1 #ifndef lint 2 static char vcid[] = "$Id: baijfact.c,v 1.39 1997/02/04 01:21:08 balay Exp balay $"; 3 #endif 4 /* 5 Factorization code for BAIJ format. 6 */ 7 8 #include "src/mat/impls/baij/seq/baij.h" 9 #include "src/vec/vecimpl.h" 10 #include "src/inline/ilu.h" 11 12 13 /* 14 The symbolic factorization code is identical to that for AIJ format, 15 except for very small changes since this is now a SeqBAIJ datastructure. 16 NOT good code reuse. 17 */ 18 #undef __FUNC__ 19 #define __FUNC__ "MatLUFactorSymbolic_SeqBAIJ" 20 int MatLUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 21 { 22 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 23 IS isicol; 24 int *r,*ic, ierr, i, n = a->mbs, *ai = a->i, *aj = a->j; 25 int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, bs = a->bs, bs2=a->bs2; 26 int *idnew, idx, row,m,fm, nnz, nzi,realloc = 0,nzbd,*im; 27 28 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 PLogInfo(A, 112 "Info:MatLUFactorSymbolic_SeqBAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 113 realloc,f,((double)ainew[n])/((double)ai[i])); 114 115 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 116 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 117 118 PetscFree(fill); 119 120 /* put together the new matrix */ 121 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,B); CHKERRQ(ierr); 122 PLogObjectParent(*B,isicol); 123 ierr = ISDestroy(isicol); CHKERRQ(ierr); 124 b = (Mat_SeqBAIJ *) (*B)->data; 125 PetscFree(b->imax); 126 b->singlemalloc = 0; 127 /* the next line frees the default space generated by the Create() */ 128 PetscFree(b->a); PetscFree(b->ilen); 129 b->a = (Scalar *) PetscMalloc((ainew[n]+1)*sizeof(Scalar)*bs2);CHKPTRQ(b->a); 130 b->j = ajnew; 131 b->i = ainew; 132 b->diag = idnew; 133 b->ilen = 0; 134 b->imax = 0; 135 b->row = isrow; 136 b->col = iscol; 137 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 138 CHKPTRQ(b->solve_work); 139 /* In b structure: Free imax, ilen, old a, old j. 140 Allocate idnew, solve_work, new a, new j */ 141 PLogObjectMemory(*B,(ainew[n]-n)*(sizeof(int)+sizeof(Scalar))); 142 b->maxnz = b->nz = ainew[n]; 143 144 (*B)->info.factor_mallocs = realloc; 145 (*B)->info.fill_ratio_given = f; 146 (*B)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[i]); 147 148 return 0; 149 } 150 151 /* ----------------------------------------------------------- */ 152 #undef __FUNC__ 153 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_N" 154 int MatLUFactorNumeric_SeqBAIJ_N(Mat A,Mat *B) 155 { 156 Mat C = *B; 157 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 158 IS iscol = b->col, isrow = b->row, isicol; 159 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 160 int *ajtmpold, *ajtmp, nz, row, bslog,*ai=a->i,*aj=a->j,k,flg; 161 int *diag_offset=b->diag,diag,bs=a->bs,bs2 = a->bs2,*v_pivots; 162 Scalar *ba = b->a,*aa = a->a; 163 register Scalar *pv,*v,*rtmp,*multiplier,*v_work,*pc,*w; 164 register int *pj; 165 166 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 167 PLogObjectParent(*B,isicol); 168 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 169 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 170 rtmp = (Scalar *) PetscMalloc(bs2*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 171 PetscMemzero(rtmp,bs2*(n+1)*sizeof(Scalar)); 172 /* generate work space needed by dense LU factorization */ 173 v_work = (Scalar *) PetscMalloc(bs*sizeof(int) + (bs+bs2)*sizeof(Scalar)); 174 CHKPTRQ(v_work); 175 multiplier = v_work + bs; 176 v_pivots = (int *) (multiplier + bs2); 177 178 /* flops in while loop */ 179 bslog = 2*bs*bs2; 180 181 for ( i=0; i<n; i++ ) { 182 nz = bi[i+1] - bi[i]; 183 ajtmp = bj + bi[i]; 184 for ( j=0; j<nz; j++ ) { 185 PetscMemzero(rtmp+bs2*ajtmp[j],bs2*sizeof(Scalar)); 186 } 187 /* load in initial (unfactored row) */ 188 nz = ai[r[i]+1] - ai[r[i]]; 189 ajtmpold = aj + ai[r[i]]; 190 v = aa + bs2*ai[r[i]]; 191 for ( j=0; j<nz; j++ ) { 192 PetscMemcpy(rtmp+bs2*ic[ajtmpold[j]],v+bs2*j,bs2*sizeof(Scalar)); 193 } 194 row = *ajtmp++; 195 while (row < i) { 196 pc = rtmp + bs2*row; 197 /* if (*pc) { */ 198 for ( flg=0,k=0; k<bs2; k++ ) { if (pc[k]!=0.0) { flg =1; break; }} 199 if (flg) { 200 pv = ba + bs2*diag_offset[row]; 201 pj = bj + diag_offset[row] + 1; 202 Kernel_A_gets_A_times_B(bs,pc,pv,multiplier); 203 nz = bi[row+1] - diag_offset[row] - 1; 204 pv += bs2; 205 for (j=0; j<nz; j++) { 206 Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); 207 } 208 PLogFlops(bslog*(nz+1)-bs); 209 } 210 row = *ajtmp++; 211 } 212 /* finished row so stick it into b->a */ 213 pv = ba + bs2*bi[i]; 214 pj = bj + bi[i]; 215 nz = bi[i+1] - bi[i]; 216 for ( j=0; j<nz; j++ ) { 217 PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(Scalar)); 218 } 219 diag = diag_offset[i] - bi[i]; 220 /* invert diagonal block */ 221 w = pv + bs2*diag; 222 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 223 } 224 225 PetscFree(rtmp); PetscFree(v_work); 226 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 227 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 228 ierr = ISDestroy(isicol); CHKERRQ(ierr); 229 C->factor = FACTOR_LU; 230 C->assembled = PETSC_TRUE; 231 PLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */ 232 return 0; 233 } 234 /* ------------------------------------------------------------*/ 235 /* 236 Version for when blocks are 5 by 5 237 */ 238 #undef __FUNC__ 239 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_5" 240 int MatLUFactorNumeric_SeqBAIJ_5(Mat A,Mat *B) 241 { 242 Mat C = *B; 243 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 244 IS iscol = b->col, isrow = b->row, isicol; 245 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 246 int *ajtmpold, *ajtmp, nz, row, v_pivots[5]; 247 int *diag_offset = b->diag,bs = 5,idx,*ai=a->i,*aj=a->j; 248 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 249 Scalar p1,p2,p3,p4,v_work[5],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 250 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 251 Scalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 252 Scalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 253 Scalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 254 Scalar *ba = b->a,*aa = a->a; 255 register int *pj; 256 257 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 258 PLogObjectParent(*B,isicol); 259 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 260 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 261 rtmp = (Scalar *) PetscMalloc(25*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 262 263 for ( i=0; i<n; i++ ) { 