1 2 #ifndef lint 3 static char vcid[] = "$Id: baijfact.c,v 1.40 1997/02/04 16:28:38 balay Exp bsmith $"; 4 #endif 5 /* 6 Factorization code for BAIJ format. 7 */ 8 9 #include "src/mat/impls/baij/seq/baij.h" 10 #include "src/vec/vecimpl.h" 11 #include "src/inline/ilu.h" 12 13 14 /* 15 The symbolic factorization code is identical to that for AIJ format, 16 except for very small changes since this is now a SeqBAIJ datastructure. 17 NOT good code reuse. 18 */ 19 #undef __FUNC__ 20 #define __FUNC__ "MatLUFactorSymbolic_SeqBAIJ" 21 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 22 { 23 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 24 IS isicol; 25 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 26 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 27 int *idnew, idx, row,m,fm, nnz, nzi,realloc = 0,nzbd,*im; 28 29 PetscValidHeaderSpecific(isrow,IS_COOKIE); 30 PetscValidHeaderSpecific(iscol,IS_COOKIE); 31 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 32 ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 33 34 /* get new row pointers */ 35 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 36 ainew[0] = 0; 37 /* don't know how many column pointers are needed so estimate */ 38 jmax = (int) (f*ai[n] + 1); 39 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 40 /* fill is a linked list of nonzeros in active row */ 41 fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 42 im = fill + n + 1; 43 /* idnew is location of diagonal in factor */ 44 idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 45 idnew[0] = 0; 46 47 for ( i=0; i<n; i++ ) { 48 /* first copy previous fill into linked list */ 49 nnz = nz = ai[r[i]+1] - ai[r[i]]; 50 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 51 ajtmp = aj + ai[r[i]]; 52 fill[n] = n; 53 while (nz--) { 54 fm = n; 55 idx = ic[*ajtmp++]; 56 do { 57 m = fm; 58 fm = fill[m]; 59 } while (fm < idx); 60 fill[m] = idx; 61 fill[idx] = fm; 62 } 63 row = fill[n]; 64 while ( row < i ) { 65 ajtmp = ajnew + idnew[row] + 1; 66 nzbd = 1 + idnew[row] - ainew[row]; 67 nz = im[row] - nzbd; 68 fm = row; 69 while (nz-- > 0) { 70 idx = *ajtmp++; 71 nzbd++; 72 if (idx == i) im[row] = nzbd; 73 do { 74 m = fm; 75 fm = fill[m]; 76 } while (fm < idx); 77 if (fm != idx) { 78 fill[m] = idx; 79 fill[idx] = fm; 80 fm = idx; 81 nnz++; 82 } 83 } 84 row = fill[row]; 85 } 86 /* copy new filled row into permanent storage */ 87 ainew[i+1] = ainew[i] + nnz; 88 if (ainew[i+1] > jmax) { 89 /* allocate a longer ajnew */ 90 int maxadd; 91 maxadd = (int) ((f*(ai[n]+1)*(n-i+5))/n); 92 if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 93 jmax += maxadd; 94 ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 95 PetscMemcpy(ajtmp,ajnew,ainew[i]*sizeof(int)); 96 PetscFree(ajnew); 97 ajnew = ajtmp; 98 realloc++; /* count how many times we realloc */ 99 } 100 ajtmp = ajnew + ainew[i]; 101 fm = fill[n]; 102 nzi = 0; 103 im[i] = nnz; 104 while (nnz--) { 105 if (fm < i) nzi++; 106 *ajtmp++ = fm; 107 fm = fill[fm]; 108 } 109 idnew[i] = ainew[i] + nzi; 110 } 111 112 PLogInfo(A, 113 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 114 realloc,f,((double)ainew[n])/((double)ai[i])); 115 116 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 117 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 118 119 PetscFree(fill); 120 121 /* put together the new matrix */ 122 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 123 PLogObjectParent(*B,isicol); 124 ierr = ISDestroy(isicol); CHKERRQ(ierr); 125 b = (Mat_SeqBAIJ *) (*B)->data; 126 PetscFree(b->imax); 127 b->singlemalloc = 0; 128 /* the next line frees the default space generated by the Create() */ 129 PetscFree(b->a); PetscFree(b->ilen); 130 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 131 b->j = ajnew; 132 b->i = ainew; 133 b->diag = idnew; 134 b->ilen = 0; 135 b->imax = 0; 136 b->row = isrow; 137 b->col = iscol; 138 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 139 CHKPTRQ(b->solve_work); 140 /* In b structure: Free imax, ilen, old a, old j. 141 Allocate idnew, solve_work, new a, new j */ 142 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 143 b->maxnz = b->nz = ainew[n]; 144 145 (*B)->info.factor_mallocs = realloc; 146 (*B)->info.fill_ratio_given = f; 147 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 148 149 return 0; 150 } 151 152 /* ----------------------------------------------------------- */ 153 #undef __FUNC__ 154 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 155 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 156 { 157 Mat C = *B; 158 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 159 IS iscol = b->col, isrow = b->row, isicol; 160 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 161 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 162 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 163 Scalar *ba = b->a,*aa = a->a; 164 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 165 register int *pj; 166 167 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 168 PLogObjectParent(*B,isicol); 169 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 170 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 171 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 172 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 173 /* generate work space needed by dense LU factorization */ 174 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 175 CHKPTRQ(v_work); 176 multiplier = v_work + bs; 177 v_pivots = (int *) (multiplier + bs2); 178 179 /* flops in while loop */ 180 bslog = 2*bs*bs2; 181 182 for ( i=0; i<n; i++ ) { 183 nz = bi[i+1] - bi[i]; 184 ajtmp = bj + bi[i]; 185 for ( j=0; j<nz; j++ ) { 186 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 187 } 188 /* load in initial (unfactored row) */ 189 nz = ai[r[i]+1] - ai[r[i]]; 190 ajtmpold = aj + ai[r[i]]; 191 v = aa + bs2*ai[r[i]]; 192 for ( j=0; j<nz; j++ ) { 193 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 194 } 195 row = *ajtmp++; 196 while (row < i) { 197 pc = rtmp + bs2*row; 198 /* if (*pc) { */ 199 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 200 if (flg) { 201 pv = ba + bs2*diag_offset[row]; 202 pj = bj + diag_offset[row] + 1; 203 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 204 nz = bi[row+1] - diag_offset[row] - 1; 205 pv += bs2; 206 for (j=0; j<nz; j++) { 207 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 208 } 209 PLogFlops(bslog*(nz+1)-bs); 210 } 211 row = *ajtmp++; 212 } 213 /* finished row so stick it into b->a */ 214 pv = ba + bs2*bi[i]; 215 pj = bj + bi[i]; 216 nz = bi[i+1] - bi[i]; 217 for ( j=0; j<nz; j++ ) { 218 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 219 } 220 diag = diag_offset[i] - bi[i]; 221 /* invert diagonal block */ 222 w = pv + bs2*diag; 223 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 224 } 225 226 PetscFree(rtmp); PetscFree(v_work); 227 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 228 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 229 ierr = ISDestroy(isicol); CHKERRQ(ierr); 230 C->factor = FACTOR_LU; 231 C->assembled = PETSC_TRUE; 232 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 233 return 0; 234 } 235 /* ------------------------------------------------------------*/ 236 /* 237 Version for when blocks are 5 by 5 238 */ 239 #undef __FUNC__ 240 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 241 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 242 { 243 Mat C = *B; 244 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 245 IS iscol = b->col, isrow = b->row, isicol; 246 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 247 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 248 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 249 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 250 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 251 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 252 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 253 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 254 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 255 Scalar *ba = b->a,*aa = a->a; 256 register int *pj; 257 258 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 259 PLogObjectParent(*B,isicol); 260 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 261 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 262 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 263 264 for ( i=0; i<n; i++ ) { 265 nz = bi[i+1] - bi[i]; 266 ajtmp = bj + bi[i]; 267 for ( j=0; j<nz; j++ ) { 268 x = rtmp+25*ajtmp[j]; 269 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 270 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 271 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 272 } 273 /* load in initial (unfactored row) */ 274 idx = r[i]; 275 nz = ai[idx+1] - ai[idx]; 276 ajtmpold = aj + ai[idx]; 277 v = aa + 25*ai[idx]; 278 for ( j=0; j<nz; j++ ) { 279 x = rtmp+25*ic[ajtmpold[j]]; 280 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 281 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 282 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 283 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 284 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 285 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 286 v += 25; 287 } 288 row = *ajtmp++; 289 while (row < i) { 290 pc = rtmp + 25*row; 291 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 292 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 293 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 294 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 295 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 296 p25 = pc[24]; 297 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 298 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 299 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 300 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 301 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 302 p24 != 0.0 || p25 != 0.0) { 303 pv = ba + 25*diag_offset[row]; 304 pj = bj + diag_offset[row] + 1; 305 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 306 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 307 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 308 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 309 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 310 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 311 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 312 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 313 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 314 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 315 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 316 317 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 318 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 319 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 320 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 321 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 322 323 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 324 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 325 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 326 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 327 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 328 329 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 330 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 331 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 332 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 333 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 334 335 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 336 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 337 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 338 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 339 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 340 341 nz = bi[row+1] - diag_offset[row] - 1; 342 pv += 25; 343 for (j=0; j<nz; j++) { 344 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 345 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 346 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 347 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 348 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 349 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 350 x = rtmp + 25*pj[j]; 351 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 352 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 353 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 354 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 355 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 356 357 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 358 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 359 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 360 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 361 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 362 363 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 364 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 365 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 366 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 367 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 368 369 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 370 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 371 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 372 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 373 