1 2 /* 3 Defines the basic matrix operations for the BAIJ (compressed row) 4 matrix storage format. 5 */ 6 #include <../src/mat/impls/baij/seq/baij.h> /*I "petscmat.h" I*/ 7 #include <petscblaslapack.h> 8 #include <petsc-private/kernels/blockinvert.h> 9 #include <petsc-private/kernels/blockmatmult.h> 10 11 #undef __FUNCT__ 12 #define __FUNCT__ "MatInvertBlockDiagonal_SeqBAIJ" 13 PetscErrorCode MatInvertBlockDiagonal_SeqBAIJ(Mat A,const PetscScalar **values) 14 { 15 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*) A->data; 16 PetscErrorCode ierr; 17 PetscInt *diag_offset,i,bs = A->rmap->bs,mbs = a->mbs,ipvt[5],bs2 = bs*bs,*v_pivots; 18 MatScalar *v = a->a,*odiag,*diag,*mdiag,work[25],*v_work; 19 PetscReal shift = 0.0; 20 21 PetscFunctionBegin; 22 if (a->idiagvalid) { 23 if (values) *values = a->idiag; 24 PetscFunctionReturn(0); 25 } 26 ierr = MatMarkDiagonal_SeqBAIJ(A);CHKERRQ(ierr); 27 diag_offset = a->diag; 28 if (!a->idiag) { 29 ierr = PetscMalloc1(2*bs2*mbs,&a->idiag);CHKERRQ(ierr); 30 ierr = PetscLogObjectMemory((PetscObject)A,2*bs2*mbs*sizeof(PetscScalar));CHKERRQ(ierr); 31 } 32 diag = a->idiag; 33 mdiag = a->idiag+bs2*mbs; 34 if (values) *values = a->idiag; 35 /* factor and invert each block */ 36 switch (bs) { 37 case 1: 38 for (i=0; i<mbs; i++) { 39 odiag = v + 1*diag_offset[i]; 40 diag[0] = odiag[0]; 41 mdiag[0] = odiag[0]; 42 diag[0] = (PetscScalar)1.0 / (diag[0] + shift); 43 diag += 1; 44 mdiag += 1; 45 } 46 break; 47 case 2: 48 for (i=0; i<mbs; i++) { 49 odiag = v + 4*diag_offset[i]; 50 diag[0] = odiag[0]; diag[1] = odiag[1]; diag[2] = odiag[2]; diag[3] = odiag[3]; 51 mdiag[0] = odiag[0]; mdiag[1] = odiag[1]; mdiag[2] = odiag[2]; mdiag[3] = odiag[3]; 52 ierr = PetscKernel_A_gets_inverse_A_2(diag,shift);CHKERRQ(ierr); 53 diag += 4; 54 mdiag += 4; 55 } 56 break; 57 case 3: 58 for (i=0; i<mbs; i++) { 59 odiag = v + 9*diag_offset[i]; 60 diag[0] = odiag[0]; diag[1] = odiag[1]; diag[2] = odiag[2]; diag[3] = odiag[3]; 61 diag[4] = odiag[4]; diag[5] = odiag[5]; diag[6] = odiag[6]; diag[7] = odiag[7]; 62 diag[8] = odiag[8]; 63 mdiag[0] = odiag[0]; mdiag[1] = odiag[1]; mdiag[2] = odiag[2]; mdiag[3] = odiag[3]; 64 mdiag[4] = odiag[4]; mdiag[5] = odiag[5]; mdiag[6] = odiag[6]; mdiag[7] = odiag[7]; 65 mdiag[8] = odiag[8]; 66 ierr = PetscKernel_A_gets_inverse_A_3(diag,shift);CHKERRQ(ierr); 67 diag += 9; 68 mdiag += 9; 69 } 70 break; 71 case 4: 72 for (i=0; i<mbs; i++) { 73 odiag = v + 16*diag_offset[i]; 74 ierr = PetscMemcpy(diag,odiag,16*sizeof(PetscScalar));CHKERRQ(ierr); 75 ierr = PetscMemcpy(mdiag,odiag,16*sizeof(PetscScalar));CHKERRQ(ierr); 76 ierr = PetscKernel_A_gets_inverse_A_4(diag,shift);CHKERRQ(ierr); 77 diag += 16; 78 mdiag += 16; 79 } 80 break; 81 case 5: 82 for (i=0; i<mbs; i++) { 83 odiag = v + 25*diag_offset[i]; 84 ierr = PetscMemcpy(diag,odiag,25*sizeof(PetscScalar));CHKERRQ(ierr); 85 ierr = PetscMemcpy(mdiag,odiag,25*sizeof(PetscScalar));CHKERRQ(ierr); 86 ierr = PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift);CHKERRQ(ierr); 87 diag += 25; 88 mdiag += 25; 89 } 90 break; 91 case 6: 92 for (i=0; i<mbs; i++) { 93 odiag = v + 36*diag_offset[i]; 94 ierr = PetscMemcpy(diag,odiag,36*sizeof(PetscScalar));CHKERRQ(ierr); 95 ierr = PetscMemcpy(mdiag,odiag,36*sizeof(PetscScalar));CHKERRQ(ierr); 96 ierr = PetscKernel_A_gets_inverse_A_6(diag,shift);CHKERRQ(ierr); 97 diag += 36; 98 mdiag += 36; 99 } 100 break; 101 case 7: 102 for (i=0; i<mbs; i++) { 103 odiag = v + 49*diag_offset[i]; 104 ierr = PetscMemcpy(diag,odiag,49*sizeof(PetscScalar));CHKERRQ(ierr); 105 ierr = PetscMemcpy(mdiag,odiag,49*sizeof(PetscScalar));CHKERRQ(ierr); 106 ierr = PetscKernel_A_gets_inverse_A_7(diag,shift);CHKERRQ(ierr); 107 diag += 49; 108 mdiag += 49; 109 } 110 break; 111 default: 112 ierr = PetscMalloc2(bs,&v_work,bs,&v_pivots);CHKERRQ(ierr); 113 for (i=0; i<mbs; i++) { 114 odiag = v + bs2*diag_offset[i]; 115 ierr = PetscMemcpy(diag,odiag,bs2*sizeof(PetscScalar));CHKERRQ(ierr); 116 ierr = PetscMemcpy(mdiag,odiag,bs2*sizeof(PetscScalar));CHKERRQ(ierr); 117 ierr = PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work);CHKERRQ(ierr); 118 diag += bs2; 119 mdiag += bs2; 120 } 121 ierr = PetscFree2(v_work,v_pivots);CHKERRQ(ierr); 122 } 123 a->idiagvalid = PETSC_TRUE; 124 PetscFunctionReturn(0); 125 } 126 127 #undef __FUNCT__ 128 #define __FUNCT__ "MatSOR_SeqBAIJ" 129 PetscErrorCode MatSOR_SeqBAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 130 { 131 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 132 PetscScalar *x,*work,*w,*workt,*t; 133 const MatScalar *v,*aa = a->a, *idiag; 134 const PetscScalar *b,*xb; 135 PetscScalar s[7], xw[7]={0}; /* avoid some compilers thinking xw is uninitialized */ 136 PetscErrorCode ierr; 137 PetscInt m = a->mbs,i,i2,nz,bs = A->rmap->bs,bs2 = bs*bs,k,j,idx,it; 138 const PetscInt *diag,*ai = a->i,*aj = a->j,*vi; 139 140 PetscFunctionBegin; 141 its = its*lits; 142 if (flag & SOR_EISENSTAT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for Eisenstat"); 143 if (its <= 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits); 144 if (fshift) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Sorry, no support for diagonal shift"); 145 if (omega != 1.0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Sorry, no support for non-trivial relaxation factor"); 146 if ((flag & SOR_APPLY_UPPER) || (flag & SOR_APPLY_LOWER)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Sorry, no support for applying upper or lower triangular parts"); 147 148 if (!a->idiagvalid) {ierr = MatInvertBlockDiagonal(A,NULL);CHKERRQ(ierr);} 149 150 if (!m) PetscFunctionReturn(0); 151 diag = a->diag; 152 idiag = a->idiag; 153 k = PetscMax(A->rmap->n,A->cmap->n); 154 if (!a->mult_work) { 155 ierr = PetscMalloc1((2*k+1),&a->mult_work);CHKERRQ(ierr); 156 } 157 work = a->mult_work; 158 t = work + k+1; 159 if (!a->sor_work) { 160 ierr = PetscMalloc1(bs,&a->sor_work);CHKERRQ(ierr); 161 } 162 w = a->sor_work; 163 164 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 165 ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr); 166 167 if (flag & SOR_ZERO_INITIAL_GUESS) { 168 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 169 switch (bs) { 170 case 1: 171 PetscKernel_v_gets_A_times_w_1(x,idiag,b); 172 t[0] = b[0]; 173 i2 = 1; 174 idiag += 1; 175 for (i=1; i<m; i++) { 176 v = aa + ai[i]; 177 vi = aj + ai[i]; 178 nz = diag[i] - ai[i]; 179 s[0] = b[i2]; 180 for (j=0; j<nz; j++) { 181 xw[0] = x[vi[j]]; 182 PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw); 183 } 184 t[i2] = s[0]; 185 PetscKernel_v_gets_A_times_w_1(xw,idiag,s); 186 x[i2] = xw[0]; 187 idiag += 1; 188 i2 += 1; 189 } 190 break; 191 case 2: 192 PetscKernel_v_gets_A_times_w_2(x,idiag,b); 193 t[0] = b[0]; t[1] = b[1]; 194 i2 = 2; 195 idiag += 4; 196 for (i=1; i<m; i++) { 197 v = aa + 4*ai[i]; 198 vi = aj + ai[i]; 199 nz = diag[i] - ai[i]; 200 s[0] = b[i2]; s[1] = b[i2+1]; 201 for (j=0; j<nz; j++) { 202 idx = 2*vi[j]; 203 it = 4*j; 204 xw[0] = x[idx]; xw[1] = x[1+idx]; 205 PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw); 206 } 207 t[i2] = s[0]; t[i2+1] = s[1]; 208 PetscKernel_v_gets_A_times_w_2(xw,idiag,s); 209 x[i2] = xw[0]; x[i2+1] = xw[1]; 210 idiag += 4; 211 i2 += 2; 212 } 213 break; 214 case 3: 215 PetscKernel_v_gets_A_times_w_3(x,idiag,b); 216 t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; 217 i2 = 3; 218 idiag += 9; 219 for (i=1; i<m; i++) { 220 v = aa + 9*ai[i]; 221 vi = aj + ai[i]; 222 nz = diag[i] - ai[i]; 223 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; 224 while (nz--) { 225 idx = 3*(*vi++); 226 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 227 PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw); 228 v += 9; 229 } 230 t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2]; 231 PetscKernel_v_gets_A_times_w_3(xw,idiag,s); 232 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; 233 idiag += 9; 234 i2 += 3; 235 } 236 break; 237 case 4: 238 PetscKernel_v_gets_A_times_w_4(x,idiag,b); 239 t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; t[3] = b[3]; 240 i2 = 4; 241 idiag += 16; 242 for (i=1; i<m; i++) { 243 v = aa + 16*ai[i]; 244 vi = aj + ai[i]; 245 nz = diag[i] - ai[i]; 246 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; 247 while (nz--) { 248 idx = 4*(*vi++); 249 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; 250 PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw); 251 v += 16; 252 } 253 t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2]; t[i2 + 3] = s[3]; 254 PetscKernel_v_gets_A_times_w_4(xw,idiag,s); 255 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; 256 idiag += 16; 257 i2 += 4; 258 } 259 break; 260 case 5: 261 PetscKernel_v_gets_A_times_w_5(x,idiag,b); 262 t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; t[3] = b[3]; t[4] = b[4]; 263 i2 = 5; 264 idiag += 25; 265 for (i=1; i<m; i++) { 266 v = aa + 25*ai[i]; 267 vi = aj + ai[i]; 268 nz = diag[i] - ai[i]; 269 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; 270 while (nz--) { 271 idx = 5*(*vi++); 272 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx]; 273 PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw); 274 v += 25; 275 } 276 t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2]; t[i2+3] = s[3]; t[i2+4] = s[4]; 277 PetscKernel_v_gets_A_times_w_5(xw,idiag,s); 278 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; 279 idiag += 25; 280 i2 += 5; 281 } 282 break; 283 case 6: 284 PetscKernel_v_gets_A_times_w_6(x,idiag,b); 285 t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; t[3] = b[3]; t[4] = b[4]; t[5] = b[5]; 286 i2 = 6; 287 idiag += 36; 288 for (i=1; i<m; i++) { 289 v = aa + 36*ai[i]; 290 vi = aj + ai[i]; 291 nz = diag[i] - ai[i]; 292 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; 293 while (nz--) { 294 idx = 6*(*vi++); 295 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 296 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; 297 PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw); 298 v += 36; 299 } 300 t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2]; 301 t[i2+3] = s[3]; t[i2+4] = s[4]; t[i2+5] = s[5]; 302 PetscKernel_v_gets_A_times_w_6(xw,idiag,s); 303 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; 304 idiag += 36; 305 i2 += 6; 306 } 307 break; 308 case 7: 309 PetscKernel_v_gets_A_times_w_7(x,idiag,b); 310 t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; 311 t[3] = b[3]; t[4] = b[4]; t[5] = b[5]; t[6] = b[6]; 312 i2 = 7; 313 idiag += 49; 314 for (i=1; i<m; i++) { 315 v = aa + 49*ai[i]; 316 vi = aj + ai[i]; 317 nz = diag[i] - ai[i]; 318 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; 319 s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; s[6] = b[i2+6]; 320 while (nz--) { 321 idx = 7*(*vi++); 322 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 323 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx]; 324 PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw); 325 v += 49; 326 } 327 t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2]; 328 