264 nz = bi[i+1] - bi[i]; 265 ajtmp = bj + bi[i]; 266 for ( j=0; j<nz; j++ ) { 267 x = rtmp+25*ajtmp[j]; 268 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 269 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 270 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 271 } 272 /* load in initial (unfactored row) */ 273 idx = r[i]; 274 nz = ai[idx+1] - ai[idx]; 275 ajtmpold = aj + ai[idx]; 276 v = aa + 25*ai[idx]; 277 for ( j=0; j<nz; j++ ) { 278 x = rtmp+25*ic[ajtmpold[j]]; 279 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 280 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 281 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 282 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 283 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 284 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 285 v += 25; 286 } 287 row = *ajtmp++; 288 while (row < i) { 289 pc = rtmp + 25*row; 290 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 291 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 292 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 293 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 294 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 295 p25 = pc[24]; 296 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 297 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 298 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 299 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 300 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 301 p24 != 0.0 || p25 != 0.0) { 302 pv = ba + 25*diag_offset[row]; 303 pj = bj + diag_offset[row] + 1; 304 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 305 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 306 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 307 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 308 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 309 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 310 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 311 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 312 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 313 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 314 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 315 316 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 317 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 318 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 319 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 320 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 321 322 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 323 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 324 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 325 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 326 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 327 328 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 329 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 330 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 331 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 332 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 333 334 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 335 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 336 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 337 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 338 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 339 340 nz = bi[row+1] - diag_offset[row] - 1; 341 pv += 25; 342 for (j=0; j<nz; j++) { 343 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 344 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 345 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 346 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 347 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 348 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 349 x = rtmp + 25*pj[j]; 350 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 351 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 352 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 353 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 354 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 355 356 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 357 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 358 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 359 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 360 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 361 362 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 363 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 364 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 365 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 366 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 367 368 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 369 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 370 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 371 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 372 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 373 374 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 375 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 376 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 377 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 378 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 379 380 pv += 25; 381 } 382 PLogFlops(250*nz+225); 383 } 384 row = *ajtmp++; 385 } 386 /* finished row so stick it into b->a */ 387 pv = ba + 25*bi[i]; 388 pj = bj + bi[i]; 389 nz = bi[i+1] - bi[i]; 390 for ( j=0; j<nz; j++ ) { 391 x = rtmp+25*pj[j]; 392 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 393 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 394 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 395 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 396 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 397 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 398 pv += 25; 399 } 400 /* invert diagonal block */ 401 w = ba + 25*diag_offset[i]; 402 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 403 } 404 405 PetscFree(rtmp); 406 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 407 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 408 ierr = ISDestroy(isicol); CHKERRQ(ierr); 409 C->factor = FACTOR_LU; 410 C->assembled = PETSC_TRUE; 411 PLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */ 412 return 0; 413 } 414 415 /* ------------------------------------------------------------*/ 416 /* 417 Version for when blocks are 4 by 4 418 */ 419 #undef __FUNC__ 420 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_4" 421 int MatLUFactorNumeric_SeqBAIJ_4(Mat A,Mat *B) 422 { 423 Mat C = *B; 424 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 425 IS iscol = b->col, isrow = b->row, isicol; 426 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 427 int *ajtmpold, *ajtmp, nz, row, v_pivots[4]; 428 int *diag_offset = b->diag,bs = 4,idx,*ai=a->i,*aj=a->j; 429 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 430 Scalar p1,p2,p3,p4,v_work[4],m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 431 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 432 Scalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12; 433 Scalar m13,m14,m15,m16; 434 Scalar *ba = b->a,*aa = a->a; 435 register int *pj; 436 437 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 438 PLogObjectParent(*B,isicol); 439 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 440 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 441 rtmp = (Scalar *) PetscMalloc(16*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 442 443 for ( i=0; i<n; i++ ) { 444 nz = bi[i+1] - bi[i]; 445 ajtmp = bj + bi[i]; 446 for ( j=0; j<nz; j++ ) { 447 x = rtmp+16*ajtmp[j]; 448 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 449 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 450 } 451 /* load in initial (unfactored row) */ 452 idx = r[i]; 453 nz = ai[idx+1] - ai[idx]; 454 ajtmpold = aj + ai[idx]; 455 v = aa + 16*ai[idx]; 456 for ( j=0; j<nz; j++ ) { 457 x = rtmp+16*ic[ajtmpold[j]]; 458 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 459 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 460 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 461 x[14] = v[14]; x[15] = v[15]; 462 v += 16; 463 } 464 row = *ajtmp++; 465 while (row < i) { 466 pc = rtmp + 16*row; 467 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 468 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 469 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 470 p15 = pc[14]; p16 = pc[15]; 471 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 472 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 473 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 474 || p16 != 0.0) { 475 pv = ba + 16*diag_offset[row]; 476 pj = bj + diag_offset[row] + 1; 477 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 478 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 479 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 480 x15 = pv[14]; x16 = pv[15]; 481 pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4; 482 pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4; 483 pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4; 484 pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4; 485 486 pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8; 487 pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8; 488 pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8; 489 pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8; 490 491 pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12; 492 pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12; 493 pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12; 494 pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12; 495 496 pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16; 497 pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16; 498 pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16; 499 pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16; 500 501 nz = bi[row+1] - diag_offset[row] - 1; 502 pv += 16; 503 for (j=0; j<nz; j++) { 504 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 505 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 506 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 507 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; 508 x = rtmp + 16*pj[j]; 509 x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4; 510 x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4; 511 x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4; 512 x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4; 513 514 x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8; 515 x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8; 516 x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8; 517 x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8; 518 519 x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12; 520 x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12; 521 x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12; 522 x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12; 523 524 x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16; 525 x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16; 526 x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16; 527 x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16; 528 529 pv += 16; 530 } 531 PLogFlops(128*nz+112); 532 } 533 row = *ajtmp++; 534 } 535 /* finished row so stick it into b->a */ 536 pv = ba + 16*bi[i]; 537 pj = bj + bi[i]; 538 nz = bi[i+1] - bi[i]; 539 for ( j=0; j<nz; j++ ) { 540 x = rtmp+16*pj[j]; 541 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 542 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 543 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 544 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; 545 pv += 16; 546 } 547 /* invert diagonal block */ 548 w = ba + 16*diag_offset[i]; 549 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 550 } 551 552 PetscFree(rtmp); 553 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 554 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 555 ierr = ISDestroy(isicol); CHKERRQ(ierr); 556 C->factor = FACTOR_LU; 557 C->assembled = PETSC_TRUE; 558 PLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 559 return 0; 560 } 561 /* ------------------------------------------------------------*/ 562 /* 563 Version for when blocks are 3 by 3 564 */ 565 #undef __FUNC__ 566 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_3" 567 int MatLUFactorNumeric_SeqBAIJ_3(Mat A,Mat *B) 568 { 569 Mat C = *B; 570 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 571 IS iscol = b->col, isrow = b->row, isicol; 572 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 573 int *ajtmpold, *ajtmp, nz, row, *ai=a->i,*aj=a->j; 574 int *diag_offset = b->diag,idx; 575 register Scalar *pv,*v,*rtmp,*pc,*w,*x; 576 Scalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 577 Scalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9; 578 Scalar *ba = b->a,*aa = a->a; 579 register