1 #ifdef PETSC_RCS_HEADER 2 static char vcid[] = "$Id: baijfact.c,v 1.49 1997/07/09 13:09:54 bsmith Exp balay $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "src/mat/impls/baij/seq/baij.h" 9 #include "src/vec/vecimpl.h" 10 #include "src/inline/ilu.h" 11 12 13 /* 14 The symbolic factorization code is identical to that for AIJ format, 15 except for very small changes since this is now a SeqBAIJ datastructure. 16 NOT good code reuse. 17 */ 18 #undef __FUNC__ 19 #define __FUNC__ "MatLUFactorSymbolic_SeqBAIJ" 20 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 21 { 22 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 23 IS isicol; 24 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 25 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 26 int *idnew, idx, row,m,fm, nnz, nzi,realloc = 0,nzbd,*im; 27 28 PetscValidHeaderSpecific(isrow,IS_COOKIE); 29 PetscValidHeaderSpecific(iscol,IS_COOKIE); 30 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 31 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 32 33 /* get new row pointers */ 34 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 35 ainew[0] = 0; 36 /* don't know how many column pointers are needed so estimate */ 37 jmax = (int) (f*ai[n] + 1); 38 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 39 /* fill is a linked list of nonzeros in active row */ 40 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 41 im = fill + n + 1; 42 /* idnew is location of diagonal in factor */ 43 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 44 idnew[0] = 0; 45 46 for ( i=0; i<n; i++ ) { 47 /* first copy previous fill into linked list */ 48 nnz = nz = ai[r[i]+1] - ai[r[i]]; 49 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 50 ajtmp = aj + ai[r[i]]; 51 fill[n] = n; 52 while (nz--) { 53 fm = n; 54 idx = ic[*ajtmp++]; 55 do { 56 m = fm; 57 fm = fill[m]; 58 } while (fm < idx); 59 fill[m] = idx; 60 fill[idx] = fm; 61 } 62 row = fill[n]; 63 while ( row < i ) { 64 ajtmp = ajnew + idnew[row] + 1; 65 nzbd = 1 + idnew[row] - ainew[row]; 66 nz = im[row] - nzbd; 67 fm = row; 68 while (nz-- > 0) { 69 idx = *ajtmp++; 70 nzbd++; 71 if (idx == i) im[row] = nzbd; 72 do { 73 m = fm; 74 fm = fill[m]; 75 } while (fm < idx); 76 if (fm != idx) { 77 fill[m] = idx; 78 fill[idx] = fm; 79 fm = idx; 80 nnz++; 81 } 82 } 83 row = fill[row]; 84 } 85 /* copy new filled row into permanent storage */ 86 ainew[i+1] = ainew[i] + nnz; 87 if (ainew[i+1] > jmax) { 88 /* allocate a longer ajnew */ 89 int maxadd; 90 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 91 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 92 jmax += maxadd; 93 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 94 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 95 PetscFree(ajnew); 96 ajnew = ajtmp; 97 realloc++; /* count how many times we realloc */ 98 } 99 ajtmp = ajnew + ainew[i]; 100 fm = fill[n]; 101 nzi = 0; 102 im[i] = nnz; 103 while (nnz--) { 104 if (fm < i) nzi++; 105 *ajtmp++ = fm; 106 fm = fill[fm]; 107 } 108 idnew[i] = ainew[i] + nzi; 109 } 110 111 if (ai[n] != 0) { 112 double af = ((double)ainew[n])/((double)ai[n]); 113 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 114 realloc,f,af); 115 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Run with -pc_lu_fill %g or use \n",af); 116 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:PCLUSetFill(pc,%g);\n",af); 117 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:for best performance.\n"); 118 } else { 119 PLogInfo(A,"Info:MatLUFactorSymbolic_SeqBAIJ:Empty matrix.\n"); 120 } 121 122 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 123 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 124 125 PetscFree(fill); 126 127 /* put together the new matrix */ 128 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 129 PLogObjectParent(*B,isicol); 130 ierr = ISDestroy(isicol); CHKERRQ(ierr); 131 b = (Mat_SeqBAIJ *) (*B)->data; 132 PetscFree(b->imax); 133 b->singlemalloc = 0; 134 /* the next line frees the default space generated by the Create() */ 135 PetscFree(b->a); PetscFree(b->ilen); 136 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 137 b->j = ajnew; 138 b->i = ainew; 139 b->diag = idnew; 140 b->ilen = 0; 141 b->imax = 0; 142 b->row = isrow; 143 b->col = iscol; 144 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 145 CHKPTRQ(b->solve_work); 146 /* In b structure: Free imax, ilen, old a, old j. 147 Allocate idnew, solve_work, new a, new j */ 148 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 149 b->maxnz = b->nz = ainew[n]; 150 151 (*B)->info.factor_mallocs = realloc; 152 (*B)->info.fill_ratio_given = f; 153 if (ai[i] != 0) { 154 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 155 } else { 156 (*B)->info.fill_ratio_needed = 0.0; 157 } 158 159 160 return 0; 161 } 162 163 /* ----------------------------------------------------------- */ 164 #undef __FUNC__ 165 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 166 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 167 { 168 Mat C = *B; 169 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 170 IS iscol = b->col, isrow = b->row, isicol; 171 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 172 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 173 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 174 Scalar *ba = b->a,*aa = a->a; 175 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 176 register int *pj; 177 178 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 179 PLogObjectParent(*B,isicol); 180 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 181 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 182 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 183 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 184 /* generate work space needed by dense LU factorization */ 185 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 186 CHKPTRQ(v_work); 187 multiplier = v_work + bs; 188 v_pivots = (int *) (multiplier + bs2); 189 190 /* flops in while loop */ 191 bslog = 2*bs*bs2; 192 193 for ( i=0; i<n; i++ ) { 194 nz = bi[i+1] - bi[i]; 195 ajtmp = bj + bi[i]; 196 for ( j=0; j<nz; j++ ) { 197 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 198 } 199 /* load in initial (unfactored row) */ 200 nz = ai[r[i]+1] - ai[r[i]]; 201 ajtmpold = aj + ai[r[i]]; 202 v = aa + bs2*ai[r[i]]; 203 for ( j=0; j<nz; j++ ) { 204 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 205 } 206 row = *ajtmp++; 207 while (row < i) { 208 pc = rtmp + bs2*row; 209 /* if (*pc) { */ 210 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 211 if (flg) { 212 pv = ba + bs2*diag_offset[row]; 213 pj = bj + diag_offset[row] + 1; 214 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 215 nz = bi[row+1] - diag_offset[row] - 1; 216 pv += bs2; 217 for (j=0; j<nz; j++) { 218 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 219 } 220 PLogFlops(bslog*(nz+1)-bs); 221 } 222 row = *ajtmp++; 223 } 224 /* finished row so stick it into b->a */ 225 pv = ba + bs2*bi[i]; 226 pj = bj + bi[i]; 227 nz = bi[i+1] - bi[i]; 228 for ( j=0; j<nz; j++ ) { 229 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 230 } 231 diag = diag_offset[i] - bi[i]; 232 /* invert diagonal block */ 233 w = pv + bs2*diag; 234 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 235 } 236 237 PetscFree(rtmp); PetscFree(v_work); 238 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 239 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 240 ierr = ISDestroy(isicol); CHKERRQ(ierr); 241 C->factor = FACTOR_LU; 242 C->assembled = PETSC_TRUE; 243 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 244 return 0; 245 } 246 /* ------------------------------------------------------------*/ 247 /* 248 Version for when blocks are 5 by 5 249 */ 250 #undef __FUNC__ 251 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 252 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 253 { 254 Mat C = *B; 255 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 256 IS iscol = b->col, isrow = b->row, isicol; 257 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 258 int *ajtmpold, *ajtmp, nz, row; 259 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 260 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 261 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 262 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 263 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 264 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 265 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 266 Scalar *ba = b->a,*aa = a->a; 267 register int *pj; 268 269 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 270 PLogObjectParent(*B,isicol); 271 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 272 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 273 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 274 275 for ( i=0; i<n; i++ ) { 276 nz = bi[i+1] - bi[i]; 277 ajtmp = bj + bi[i]; 278 for ( j=0; j<nz; j++ ) { 279 x = rtmp+25*ajtmp[j]; 280 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 281 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 282 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 283 } 284 /* load in initial (unfactored row) */ 285 idx = r[i]; 286 nz = ai[idx+1] - ai[idx]; 287 ajtmpold = aj + ai[idx]; 288 v = aa + 25*ai[idx]; 289 for ( j=0; j<nz; j++ ) { 290 x = rtmp+25*ic[ajtmpold[j]]; 291 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 292 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 293 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 294 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 295 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 296 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 297 v += 25; 298 } 299 row = *ajtmp++; 300 while (row < i) { 301 pc = rtmp + 25*row; 302 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 303 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 304 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 305 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 306 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 307 p25 = pc[24]; 308 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 309 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 310 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 311 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 312 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 313 p24 != 0.0 || p25 != 0.0) { 314 pv = ba + 25*diag_offset[row]; 315 pj = bj + diag_offset[row] + 1; 316 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 317 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 318 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 319 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 320 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 321 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 322 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 323 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 324 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 325 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 326 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 327 328 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 329 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 330 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 331 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 332 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 333 334 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 335 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 336 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 337 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 338 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 339 340 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 341 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 342 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 343 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 344 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 345 346 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 347 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 348 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 349 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 350 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 351 352 nz = bi[row+1] - diag_offset[row] - 1; 353 pv += 25; 354 for (j=0; j<nz; j++) { 355 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 356 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 357 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 358 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 359 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 360 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 361 x = rtmp + 25*pj[j]; 362 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 363 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 364 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 365 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 366 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 367 368 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 369 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 370 