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 374 375 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 376 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 377 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 378 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 379 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 380 381 pv += 25; 382 } 383 PLogFlops(250*nz+225); 384 } 385 row = *ajtmp++; 386 } 387 /* finished row so stick it into b->a */ 388 pv = ba + 25*bi[i]; 389 pj = bj + bi[i]; 390 nz = bi[i+1] - bi[i]; 391 for ( j=0; j<nz; j++ ) { 392 x = rtmp+25*pj[j]; 393 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 394 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 395 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 396 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 397 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 398 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 399 pv += 25; 400 } 401 /* invert diagonal block */ 402 w = ba + 25*diag_offset[i]; 403 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 404 } 405 406 PetscFree(rtmp); 407 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 408 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 409 ierr = ISDestroy(isicol); CHKERRQ(ierr); 410 C->factor = FACTOR_LU; 411 C->assembled = PETSC_TRUE; 412 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 413 return 0; 414 } 415 416 /* ------------------------------------------------------------*/ 417 /* 418 Version for when blocks are 4 by 4 419 */ 420 #undef __FUNC__ 421 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 422 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 423 { 424 Mat C = *B; 425 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 426 IS iscol = b->col, isrow = b->row, isicol; 427 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 428 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 429 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 430 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 431 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 432 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 433 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 434 Scalar m13,m14,m15,m16; 435 Scalar *ba = b->a,*aa = a->a; 436 register int *pj; 437 438 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 439 PLogObjectParent(*B,isicol); 440 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 441 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 442 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 443 444 for ( i=0; i<n; i++ ) { 445 nz = bi[i+1] - bi[i]; 446 ajtmp = bj + bi[i]; 447 for ( j=0; j<nz; j++ ) { 448 x = rtmp+16*ajtmp[j]; 449 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 450 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 451 } 452 /* load in initial (unfactored row) */ 453 idx = r[i]; 454 nz = ai[idx+1] - ai[idx]; 455 ajtmpold = aj + ai[idx]; 456 v = aa + 16*ai[idx]; 457 for ( j=0; j<nz; j++ ) { 458 x = rtmp+16*ic[ajtmpold[j]]; 459 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 460 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 461 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 462 x[14] = v[14]; x[15] = v[15]; 463 v += 16; 464 } 465 row = *ajtmp++; 466 while (row < i) { 467 pc = rtmp + 16*row; 468 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 469 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 470 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 471 p15 = pc[14]; p16 = pc[15]; 472 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 473 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 474 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 475 || p16 != 0.0) { 476 pv = ba + 16*diag_offset[row]; 477 pj = bj + diag_offset[row] + 1; 478 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 479 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 480 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 481 x15 = pv[14]; x16 = pv[15]; 482 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 483 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 484 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 485 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 486 487 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 488 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 489 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 490 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 491 492 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 493 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 494 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 495 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 496 497 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 498 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 499 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 500 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 501 502 nz = bi[row+1] - diag_offset[row] - 1; 503 pv += 16; 504 for (j=0; j<nz; j++) { 505 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 506 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 507 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 508 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 509 x = rtmp + 16*pj[j]; 510 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 511 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 512 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 513 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 514 515 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 516 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 517 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 518 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 519 520 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 521 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 522 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 523 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 524 525 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 526 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 527 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 528 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 529 530 pv += 16; 531 } 532 PLogFlops(128*nz+112); 533 } 534 row = *ajtmp++; 535 } 536 /* finished row so stick it into b->a */ 537 pv = ba + 16*bi[i]; 538 pj = bj + bi[i]; 539 nz = bi[i+1] - bi[i]; 540 for ( j=0; j<nz; j++ ) { 541 x = rtmp+16*pj[j]; 542 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 543 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 544 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 545 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 546 pv += 16; 547 } 548 /* invert diagonal block */ 549 w = ba + 16*diag_offset[i]; 550 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 551 } 552 553 PetscFree(rtmp); 554 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 555 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 556 ierr = ISDestroy(isicol); CHKERRQ(ierr); 557 C->factor = FACTOR_LU; 558 C->assembled = PETSC_TRUE; 559 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 560 return 0; 561 } 562 /* ------------------------------------------------------------*/ 563 /* 564 Version for when blocks are 3 by 3 565 */ 566 #undef __FUNC__ 567 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 568 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 569 { 570 Mat C = *B; 571 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 572 IS iscol = b->col, isrow = b->row, isicol; 573 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 574 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 575 int *diag_offset = b->diag,idx; 576 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 577 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 578 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 579 Scalar *ba = b->a,*aa = a->a; 580 register int *pj; 581 582 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 583 PLogObjectParent(*B,isicol); 584 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 585 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 586 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 587 588 for ( i=0; i<n; i++ ) { 589 nz = bi[i+1] - bi[i]; 590 ajtmp = bj + bi[i]; 591 for ( j=0; j<nz; j++ ) { 592 x = rtmp + 9*ajtmp[j]; 593 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 594 } 595 /* load in initial (unfactored row) */ 596 idx = r[i]; 597 nz = ai[idx+1] - ai[idx]; 598 ajtmpold = aj + ai[idx]; 599 v = aa + 9*ai[idx]; 600 for ( j=0; j<nz; j++ ) { 601 x = rtmp + 9*ic[ajtmpold[j]]; 602 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 603 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 604 v += 9; 605 } 606 row = *ajtmp++; 607 while (row < i) { 608 pc = rtmp + 9*row; 609 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 610 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 611 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 612 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 613 pv = ba + 9*diag_offset[row]; 614 pj = bj + diag_offset[row] + 1; 615 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 616 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 617 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 618 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 619 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 620 621 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 622 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 623 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 624 625 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 626 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 627 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 628 nz = bi[row+1] - diag_offset[row] - 1; 629 pv += 9; 630 for (j=0; j<nz; j++) { 631 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 632 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 633 x = rtmp + 9*pj[j]; 634 x[0] -= m1*x1 + m4*x2 + m7*x3; 635 x[1] -= m2*x1 + m5*x2 + m8*x3; 636 x[2] -= m3*x1 + m6*x2 + m9*x3; 637 638 x[3] -= m1*x4 + m4*x5 + m7*x6; 639 x[4] -= m2*x4 + m5*x5 + m8*x6; 640 x[5] -= m3*x4 + m6*x5 + m9*x6; 641 642 x[6] -= m1*x7 + m4*x8 + m7*x9; 643 x[7] -= m2*x7 + m5*x8 + m8*x9; 644 x[8] -= m3*x7 + m6*x8 + m9*x9; 645 pv += 9; 646 } 647 PLogFlops(54*nz+36); 648 } 649 row = *ajtmp++; 650 } 651 /* finished row so stick it into b->a */ 652 pv = ba + 9*bi[i]; 653 pj = bj + bi[i]; 654 nz = bi[i+1] - bi[i]; 655 for ( j=0; j<nz; j++ ) { 656 x = rtmp + 9*pj[j]; 657 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 658 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 659 pv += 9; 660 } 661 /* invert diagonal block */ 662 w = ba + 9*diag_offset[i]; 663 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 664 } 665 666 PetscFree(rtmp); 667 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 668 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 669 ierr = ISDestroy(isicol); CHKERRQ(ierr); 670 C->factor = FACTOR_LU; 671 C->assembled = PETSC_TRUE; 672 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 673 return 0; 674 } 675 676 /* ------------------------------------------------------------*/ 677 /* 678 Version for when blocks are 2 by 2 679 */ 680 #undef __FUNC__ 681 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 682 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 683 { 684 Mat C = *B; 685 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 686 IS iscol = b->col, isrow = b->row, isicol; 687 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 688 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 689 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 690 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 691 Scalar p1,p2,p3,p4,v_work[2]; 692 Scalar *ba = b->a,*aa = a->a; 693 register int *pj; 694 695 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 696 PLogObjectParent(*B,isicol); 697 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 698 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 699 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 700 701 for ( i=0; i<n; i++ ) { 702 nz = bi[i+1] - bi[i]; 703 ajtmp = bj + bi[i]; 704 for ( j=0; j<nz; j++ ) { 705 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 706 } 707 /* load in initial (unfactored row) */ 708 idx = r[i]; 709 nz = ai[idx+1] - ai[idx]; 710 ajtmpold = aj + ai[idx]; 711 v = aa + 4*ai[idx]; 712 for ( j=0; j<nz; j++ ) { 713 x = rtmp+4*ic[ajtmpold[j]]; 714 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 715 v += 4; 716 } 717 row = *ajtmp++; 718 while (row < i) { 719 pc = rtmp + 4*row; 720 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 721 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 722 pv = ba + 4*diag_offset[row]; 723 pj = bj + diag_offset[row] + 1; 724 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 725 pc[0] = m1 = p1*x1 + p3*x2; 726 pc[1] = m2 = p2*x1 + p4*x2; 727 pc[2] = m3 = p1*x3 + p3*x4; 728 pc[3] = m4 = p2*x3 + p4*x4; 729 nz = bi[row+1] - diag_offset[row] - 1; 730 pv += 4; 731 for (j=0; j<nz; j++) { 732 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 733 x = rtmp + 4*pj[j]; 734 x[0] -= m1*x1 + m3*x2; 735 x[1] -= m2*x1 + m4*x2; 736 x[2] -= m1*x3 + m3*x4; 737 x[3] -= m2*x3 + m4*x4; 738 pv += 4; 739 } 740 PLogFlops(16*nz+12); 741 } 742 row = *ajtmp++; 743 } 744 /* finished row so stick it into b->a */ 745 pv = ba + 4*bi[i]; 746 pj = bj + bi[i]; 747 nz = bi[i+1] - bi[i]; 748 for ( j=0; j<nz; j++ ) { 749 x = rtmp+4*pj[j]; 750 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 751 pv += 4; 752 } 753 /* invert diagonal block */ 754 w = ba + 4*diag_offset[i]; 755 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 756 } 757 758 PetscFree(rtmp); 759 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 760 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 761 ierr = ISDestroy(isicol); CHKERRQ(ierr); 762 C->factor = FACTOR_LU; 763 C->assembled = PETSC_TRUE; 764 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 765 return 0; 766 } 767 768 /* ----------------------------------------------------------- */ 769 /* 770 Version for when blocks are 1 by 1. 