int *pj; 580 581 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 582 PLogObjectParent(*B,isicol); 583 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 584 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 585 rtmp = (Scalar *) PetscMalloc(9*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 586 587 for ( i=0; i<n; i++ ) { 588 nz = bi[i+1] - bi[i]; 589 ajtmp = bj + bi[i]; 590 for ( j=0; j<nz; j++ ) { 591 x = rtmp + 9*ajtmp[j]; 592 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 593 } 594 /* load in initial (unfactored row) */ 595 idx = r[i]; 596 nz = ai[idx+1] - ai[idx]; 597 ajtmpold = aj + ai[idx]; 598 v = aa + 9*ai[idx]; 599 for ( j=0; j<nz; j++ ) { 600 x = rtmp + 9*ic[ajtmpold[j]]; 601 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 602 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 603 v += 9; 604 } 605 row = *ajtmp++; 606 while (row < i) { 607 pc = rtmp + 9*row; 608 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 609 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 610 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 611 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) { 612 pv = ba + 9*diag_offset[row]; 613 pj = bj + diag_offset[row] + 1; 614 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 615 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 616 pc[0] = m1 = p1*x1 + p4*x2 + p7*x3; 617 pc[1] = m2 = p2*x1 + p5*x2 + p8*x3; 618 pc[2] = m3 = p3*x1 + p6*x2 + p9*x3; 619 620 pc[3] = m4 = p1*x4 + p4*x5 + p7*x6; 621 pc[4] = m5 = p2*x4 + p5*x5 + p8*x6; 622 pc[5] = m6 = p3*x4 + p6*x5 + p9*x6; 623 624 pc[6] = m7 = p1*x7 + p4*x8 + p7*x9; 625 pc[7] = m8 = p2*x7 + p5*x8 + p8*x9; 626 pc[8] = m9 = p3*x7 + p6*x8 + p9*x9; 627 nz = bi[row+1] - diag_offset[row] - 1; 628 pv += 9; 629 for (j=0; j<nz; j++) { 630 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 631 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 632 x = rtmp + 9*pj[j]; 633 x[0] -= m1*x1 + m4*x2 + m7*x3; 634 x[1] -= m2*x1 + m5*x2 + m8*x3; 635 x[2] -= m3*x1 + m6*x2 + m9*x3; 636 637 x[3] -= m1*x4 + m4*x5 + m7*x6; 638 x[4] -= m2*x4 + m5*x5 + m8*x6; 639 x[5] -= m3*x4 + m6*x5 + m9*x6; 640 641 x[6] -= m1*x7 + m4*x8 + m7*x9; 642 x[7] -= m2*x7 + m5*x8 + m8*x9; 643 x[8] -= m3*x7 + m6*x8 + m9*x9; 644 pv += 9; 645 } 646 PLogFlops(54*nz+36); 647 } 648 row = *ajtmp++; 649 } 650 /* finished row so stick it into b->a */ 651 pv = ba + 9*bi[i]; 652 pj = bj + bi[i]; 653 nz = bi[i+1] - bi[i]; 654 for ( j=0; j<nz; j++ ) { 655 x = rtmp + 9*pj[j]; 656 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 657 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 658 pv += 9; 659 } 660 /* invert diagonal block */ 661 w = ba + 9*diag_offset[i]; 662 ierr = Kernel_A_gets_inverse_A_3(w); CHKERRQ(ierr); 663 } 664 665 PetscFree(rtmp); 666 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 667 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 668 ierr = ISDestroy(isicol); CHKERRQ(ierr); 669 C->factor = FACTOR_LU; 670 C->assembled = PETSC_TRUE; 671 PLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */ 672 return 0; 673 } 674 675 /* ------------------------------------------------------------*/ 676 /* 677 Version for when blocks are 2 by 2 678 */ 679 #undef __FUNC__ 680 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_2" 681 int MatLUFactorNumeric_SeqBAIJ_2(Mat A,Mat *B) 682 { 683 Mat C = *B; 684 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*b = (Mat_SeqBAIJ *)C->data; 685 IS iscol = b->col, isrow = b->row, isicol; 686 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 687 int *ajtmpold, *ajtmp, nz, row, v_pivots[2]; 688 int *diag_offset=b->diag,bs = 2,idx,*ai=a->i,*aj=a->j; 689 register Scalar *pv,*v,*rtmp,m1,m2,m3,m4,*pc,*w,*x,x1,x2,x3,x4; 690 Scalar p1,p2,p3,p4,v_work[2]; 691 Scalar *ba = b->a,*aa = a->a; 692 register int *pj; 693 694 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 695 PLogObjectParent(*B,isicol); 696 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 697 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 698 rtmp = (Scalar *) PetscMalloc(4*(n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 699 700 for ( i=0; i<n; i++ ) { 701 nz = bi[i+1] - bi[i]; 702 ajtmp = bj + bi[i]; 703 for ( j=0; j<nz; j++ ) { 704 x = rtmp+4*ajtmp[j]; x[0] = x[1] = x[2] = x[3] = 0.0; 705 } 706 /* load in initial (unfactored row) */ 707 idx = r[i]; 708 nz = ai[idx+1] - ai[idx]; 709 ajtmpold = aj + ai[idx]; 710 v = aa + 4*ai[idx]; 711 for ( j=0; j<nz; j++ ) { 712 x = rtmp+4*ic[ajtmpold[j]]; 713 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 714 v += 4; 715 } 716 row = *ajtmp++; 717 while (row < i) { 718 pc = rtmp + 4*row; 719 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 720 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0) { 721 pv = ba + 4*diag_offset[row]; 722 pj = bj + diag_offset[row] + 1; 723 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 724 pc[0] = m1 = p1*x1 + p3*x2; 725 pc[1] = m2 = p2*x1 + p4*x2; 726 pc[2] = m3 = p1*x3 + p3*x4; 727 pc[3] = m4 = p2*x3 + p4*x4; 728 nz = bi[row+1] - diag_offset[row] - 1; 729 pv += 4; 730 for (j=0; j<nz; j++) { 731 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 732 x = rtmp + 4*pj[j]; 733 x[0] -= m1*x1 + m3*x2; 734 x[1] -= m2*x1 + m4*x2; 735 x[2] -= m1*x3 + m3*x4; 736 x[3] -= m2*x3 + m4*x4; 737 pv += 4; 738 } 739 PLogFlops(16*nz+12); 740 } 741 row = *ajtmp++; 742 } 743 /* finished row so stick it into b->a */ 744 pv = ba + 4*bi[i]; 745 pj = bj + bi[i]; 746 nz = bi[i+1] - bi[i]; 747 for ( j=0; j<nz; j++ ) { 748 x = rtmp+4*pj[j]; 749 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 750 pv += 4; 751 } 752 /* invert diagonal block */ 753 w = ba + 4*diag_offset[i]; 754 Kernel_A_gets_inverse_A(bs,w,v_pivots,v_work); 755 } 756 757 PetscFree(rtmp); 758 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 759 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 760 ierr = ISDestroy(isicol); CHKERRQ(ierr); 761 C->factor = FACTOR_LU; 762 C->assembled = PETSC_TRUE; 763 PLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */ 764 return 0; 765 } 766 767 /* ----------------------------------------------------------- */ 768 /* 769 Version for when blocks are 1 by 1. 