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 371 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 372 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 373 374 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 375 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 376 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 377 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 378 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 379 380 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 381 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 382 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 383 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 384 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 385 386 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 387 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 388 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 389 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 390 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 391 392 pv += 25; 393 } 394 PLogFlops(250*nz+225); 395 } 396 row = *ajtmp++; 397 } 398 /* finished row so stick it into b->a */ 399 pv = ba + 25*bi[i]; 400 pj = bj + bi[i]; 401 nz = bi[i+1] - bi[i]; 402 for ( j=0; j<nz; j++ ) { 403 x = rtmp+25*pj[j]; 404 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 405 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 406 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 407 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 408 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 409 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 410 pv += 25; 411 } 412 /* invert diagonal block */ 413 w = ba + 25*diag_offset[i]; 414 ierr = Kernel_A_gets_inverse_A_5(w); CHKERRQ(ierr); 415 } 416 417 PetscFree(rtmp); 418 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 419 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 420 ierr = ISDestroy(isicol); CHKERRQ(ierr); 421 C->factor = FACTOR_LU; 422 C->assembled = PETSC_TRUE; 423 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 424 return 0; 425 } 426 427 /* ------------------------------------------------------------*/ 428 /* 429 Version for when blocks are 4 by 4 430 */ 431 #undef __FUNC__ 432 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 433 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 434 { 435 Mat C = *B; 436 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 437 IS iscol = b->col, isrow = b->row, isicol; 438 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 439 int *ajtmpold, *ajtmp, nz, row; 440 int *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j; 441 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 442 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 443 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 444 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 445 Scalar m13,m14,m15,m16; 446 Scalar *ba = b->a,*aa = a->a; 447 register int *pj; 448 449 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 450 PLogObjectParent(*B,isicol); 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 ierr = ISDestroy(isicol); CHKERRQ(ierr); 568 C->factor = FACTOR_LU; 569 C->assembled = PETSC_TRUE; 570 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 571 return 0; 572 } 573 /* ------------------------------------------------------------*/ 574 /* 575 Version for when blocks are 3 by 3 576 */ 577 #undef __FUNC__ 578 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 579 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 580 { 581 Mat C = *B; 582 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 583 IS iscol = b->col, isrow = b->row, isicol; 584 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 585 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 586 int *diag_offset = b->diag,idx; 587 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 588 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 589 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 590 Scalar *ba = b->a,*aa = a->a; 591 register int *pj; 592 593 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 594 PLogObjectParent(*B,isicol); 595 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 596 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 597 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 598 599 for ( i=0; i<n; i++ ) { 600 nz = bi[i+1] - bi[i]; 601 ajtmp = bj + bi[i]; 602 for ( j=0; j<nz; j++ ) { 603 x = rtmp + 9*ajtmp[j]; 604 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 605 } 606 /* load in initial (unfactored row) */ 607 idx = r[i]; 608 nz = ai[idx+1] - ai[idx]; 609 ajtmpold = aj + ai[idx]; 610 v = aa + 9*ai[idx]; 611 for ( j=0; j<nz; j++ ) { 612 x = rtmp + 9*ic[ajtmpold[j]]; 613 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 614 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 615 v += 9; 616 } 617 row = *ajtmp++; 618 while (row < i) { 619 pc = rtmp + 9*row; 620 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 621 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 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) { 624 pv = ba + 9*diag_offset[row]; 625 pj = bj + diag_offset[row] + 1; 626 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 627 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 628 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 629 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 630 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 631 632 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 633 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 634 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 635 636 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 637 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 638 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 639 nz = bi[row+1] - diag_offset[row] - 1; 640 pv += 9; 641 for (j=0; j<nz; j++) { 642 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 643 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 644 x = rtmp + 9*pj[j]; 645 x[0] -= m1*x1 + m4*x2 + m7*x3; 646 x[1] -= m2*x1 + m5*x2 + m8*x3; 647 x[2] -= m3*x1 + m6*x2 + m9*x3; 648 649 x[3] -= m1*x4 + m4*x5 + m7*x6; 650 x[4] -= m2*x4 + m5*x5 + m8*x6; 651 x[5] -= m3*x4 + m6*x5 + m9*x6; 652 653 x[6] -= m1*x7 + m4*x8 + m7*x9; 654 x[7] -= m2*x7 + m5*x8 + m8*x9; 655 x[8] -= m3*x7 + m6*x8 + m9*x9; 656 pv += 9; 657 } 658 PLogFlops(54*nz+36); 659 } 660 row = *ajtmp++; 661 } 662 /* finished row so stick it into b->a */ 663 pv = ba + 9*bi[i]; 664 pj = bj + bi[i]; 665 nz = bi[i+1] - bi[i]; 666 for ( j=0; j<nz; j++ ) { 667 x = rtmp + 9*pj[j]; 668 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 669 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 670 pv += 9; 671 } 672 /* invert diagonal block */ 673 w = ba + 9*diag_offset[i]; 674 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 675 } 676 677 PetscFree(rtmp); 678 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 679 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 680 ierr = ISDestroy(isicol); CHKERRQ(ierr); 681 C->factor = FACTOR_LU; 682 C->assembled = PETSC_TRUE; 683 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 684 return 0; 685 } 686 687 /* ------------------------------------------------------------*/ 688 /* 689 Version for when blocks are 2 by 2 690 */ 691 #undef __FUNC__ 692 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 693 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 694 { 695 Mat C = *B; 696 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 697 IS iscol = b->col, isrow = b->row, isicol; 698 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 699 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 700 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 701 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 702 Scalar p1,p2,p3,p4,v_work[2]; 703 Scalar *ba = b->a,*aa = a->a; 704 register int *pj; 705 706 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 707 PLogObjectParent(*B,isicol); 708 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 709 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 710 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 711 712 for ( i=0; i<n; i++ ) { 713 nz = bi[i+1] - bi[i]; 714 ajtmp = bj + bi[i]; 715 for ( j=0; j<nz; j++ ) { 716 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 717 } 718 /* load in initial (unfactored row) */ 719 idx = r[i]; 720 nz = ai[idx+1] - ai[idx]; 721 ajtmpold = aj + ai[idx]; 722 v = aa + 4*ai[idx]; 723 for ( j=0; j<nz; j++ ) { 724 x = rtmp+4*ic[ajtmpold[j]]; 725 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 726 v += 4; 727 } 728 row = *ajtmp++; 729 while (row < i) { 730 pc = rtmp + 4*row; 731 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 732 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 733 pv = ba + 4*diag_offset[row]; 734 pj = bj + diag_offset[row] + 1; 735 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 736 pc[0] = m1 = p1*x1 + p3*x2; 737 pc[1] = m2 = p2*x1 + p4*x2; 738 pc[2] = m3 = p1*x3 + p3*x4; 739 pc[3] = m4 = p2*x3 + p4*x4; 740 nz = bi[row+1] - diag_offset[row] - 1; 741 pv += 4; 742 for (j=0; j<nz; j++) { 743 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 744 x = rtmp + 4*pj[j]; 745 x[0] -= m1*x1 + m3*x2; 746 x[1] -= m2*x1 + m4*x2; 747 x[2] -= m1*x3 + m3*x4; 748 x[3] -= m2*x3 + m4*x4; 749 pv += 4; 750 } 751 PLogFlops(16*nz+12); 752 } 753 row = *ajtmp++; 754 } 755 /* finished row so stick it into b->a */ 756 pv = ba + 4*bi[i]; 757 pj = bj + bi[i]; 758 nz = bi[i+1] - bi[i]; 759 for ( j=0; j<nz; j++ ) { 760 x = rtmp+4*pj[j]; 761 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 762 pv += 4; 763 } 764 /* invert diagonal block */ 765 w = ba + 4*diag_offset[i]; 766 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 767 } 768 769 PetscFree(rtmp); 770 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 771 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 772 ierr = ISDestroy(isicol); CHKERRQ(ierr); 773 C->factor = FACTOR_LU; 774 C->assembled = PETSC_TRUE; 775 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 776 return 0; 777 } 778 779 /* ----------------------------------------------------------- */ 780 /* 781 Version for when blocks are 1 by 1. 782 */ 783 #undef __FUNC__ 784 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 785 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 786 { 787 Mat C = *B; 788 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 789 IS iscol = b->col, isrow = b->row, isicol; 790 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 791 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 792 int *diag_offset = b->diag,diag; 793 register Scalar *pv,*v,*rtmp,multiplier,*pc; 794 Scalar *ba = b->a,*aa = a->a; 795 register int *pj; 796 797 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 798 PLogObjectParent(*B,isicol); 799 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 800 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 801 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 802 803 for ( i=0; i<n; i++ ) { 804 nz = bi[i+1] - bi[i]; 805 ajtmp = bj + bi[i]; 806 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 807 808 /* load in initial (unfactored row) */ 809 nz = ai[r[i]+1] - ai[r[i]]; 810 ajtmpold = aj + ai[r[i]]; 811 v = aa + ai[r[i]]; 812 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 813 814 row = *ajtmp++; 815 while (row < i) { 816 pc = rtmp + row; 817 if (*pc != 0.0) { 818 pv = ba + diag_offset[row]; 819 pj = bj + diag_offset[row] + 1; 820 multiplier = *pc * *pv++; 821 *pc = multiplier; 822 nz = bi[row+1] - diag_offset[row] - 1; 823 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 824 PLogFlops(1+2*nz); 825 } 826 row = *ajtmp++; 827 } 828 /* finished row so stick it into b->a */ 829 pv = ba + bi[i]; 830 pj = bj + bi[i]; 831 nz = bi[i+1] - bi[i]; 832 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 833 diag = diag_offset[i] - bi[i]; 834 /* check pivot entry for current row */ 835 if (pv[diag] == 0.0) { 836 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 837 } 838 pv[diag] = 1.0/pv[diag]; 839 } 840 841 PetscFree(rtmp); 842 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 843 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 844 ierr = ISDestroy(isicol); CHKERRQ(ierr); 845 C->factor = FACTOR_LU; 846 C->assembled = PETSC_TRUE; 847 PLogFlops(b->n); 848 return 0; 849 } 850 851 /* ----------------------------------------------------------- */ 852 #undef __FUNC__ 853 #define __FUNC__ "MatLUFactor_SeqBAIJ" 854 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 855 { 856 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 857 int ierr; 858 Mat C; 859 860 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 861 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 862 863 /* free all the data structures from mat */ 864 PetscFree(mat->a); 865 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 866 if (mat->diag) PetscFree(mat->diag); 867 if (mat->ilen) PetscFree(mat->ilen); 868 if (mat->imax) PetscFree(mat->imax); 869 if (mat->solve_work) PetscFree(mat->solve_work); 870 if (mat->mult_work) PetscFree(mat->mult_work); 871 