t[i2+3] = s[3]; t[i2+4] = s[4]; t[i2+5] = s[5]; t[i2+6] = s[6]; 329 PetscKernel_v_gets_A_times_w_7(xw,idiag,s); 330 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; 331 x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; x[i2+6] = xw[6]; 332 idiag += 49; 333 i2 += 7; 334 } 335 break; 336 default: 337 PetscKernel_w_gets_Ar_times_v(bs,bs,b,idiag,x); 338 ierr = PetscMemcpy(t,b,bs*sizeof(PetscScalar));CHKERRQ(ierr); 339 i2 = bs; 340 idiag += bs2; 341 for (i=1; i<m; i++) { 342 v = aa + bs2*ai[i]; 343 vi = aj + ai[i]; 344 nz = diag[i] - ai[i]; 345 346 ierr = PetscMemcpy(w,b+i2,bs*sizeof(PetscScalar));CHKERRQ(ierr); 347 /* copy all rows of x that are needed into contiguous space */ 348 workt = work; 349 for (j=0; j<nz; j++) { 350 ierr = PetscMemcpy(workt,x + bs*(*vi++),bs*sizeof(PetscScalar));CHKERRQ(ierr); 351 workt += bs; 352 } 353 PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work); 354 ierr = PetscMemcpy(t+i2,w,bs*sizeof(PetscScalar));CHKERRQ(ierr); 355 PetscKernel_w_gets_Ar_times_v(bs,bs,w,idiag,x+i2); 356 357 idiag += bs2; 358 i2 += bs; 359 } 360 break; 361 } 362 /* for logging purposes assume number of nonzero in lower half is 1/2 of total */ 363 ierr = PetscLogFlops(1.0*bs2*a->nz);CHKERRQ(ierr); 364 xb = t; 365 } 366 else xb = b; 367 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 368 idiag = a->idiag+bs2*(a->mbs-1); 369 i2 = bs * (m-1); 370 switch (bs) { 371 case 1: 372 s[0] = xb[i2]; 373 PetscKernel_v_gets_A_times_w_1(xw,idiag,s); 374 x[i2] = xw[0]; 375 i2 -= 1; 376 for (i=m-2; i>=0; i--) { 377 v = aa + (diag[i]+1); 378 vi = aj + diag[i] + 1; 379 nz = ai[i+1] - diag[i] - 1; 380 s[0] = xb[i2]; 381 for (j=0; j<nz; j++) { 382 xw[0] = x[vi[j]]; 383 PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw); 384 } 385 PetscKernel_v_gets_A_times_w_1(xw,idiag,s); 386 x[i2] = xw[0]; 387 idiag -= 1; 388 i2 -= 1; 389 } 390 break; 391 case 2: 392 s[0] = xb[i2]; s[1] = xb[i2+1]; 393 PetscKernel_v_gets_A_times_w_2(xw,idiag,s); 394 x[i2] = xw[0]; x[i2+1] = xw[1]; 395 i2 -= 2; 396 idiag -= 4; 397 for (i=m-2; i>=0; i--) { 398 v = aa + 4*(diag[i] + 1); 399 vi = aj + diag[i] + 1; 400 nz = ai[i+1] - diag[i] - 1; 401 s[0] = xb[i2]; s[1] = xb[i2+1]; 402 for (j=0; j<nz; j++) { 403 idx = 2*vi[j]; 404 it = 4*j; 405 xw[0] = x[idx]; xw[1] = x[1+idx]; 406 PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw); 407 } 408 PetscKernel_v_gets_A_times_w_2(xw,idiag,s); 409 x[i2] = xw[0]; x[i2+1] = xw[1]; 410 idiag -= 4; 411 i2 -= 2; 412 } 413 break; 414 case 3: 415 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; 416 PetscKernel_v_gets_A_times_w_3(xw,idiag,s); 417 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; 418 i2 -= 3; 419 idiag -= 9; 420 for (i=m-2; i>=0; i--) { 421 v = aa + 9*(diag[i]+1); 422 vi = aj + diag[i] + 1; 423 nz = ai[i+1] - diag[i] - 1; 424 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; 425 while (nz--) { 426 idx = 3*(*vi++); 427 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 428 PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw); 429 v += 9; 430 } 431 PetscKernel_v_gets_A_times_w_3(xw,idiag,s); 432 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; 433 idiag -= 9; 434 i2 -= 3; 435 } 436 break; 437 case 4: 438 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; 439 PetscKernel_v_gets_A_times_w_4(xw,idiag,s); 440 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; 441 i2 -= 4; 442 idiag -= 16; 443 for (i=m-2; i>=0; i--) { 444 v = aa + 16*(diag[i]+1); 445 vi = aj + diag[i] + 1; 446 nz = ai[i+1] - diag[i] - 1; 447 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; 448 while (nz--) { 449 idx = 4*(*vi++); 450 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; 451 PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw); 452 v += 16; 453 } 454 PetscKernel_v_gets_A_times_w_4(xw,idiag,s); 455 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; 456 idiag -= 16; 457 i2 -= 4; 458 } 459 break; 460 case 5: 461 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4]; 462 PetscKernel_v_gets_A_times_w_5(xw,idiag,s); 463 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; 464 i2 -= 5; 465 idiag -= 25; 466 for (i=m-2; i>=0; i--) { 467 v = aa + 25*(diag[i]+1); 468 vi = aj + diag[i] + 1; 469 nz = ai[i+1] - diag[i] - 1; 470 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4]; 471 while (nz--) { 472 idx = 5*(*vi++); 473 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx]; 474 PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw); 475 v += 25; 476 } 477 PetscKernel_v_gets_A_times_w_5(xw,idiag,s); 478 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; 479 idiag -= 25; 480 i2 -= 5; 481 } 482 break; 483 case 6: 484 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5]; 485 PetscKernel_v_gets_A_times_w_6(xw,idiag,s); 486 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; 487 i2 -= 6; 488 idiag -= 36; 489 for (i=m-2; i>=0; i--) { 490 v = aa + 36*(diag[i]+1); 491 vi = aj + diag[i] + 1; 492 nz = ai[i+1] - diag[i] - 1; 493 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5]; 494 while (nz--) { 495 idx = 6*(*vi++); 496 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 497 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; 498 PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw); 499 v += 36; 500 } 501 PetscKernel_v_gets_A_times_w_6(xw,idiag,s); 502 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; 503 idiag -= 36; 504 i2 -= 6; 505 } 506 break; 507 case 7: 508 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; 509 