770 */ 771 #undef __FUNC__ 772 #define __FUNC__ "MatLUFactorNumeric_SeqBAIJ_1" 773 int MatLUFactorNumeric_SeqBAIJ_1(Mat A,Mat *B) 774 { 775 Mat C = *B; 776 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b = (Mat_SeqBAIJ *)C->data; 777 IS iscol = b->col, isrow = b->row, isicol; 778 int *r,*ic, ierr, i, j, n = a->mbs, *bi = b->i, *bj = b->j; 779 int *ajtmpold, *ajtmp, nz, row,*ai = a->i,*aj = a->j; 780 int *diag_offset = b->diag,diag; 781 register Scalar *pv,*v,*rtmp,multiplier,*pc; 782 Scalar *ba = b->a,*aa = a->a; 783 register int *pj; 784 785 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 786 PLogObjectParent(*B,isicol); 787 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 788 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 789 rtmp = (Scalar *) PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 790 791 for ( i=0; i<n; i++ ) { 792 nz = bi[i+1] - bi[i]; 793 ajtmp = bj + bi[i]; 794 for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]] = 0.0; 795 796 /* load in initial (unfactored row) */ 797 nz = ai[r[i]+1] - ai[r[i]]; 798 ajtmpold = aj + ai[r[i]]; 799 v = aa + ai[r[i]]; 800 for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]]] = v[j]; 801 802 row = *ajtmp++; 803 while (row < i) { 804 pc = rtmp + row; 805 if (*pc != 0.0) { 806 pv = ba + diag_offset[row]; 807 pj = bj + diag_offset[row] + 1; 808 multiplier = *pc * *pv++; 809 *pc = multiplier; 810 nz = bi[row+1] - diag_offset[row] - 1; 811 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 812 PLogFlops(1+2*nz); 813 } 814 row = *ajtmp++; 815 } 816 /* finished row so stick it into b->a */ 817 pv = ba + bi[i]; 818 pj = bj + bi[i]; 819 nz = bi[i+1] - bi[i]; 820 for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]];} 821 diag = diag_offset[i] - bi[i]; 822 /* check pivot entry for current row */ 823 if (pv[diag] == 0.0) { 824 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"Zero pivot"); 825 } 826 pv[diag] = 1.0/pv[diag]; 827 } 828 829 PetscFree(rtmp); 830 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 831 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 832 ierr = ISDestroy(isicol); CHKERRQ(ierr); 833 C->factor = FACTOR_LU; 834 C->assembled = PETSC_TRUE; 835 PLogFlops(b->n); 836 return 0; 837 } 838 839 /* ----------------------------------------------------------- */ 840 #undef __FUNC__ 841 #define __FUNC__ "MatLUFactor_SeqBAIJ" 842 int MatLUFactor_SeqBAIJ(Mat A,IS row,IS col,double f) 843 { 844 Mat_SeqBAIJ *mat = (Mat_SeqBAIJ *) A->data; 845 int ierr; 846 Mat C; 847 848 ierr = MatLUFactorSymbolic(A,row,col,f,&C); CHKERRQ(ierr); 849 ierr = MatLUFactorNumeric(A,&C); CHKERRQ(ierr); 850 851 /* free all the data structures from mat */ 852 PetscFree(mat->a); 853 if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 854 if (mat->diag) PetscFree(mat->diag); 855 if (mat->ilen) PetscFree(mat->ilen); 856 if (mat->imax) PetscFree(mat->imax); 857 if (mat->solve_work) PetscFree(mat->solve_work); 858 if (mat->mult_work) PetscFree(mat->mult_work); 859 PetscFree(mat); 860 861 PetscMemcpy(A,C,sizeof(struct _Mat)); 862 PetscHeaderDestroy(C); 863 return 0; 864 } 865 /* ----------------------------------------------------------- */ 866 #undef __FUNC__ 867 #define __FUNC__ "MatSolve_SeqBAIJ_N" 868 int MatSolve_SeqBAIJ_N(Mat A,Vec bb,Vec xx) 869 { 870 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 871 IS iscol=a->col,isrow=a->row; 872 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j; 873 int nz,bs=a->bs,bs2=a->bs2,*rout,*cout; 874 Scalar *aa=a->a,*sum; 875 register Scalar *x,*b,*lsum,*tmp,*v; 876 877 VecGetArray_Fast(bb,b); 878 VecGetArray_Fast(xx,x); 879 tmp = a->solve_work; 880 881 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 882 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 883 884 /* forward solve the lower triangular */ 885 PetscMemcpy(tmp,b + bs*(*r++), bs*sizeof(Scalar)); 886 for ( i=1; i<n; i++ ) { 887 v = aa + bs2*ai[i]; 888 vi = aj + ai[i]; 889 nz = a->diag[i] - ai[i]; 890 sum = tmp + bs*i; 891 PetscMemcpy(sum,b+bs*(*r++),bs*sizeof(Scalar)); 892 while (nz--) { 893 Kernel_v_gets_v_minus_A_times_w(bs,sum,v,tmp+bs*(*vi++)); 894 v += bs2; 895 } 896 } 897 /* backward solve the upper triangular */ 898 lsum = a->solve_work + a->n; 899 for ( i=n-1; i>=0; i-- ){ 900 v = aa + bs2*(a->diag[i] + 1); 901 vi = aj + a->diag[i] + 1; 902 nz = ai[i+1] - a->diag[i] - 1; 903 PetscMemcpy(lsum,tmp+i*bs,bs*sizeof(Scalar)); 904 while (nz--) { 905 Kernel_v_gets_v_minus_A_times_w(bs,lsum,v,tmp+bs*(*vi++)); 906 v += bs2; 907 } 908 Kernel_w_gets_A_times_v(bs,lsum,aa+bs2*a->diag[i],tmp+i*bs); 909 PetscMemcpy(x + bs*(*c--),tmp+i*bs,bs*sizeof(Scalar)); 910 } 911 912 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 913 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 914 VecRestoreArray_Fast(bb,b); 915 VecRestoreArray_Fast(xx,x); 916 PLogFlops(2*(a->bs2)*(a->nz) - a->n); 917 return 0; 918 } 919 920 #undef __FUNC__ 921 #define __FUNC__ "MatSolve_SeqBAIJ_7" 922 int MatSolve_SeqBAIJ_7(Mat A,Vec bb,Vec xx) 923 { 924 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 925 IS iscol=a->col,isrow=a->row; 926 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 927 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,sum6,sum7,x1,x2,x3,x4,x5,x6,x7; 928 register Scalar *x,*b,*tmp,*v; 929 930 VecGetArray_Fast(bb,b); 931 VecGetArray_Fast(xx,x); 932 tmp = a->solve_work; 933 934 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 935 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 936 937 /* forward solve the lower triangular */ 938 idx = 7*(*r++); 939 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 940 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 941 tmp[5] = b[5+idx]; tmp[6] = b[6+idx]; 942 943 for ( i=1; i<n; i++ ) { 944 v = aa + 49*ai[i]; 945 vi = aj + ai[i]; 946 nz = a->diag[i] - ai[i]; 947 idx = 7*(*r++); 948 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 949 sum5 = b[4+idx];sum6 = b[5+idx];sum7 = b[6+idx]; 950 while (nz--) { 951 idx = 7*(*vi++); 952 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 953 x4 = tmp[3+idx];x5 = tmp[4+idx]; 954 x6 = tmp[5+idx];x7 = tmp[6+idx]; 955 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 956 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 957 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 958 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 959 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 960 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 961 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 962 v += 49; 963 } 964 idx = 7*i; 965 tmp[idx] = sum1;tmp[1+idx] = sum2; 966 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 967 tmp[5+idx] = sum6;tmp[6+idx] = sum7; 968 } 969 /* backward solve the upper triangular */ 970 for ( i=n-1; i>=0; i-- ){ 971 v = aa + 49*a->diag[i] + 49; 972 vi = aj + a->diag[i] + 1; 973 nz = ai[i+1] - a->diag[i] - 1; 974 idt = 7*i; 975 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 