PetscFree(mat); 872 873 PetscMemcpy(A,C,sizeof(struct _p_Mat)); 874 PetscHeaderDestroy(C); 875 return 0; 876 } 877 /* ----------------------------------------------------------- */ 878 #undef __FUNC__ 879 #define __FUNC__ "MatSolve_SeqBAIJ_N" 880 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 881 { 882 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 883 IS iscol=a->col,isrow=a->row; 884 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 885 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 886 Scalar *aa=a->a,*sum; 887 register Scalar *x,*b,*lsum,*tmp,*v; 888 889 VecGetArray_Fast(bb,b); 890 VecGetArray_Fast(xx,x); 891 tmp = a->solve_work; 892 893 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 894 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 895 896 /* forward solve the lower triangular */ 897 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 898 for ( i=1; i<n; i++ ) { 899 v = aa + bs2*ai[i]; 900 vi = aj + ai[i]; 901 nz = a->diag[i] - ai[i]; 902 sum = tmp + bs*i; 903 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 904 while (nz--) { 905 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 906 v += bs2; 907 } 908 } 909 /* backward solve the upper triangular */ 910 lsum = a->solve_work + a->n; 911 for ( i=n-1; i>=0; i-- ){ 912 v = aa + bs2*(a->diag[i] + 1); 913 vi = aj + a->diag[i] + 1; 914 nz = ai[i+1] - a->diag[i] - 1; 915 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 916 while (nz--) { 917 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 918 v += bs2; 919 } 920 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 921 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 922 } 923 924 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 925 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 926 VecRestoreArray_Fast(bb,b); 927 VecRestoreArray_Fast(xx,x); 928 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 929 return 0; 930 } 931 932 #undef __FUNC__ 933 #define __FUNC__ "MatSolve_SeqBAIJ_7" 934 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 935 { 936 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 937 IS iscol=a->col,isrow=a->row; 938 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 939 int *diag = a->diag; 940 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 941 register Scalar *x,*b,*tmp,*v; 942 943 VecGetArray_Fast(bb,b); 944 VecGetArray_Fast(xx,x); 945 tmp = a->solve_work; 946 947 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 948 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 949 950 /* forward solve the lower triangular */ 951 idx = 7*(*r++); 952 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 953 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 954 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 955 956 for ( i=1; i<n; i++ ) { 957 v = aa + 49*ai[i]; 958 vi = aj + ai[i]; 959 nz = diag[i] - ai[i]; 960 idx = 7*(*r++); 961 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 962 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 963 while (nz--) { 964 idx = 7*(*vi++); 965 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 966 x4 = tmp[3+idx];x5 = tmp[4+idx]; 967 x6 = tmp[5+idx];x7 = tmp[6+idx]; 968 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 969 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 970 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 971 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 972 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 973 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 974 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 975 v += 49; 976 } 977 idx = 7*i; 978 tmp[idx] = sum1;tmp[1+idx] = sum2; 979 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 980 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 981 } 982 /* backward solve the upper triangular */ 983 for ( i=n-1; i>=0; i-- ){ 984 v = aa + 49*diag[i] + 49; 985 vi = aj + diag[i] + 1; 986 nz = ai[i+1] - diag[i] - 1; 987 idt = 7*i; 988 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 989 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 990 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 991 while (nz--) { 992 idx = 7*(*vi++); 993 x1 = tmp[idx]; x2 = tmp[1+idx]; 994 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 995 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 996 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 997 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 998 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 999 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 1000 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 1001 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 1002 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 1003 v += 49; 1004 } 1005 idc = 7*(*c--); 1006 v = aa + 49*diag[i]; 1007 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 1008 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 1009 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 1010 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 1011 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 1012 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1013 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1014 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1015 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1016 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1017 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1018 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1019 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1020 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1021 } 1022 1023 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1024 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1025 VecRestoreArray_Fast(bb,b); 1026 VecRestoreArray_Fast(xx,x); 1027 PLogFlops(2*49*(a->nz) - a->n); 1028 return 0; 1029 } 1030 1031 #undef __FUNC__ 1032 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1033 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1034 { 1035 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1036 IS iscol=a->col,isrow=a->row; 1037 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1038 int *diag = a->diag; 1039 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1040 register Scalar *x,*b,*tmp,*v; 1041 1042 VecGetArray_Fast(bb,b); 1043 VecGetArray_Fast(xx,x); 1044 tmp = a->solve_work; 1045 1046 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1047 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1048 1049 /* forward solve the lower triangular */ 1050 idx = 5*(*r++); 1051 