771 */ 772 #undef __FUNC__ 773 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 774 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 775 { 776 Mat C = *B; 777 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 778 IS iscol = b->col, isrow = b->row, isicol; 779 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 780 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 781 int *diag_offset = b->diag,diag; 782 register Scalar *pv,*v,*rtmp,multiplier,*pc; 783 Scalar *ba = b->a,*aa = a->a; 784 register int *pj; 785 786 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 787 PLogObjectParent(*B,isicol); 788 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 789 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 790 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 791 792 for ( i=0; i<n; i++ ) { 793 nz = bi[i+1] - bi[i]; 794 ajtmp = bj + bi[i]; 795 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 796 797 /* load in initial (unfactored row) */ 798 nz = ai[r[i]+1] - ai[r[i]]; 799 ajtmpold = aj + ai[r[i]]; 800 v = aa + ai[r[i]]; 801 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 802 803 row = *ajtmp++; 804 while (row < i) { 805 pc = rtmp + row; 806 if (*pc != 0.0) { 807 pv = ba + diag_offset[row]; 808 pj = bj + diag_offset[row] + 1; 809 multiplier = *pc * *pv++; 810 *pc = multiplier; 811 nz = bi[row+1] - diag_offset[row] - 1; 812 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 813 PLogFlops(1+2*nz); 814 } 815 row = *ajtmp++; 816 } 817 /* finished row so stick it into b->a */ 818 pv = ba + bi[i]; 819 pj = bj + bi[i]; 820 nz = bi[i+1] - bi[i]; 821 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 822 diag = diag_offset[i] - bi[i]; 823 /* check pivot entry for current row */ 824 if (pv[diag] == 0.0) { 825 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 826 } 827 pv[diag] = 1.0/pv[diag]; 828 } 829 830 PetscFree(rtmp); 831 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 832 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 833 ierr = ISDestroy(isicol); CHKERRQ(ierr); 834 C->factor = FACTOR_LU; 835 C->assembled = PETSC_TRUE; 836 PLogFlops(b->n); 837 return 0; 838 } 839 840 /* ----------------------------------------------------------- */ 841 #undef __FUNC__ 842 #define __FUNC__ "MatLUFactor_SeqBAIJ" 843 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 844 { 845 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 846 int ierr; 847 Mat C; 848 849 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 850 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 851 852 /* free all the data structures from mat */ 853 PetscFree(mat->a); 854 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 855 if (mat->diag) PetscFree(mat->diag); 856 if (mat->ilen) PetscFree(mat->ilen); 857 if (mat->imax) PetscFree(mat->imax); 858 if (mat->solve_work) PetscFree(mat->solve_work); 859 if (mat->mult_work) PetscFree(mat->mult_work); 860 PetscFree(mat); 861 862 PetscMemcpy(A,C,sizeof(struct _Mat)); 863 PetscHeaderDestroy(C); 864 return 0; 865 } 866 /* ----------------------------------------------------------- */ 867 #undef __FUNC__ 868 #define __FUNC__ "MatSolve_SeqBAIJ_N" 869 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 870 { 871 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 872 IS iscol=a->col,isrow=a->row; 873 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 874 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 875 Scalar *aa=a->a,*sum; 876 register Scalar *x,*b,*lsum,*tmp,*v; 877 878 VecGetArray_Fast(bb,b); 879 VecGetArray_Fast(xx,x); 880 tmp = a->solve_work; 881 882 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 883 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 884 885 /* forward solve the lower triangular */ 886 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 887 for ( i=1; i<n; i++ ) { 888 v = aa + bs2*ai[i]; 889 vi = aj + ai[i]; 890 nz = a->diag[i] - ai[i]; 891 sum = tmp + bs*i; 892 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 893 while (nz--) { 894 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 895 v += bs2; 896 } 897 } 898 /* backward solve the upper triangular */ 899 lsum = a->solve_work + a->n; 900 for ( i=n-1; i>=0; i-- ){ 901 v = aa + bs2*(a->diag[i] + 1); 902 vi = aj + a->diag[i] + 1; 903 nz = ai[i+1] - a->diag[i] - 1; 904 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 905 while (nz--) { 906 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 907 v += bs2; 908 } 909 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 910 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 911 } 912 913 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 914 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 915 VecRestoreArray_Fast(bb,b); 916 VecRestoreArray_Fast(xx,x); 917 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 918 return 0; 919 } 920 921 #undef __FUNC__ 922 #define __FUNC__ "MatSolve_SeqBAIJ_7" 923 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 924 { 925 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 926 IS iscol=a->col,isrow=a->row; 927 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 928 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 929 register Scalar *x,*b,*tmp,*v; 930 931 VecGetArray_Fast(bb,b); 932 VecGetArray_Fast(xx,x); 933 tmp = a->solve_work; 934 935 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 936 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 937 938 /* forward solve the lower triangular */ 939 idx = 7*(*r++); 940 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 941 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 942 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 943 944 for ( i=1; i<n; i++ ) { 945 v = aa + 49*ai[i]; 946 vi = aj + ai[i]; 947 nz = a->diag[i] - ai[i]; 948 idx = 7*(*r++); 949 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 950 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 951 while (nz--) { 952 idx = 7*(*vi++); 953 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 