976 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 977 sum6 = tmp[5+idt];sum7 = tmp[6+idt]; 978 while (nz--) { 979 idx = 7*(*vi++); 980 x1 = tmp[idx]; x2 = tmp[1+idx]; 981 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 982 x6 = tmp[5+idx]; x7 = tmp[6+idx]; 983 sum1 -= v[0]*x1 + v[7]*x2 + v[14]*x3 + v[21]*x4 + v[28]*x5 + v[35]*x6 + v[42]*x7; 984 sum2 -= v[1]*x1 + v[8]*x2 + v[15]*x3 + v[22]*x4 + v[29]*x5 + v[36]*x6 + v[43]*x7; 985 sum3 -= v[2]*x1 + v[9]*x2 + v[16]*x3 + v[23]*x4 + v[30]*x5 + v[37]*x6 + v[44]*x7; 986 sum4 -= v[3]*x1 + v[10]*x2 + v[17]*x3 + v[24]*x4 + v[31]*x5 + v[38]*x6 + v[45]*x7; 987 sum5 -= v[4]*x1 + v[11]*x2 + v[18]*x3 + v[25]*x4 + v[32]*x5 + v[39]*x6 + v[46]*x7; 988 sum6 -= v[5]*x1 + v[12]*x2 + v[19]*x3 + v[26]*x4 + v[33]*x5 + v[40]*x6 + v[47]*x7; 989 sum7 -= v[6]*x1 + v[13]*x2 + v[20]*x3 + v[27]*x4 + v[34]*x5 + v[41]*x6 + v[48]*x7; 990 v += 49; 991 } 992 idc = 7*(*c--); 993 v = aa + 49*a->diag[i]; 994 x[idc] = tmp[idt] = v[0]*sum1+v[7]*sum2+v[14]*sum3+ 995 v[21]*sum4+v[28]*sum5+v[35]*sum6+v[42]*sum7; 996 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[8]*sum2+v[15]*sum3+ 997 v[22]*sum4+v[29]*sum5+v[36]*sum6+v[43]*sum7; 998 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[9]*sum2+v[16]*sum3+ 999 v[23]*sum4+v[30]*sum5+v[37]*sum6+v[44]*sum7; 1000 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[10]*sum2+v[17]*sum3+ 1001 v[24]*sum4+v[31]*sum5+v[38]*sum6+v[45]*sum7; 1002 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[11]*sum2+v[18]*sum3+ 1003 v[25]*sum4+v[32]*sum5+v[39]*sum6+v[46]*sum7; 1004 x[5+idc] = tmp[5+idt] = v[5]*sum1+v[12]*sum2+v[19]*sum3+ 1005 v[26]*sum4+v[33]*sum5+v[40]*sum6+v[47]*sum7; 1006 x[6+idc] = tmp[6+idt] = v[6]*sum1+v[13]*sum2+v[20]*sum3+ 1007 v[27]*sum4+v[34]*sum5+v[41]*sum6+v[48]*sum7; 1008 } 1009 1010 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1011 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1012 VecRestoreArray_Fast(bb,b); 1013 VecRestoreArray_Fast(xx,x); 1014 PLogFlops(2*49*(a->nz) - a->n); 1015 return 0; 1016 } 1017 1018 #undef __FUNC__ 1019 #define __FUNC__ "MatSolve_SeqBAIJ_5" 1020 int MatSolve_SeqBAIJ_5(Mat A,Vec bb,Vec xx) 1021 { 1022 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1023 IS iscol=a->col,isrow=a->row; 1024 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1025 Scalar *aa=a->a,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5; 1026 register Scalar *x,*b,*tmp,*v; 1027 1028 VecGetArray_Fast(bb,b); 1029 VecGetArray_Fast(xx,x); 1030 tmp = a->solve_work; 1031 1032 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1033 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1034 1035 /* forward solve the lower triangular */ 1036 idx = 5*(*r++); 1037 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1038 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; tmp[4] = b[4+idx]; 1039 for ( i=1; i<n; i++ ) { 1040 v = aa + 25*ai[i]; 1041 vi = aj + ai[i]; 1042 nz = a->diag[i] - ai[i]; 1043 idx = 5*(*r++); 1044 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1045 sum5 = b[4+idx]; 1046 while (nz--) { 1047 idx = 5*(*vi++); 1048 x1 = tmp[idx]; x2 = tmp[1+idx];x3 = tmp[2+idx]; 1049 x4 = tmp[3+idx];x5 = tmp[4+idx]; 1050 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1051 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1052 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1053 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1054 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1055 v += 25; 1056 } 1057 idx = 5*i; 1058 tmp[idx] = sum1;tmp[1+idx] = sum2; 1059 tmp[2+idx] = sum3;tmp[3+idx] = sum4; tmp[4+idx] = sum5; 1060 } 1061 /* backward solve the upper triangular */ 1062 for ( i=n-1; i>=0; i-- ){ 1063 v = aa + 25*a->diag[i] + 25; 1064 vi = aj + a->diag[i] + 1; 1065 nz = ai[i+1] - a->diag[i] - 1; 1066 idt = 5*i; 1067 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1068 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; sum5 = tmp[4+idt]; 1069 while (nz--) { 1070 idx = 5*(*vi++); 1071 x1 = tmp[idx]; x2 = tmp[1+idx]; 1072 x3 = tmp[2+idx]; x4 = tmp[3+idx]; x5 = tmp[4+idx]; 1073 sum1 -= v[0]*x1 + v[5]*x2 + v[10]*x3 + v[15]*x4 + v[20]*x5; 1074 sum2 -= v[1]*x1 + v[6]*x2 + v[11]*x3 + v[16]*x4 + v[21]*x5; 1075 sum3 -= v[2]*x1 + v[7]*x2 + v[12]*x3 + v[17]*x4 + v[22]*x5; 1076 sum4 -= v[3]*x1 + v[8]*x2 + v[13]*x3 + v[18]*x4 + v[23]*x5; 1077 sum5 -= v[4]*x1 + v[9]*x2 + v[14]*x3 + v[19]*x4 + v[24]*x5; 1078 v += 25; 1079 } 1080 idc = 5*(*c--); 1081 v = aa + 25*a->diag[i]; 1082 x[idc] = tmp[idt] = v[0]*sum1+v[5]*sum2+v[10]*sum3+ 1083 v[15]*sum4+v[20]*sum5; 1084 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[6]*sum2+v[11]*sum3+ 1085 v[16]*sum4+v[21]*sum5; 1086 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[7]*sum2+v[12]*sum3+ 1087 v[17]*sum4+v[22]*sum5; 1088 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[8]*sum2+v[13]*sum3+ 1089 v[18]*sum4+v[23]*sum5; 1090 x[4+idc] = tmp[4+idt] = v[4]*sum1+v[9]*sum2+v[14]*sum3+ 1091 v[19]*sum4+v[24]*sum5; 1092 } 1093 1094 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1095 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1096 VecRestoreArray_Fast(bb,b); 1097 VecRestoreArray_Fast(xx,x); 1098 PLogFlops(2*25*(a->nz) - a->n); 1099 return 0; 1100 } 1101 1102 #undef __FUNC__ 1103 #define __FUNC__ "MatSolve_SeqBAIJ_4" 1104 int MatSolve_SeqBAIJ_4(Mat A,Vec bb,Vec xx) 1105 { 1106 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1107 IS iscol=a->col,isrow=a->row; 1108 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1109 Scalar *aa=a->a,sum1,sum2,sum3,sum4,x1,x2,x3,x4; 1110 register Scalar *x,*b,*tmp,*v; 1111 1112 VecGetArray_Fast(bb,b); 1113 VecGetArray_Fast(xx,x); 1114 tmp = a->solve_work; 1115 1116 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1117 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1118 1119 /* forward solve the lower triangular */ 1120 idx = 4*(*r++); 1121 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1122 tmp[2] = b[2+idx]; tmp[3] = b[3+idx]; 1123 for ( i=1; i<n; i++ ) { 1124 v = aa + 16*ai[i]; 1125 vi = aj + ai[i]; 1126 nz = a->diag[i] - ai[i]; 1127 idx = 4*(*r++); 1128 sum1 = b[idx];sum2 = b[1+idx];sum3 = b[2+idx];sum4 = b[3+idx]; 1129 while (nz--) { 1130 idx = 4*(*vi++); 1131 x1 = tmp[idx];x2 = tmp[1+idx];x3 = tmp[2+idx];x4 = tmp[3+idx]; 1132 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1133 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1134 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1135 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1136 v += 16; 1137 } 1138 idx = 4*i; 