s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5]; s[6] = xb[i2+6]; 510 PetscKernel_v_gets_A_times_w_7(x,idiag,b); 511 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; 512 x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; x[i2+6] = xw[6]; 513 i2 -= 7; 514 idiag -= 49; 515 for (i=m-2; i>=0; i--) { 516 v = aa + 49*(diag[i]+1); 517 vi = aj + diag[i] + 1; 518 nz = ai[i+1] - diag[i] - 1; 519 s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; 520 s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5]; s[6] = xb[i2+6]; 521 while (nz--) { 522 idx = 7*(*vi++); 523 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 524 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx]; 525 PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw); 526 v += 49; 527 } 528 PetscKernel_v_gets_A_times_w_7(xw,idiag,s); 529 x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; 530 x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; x[i2+6] = xw[6]; 531 idiag -= 49; 532 i2 -= 7; 533 } 534 break; 535 default: 536 ierr = PetscMemcpy(w,xb+i2,bs*sizeof(PetscScalar));CHKERRQ(ierr); 537 PetscKernel_w_gets_Ar_times_v(bs,bs,w,idiag,x+i2); 538 i2 -= bs; 539 idiag -= bs2; 540 for (i=m-2; i>=0; i--) { 541 v = aa + bs2*(diag[i]+1); 542 vi = aj + diag[i] + 1; 543 nz = ai[i+1] - diag[i] - 1; 544 545 ierr = PetscMemcpy(w,xb+i2,bs*sizeof(PetscScalar));CHKERRQ(ierr); 546 /* copy all rows of x that are needed into contiguous space */ 547 workt = work; 548 for (j=0; j<nz; j++) { 549 ierr = PetscMemcpy(workt,x + bs*(*vi++),bs*sizeof(PetscScalar));CHKERRQ(ierr); 550 workt += bs; 551 } 552 PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work); 553 PetscKernel_w_gets_Ar_times_v(bs,bs,w,idiag,x+i2); 554 555 idiag -= bs2; 556 i2 -= bs; 557 } 558 break; 559 } 560 ierr = PetscLogFlops(1.0*bs2*(a->nz));CHKERRQ(ierr); 561 } 562 its--; 563 } 564 while (its--) { 565 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 566 idiag = a->idiag; 567 i2 = 0; 568 switch (bs) { 569 case 1: 570 for (i=0; i<m; i++) { 571 v = aa + ai[i]; 572 vi = aj + ai[i]; 573 nz = ai[i+1] - ai[i]; 574 s[0] = b[i2]; 575 for (j=0; j<nz; j++) { 576 xw[0] = x[vi[j]]; 577 PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw); 578 } 579 PetscKernel_v_gets_A_times_w_1(xw,idiag,s); 580 x[i2] += xw[0]; 581 idiag += 1; 582 i2 += 1; 583 } 584 break; 585 case 2: 586 for (i=0; i<m; i++) { 587 v = aa + 4*ai[i]; 588 vi = aj + ai[i]; 589 nz = ai[i+1] - ai[i]; 590 s[0] = b[i2]; s[1] = b[i2+1]; 591 for (j=0; j<nz; j++) { 592 idx = 2*vi[j]; 593 it = 4*j; 594 xw[0] = x[idx]; xw[1] = x[1+idx]; 595 PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw); 596 } 597 PetscKernel_v_gets_A_times_w_2(xw,idiag,s); 598 x[i2] += xw[0]; x[i2+1] += xw[1]; 599 idiag += 4; 600 i2 += 2; 601 } 602 break; 603 case 3: 604 for (i=0; i<m; i++) { 605 v = aa + 9*ai[i]; 606 vi = aj + ai[i]; 607 nz = ai[i+1] - ai[i]; 608 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; 609 while (nz--) { 610 idx = 3*(*vi++); 611 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 612 PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw); 613 v += 9; 614 } 615 PetscKernel_v_gets_A_times_w_3(xw,idiag,s); 616 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; 617 idiag += 9; 618 i2 += 3; 619 } 620 break; 621 case 4: 622 for (i=0; i<m; i++) { 623 v = aa + 16*ai[i]; 624 vi = aj + ai[i]; 625 nz = ai[i+1] - ai[i]; 626 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; 627 while (nz--) { 628 idx = 4*(*vi++); 629 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; 630 PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw); 631 v += 16; 632 } 633 PetscKernel_v_gets_A_times_w_4(xw,idiag,s); 634 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3]; 635 idiag += 16; 636 i2 += 4; 637 } 638 break; 639 case 5: 640 for (i=0; i<m; i++) { 641 v = aa + 25*ai[i]; 642 vi = aj + ai[i]; 643 nz = ai[i+1] - ai[i]; 644 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; 645 while (nz--) { 646 idx = 5*(*vi++); 647 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx]; 648 PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw); 649 v += 25; 650 } 651 PetscKernel_v_gets_A_times_w_5(xw,idiag,s); 652 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3]; x[i2+4] += xw[4]; 653 idiag += 25; 654 i2 += 5; 655 } 656 break; 657 case 6: 658 for (i=0; i<m; i++) { 659 v = aa + 36*ai[i]; 660 vi = aj + ai[i]; 661 nz = ai[i+1] - ai[i]; 662 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; 663 while (nz--) { 664 idx = 6*(*vi++); 665 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 666 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; 667 PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw); 668 v += 36; 669 } 670 PetscKernel_v_gets_A_times_w_6(xw,idiag,s); 671 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; 672 x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5]; 673 idiag += 36; 674 i2 += 6; 675 } 676 break; 677 case 7: 678 for (i=0; i<m; i++) { 679 v = aa + 49*ai[i]; 680 vi = aj + ai[i]; 681 nz = ai[i+1] - ai[i]; 682 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; 683 s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; s[6] = b[i2+6]; 684 while (nz--) { 685 idx = 7*(*vi++); 686 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 687 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx]; 688 PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw); 689 v += 49; 690 } 691 PetscKernel_v_gets_A_times_w_7(xw,idiag,s); 692 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; 693 