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1052 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1053 for ( i=1; i<n; i++ ) { 1054 v = aa + 25*ai[i]; 1055 vi = aj + ai[i]; 1056 nz = diag[i] - ai[i]; 1057 idx = 5*(*r++); 1058 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1059 sum5 = b[4+idx]; 1060 while (nz--) { 1061 idx = 5*(*vi++); 1062 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1063 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1064 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1065 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1066 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1067 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1068 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1069 v += 25; 1070 } 1071 idx = 5*i; 1072 tmp[idx] = sum1;tmp[1+idx] = sum2; 1073 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1074 } 1075 /* backward solve the upper triangular */ 1076 for ( i=n-1; i>=0; i-- ){ 1077 v = aa + 25*diag[i] + 25; 1078 vi = aj + diag[i] + 1; 1079 nz = ai[i+1] - diag[i] - 1; 1080 idt = 5*i; 1081 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1082 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1083 while (nz--) { 1084 idx = 5*(*vi++); 1085 x1 = tmp[idx]; x2 = tmp[1+idx]; 1086 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1087 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1088 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1089 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1090 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1091 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1092 v += 25; 1093 } 1094 idc = 5*(*c--); 1095 v = aa + 25*diag[i]; 1096 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1097 v[15]*sum4+v[20]*sum5; 1098 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1099 v[16]*sum4+v[21]*sum5; 1100 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1101 v[17]*sum4+v[22]*sum5; 1102 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1103 v[18]*sum4+v[23]*sum5; 1104 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1105 v[19]*sum4+v[24]*sum5; 1106 } 1107 1108 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1109 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1110 VecRestoreArray_Fast(bb,b); 1111 VecRestoreArray_Fast(xx,x); 1112 PLogFlops(2*25*(a->nz) - a->n); 1113 return 0; 1114 } 1115 1116 #undef __FUNC__ 1117 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1118 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1119 { 1120 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1121 IS iscol=a->col,isrow=a->row; 1122 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1123 int *diag = a->diag; 1124 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1125 register Scalar *x,*b,*tmp,*v; 1126 1127 VecGetArray_Fast(bb,b); 1128 VecGetArray_Fast(xx,x); 1129 tmp = a->solve_work; 1130 1131 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1132 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1133 1134 /* forward solve the lower triangular */ 1135 idx = 4*(*r++); 1136 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1137 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1138 for ( i=1; i<n; i++ ) { 1139 v = aa + 16*ai[i]; 1140 vi = aj + ai[i]; 1141 nz = diag[i] - ai[i]; 1142 idx = 4*(*r++); 1143 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1144 while (nz--) { 1145 idx = 4*(*vi++); 1146 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1147 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1148 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1149 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1150 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1151 v += 16; 1152 } 1153 idx = 4*i; 1154 tmp[idx] = sum1;tmp[1+idx] = sum2; 1155 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1156 } 1157 /* backward solve the upper triangular */ 1158 for ( i=n-1; i>=0; i-- ){ 1159 v = aa + 16*diag[i] + 16; 1160 vi = aj + diag[i] + 1; 1161 nz = ai[i+1] - diag[i] - 1; 1162 idt = 4*i; 1163 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1164 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1165 while (nz--) { 1166 idx = 4*(*vi++); 1167 x1 = tmp[idx]; x2 = tmp[1+idx]; 1168 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1169 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1170 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1171 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1172 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1173 v += 16; 1174 } 1175 idc = 4*(*c--); 1176 v = aa + 16*diag[i]; 1177 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1178 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1179 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1180 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1181 } 1182 1183 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1184 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1185 VecRestoreArray_Fast(bb,b); 1186 VecRestoreArray_Fast(xx,x); 1187 PLogFlops(2*16*(a->nz) - a->n); 1188 return 0; 1189 } 1190 1191 /* 1192 Special case where the matrix was ILU(0) factored in the natural 1193 ordering. This eliminates the need for the column and row permutation. 1194 */ 1195 #undef __FUNC__ 1196 #define __FUNC__ "MatSolve_SeqBAIJ_4_NaturalOrdering" 1197 int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat A,Vec bb,Vec xx) 1198 { 1199 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data; 1200 int i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt; 1201 int *diag = a->diag,jdx; 1202 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1203 register Scalar *x,*b,*v; 1204 1205 VecGetArray_Fast(bb,b); 1206 VecGetArray_Fast(xx,x); 1207 1208 /* forward solve the lower triangular */ 1209 idx = 0; 1210 x[0] = b[idx]; x[1] = b[1+idx]; x[2] = b[2+idx]; x[3] = b[3+idx]; 1211 for ( i=1; i<n; i++ ) { 1212 v = aa + 16*ai[i]; 1213 vi = aj + ai[i]; 1214 nz = diag[i] - ai[i]; 1215 idx = 4*i; 1216 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1217 while (nz--) { 1218 jdx = 4*(*vi++); 1219 x1 = x[jdx];x2 = x[1+jdx];x3 = x[2+jdx];x4 = x[3+jdx]; 1220 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1221 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1222 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1223 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1224 v += 16; 1225 } 1226 x[idx] = sum1; 1227 x[1+idx] = sum2; 1228 x[2+idx] = sum3; 1229 x[3+idx] = sum4; 1230 } 1231 /* backward solve the upper triangular */ 1232 for ( i=n-1; i>=0; i-- ){ 1233 v = aa + 16*diag[i] + 16; 