1139 tmp[idx] = sum1;tmp[1+idx] = sum2; 1140 tmp[2+idx] = sum3;tmp[3+idx] = sum4; 1141 } 1142 /* backward solve the upper triangular */ 1143 for ( i=n-1; i>=0; i-- ){ 1144 v = aa + 16*a->diag[i] + 16; 1145 vi = aj + a->diag[i] + 1; 1146 nz = ai[i+1] - a->diag[i] - 1; 1147 idt = 4*i; 1148 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1149 sum3 = tmp[2+idt];sum4 = tmp[3+idt]; 1150 while (nz--) { 1151 idx = 4*(*vi++); 1152 x1 = tmp[idx]; x2 = tmp[1+idx]; 1153 x3 = tmp[2+idx]; x4 = tmp[3+idx]; 1154 sum1 -= v[0]*x1 + v[4]*x2 + v[8]*x3 + v[12]*x4; 1155 sum2 -= v[1]*x1 + v[5]*x2 + v[9]*x3 + v[13]*x4; 1156 sum3 -= v[2]*x1 + v[6]*x2 + v[10]*x3 + v[14]*x4; 1157 sum4 -= v[3]*x1 + v[7]*x2 + v[11]*x3 + v[15]*x4; 1158 v += 16; 1159 } 1160 idc = 4*(*c--); 1161 v = aa + 16*a->diag[i]; 1162 x[idc] = tmp[idt] = v[0]*sum1+v[4]*sum2+v[8]*sum3+v[12]*sum4; 1163 x[1+idc] = tmp[1+idt] = v[1]*sum1+v[5]*sum2+v[9]*sum3+v[13]*sum4; 1164 x[2+idc] = tmp[2+idt] = v[2]*sum1+v[6]*sum2+v[10]*sum3+v[14]*sum4; 1165 x[3+idc] = tmp[3+idt] = v[3]*sum1+v[7]*sum2+v[11]*sum3+v[15]*sum4; 1166 } 1167 1168 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1169 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1170 VecRestoreArray_Fast(bb,b); 1171 VecRestoreArray_Fast(xx,x); 1172 PLogFlops(2*16*(a->nz) - a->n); 1173 return 0; 1174 } 1175 1176 1177 #undef __FUNC__ 1178 #define __FUNC__ "MatSolve_SeqBAIJ_3" 1179 int MatSolve_SeqBAIJ_3(Mat A,Vec bb,Vec xx) 1180 { 1181 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1182 IS iscol=a->col,isrow=a->row; 1183 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1184 Scalar *aa=a->a,sum1,sum2,sum3,x1,x2,x3; 1185 register Scalar *x,*b,*tmp,*v; 1186 1187 VecGetArray_Fast(bb,b); 1188 VecGetArray_Fast(xx,x); 1189 tmp = a->solve_work; 1190 1191 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1192 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1193 1194 /* forward solve the lower triangular */ 1195 idx = 3*(*r++); 1196 tmp[0] = b[idx]; tmp[1] = b[1+idx]; tmp[2] = b[2+idx]; 1197 for ( i=1; i<n; i++ ) { 1198 v = aa + 9*ai[i]; 1199 vi = aj + ai[i]; 1200 nz = a->diag[i] - ai[i]; 1201 idx = 3*(*r++); 1202 sum1 = b[idx]; sum2 = b[1+idx]; sum3 = b[2+idx]; 1203 while (nz--) { 1204 idx = 3*(*vi++); 1205 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1206 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1207 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1208 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1209 v += 9; 1210 } 1211 idx = 3*i; 1212 tmp[idx] = sum1; tmp[1+idx] = sum2; tmp[2+idx] = sum3; 1213 } 1214 /* backward solve the upper triangular */ 1215 for ( i=n-1; i>=0; i-- ){ 1216 v = aa + 9*a->diag[i] + 9; 1217 vi = aj + a->diag[i] + 1; 1218 nz = ai[i+1] - a->diag[i] - 1; 1219 idt = 3*i; 1220 sum1 = tmp[idt]; sum2 = tmp[1+idt]; sum3 = tmp[2+idt]; 1221 while (nz--) { 1222 idx = 3*(*vi++); 1223 x1 = tmp[idx]; x2 = tmp[1+idx]; x3 = tmp[2+idx]; 1224 sum1 -= v[0]*x1 + v[3]*x2 + v[6]*x3; 1225 sum2 -= v[1]*x1 + v[4]*x2 + v[7]*x3; 1226 sum3 -= v[2]*x1 + v[5]*x2 + v[8]*x3; 1227 v += 9; 1228 } 1229 idc = 3*(*c--); 1230 v = aa + 9*a->diag[i]; 1231 x[idc] = tmp[idt] = v[0]*sum1 + v[3]*sum2 + v[6]*sum3; 1232 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[4]*sum2 + v[7]*sum3; 1233 x[2+idc] = tmp[2+idt] = v[2]*sum1 + v[5]*sum2 + v[8]*sum3; 1234 } 1235 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1236 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1237 VecRestoreArray_Fast(bb,b); 1238 VecRestoreArray_Fast(xx,x); 1239 PLogFlops(2*9*(a->nz) - a->n); 1240 return 0; 1241 } 1242 1243 #undef __FUNC__ 1244 #define __FUNC__ "MatSolve_SeqBAIJ_2" 1245 int MatSolve_SeqBAIJ_2(Mat A,Vec bb,Vec xx) 1246 { 1247 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1248 IS iscol=a->col,isrow=a->row; 1249 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,idx,idt,idc,*rout,*cout; 1250 Scalar *aa=a->a,sum1,sum2,x1,x2; 1251 register Scalar *x,*b,*tmp,*v; 1252 1253 VecGetArray_Fast(bb,b); 1254 VecGetArray_Fast(xx,x); 1255 tmp = a->solve_work; 1256 1257 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1258 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1259 1260 /* forward solve the lower triangular */ 1261 idx = 2*(*r++); 1262 tmp[0] = b[idx]; tmp[1] = b[1+idx]; 1263 for ( i=1; i<n; i++ ) { 1264 v = aa + 4*ai[i]; 1265 vi = aj + ai[i]; 1266 nz = a->diag[i] - ai[i]; 1267 idx = 2*(*r++); 1268 sum1 = b[idx]; sum2 = b[1+idx]; 1269 while (nz--) { 1270 idx = 2*(*vi++); 1271 x1 = tmp[idx]; x2 = tmp[1+idx]; 1272 sum1 -= v[0]*x1 + v[2]*x2; 1273 sum2 -= v[1]*x1 + v[3]*x2; 1274 v += 4; 1275 } 1276 idx = 2*i; 1277 tmp[idx] = sum1; tmp[1+idx] = sum2; 1278 } 1279 /* backward solve the upper triangular */ 1280 for ( i=n-1; i>=0; i-- ){ 1281 v = aa + 4*a->diag[i] + 4; 1282 vi = aj + a->diag[i] + 1; 1283 nz = ai[i+1] - a->diag[i] - 1; 1284 idt = 2*i; 1285 sum1 = tmp[idt]; sum2 = tmp[1+idt]; 1286 while (nz--) { 1287 idx = 2*(*vi++); 1288 x1 = tmp[idx]; x2 = tmp[1+idx]; 1289 sum1 -= v[0]*x1 + v[2]*x2; 1290 sum2 -= v[1]*x1 + v[3]*x2; 1291 v += 4; 1292 } 1293 idc = 2*(*c--); 1294 v = aa + 4*a->diag[i]; 1295 x[idc] = tmp[idt] = v[0]*sum1 + v[2]*sum2; 1296 x[1+idc] = tmp[1+idt] = v[1]*sum1 + v[3]*sum2; 1297 } 1298 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1299 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1300 VecRestoreArray_Fast(bb,b); 1301 VecRestoreArray_Fast(xx,x); 1302 PLogFlops(2*4*(a->nz) - a->n); 1303 return 0; 1304 } 1305 1306 1307 #undef __FUNC__ 1308 #define __FUNC__ "MatSolve_SeqBAIJ_1" 1309 int MatSolve_SeqBAIJ_1(Mat A,Vec bb,Vec xx) 1310 { 1311 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data; 1312 IS iscol=a->col,isrow=a->row; 1313 int *r,*c,ierr,i,n=a->mbs,*vi,*ai=a->i,*aj=a->j,nz,*rout,*cout; 1314 Scalar *aa=a->a,sum1; 1315 register Scalar *x,*b,*tmp,*v; 1316 1317 if (!n) return 0; 1318 1319 VecGetArray_Fast(bb,b); 1320 VecGetArray_Fast(xx,x); 1321 tmp = a->solve_work; 1322 1323 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1324 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1325 1326 /* forward solve the lower triangular */ 1327 tmp[0] = b[*r++]; 1328 for ( i=1; i<n; i++ ) { 1329 v = aa + ai[i]; 1330 vi = aj + ai[i]; 1331 nz = a->diag[i] - ai[i]; 1332 sum1 = b[*r++]; 1333 while (nz--) { 1334 sum1 -= (*v++)*tmp[*vi++]; 1335 } 1336 tmp[i] = sum1; 1337 } 1338 /* backward solve the upper triangular */ 1339 for ( i=n-1; i>=0; i-- ){ 1340 v = aa + a->diag[i] + 1; 1341 vi = aj + a->diag[i] + 1; 1342 nz = ai[i+1] - a->diag[i] - 1; 1343 sum1 = tmp[i]; 1344 while (nz--) { 1345 sum1 -= (*v++)*tmp[*vi++]; 1346 } 1347 x[*c--] = tmp[i] = aa[a->diag[i]]*sum1; 1348 } 1349 1350 ierr = ISRestoreIndices(isrow,&rout); CHKERRQ(ierr); 1351 ierr = ISRestoreIndices(iscol,&cout); CHKERRQ(ierr); 1352 VecRestoreArray_Fast(bb,b); 1353 VecRestoreArray_Fast(xx,x); 1354 PLogFlops(2*1*(a->nz) - a->n); 1355 return 0; 1356 } 1357 1358 /* ----------------------------------------------------------------*/ 1359 /* 1360 This code is virtually identical to MatILUFactorSymbolic_SeqAIJ 1361 except that the data structure of Mat_SeqAIJ is slightly different. 