x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5]; x[i2+6] += xw[6]; 694 idiag += 49; 695 i2 += 7; 696 } 697 break; 698 default: 699 for (i=0; i<m; i++) { 700 v = aa + bs2*ai[i]; 701 vi = aj + ai[i]; 702 nz = ai[i+1] - ai[i]; 703 704 ierr = PetscMemcpy(w,b+i2,bs*sizeof(PetscScalar));CHKERRQ(ierr); 705 /* copy all rows of x that are needed into contiguous space */ 706 workt = work; 707 for (j=0; j<nz; j++) { 708 ierr = PetscMemcpy(workt,x + bs*(*vi++),bs*sizeof(PetscScalar));CHKERRQ(ierr); 709 workt += bs; 710 } 711 PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work); 712 PetscKernel_w_gets_w_plus_Ar_times_v(bs,bs,w,idiag,x+i2); 713 714 idiag += bs2; 715 i2 += bs; 716 } 717 break; 718 } 719 ierr = PetscLogFlops(2.0*bs2*a->nz);CHKERRQ(ierr); 720 } 721 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 722 idiag = a->idiag+bs2*(a->mbs-1); 723 i2 = bs * (m-1); 724 switch (bs) { 725 case 1: 726 for (i=m-1; i>=0; i--) { 727 v = aa + ai[i]; 728 vi = aj + ai[i]; 729 nz = ai[i+1] - ai[i]; 730 s[0] = b[i2]; 731 for (j=0; j<nz; j++) { 732 xw[0] = x[vi[j]]; 733 PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw); 734 } 735 PetscKernel_v_gets_A_times_w_1(xw,idiag,s); 736 x[i2] += xw[0]; 737 idiag -= 1; 738 i2 -= 1; 739 } 740 break; 741 case 2: 742 for (i=m-1; i>=0; i--) { 743 v = aa + 4*ai[i]; 744 vi = aj + ai[i]; 745 nz = ai[i+1] - ai[i]; 746 s[0] = b[i2]; s[1] = b[i2+1]; 747 for (j=0; j<nz; j++) { 748 idx = 2*vi[j]; 749 it = 4*j; 750 xw[0] = x[idx]; xw[1] = x[1+idx]; 751 PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw); 752 } 753 PetscKernel_v_gets_A_times_w_2(xw,idiag,s); 754 x[i2] += xw[0]; x[i2+1] += xw[1]; 755 idiag -= 4; 756 i2 -= 2; 757 } 758 break; 759 case 3: 760 for (i=m-1; i>=0; i--) { 761 v = aa + 9*ai[i]; 762 vi = aj + ai[i]; 763 nz = ai[i+1] - ai[i]; 764 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; 765 while (nz--) { 766 idx = 3*(*vi++); 767 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 768 PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw); 769 v += 9; 770 } 771 PetscKernel_v_gets_A_times_w_3(xw,idiag,s); 772 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; 773 idiag -= 9; 774 i2 -= 3; 775 } 776 break; 777 case 4: 778 for (i=m-1; i>=0; i--) { 779 v = aa + 16*ai[i]; 780 vi = aj + ai[i]; 781 nz = ai[i+1] - ai[i]; 782 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; 783 while (nz--) { 784 idx = 4*(*vi++); 785 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; 786 PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw); 787 v += 16; 788 } 789 PetscKernel_v_gets_A_times_w_4(xw,idiag,s); 790 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3]; 791 idiag -= 16; 792 i2 -= 4; 793 } 794 break; 795 case 5: 796 for (i=m-1; i>=0; i--) { 797 v = aa + 25*ai[i]; 798 vi = aj + ai[i]; 799 nz = ai[i+1] - ai[i]; 800 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; 801 while (nz--) { 802 idx = 5*(*vi++); 803 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx]; 804 PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw); 805 v += 25; 806 } 807 PetscKernel_v_gets_A_times_w_5(xw,idiag,s); 808 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3]; x[i2+4] += xw[4]; 809 idiag -= 25; 810 i2 -= 5; 811 } 812 break; 813 case 6: 814 for (i=m-1; i>=0; i--) { 815 v = aa + 36*ai[i]; 816 vi = aj + ai[i]; 817 nz = ai[i+1] - ai[i]; 818 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; 819 while (nz--) { 820 idx = 6*(*vi++); 821 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 822 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; 823 PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw); 824 v += 36; 825 } 826 PetscKernel_v_gets_A_times_w_6(xw,idiag,s); 827 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; 828 x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5]; 829 idiag -= 36; 830 i2 -= 6; 831 } 832 break; 833 case 7: 834 for (i=m-1; i>=0; i--) { 835 v = aa + 49*ai[i]; 836 vi = aj + ai[i]; 837 nz = ai[i+1] - ai[i]; 838 s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; 839 s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; s[6] = b[i2+6]; 840 while (nz--) { 841 idx = 7*(*vi++); 842 xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; 843 xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx]; 844 PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw); 845 v += 49; 846 } 847 PetscKernel_v_gets_A_times_w_7(xw,idiag,s); 848 x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; 849 x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5]; x[i2+6] += xw[6]; 850 idiag -= 49; 851 i2 -= 7; 852 } 853 break; 854 default: 855 for (i=m-1; i>=0; i--) { 856 v = aa + bs2*ai[i]; 857 vi = aj + ai[i]; 858 nz = ai[i+1] - ai[i]; 859 860 ierr = PetscMemcpy(w,b+i2,bs*sizeof(PetscScalar));CHKERRQ(ierr); 861 /* copy all rows of x that are needed into contiguous space */ 862 workt = work; 863 for (j=0; j<nz; j++) { 864 ierr = PetscMemcpy(workt,x + bs*(*vi++),bs*sizeof(PetscScalar));CHKERRQ(ierr); 865 workt += bs; 866 } 867 PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work); 868 PetscKernel_w_gets_w_plus_Ar_times_v(bs,bs,w,idiag,x+i2); 869 870 idiag -= bs2; 871 i2 -= bs; 872 } 873 break; 874 } 875 ierr = PetscLogFlops(2.0*bs2*(a->nz));CHKERRQ(ierr); 876 } 877 } 878 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 879 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 880 PetscFunctionReturn(0); 881 } 882 883 884 /* 885 Special version for direct calls from Fortran (Used in PETSc-fun3d) 886 */ 887 #if