1234 vi = aj + diag[i] + 1; 1235 nz = ai[i+1] - diag[i] - 1; 1236 idt = 4*i; 1237 sum1 = x[idt]; sum2 = x[1+idt]; 1238 sum3 = x[2+idt];sum4 = x[3+idt]; 1239 while (nz--) { 1240 idx = 4*(*vi++); 1241 x1 = x[idx]; x2 = x[1+idx];x3 = x[2+idx]; x4 = x[3+idx]; 1242 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1243 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1244 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1245 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1246 v += 16; 1247 } 1248 v = aa + 16*diag[i]; 1249 x[idt] = v[0]*sum1 + v[4]*sum2 + v[8]*sum3 + v[12]*sum4; 1250 x[1+idt] = v[1]*sum1 + v[5]*sum2 + v[9]*sum3 + v[13]*sum4; 1251 x[2+idt] = v[2]*sum1 + v[6]*sum2 + v[10]*sum3 + v[14]*sum4; 1252 x[3+idt] = v[3]*sum1 + v[7]*sum2 + v[11]*sum3 + v[15]*sum4; 1253 } 1254 1255 VecRestoreArray_Fast(bb,b); 1256 VecRestoreArray_Fast(xx,x); 1257 PLogFlops(2*16*(a->nz) - a->n); 1258 return 0; 1259 } 1260 1261 1262 #undef __FUNC__ 1263 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1264 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1265 { 1266 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1267 IS iscol=a->col,isrow=a->row; 1268 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1269 int *diag = a->diag; 1270 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1271 register Scalar *x,*b,*tmp,*v; 1272 1273 VecGetArray_Fast(bb,b); 1274 VecGetArray_Fast(xx,x); 1275 tmp = a->solve_work; 1276 1277 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1278 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1279 1280 /* forward solve the lower triangular */ 1281 idx = 3*(*r++); 1282 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1283 for ( i=1; i<n; i++ ) { 1284 v = aa + 9*ai[i]; 1285 vi = aj + ai[i]; 1286 nz = diag[i] - ai[i]; 1287 idx = 3*(*r++); 1288 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1289 while (nz--) { 1290 idx = 3*(*vi++); 1291 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1292 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1293 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1294 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1295 v += 9; 1296 } 1297 idx = 3*i; 1298 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1299 } 1300 /* backward solve the upper triangular */ 1301 for ( i=n-1; i>=0; i-- ){ 1302 v = aa + 9*diag[i] + 9; 1303 vi = aj + diag[i] + 1; 1304 nz = ai[i+1] - diag[i] - 1; 1305 idt = 3*i; 1306 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1307 while (nz--) { 1308 idx = 3*(*vi++); 1309 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1310 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1311 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1312 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1313 v += 9; 1314 } 1315 idc = 3*(*c--); 1316 v = aa + 9*diag[i]; 1317 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1318 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1319 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1320 } 1321 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1322 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1323 VecRestoreArray_Fast(bb,b); 1324 VecRestoreArray_Fast(xx,x); 1325 PLogFlops(2*9*(a->nz) - a->n); 1326 return 0; 1327 } 1328 1329 #undef __FUNC__ 1330 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1331 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1332 { 1333 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1334 IS iscol=a->col,isrow=a->row; 1335 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1336 int *diag = a->diag; 1337 Scalar *aa=a->a,sum1,sum2,x1,x2; 1338 register Scalar *x,*b,*tmp,*v; 1339 1340 VecGetArray_Fast(bb,b); 1341 VecGetArray_Fast(xx,x); 1342 tmp = a->solve_work; 1343 1344 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1345 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1346 1347 /* forward solve the lower triangular */ 1348 idx = 2*(*r++); 1349 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1350 for ( i=1; i<n; i++ ) { 1351 v = aa + 4*ai[i]; 1352 vi = aj + ai[i]; 1353 nz = diag[i] - ai[i]; 1354 idx = 2*(*r++); 1355 sum1 = b[idx]; sum2 = b[1+idx]; 1356 while (nz--) { 1357 idx = 2*(*vi++); 1358 x1 = tmp[idx]; x2 = tmp[1+idx]; 1359 sum1 -= v[0]*x1 + v[2]*x2; 1360 sum2 -= v[1]*x1 + v[3]*x2; 1361 v += 4; 1362 } 1363 idx = 2*i; 1364 tmp[idx] = sum1; tmp[1+idx] = sum2; 1365 } 1366 /* backward solve the upper triangular */ 1367 for ( i=n-1; i>=0; i-- ){ 1368 v = aa + 4*diag[i] + 4; 1369 vi = aj + diag[i] + 1; 1370 nz = ai[i+1] - diag[i] - 1; 1371 idt = 2*i; 1372 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1373 while (nz--) { 1374 idx = 2*(*vi++); 1375 x1 = tmp[idx]; x2 = tmp[1+idx]; 1376 sum1 -= v[0]*x1 + v[2]*x2; 1377 sum2 -= v[1]*x1 + v[3]*x2; 1378 v += 4; 1379 } 1380 idc = 2*(*c--); 1381 v = aa + 4*diag[i]; 1382 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1383 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1384 } 1385 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1386 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1387 VecRestoreArray_Fast(bb,b); 1388 VecRestoreArray_Fast(xx,x); 1389 PLogFlops(2*4*(a->nz) - a->n); 1390 return 0; 1391 } 1392 1393 1394 #undef __FUNC__ 1395 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1396 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1397 { 1398 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1399 IS iscol=a->col,isrow=a->row; 1400 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1401 int *diag = a->diag; 1402 Scalar *aa=a->a,sum1; 1403 register Scalar *x,*b,*tmp,*v; 1404 1405 if (!n) return 0; 1406 1407 VecGetArray_Fast(bb,b); 1408 VecGetArray_Fast(xx,x); 1409 tmp = a->solve_work; 1410 1411 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1412 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1413 1414 /* forward solve the lower triangular */ 1415 tmp[0] = b[*r++]; 1416 for ( i=1; i<n; i++ ) { 1417 v = aa + ai[i]; 1418 vi = aj + ai[i]; 1419 nz = diag[i] - ai[i]; 1420 sum1 = b[*r++]; 1421 while (nz--) { 1422 sum1 -= (*v++)*tmp[*vi++]; 1423 } 1424 tmp[i] = sum1; 1425 } 1426 /* backward solve the upper triangular */ 1427 for ( i=n-1; i>=0; i-- ){ 1428 v = aa + diag[i] + 1; 1429 vi = aj + diag[i] + 1; 1430 nz = ai[i+1] - diag[i] - 1; 1431 sum1 = tmp[i]; 1432 while (nz--) { 1433 sum1 -= (*v++)*tmp[*vi++]; 1434 } 1435 x[*c--] = tmp[i] = aa[diag[i]]*sum1; 1436 } 1437 1438 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1439 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1440 VecRestoreArray_Fast(bb,b); 1441 VecRestoreArray_Fast(xx,x); 1442 PLogFlops(2*1*(a->nz) - a->n); 1443 return 0; 1444 } 1445 1446 extern int MatSolve_SeqBAIJ_4_NaturalOrdering(Mat,Vec,Vec); 1447 /* ----------------------------------------------------------------*/ 1448 /* 1449 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1450 except that the data structure of Mat_SeqAIJ is slightly different. 