1362 Not a good example of code reuse. 1363 */ 1364 #undef __FUNC__ 1365 #define __FUNC__ "MatILUFactorSymbolic_SeqBAIJ" 1366 int MatILUFactorSymbolic_SeqBAIJ(Mat A,IS isrow,IS iscol,double f,int levels, 1367 Mat *fact) 1368 { 1369 Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data, *b; 1370 IS isicol; 1371 int *r,*ic, ierr, prow, n = a->mbs, *ai = a->i, *aj = a->j; 1372 int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 1373 int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 1374 int incrlev,nnz,i,bs = a->bs,bs2 = a->bs2; 1375 PetscTruth col_identity, row_identity; 1376 1377 /* special case that simply copies fill pattern */ 1378 PetscValidHeaderSpecific(isrow,IS_COOKIE); 1379 PetscValidHeaderSpecific(iscol,IS_COOKIE); 1380 ISIdentity(isrow,&row_identity); ISIdentity(iscol,&col_identity); 1381 if (levels == 0 && row_identity && col_identity) { 1382 ierr = MatConvertSameType_SeqBAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 1383 (*fact)->factor = FACTOR_LU; 1384 b = (Mat_SeqBAIJ *) (*fact)->data; 1385 if (!b->diag) { 1386 ierr = MatMarkDiag_SeqBAIJ(*fact); CHKERRQ(ierr); 1387 } 1388 b->row = isrow; 1389 b->col = iscol; 1390 b->solve_work = (Scalar *) PetscMalloc((b->m+1+b->bs)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1391 return 0; 1392 } 1393 1394 ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1395 ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 1396 ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1397 1398 /* get new row pointers */ 1399 ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 1400 ainew[0] = 0; 1401 /* don't know how many column pointers are needed so estimate */ 1402 jmax = (int) (f*ai[n] + 1); 1403 ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 1404 /* ajfill is level of fill for each fill entry */ 1405 ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 1406 /* fill is a linked list of nonzeros in active row */ 1407 fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 1408 /* im is level for each filled value */ 1409 im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 1410 /* dloc is location of diagonal in factor */ 1411 dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 1412 dloc[0] = 0; 1413 for ( prow=0; prow<n; prow++ ) { 1414 /* first copy previous fill into linked list */ 1415 nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 1416 if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,1,"Empty row in matrix"); 1417 xi = aj + ai[r[prow]]; 1418 fill[n] = n; 1419 while (nz--) { 1420 fm = n; 1421 idx = ic[*xi++]; 1422 do { 1423 m = fm; 1424 fm = fill[m]; 1425 } while (fm < idx); 1426 fill[m] = idx; 1427 fill[idx] = fm; 1428 im[idx] = 0; 1429 } 1430 nzi = 0; 1431 row = fill[n]; 1432 while ( row < prow ) { 1433 incrlev = im[row] + 1; 1434 nz = dloc[row]; 1435 xi = ajnew + ainew[row] + nz; 1436 flev = ajfill + ainew[row] + nz + 1; 1437 nnz = ainew[row+1] - ainew[row] - nz - 1; 1438 if (*xi++ != row) { 1439 SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,0,"zero pivot"); 1440 } 1441 fm = row; 1442 while (nnz-- > 0) { 1443 idx = *xi++; 1444 if (*flev + incrlev > levels) { 1445 flev++; 1446 continue; 1447 } 1448 do { 1449 m = fm; 1450 fm = fill[m]; 1451 } while (fm < idx); 1452 if (fm != idx) { 1453 im[idx] = *flev + incrlev; 1454 fill[m] = idx; 1455 fill[idx] = fm; 1456 fm = idx; 1457 nzf++; 1458 } 1459 else { 1460 if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 1461 } 1462 flev++; 1463 } 1464 row = fill[row]; 1465 nzi++; 1466 } 1467 /* copy new filled row into permanent storage */ 1468 ainew[prow+1] = ainew[prow] + nzf; 1469 if (ainew[prow+1] > jmax) { 1470 /* allocate a longer ajnew */ 1471 int maxadd; 1472 maxadd = (int) (((f*ai[n]+1)*(n-prow+5))/n); 1473 if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1474 jmax += maxadd; 1475 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1476 PetscMemcpy(xi,ajnew,ainew[prow]*sizeof(int)); 1477 PetscFree(ajnew); 1478 ajnew = xi; 1479 /* allocate a longer ajfill */ 1480 xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 1481 PetscMemcpy(xi,ajfill,ainew[prow]*sizeof(int)); 1482 PetscFree(ajfill); 1483 ajfill = xi; 1484 realloc++; 1485 } 1486 xi = ajnew + ainew[prow]; 1487 flev = ajfill + ainew[prow]; 1488 dloc[prow] = nzi; 1489 fm = fill[n]; 1490 while (nzf--) { 1491 *xi++ = fm; 1492 *flev++ = im[fm]; 1493 fm = fill[fm]; 1494 } 1495 } 1496 PetscFree(ajfill); 1497 ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 1498 ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1499 ierr = ISDestroy(isicol); CHKERRQ(ierr); 1500 PetscFree(fill); PetscFree(im); 1501 1502 PLogInfo(A, 1503 "Info:MatILUFactorSymbolic_SeqBAIJ:Realloc %d Fill ratio:given %g needed %g\n", 1504 realloc,f,((double)ainew[n])/((double)ai[prow])); 1505 1506 /* put together the new matrix */ 1507 ierr = MatCreateSeqBAIJ(A->comm,bs,bs*n,bs*n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1508 b = (Mat_SeqBAIJ *) (*fact)->data; 1509 PetscFree(b->imax); 1510 b->singlemalloc = 0; 1511 len = bs2*ainew[n]*sizeof(Scalar); 1512 /* the next line frees the default space generated by the Create() */ 1513 PetscFree(b->a); PetscFree(b->ilen); 1514 b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 1515 b->j = ajnew; 1516 b->i = ainew; 1517 for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 1518 b->diag = dloc; 1519 b->ilen = 0; 1520 b->imax = 0; 1521 b->row = isrow; 1522 b->col = iscol; 1523 b->solve_work = (Scalar *) PetscMalloc( (bs*n+bs)*sizeof(Scalar)); 1524 CHKPTRQ(b->solve_work); 1525 /* In b structure: Free imax, ilen, old a, old j. 1526 Allocate dloc, solve_work, new a, new j */ 1527 PLogObjectMemory(*fact,(ainew[n]-n)*(sizeof(int))+bs2*ainew[n]*sizeof(Scalar)); 1528 b->maxnz = b->nz = ainew[n]; 1529 (*fact)->factor = FACTOR_LU; 1530 1531 (*fact)->info.factor_mallocs = realloc; 1532 (*fact)->info.fill_ratio_given = f; 1533 (*fact)->info.fill_ratio_needed = ((double)ainew[n])/((double)ai[prow]); 1534 1535 return 0; 1536 } 1537 1538 1539 1540 1541