defined(PETSC_HAVE_FORTRAN_CAPS) 888 #define matsetvaluesblocked4_ MATSETVALUESBLOCKED4 889 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 890 #define matsetvaluesblocked4_ matsetvaluesblocked4 891 #endif 892 893 #undef __FUNCT__ 894 #define __FUNCT__ "matsetvaluesblocked4_" 895 PETSC_EXTERN void matsetvaluesblocked4_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[]) 896 { 897 Mat A = *AA; 898 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 899 PetscInt *rp,k,low,high,t,ii,jj,row,nrow,i,col,l,N,m = *mm,n = *nn; 900 PetscInt *ai =a->i,*ailen=a->ilen; 901 PetscInt *aj =a->j,stepval,lastcol = -1; 902 const PetscScalar *value = v; 903 MatScalar *ap,*aa = a->a,*bap; 904 905 PetscFunctionBegin; 906 if (A->rmap->bs != 4) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Can only be called with a block size of 4"); 907 stepval = (n-1)*4; 908 for (k=0; k<m; k++) { /* loop over added rows */ 909 row = im[k]; 910 rp = aj + ai[row]; 911 ap = aa + 16*ai[row]; 912 nrow = ailen[row]; 913 low = 0; 914 high = nrow; 915 for (l=0; l<n; l++) { /* loop over added columns */ 916 col = in[l]; 917 if (col <= lastcol) low = 0; 918 else high = nrow; 919 lastcol = col; 920 value = v + k*(stepval+4 + l)*4; 921 while (high-low > 7) { 922 t = (low+high)/2; 923 if (rp[t] > col) high = t; 924 else low = t; 925 } 926 for (i=low; i<high; i++) { 927 if (rp[i] > col) break; 928 if (rp[i] == col) { 929 bap = ap + 16*i; 930 for (ii=0; ii<4; ii++,value+=stepval) { 931 for (jj=ii; jj<16; jj+=4) { 932 bap[jj] += *value++; 933 } 934 } 935 goto noinsert2; 936 } 937 } 938 N = nrow++ - 1; 939 high++; /* added new column index thus must search to one higher than before */ 940 /* shift up all the later entries in this row */ 941 for (ii=N; ii>=i; ii--) { 942 rp[ii+1] = rp[ii]; 943 PetscMemcpy(ap+16*(ii+1),ap+16*(ii),16*sizeof(MatScalar)); 944 } 945 if (N >= i) { 946 PetscMemzero(ap+16*i,16*sizeof(MatScalar)); 947 } 948 rp[i] = col; 949 bap = ap + 16*i; 950 for (ii=0; ii<4; ii++,value+=stepval) { 951 for (jj=ii; jj<16; jj+=4) { 952 bap[jj] = *value++; 953 } 954 } 955 noinsert2:; 956 low = i; 957 } 958 ailen[row] = nrow; 959 } 960 PetscFunctionReturnVoid(); 961 } 962 963 #if defined(PETSC_HAVE_FORTRAN_CAPS) 964 #define matsetvalues4_ MATSETVALUES4 965 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 966 #define matsetvalues4_ matsetvalues4 967 #endif 968 969 #undef __FUNCT__ 970 #define __FUNCT__ "MatSetValues4_" 971 PETSC_EXTERN void matsetvalues4_(Mat *AA,PetscInt *mm,PetscInt *im,PetscInt *nn,PetscInt *in,PetscScalar *v) 972 { 973 Mat A = *AA; 974 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 975 PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,N,n = *nn,m = *mm; 976 PetscInt *ai=a->i,*ailen=a->ilen; 977 PetscInt *aj=a->j,brow,bcol; 978 PetscInt ridx,cidx,lastcol = -1; 979 MatScalar *ap,value,*aa=a->a,*bap; 980 981 PetscFunctionBegin; 982 for (k=0; k<m; k++) { /* loop over added rows */ 983 row = im[k]; brow = row/4; 984 rp = aj + ai[brow]; 985 ap = aa + 16*ai[brow]; 986 nrow = ailen[brow]; 987 low = 0; 988 high = nrow; 989 for (l=0; l<n; l++) { /* loop over added columns */ 990 col = in[l]; bcol = col/4; 991 ridx = row % 4; cidx = col % 4; 992 value = v[l + k*n]; 993 if (col <= lastcol) low = 0; 994 else high = nrow; 995 lastcol = col; 996 while (high-low > 7) { 997 t = (low+high)/2; 998 if (rp[t] > bcol) high = t; 999 else low = t; 1000 } 1001 for (i=low; i<high; i++) { 1002 if (rp[i] > bcol) break; 1003 if (rp[i] == bcol) { 1004 bap = ap + 16*i + 4*cidx + ridx; 1005 *bap += value; 1006 goto noinsert1; 1007 } 1008 } 1009 N = nrow++ - 1; 1010 high++; /* added new column thus must search to one higher than before */ 1011 /* shift up all the later entries in this row */ 1012 for (ii=N; ii>=i; ii--) { 1013 rp[ii+1] = rp[ii]; 1014 PetscMemcpy(ap+16*(ii+1),ap+16*(ii),16*sizeof(MatScalar)); 1015 } 1016 if (N>=i) { 1017 PetscMemzero(ap+16*i,16*sizeof(MatScalar)); 1018 } 1019 rp[i] = bcol; 1020 ap[16*i + 4*cidx + ridx] = value; 1021 noinsert1:; 1022 low = i; 1023 } 1024 ailen[brow] = nrow; 1025 } 1026 PetscFunctionReturnVoid(); 1027 } 1028 1029 /* 1030 Checks for missing diagonals 1031 */ 1032 #undef __FUNCT__ 1033 #define __FUNCT__ "MatMissingDiagonal_SeqBAIJ" 1034 PetscErrorCode MatMissingDiagonal_SeqBAIJ(Mat A,PetscBool *missing,PetscInt *d) 1035 { 1036 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1037 PetscErrorCode ierr; 1038 PetscInt *diag,*ii = a->i,i; 1039 1040 PetscFunctionBegin; 1041 ierr = MatMarkDiagonal_SeqBAIJ(A);CHKERRQ(ierr); 1042 *missing = PETSC_FALSE; 1043 if (A->rmap->n > 0 && !ii) { 1044 *missing = PETSC_TRUE; 1045 if (d) *d = 0; 1046 PetscInfo(A,"Matrix has no entries therefore is missing diagonal"); 1047 } else { 1048 diag = a->diag; 1049 for (i=0; i<a->mbs; i++) { 1050 if (diag[i] >= ii[i+1]) { 1051 *missing = PETSC_TRUE; 1052 if (d) *d = i; 1053 PetscInfo1(A,"Matrix is missing block diagonal number %D",i); 1054 break; 1055 } 1056 } 1057 } 1058 PetscFunctionReturn(0); 1059 } 1060 1061 #undef __FUNCT__ 1062 #define __FUNCT__ "MatMarkDiagonal_SeqBAIJ" 1063 PetscErrorCode MatMarkDiagonal_SeqBAIJ(Mat A) 1064 { 1065 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1066 PetscErrorCode ierr; 1067 PetscInt i,j,m = a->mbs; 1068 1069 PetscFunctionBegin; 1070 if (!a->diag) { 1071 ierr = PetscMalloc1(m,&a->diag);CHKERRQ(ierr); 1072 ierr = PetscLogObjectMemory((PetscObject)A,m*sizeof(PetscInt));CHKERRQ(ierr); 1073 a->free_diag = PETSC_TRUE; 1074 } 1075 for (i=0; i<m; i++) { 1076 a->diag[i] = a->i[i+1]; 