1451 Not a good example of code reuse. 1452 */ 1453 #undef __FUNC__ 1454 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1455 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1456 Mat *fact) 1457 { 1458 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1459 IS isicol; 1460 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1461 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1462 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1463 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1464 PetscTruth col_identity, row_identity; 1465 1466 /* special case that simply copies fill pattern */ 1467 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1468 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1469 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1470 if (levels == 0 && row_identity && col_identity) { 1471 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1472 (*fact)->factor = FACTOR_LU; 1473 b = (Mat_SeqBAIJ *) (*fact)->data; 1474 if (!b->diag) { 1475 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1476 } 1477 b->row = isrow; 1478 b->col = iscol; 1479 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1480 /* 1481 Blocksize 4 has a special faster solver for ILU(0) factorization 1482 with natural ordering 1483 */ 1484 if (b->bs == 4) { 1485 (*fact)->ops.solve = MatSolve_SeqBAIJ_4_NaturalOrdering; 1486 } 1487 return 0; 1488 } 1489 1490 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1491 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1492 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1493 1494 /* get new row pointers */ 1495 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1496 ainew[0] = 0; 1497 /* don't know how many column pointers are needed so estimate */ 1498 jmax = (int) (f*ai[n] + 1); 1499 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1500 /* ajfill is level of fill for each fill entry */ 1501 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1502 /* fill is a linked list of nonzeros in active row */ 1503 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1504 /* im is level for each filled value */ 1505 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1506 /* dloc is location of diagonal in factor */ 1507 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1508 dloc[0] = 0; 1509 for ( prow=0; prow<n; prow++ ) { 1510 /* first copy previous fill into linked list */ 1511 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1512 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1513 xi = aj + ai[r[prow]]; 1514 fill[n] = n; 1515 while (nz--) { 1516 fm = n; 1517 idx = ic[*xi++]; 1518 do { 1519 m = fm; 1520 fm = fill[m]; 1521 } while (fm < idx); 1522 fill[m] = idx; 1523 fill[idx] = fm; 1524 im[idx] = 0; 1525 } 1526 nzi = 0; 1527 row = fill[n]; 1528 while ( row < prow ) { 1529 incrlev = im[row] + 1; 1530 nz = dloc[row]; 1531 xi = ajnew + ainew[row] + nz; 1532 flev = ajfill + ainew[row] + nz + 1; 1533 nnz = ainew[row+1] - ainew[row] - nz - 1; 1534 if (*xi++ != row) { 1535 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot: try running with -pc_ilu_nonzeros_along_diagonal"); 1536 } 1537 fm = row; 1538 while (nnz-- > 0) { 1539 idx = *xi++; 1540 if (*flev + incrlev > levels) { 1541 flev++; 1542 continue; 1543 } 1544 do { 1545 m = fm; 1546 fm = fill[m]; 1547 } while (fm < idx); 1548 if (fm != idx) { 1549 im[idx] = *flev + incrlev; 1550 fill[m] = idx; 1551 fill[idx] = fm; 1552 fm = idx; 1553 nzf++; 1554 } 1555 else { 1556 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1557 } 1558 flev++; 1559 } 1560 row = fill[row]; 1561 nzi++; 1562 } 1563 /* copy new filled row into permanent storage */ 1564 ainew[prow+1] = ainew[prow] + nzf; 1565 if (ainew[prow+1] > jmax) { 1566 /* allocate a longer ajnew */ 1567 int maxadd; 1568 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1569 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1570 jmax += maxadd; 1571 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1572 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1573 PetscFree(ajnew); 1574 ajnew = xi; 1575 /* allocate a longer ajfill */ 1576 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1577 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1578 PetscFree(ajfill); 1579 ajfill = xi; 1580 realloc++; 1581 } 1582 xi = ajnew + ainew[prow]; 1583 flev = ajfill + ainew[prow]; 1584 dloc[prow] = nzi; 1585 fm = fill[n]; 1586 while (nzf--) { 1587 *xi++ = fm; 1588 *flev++ = im[fm]; 1589 fm = fill[fm]; 1590 } 1591 } 1592 PetscFree(ajfill); 1593 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1594 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1595 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1596 PetscFree(fill); PetscFree(im); 1597 1598 { 1599 double af = ((double)ainew[n])/((double)ai[n]); 1600 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 1601 realloc,f,af); 1602 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:Run with -pc_ilu_fill %g or use \n",af); 1603 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:PCILUSetFill(pc,%g);\n",af); 1604 PLogInfo(A,"Info:MatILUFactorSymbolic_SeqBAIJ:for best performance.\n"); 1605 } 1606 1607 /* put together the new matrix */ 1608 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1609 b = (Mat_SeqBAIJ *) (*fact)->data; 1610 PetscFree(b->imax); 1611 b->singlemalloc = 0; 1612 len = bs2*ainew[n]*sizeof(Scalar); 1613 /* the next line frees the default space generated by the Create() */ 1614 PetscFree(b->a); PetscFree(b->ilen); 1615 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1616 b->j = ajnew; 1617 b->i = ainew; 1618 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1619 b->diag = dloc; 1620 b->ilen = 0; 1621 b->imax = 0; 1622 b->row = isrow; 1623 b->col = iscol; 1624 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1625 CHKPTRQ(b->solve_work); 1626 /* In b structure: Free imax, ilen, old a, old j. 1627 Allocate dloc, solve_work, new a, new j */ 1628 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1629 b->maxnz = b->nz = ainew[n]; 1630 (*fact)->factor = FACTOR_LU; 1631 1632 (*fact)->info.factor_mallocs = realloc; 1633 (*fact)->info.fill_ratio_given = f; 1634 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1635 1636 return 0; 1637 } 1638 1639 1640 1641 1642