954 x4 = tmp[3+idx];x5 = tmp[4+idx]; 955 x6 = tmp[5+idx];x7 = tmp[6+idx]; 956 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 957 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 958 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 959 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 960 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 961 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 962 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 963 v += 49; 964 } 965 idx = 7*i; 966 tmp[idx] = sum1;tmp[1+idx] = sum2; 967 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 968 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 969 } 970 /* backward solve the upper triangular */ 971 for ( i=n-1; i>=0; i-- ){ 972 v = aa + 49*a->diag[i] + 49; 973 vi = aj + a->diag[i] + 1; 974 nz = ai[i+1] - a->diag[i] - 1; 975 idt = 7*i; 976 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 977 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 978 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 979 while (nz--) { 980 idx = 7*(*vi++); 981 x1 = tmp[idx]; x2 = tmp[1+idx]; 982 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 983 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 984 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 985 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 986 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 987 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 988 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 989 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 990 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 991 v += 49; 992 } 993 idc = 7*(*c--); 994 v = aa + 49*a->diag[i]; 995 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 996 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 997 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 998 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 999 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 1000 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1001 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1002 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1003 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1004 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1005 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1006 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1007 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1008 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1009 } 1010 1011 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1012 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1013 VecRestoreArray_Fast(bb,b); 1014 VecRestoreArray_Fast(xx,x); 1015 PLogFlops(2*49*(a->nz) - a->n); 1016 return 0; 1017 } 1018 1019 #undef __FUNC__ 1020 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1021 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1022 { 1023 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1024 IS iscol=a->col,isrow=a->row; 1025 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1026 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1027 register Scalar *x,*b,*tmp,*v; 1028 1029 VecGetArray_Fast(bb,b); 1030 VecGetArray_Fast(xx,x); 1031 tmp = a->solve_work; 1032 1033 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1034 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1035 1036 /* forward solve the lower triangular */ 1037 idx = 5*(*r++); 1038 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1039 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1040 for ( i=1; i<n; i++ ) { 1041 v = aa + 25*ai[i]; 1042 vi = aj + ai[i]; 1043 nz = a->diag[i] - ai[i]; 1044 idx = 5*(*r++); 1045 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1046 sum5 = b[4+idx]; 1047 while (nz--) { 1048 idx = 5*(*vi++); 1049 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1050 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1051 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1052 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1053 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1054 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1055 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1056 v += 25; 1057 } 1058 idx = 5*i; 1059 tmp[idx] = sum1;tmp[1+idx] = sum2; 1060 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1061 } 1062 /* backward solve the upper triangular */ 1063 for ( i=n-1; i>=0; i-- ){ 1064 v = aa + 25*a->diag[i] + 25; 1065 vi = aj + a->diag[i] + 1; 1066 nz = ai[i+1] - a->diag[i] - 1; 1067 idt = 5*i; 1068 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1069 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1070 while (nz--) { 1071 idx = 5*(*vi++); 1072 x1 = tmp[idx]; x2 = tmp[1+idx]; 1073 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1074 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1075 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1076 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1077 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1078 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1079 v += 25; 1080 } 1081 idc = 5*(*c--); 1082 v = aa + 25*a->diag[i]; 1083 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1084 v[15]*sum4+v[20]*sum5; 1085 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1086 v[16]*sum4+v[21]*sum5; 1087 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1088 v[17]*sum4+v[22]*sum5; 1089 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1090 v[18]*sum4+v[23]*sum5; 1091 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1092 v[19]*sum4+v[24]*sum5; 1093 } 1094 1095 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1096 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1097 VecRestoreArray_Fast(bb,b); 1098 VecRestoreArray_Fast(xx,x); 1099 PLogFlops(2*25*(a->nz) - a->n); 1100 return 0; 1101 } 1102 1103 #undef __FUNC__ 1104 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1105 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1106 { 1107 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1108 IS iscol=a->col,isrow=a->row; 1109 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1110 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1111 register Scalar *x,*b,*tmp,*v; 1112 1113 VecGetArray_Fast(bb,b); 1114 VecGetArray_Fast(xx,x); 1115 tmp = a->solve_work; 1116 1117 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1118 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1119 1120 /* forward solve the lower triangular */ 1121 idx = 4*(*r++); 1122 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1123 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1124 for ( i=1; i<n; i++ ) { 1125 v = aa + 16*ai[i]; 1126 vi = aj + ai[i]; 1127 nz = a->diag[i] - ai[i]; 1128 idx = 4*(*r++); 1129 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1130 while (nz--) { 1131 idx = 4*(*vi++); 1132 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1133 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1134 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1135 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1136 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1137 v += 16; 1138 } 1139 idx = 4*i; 1140 tmp[idx] = sum1;tmp[1+idx] = sum2; 1141 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1142 } 1143 /* backward solve the upper triangular */ 1144 for ( i=n-1; i>=0; i-- ){ 1145 v = aa + 16*a->diag[i] + 16; 1146 vi = aj + a->diag[i] + 1; 1147 nz = ai[i+1] - a->diag[i] - 1; 1148 idt = 4*i; 1149 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1150 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1151 while (nz--) { 1152 idx = 4*(*vi++); 1153 x1 = tmp[idx]; x2 = tmp[1+idx]; 1154 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1155 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1156 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1157 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1158 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1159 v += 16; 1160 } 1161 idc = 4*(*c--); 1162 v = aa + 16*a->diag[i]; 1163 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1164 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1165 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1166 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1167 } 1168 1169 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1170 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1171 VecRestoreArray_Fast(bb,b); 1172 VecRestoreArray_Fast(xx,x); 1173 PLogFlops(2*16*(a->nz) - a->n); 1174 return 0; 1175 } 1176 1177 1178 #undef __FUNC__ 1179 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1180 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1181 { 1182 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1183 IS iscol=a->col,isrow=a->row; 1184 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1185 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1186 register Scalar *x,*b,*tmp,*v; 1187 1188 VecGetArray_Fast(bb,b); 1189 VecGetArray_Fast(xx,x); 1190 tmp = a->solve_work; 1191 1192 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1193 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1194 1195 /* forward solve the lower triangular */ 1196 idx = 3*(*r++); 1197 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1198 for ( i=1; i<n; i++ ) { 1199 v = aa + 9*ai[i]; 1200 vi = aj + ai[i]; 1201 nz = a->diag[i] - ai[i]; 1202 idx = 3*(*r++); 1203 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1204 while (nz--) { 1205 idx = 3*(*vi++); 1206 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1207 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1208 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1209 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1210 v += 9; 1211 } 1212 idx = 3*i; 1213 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1214 } 1215 /* backward solve the upper triangular */ 1216 for ( i=n-1; i>=0; i-- ){ 1217 v = aa + 9*a->diag[i] + 9; 1218 vi = aj + a->diag[i] + 1; 1219 nz = ai[i+1] - a->diag[i] - 1; 1220 idt = 3*i; 1221 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1222 while (nz--) { 1223 idx = 3*(*vi++); 1224 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1225 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1226 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1227 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1228 v += 9; 1229 } 1230 idc = 3*(*c--); 1231 v = aa + 9*a->diag[i]; 1232 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1233 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1234 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1235 } 1236 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1237 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1238 VecRestoreArray_Fast(bb,b); 1239 VecRestoreArray_Fast(xx,x); 1240 PLogFlops(2*9*(a->nz) - a->n); 1241 return 0; 1242 } 1243 1244 #undef __FUNC__ 1245 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1246 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1247 { 1248 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1249 IS iscol=a->col,isrow=a->row; 1250 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1251 Scalar *aa=a->a,sum1,sum2,x1,x2; 1252 register Scalar *x,*b,*tmp,*v; 1253 1254 VecGetArray_Fast(bb,b); 1255 VecGetArray_Fast(xx,x); 1256 tmp = a->solve_work; 1257 1258 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1259 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1260 1261 /* forward solve the lower triangular */ 1262 idx = 2*(*r++); 1263 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1264 for ( i=1; i<n; i++ ) { 1265 v = aa + 4*ai[i]; 1266 vi = aj + ai[i]; 1267 nz = a->diag[i] - ai[i]; 1268 idx = 2*(*r++); 1269 sum1 = b[idx]; sum2 = b[1+idx]; 1270 while (nz--) { 1271 idx = 2*(*vi++); 1272 x1 = tmp[idx]; x2 = tmp[1+idx]; 1273 sum1 -= v[0]*x1 + v[2]*x2; 1274 sum2 -= v[1]*x1 + v[3]*x2; 1275 v += 4; 1276 } 1277 idx = 2*i; 1278 tmp[idx] = sum1; tmp[1+idx] = sum2; 1279 } 1280 /* backward solve the upper triangular */ 1281 for ( i=n-1; i>=0; i-- ){ 1282 v = aa + 4*a->diag[i] + 4; 1283 vi = aj + a->diag[i] + 1; 1284 nz = ai[i+1] - a->diag[i] - 1; 1285 idt = 2*i; 1286 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1287 while (nz--) { 1288 idx = 2*(*vi++); 1289 x1 = tmp[idx]; x2 = tmp[1+idx]; 1290 sum1 -= v[0]*x1 + v[2]*x2; 1291 sum2 -= v[1]*x1 + v[3]*x2; 1292 v += 4; 1293 } 1294 idc = 2*(*c--); 1295 v = aa + 4*a->diag[i]; 1296 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1297 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1298 } 1299 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1300 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1301 VecRestoreArray_Fast(bb,b); 1302 VecRestoreArray_Fast(xx,x); 1303 PLogFlops(2*4*(a->nz) - a->n); 1304 return 0; 1305 } 1306 1307 1308 #undef __FUNC__ 1309 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1310 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1311 { 1312 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1313 IS iscol=a->col,isrow=a->row; 1314 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1315 Scalar *aa=a->a,sum1; 1316 register Scalar *x,*b,*tmp,*v; 1317 1318 if (!n) return 0; 1319 1320 VecGetArray_Fast(bb,b); 1321 VecGetArray_Fast(xx,x); 1322 tmp = a->solve_work; 1323 1324 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1325 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1326 1327 /* forward solve the lower triangular */ 1328 tmp[0] = b[*r++]; 1329 for ( i=1; i<n; i++ ) { 1330 v = aa + ai[i]; 1331 vi = aj + ai[i]; 1332 nz = a->diag[i] - ai[i]; 1333 sum1 = b[*r++]; 1334 while (nz--) { 1335 sum1 -= (*v++)*tmp[*vi++]; 1336 } 1337 tmp[i] = sum1; 1338 } 1339 /* backward solve the upper triangular */ 1340 for ( i=n-1; i>=0; i-- ){ 1341 v = aa + a->diag[i] + 1; 1342 vi = aj + a->diag[i] + 1; 1343 nz = ai[i+1] - a->diag[i] - 1; 1344 sum1 = tmp[i]; 1345 while (nz--) { 1346 sum1 -= (*v++)*tmp[*vi++]; 1347 } 1348 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1349 } 1350 1351 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1352 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1353 VecRestoreArray_Fast(bb,b); 1354 VecRestoreArray_Fast(xx,x); 1355 PLogFlops(2*1*(a->nz) - a->n); 1356 return 0; 1357 } 1358 1359 /* ----------------------------------------------------------------*/ 1360 /* 1361 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1362 except that the data structure of Mat_SeqAIJ is slightly different. 