1077 for (j=a->i[i]; j<a->i[i+1]; j++) { 1078 if (a->j[j] == i) { 1079 a->diag[i] = j; 1080 break; 1081 } 1082 } 1083 } 1084 PetscFunctionReturn(0); 1085 } 1086 1087 1088 #undef __FUNCT__ 1089 #define __FUNCT__ "MatGetRowIJ_SeqBAIJ" 1090 static PetscErrorCode MatGetRowIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool blockcompressed,PetscInt *nn,const PetscInt *inia[],const PetscInt *inja[],PetscBool *done) 1091 { 1092 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1093 PetscErrorCode ierr; 1094 PetscInt i,j,n = a->mbs,nz = a->i[n],*tia,*tja,bs = A->rmap->bs,k,l,cnt; 1095 PetscInt **ia = (PetscInt**)inia,**ja = (PetscInt**)inja; 1096 1097 PetscFunctionBegin; 1098 *nn = n; 1099 if (!ia) PetscFunctionReturn(0); 1100 if (symmetric) { 1101 ierr = MatToSymmetricIJ_SeqAIJ(n,a->i,a->j,0,0,&tia,&tja);CHKERRQ(ierr); 1102 nz = tia[n]; 1103 } else { 1104 tia = a->i; tja = a->j; 1105 } 1106 1107 if (!blockcompressed && bs > 1) { 1108 (*nn) *= bs; 1109 /* malloc & create the natural set of indices */ 1110 ierr = PetscMalloc1((n+1)*bs,ia);CHKERRQ(ierr); 1111 if (n) { 1112 (*ia)[0] = 0; 1113 for (j=1; j<bs; j++) { 1114 (*ia)[j] = (tia[1]-tia[0])*bs+(*ia)[j-1]; 1115 } 1116 } 1117 1118 for (i=1; i<n; i++) { 1119 (*ia)[i*bs] = (tia[i]-tia[i-1])*bs + (*ia)[i*bs-1]; 1120 for (j=1; j<bs; j++) { 1121 (*ia)[i*bs+j] = (tia[i+1]-tia[i])*bs + (*ia)[i*bs+j-1]; 1122 } 1123 } 1124 if (n) { 1125 (*ia)[n*bs] = (tia[n]-tia[n-1])*bs + (*ia)[n*bs-1]; 1126 } 1127 1128 if (inja) { 1129 ierr = PetscMalloc1(nz*bs*bs,ja);CHKERRQ(ierr); 1130 cnt = 0; 1131 for (i=0; i<n; i++) { 1132 for (j=0; j<bs; j++) { 1133 for (k=tia[i]; k<tia[i+1]; k++) { 1134 for (l=0; l<bs; l++) { 1135 (*ja)[cnt++] = bs*tja[k] + l; 1136 } 1137 } 1138 } 1139 } 1140 } 1141 1142 if (symmetric) { /* deallocate memory allocated in MatToSymmetricIJ_SeqAIJ() */ 1143 ierr = PetscFree(tia);CHKERRQ(ierr); 1144 ierr = PetscFree(tja);CHKERRQ(ierr); 1145 } 1146 } else if (oshift == 1) { 1147 if (symmetric) { 1148 nz = tia[A->rmap->n/bs]; 1149 /* add 1 to i and j indices */ 1150 for (i=0; i<A->rmap->n/bs+1; i++) tia[i] = tia[i] + 1; 1151 *ia = tia; 1152 if (ja) { 1153 for (i=0; i<nz; i++) tja[i] = tja[i] + 1; 1154 *ja = tja; 1155 } 1156 } else { 1157 nz = a->i[A->rmap->n/bs]; 1158 /* malloc space and add 1 to i and j indices */ 1159 ierr = PetscMalloc1((A->rmap->n/bs+1),ia);CHKERRQ(ierr); 1160 for (i=0; i<A->rmap->n/bs+1; i++) (*ia)[i] = a->i[i] + 1; 1161 if (ja) { 1162 ierr = PetscMalloc1(nz,ja);CHKERRQ(ierr); 1163 for (i=0; i<nz; i++) (*ja)[i] = a->j[i] + 1; 1164 } 1165 } 1166 } else { 1167 *ia = tia; 1168 if (ja) *ja = tja; 1169 } 1170 PetscFunctionReturn(0); 1171 } 1172 1173 #undef __FUNCT__ 1174 #define __FUNCT__ "MatRestoreRowIJ_SeqBAIJ" 1175 static PetscErrorCode MatRestoreRowIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool blockcompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 1176 { 1177 PetscErrorCode ierr; 1178 1179 PetscFunctionBegin; 1180 if (!ia) PetscFunctionReturn(0); 1181 if ((!blockcompressed && A->rmap->bs > 1) || (symmetric || oshift == 1)) { 1182 ierr = PetscFree(*ia);CHKERRQ(ierr); 1183 if (ja) {ierr = PetscFree(*ja);CHKERRQ(ierr);} 1184 } 1185 PetscFunctionReturn(0); 1186 } 1187 1188 #undef __FUNCT__ 1189 #define __FUNCT__ "MatDestroy_SeqBAIJ" 1190 PetscErrorCode MatDestroy_SeqBAIJ(Mat A) 1191 { 1192 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1193 PetscErrorCode ierr; 1194 1195 PetscFunctionBegin; 1196 #if defined(PETSC_USE_LOG) 1197 PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->N,A->cmap->n,a->nz); 1198 #endif 1199 ierr = MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);CHKERRQ(ierr); 1200 ierr = ISDestroy(&a->row);CHKERRQ(ierr); 1201 ierr = ISDestroy(&a->col);CHKERRQ(ierr); 1202 if (a->free_diag) {ierr = PetscFree(a->diag);CHKERRQ(ierr);} 1203 ierr = PetscFree(a->idiag);CHKERRQ(ierr); 1204 if (a->free_imax_ilen) {ierr = PetscFree2(a->imax,a->ilen);CHKERRQ(ierr);} 1205 ierr = PetscFree(a->solve_work);CHKERRQ(ierr); 1206 ierr = PetscFree(a->mult_work);CHKERRQ(ierr); 1207 ierr = PetscFree(a->sor_work);CHKERRQ(ierr); 1208 ierr = ISDestroy(&a->icol);CHKERRQ(ierr); 1209 ierr = PetscFree(a->saved_values);CHKERRQ(ierr); 1210 ierr = PetscFree(a->xtoy);CHKERRQ(ierr); 1211 ierr = PetscFree2(a->compressedrow.i,a->compressedrow.rindex);CHKERRQ(ierr); 1212 1213 ierr = MatDestroy(&a->sbaijMat);CHKERRQ(ierr); 1214 ierr = MatDestroy(&a->parent);CHKERRQ(ierr); 1215 ierr = PetscFree(A->data);CHKERRQ(ierr); 1216 1217 ierr = PetscObjectChangeTypeName((PetscObject)A,0);CHKERRQ(ierr); 1218 ierr = PetscObjectComposeFunction((PetscObject)A,"MatInvertBlockDiagonal_C",NULL);CHKERRQ(ierr); 1219 ierr = PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1220 ierr = PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1221 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetColumnIndices_C",NULL);CHKERRQ(ierr); 1222 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqaij_C",NULL);CHKERRQ(ierr); 1223 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqsbaij_C",NULL);CHKERRQ(ierr); 1224 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1225 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1226 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqbstrm_C",NULL);CHKERRQ(ierr); 1227 ierr = PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1228 PetscFunctionReturn(0); 1229 } 1230 1231 #undef __FUNCT__ 1232 #define __FUNCT__ "MatSetOption_SeqBAIJ" 1233 PetscErrorCode MatSetOption_SeqBAIJ(Mat A,MatOption op,PetscBool flg) 1234 { 1235 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1236 PetscErrorCode ierr; 1237 1238 PetscFunctionBegin; 1239 switch (op) { 1240 case MAT_ROW_ORIENTED: 1241 a->roworiented = flg; 1242 break; 1243 case MAT_KEEP_NONZERO_PATTERN: 1244 a->keepnonzeropattern = flg; 1245 break; 1246 case MAT_NEW_NONZERO_LOCATIONS: 1247 a->nonew = (flg ? 