 1411 tmp = a->solve_work; 1412 1413 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1414 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1415 1416 /* forward solve the lower triangular */ 1417 tmp[0] = b[*r++]; 1418 for ( i=1; i<n; i++ ) { 1419 v = aa + ai[i]; 1420 vi = aj + ai[i]; 1421 nz = diag[i] - ai[i]; 1422 sum1 = b[*r++]; 1423 while (nz--) { 1424 sum1 -= (*v++)*tmp[*vi++]; 1425 } 1426 tmp[i] = sum1; 1427 } 1428 /* backward solve the upper triangular */ 1429 for ( i=n-1; i>=0; i-- ){ 1430 v = aa + diag[i] + 1; 1431 vi = aj + diag[i] + 1; 1432 nz = ai[i+1] - diag[i] - 1; 1433 sum1 = tmp[i]; 1434 while (nz--) { 1435 sum1 -= (*v++)*tmp[*vi++]; 1436 } 1437 x[*c--] = tmp[i] = aa[diag[i]]*sum1; 1438 } 1439 1440 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1441 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1442 VecRestoreArray_Fast(bb,b); 1443 VecRestoreArray_Fast(xx,x); 1444 PLogFlops(2*1*(a->nz) - a->n); 1445 return 0; 1446 } 1447 1448 extern int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat,Vec,Vec); 1449 /* ----------------------------------------------------------------*/ 1450 /* 1451 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1452 except that the data structure of Mat_SeqAIJ is slightly different. 1453 Not a good example of code reuse. 1454 */ 1455 #undef __FUNC__ 1456 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1457 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1458 Mat *fact) 1459 { 1460 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1461 IS isicol; 1462 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1463 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1464 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1465 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1466 PetscTruth col_identity, row_identity; 1467 1468 /* special case that simply copies fill pattern */ 1469 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1470 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1471 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1472 if (levels == 0 && row_identity && col_identity) { 1473 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1474 (*fact)->factor = FACTOR_LU; 1475 b = (Mat_SeqBAIJ *) (*fact)->data; 1476 if (!b->diag) { 1477 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1478 } 1479 b->row = isrow; 1480 b->col = iscol; 1481 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1482 /* 1483 Blocksize 4 has a special faster solver for ILU(0) factorization 1484 with natural ordering 1485 */ 1486 if (b->bs == 4) { 1487 (*fact)->ops.solve = MatSolve_SeqBAIJ_4_NaturalOrdering; 1488 } 1489 return 0; 1490 } 1491 1492 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1493 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1494 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1495 1496 /* get new row pointers */ 1497 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1498 ainew[0] = 0; 1499 /* don't know how many column pointers are needed so estimate */ 1500 jmax = (int) (f*ai[n] + 1); 1501 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1502 /* ajfill is level of fill for each fill entry */ 1503 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1504 /* fill is a linked list of nonzeros in active row */ 1505 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1506 /* im is level for each filled value */ 1507 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1508 /* dloc is location of diagonal in factor */ 1509 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1510 dloc[0] = 0; 1511 for ( prow=0; prow<n; prow++ ) { 1512 /* first copy previous fill into linked list */ 1513 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1514 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1515 xi = aj + ai[r[prow]]; 1516 fill[n] = n; 1517 while (nz--) { 1518 fm = n; 1519 idx = ic[*xi++]; 1520 do { 1521 m = fm; 1522 fm = fill[m]; 1523 } while (fm < idx); 1524 fill[m] = idx; 1525 fill[idx] = fm; 1526 im[idx] = 0; 1527 } 1528 nzi = 0; 1529 row = fill[n]; 1530 while ( row < prow ) { 1531 incrlev = im[row] + 1; 1532 nz = dloc[row]; 1533 xi = ajnew + ainew[row] + nz; 1534 flev = ajfill + ainew[row] + nz + 1; 1535 nnz = ainew[row+1] - ainew[row] - nz - 1; 1536 if (*xi++ != row) { 1537 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot: try running with -pc_ilu_nonzeros_along_diagonal"); 1538 } 1539 fm = row; 1540 while (nnz-- > 0) { 1541 idx = *xi++; 1542 if (*flev + incrlev > levels) { 1543 flev++; 1544 continue; 1545 } 1546 do { 1547 m = fm; 1548 fm = fill[m]; 1549 } while (fm < idx); 1550 if (fm != idx) { 1551 im[idx] = *flev + incrlev; 1552 fill[m] = idx; 1553 fill[idx] = fm; 1554 fm = idx; 1555 nzf++; 1556 } 1557 else { 1558 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1559 } 1560 flev++; 1561 } 1562 row = fill[row]; 1563 nzi++; 1564 } 1565 /* copy new filled row into permanent storage */ 1566 ainew[prow+1] = ainew[prow] + nzf; 1567 if (ainew[prow+1] > jmax) { 1568 /* allocate a longer ajnew */ 1569 int maxadd; 1570 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1571 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1572 jmax += maxadd; 1573 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1574 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1575 PetscFree(ajnew); 1576 ajnew = xi; 1577 /* allocate a longer ajfill */ 1578 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1579 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1580 PetscFree(ajfill); 1581 ajfill = xi; 1582 realloc++; 1583 } 1584 xi = ajnew + ainew[prow]; 1585 flev = ajfill + ainew[prow]; 1586 dloc[prow] = nzi; 1587 fm = fill[n]; 1588 while (nzf--) { 1589 *xi++ = fm; 1590 *flev++ = im[fm]; 1591 fm = fill[fm]; 1592 } 1593 } 1594 PetscFree(ajfill); 1595 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1596 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1597 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1598 PetscFree(fill); PetscFree(im); 1599 1600 { 1601 double af = ((double)ainew[n])/((double)ai[n]); 1602 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 1603 realloc,f,af); 1604 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Run with -pc_ilu_fill %g or use \n",af); 1605 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:PCILUSetFill(pc,%g);\n",af); 1606 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:for best performance.\n"); 1607 } 1608 1609 /* put together the new matrix */ 1610 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1611 b = (Mat_SeqBAIJ *) (*fact)->data; 1612 PetscFree(b->imax); 1613 b->singlemalloc = 0; 1614 len = bs2*ainew[n]*sizeof(Scalar); 1615 /* the next line frees the default space generated by the Create() */ 1616 PetscFree(b->a); PetscFree(b->ilen); 1617 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1618 b->j = ajnew; 1619 b->i = ainew; 1620 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1621 b->diag = dloc; 1622 b->ilen = 0; 1623 b->imax = 0; 1624 b->row = isrow; 1625 b->col = iscol; 1626 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1627 CHKPTRQ(b->solve_work); 1628 /* In b structure: Free imax, ilen, old a, old j. 1629 Allocate dloc, solve_work, new a, new j */ 1630 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1631 b->maxnz = b->nz = ainew[n]; 1632 (*fact)->factor = FACTOR_LU; 1633 1634 (*fact)->info.factor_mallocs = realloc; 1635 (*fact)->info.fill_ratio_given = f; 1636 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1637 1638 return 0; 1639 } 1640 1641 1642 1643 1644