1363 Not a good example of code reuse. 1364 */ 1365 #undef __FUNC__ 1366 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1367 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1368 Mat *fact) 1369 { 1370 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1371 IS isicol; 1372 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1373 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1374 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1375 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1376 PetscTruth col_identity, row_identity; 1377 1378 /* special case that simply copies fill pattern */ 1379 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1380 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1381 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1382 if (levels == 0 && row_identity && col_identity) { 1383 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1384 (*fact)->factor = FACTOR_LU; 1385 b = (Mat_SeqBAIJ *) (*fact)->data; 1386 if (!b->diag) { 1387 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1388 } 1389 b->row = isrow; 1390 b->col = iscol; 1391 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1392 return 0; 1393 } 1394 1395 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1396 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1397 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1398 1399 /* get new row pointers */ 1400 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1401 ainew[0] = 0; 1402 /* don't know how many column pointers are needed so estimate */ 1403 jmax = (int) (f*ai[n] + 1); 1404 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1405 /* ajfill is level of fill for each fill entry */ 1406 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1407 /* fill is a linked list of nonzeros in active row */ 1408 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1409 /* im is level for each filled value */ 1410 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1411 /* dloc is location of diagonal in factor */ 1412 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1413 dloc[0] = 0; 1414 for ( prow=0; prow<n; prow++ ) { 1415 /* first copy previous fill into linked list */ 1416 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1417 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1418 xi = aj + ai[r[prow]]; 1419 fill[n] = n; 1420 while (nz--) { 1421 fm = n; 1422 idx = ic[*xi++]; 1423 do { 1424 m = fm; 1425 fm = fill[m]; 1426 } while (fm < idx); 1427 fill[m] = idx; 1428 fill[idx] = fm; 1429 im[idx] = 0; 1430 } 1431 nzi = 0; 1432 row = fill[n]; 1433 while ( row < prow ) { 1434 incrlev = im[row] + 1; 1435 nz = dloc[row]; 1436 xi = ajnew + ainew[row] + nz; 1437 flev = ajfill + ainew[row] + nz + 1; 1438 nnz = ainew[row+1] - ainew[row] - nz - 1; 1439 if (*xi++ != row) { 1440 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"zero pivot"); 1441 } 1442 fm = row; 1443 while (nnz-- > 0) { 1444 idx = *xi++; 1445 if (*flev + incrlev > levels) { 1446 flev++; 1447 continue; 1448 } 1449 do { 1450 m = fm; 1451 fm = fill[m]; 1452 } while (fm < idx); 1453 if (fm != idx) { 1454 im[idx] = *flev + incrlev; 1455 fill[m] = idx; 1456 fill[idx] = fm; 1457 fm = idx; 1458 nzf++; 1459 } 1460 else { 1461 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1462 } 1463 flev++; 1464 } 1465 row = fill[row]; 1466 nzi++; 1467 } 1468 /* copy new filled row into permanent storage */ 1469 ainew[prow+1] = ainew[prow] + nzf; 1470 if (ainew[prow+1] > jmax) { 1471 /* allocate a longer ajnew */ 1472 int maxadd; 1473 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1474 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1475 jmax += maxadd; 1476 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1477 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1478 PetscFree(ajnew); 1479 ajnew = xi; 1480 /* allocate a longer ajfill */ 1481 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1482 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1483 PetscFree(ajfill); 1484 ajfill = xi; 1485 realloc++; 1486 } 1487 xi = ajnew + ainew[prow]; 1488 flev = ajfill + ainew[prow]; 1489 dloc[prow] = nzi; 1490 fm = fill[n]; 1491 while (nzf--) { 1492 *xi++ = fm; 1493 *flev++ = im[fm]; 1494 fm = fill[fm]; 1495 } 1496 } 1497 PetscFree(ajfill); 1498 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1499 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1500 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1501 PetscFree(fill); PetscFree(im); 1502 1503 PLogInfo(A, 1504 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1505 realloc,f,((double)ainew[n])/((double)ai[prow])); 1506 1507 /* put together the new matrix */ 1508 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1509 b = (Mat_SeqBAIJ *) (*fact)->data; 1510 PetscFree(b->imax); 1511 b->singlemalloc = 0; 1512 len = bs2*ainew[n]*sizeof(Scalar); 1513 /* the next line frees the default space generated by the Create() */ 1514 PetscFree(b->a); PetscFree(b->ilen); 1515 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1516 b->j = ajnew; 1517 b->i = ainew; 1518 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1519 b->diag = dloc; 1520 b->ilen = 0; 1521 b->imax = 0; 1522 b->row = isrow; 1523 b->col = iscol; 1524 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1525 CHKPTRQ(b->solve_work); 1526 /* In b structure: Free imax, ilen, old a, old j. 1527 Allocate dloc, solve_work, new a, new j */ 1528 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1529 b->maxnz = b->nz = ainew[n]; 1530 (*fact)->factor = FACTOR_LU; 1531 1532 (*fact)->info.factor_mallocs = realloc; 1533 (*fact)->info.fill_ratio_given = f; 1534 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1535 1536 return 0; 1537 } 1538 1539 1540 1541 1542