0 : 1); 1248 break; 1249 case MAT_NEW_NONZERO_LOCATION_ERR: 1250 a->nonew = (flg ? -1 : 0); 1251 break; 1252 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1253 a->nonew = (flg ? -2 : 0); 1254 break; 1255 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1256 a->nounused = (flg ? -1 : 0); 1257 break; 1258 case MAT_NEW_DIAGONALS: 1259 case MAT_IGNORE_OFF_PROC_ENTRIES: 1260 case MAT_USE_HASH_TABLE: 1261 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1262 break; 1263 case MAT_SPD: 1264 case MAT_SYMMETRIC: 1265 case MAT_STRUCTURALLY_SYMMETRIC: 1266 case MAT_HERMITIAN: 1267 case MAT_SYMMETRY_ETERNAL: 1268 /* These options are handled directly by MatSetOption() */ 1269 break; 1270 default: 1271 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1272 } 1273 PetscFunctionReturn(0); 1274 } 1275 1276 /* used for both SeqBAIJ and SeqSBAIJ matrices */ 1277 #undef __FUNCT__ 1278 #define __FUNCT__ "MatGetRow_SeqBAIJ_private" 1279 PetscErrorCode MatGetRow_SeqBAIJ_private(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v,PetscInt *ai,PetscInt *aj,PetscScalar *aa) 1280 { 1281 PetscErrorCode ierr; 1282 PetscInt itmp,i,j,k,M,bn,bp,*idx_i,bs,bs2; 1283 MatScalar *aa_i; 1284 PetscScalar *v_i; 1285 1286 PetscFunctionBegin; 1287 bs = A->rmap->bs; 1288 bs2 = bs*bs; 1289 if (row < 0 || row >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range", row); 1290 1291 bn = row/bs; /* Block number */ 1292 bp = row % bs; /* Block Position */ 1293 M = ai[bn+1] - ai[bn]; 1294 *nz = bs*M; 1295 1296 if (v) { 1297 *v = 0; 1298 if (*nz) { 1299 ierr = PetscMalloc1((*nz),v);CHKERRQ(ierr); 1300 for (i=0; i<M; i++) { /* for each block in the block row */ 1301 v_i = *v + i*bs; 1302 aa_i = aa + bs2*(ai[bn] + i); 1303 for (j=bp,k=0; j<bs2; j+=bs,k++) v_i[k] = aa_i[j]; 1304 } 1305 } 1306 } 1307 1308 if (idx) { 1309 *idx = 0; 1310 if (*nz) { 1311 ierr = PetscMalloc1((*nz),idx);CHKERRQ(ierr); 1312 for (i=0; i<M; i++) { /* for each block in the block row */ 1313 idx_i = *idx + i*bs; 1314 itmp = bs*aj[ai[bn] + i]; 1315 for (j=0; j<bs; j++) idx_i[j] = itmp++; 1316 } 1317 } 1318 } 1319 PetscFunctionReturn(0); 1320 } 1321 1322 #undef __FUNCT__ 1323 #define __FUNCT__ "MatGetRow_SeqBAIJ" 1324 PetscErrorCode MatGetRow_SeqBAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1325 { 1326 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1327 PetscErrorCode ierr; 1328 1329 PetscFunctionBegin; 1330 ierr = MatGetRow_SeqBAIJ_private(A,row,nz,idx,v,a->i,a->j,a->a);CHKERRQ(ierr); 1331 PetscFunctionReturn(0); 1332 } 1333 1334 #undef __FUNCT__ 1335 #define __FUNCT__ "MatRestoreRow_SeqBAIJ" 1336 PetscErrorCode MatRestoreRow_SeqBAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1337 { 1338 PetscErrorCode ierr; 1339 1340 PetscFunctionBegin; 1341 if (idx) {ierr = PetscFree(*idx);CHKERRQ(ierr);} 1342 if (v) {ierr = PetscFree(*v);CHKERRQ(ierr);} 1343 PetscFunctionReturn(0); 1344 } 1345 1346 extern PetscErrorCode MatSetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode); 1347 1348 #undef __FUNCT__ 1349 #define __FUNCT__ "MatTranspose_SeqBAIJ" 1350 PetscErrorCode MatTranspose_SeqBAIJ(Mat A,MatReuse reuse,Mat *B) 1351 { 1352 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data; 1353 Mat C; 1354 PetscErrorCode ierr; 1355 PetscInt i,j,k,*aj=a->j,*ai=a->i,bs=A->rmap->bs,mbs=a->mbs,nbs=a->nbs,len,*col; 1356 PetscInt *rows,*cols,bs2=a->bs2; 1357 MatScalar *array; 1358 1359 PetscFunctionBegin; 1360 if (reuse == MAT_REUSE_MATRIX && A == *B && mbs != nbs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1361 if (reuse == MAT_INITIAL_MATRIX || A == *B) { 1362 ierr = PetscCalloc1((1+nbs),&col);CHKERRQ(ierr); 1363 1364 for (i=0; i<ai[mbs]; i++) col[aj[i]] += 1; 1365 ierr = MatCreate(PetscObjectComm((PetscObject)A),&C);CHKERRQ(ierr); 1366 ierr = MatSetSizes(C,A->cmap->n,A->rmap->N,A->cmap->n,A->rmap->N);CHKERRQ(ierr); 1367 ierr = MatSetType(C,((PetscObject)A)->type_name);CHKERRQ(ierr); 1368 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(C,bs,0,col);CHKERRQ(ierr); 1369 ierr = PetscFree(col);CHKERRQ(ierr); 1370 } else { 1371 C = *B; 1372 } 1373 1374 array = a->a; 1375 ierr = PetscMalloc2(bs,&rows,bs,&cols);CHKERRQ(ierr); 1376 for (i=0; i<mbs; i++) { 1377 cols[0] = i*bs; 1378 for (k=1; k<bs; k++) cols[k] = cols[k-1] + 1; 1379 len = ai[i+1] - ai[i]; 1380 for (j=0; j<len; j++) { 1381 rows[0] = (*aj++)*bs; 1382 for (k=1; k<bs; k++) rows[k] = rows[k-1] + 1; 1383 ierr = MatSetValues_SeqBAIJ(C,bs,rows,bs,cols,array,INSERT_VALUES);CHKERRQ(ierr); 1384 array += bs2; 1385 } 1386 } 1387 ierr = PetscFree2(rows,cols);CHKERRQ(ierr); 1388 1389 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1390 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1391 1392 if (reuse == MAT_INITIAL_MATRIX || *B != A) { 1393 *B = C; 1394 } else { 1395 ierr = MatHeaderMerge(A,C);CHKERRQ(ierr); 1396 } 1397 PetscFunctionReturn(0); 1398 } 1399 1400 #undef __FUNCT__ 1401 #define __FUNCT__ "MatIsTranspose_SeqBAIJ" 1402 PetscErrorCode MatIsTranspose_SeqBAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f) 1403 { 1404 PetscErrorCode ierr; 1405 Mat Btrans; 1406 1407 PetscFunctionBegin; 1408 *f = PETSC_FALSE; 1409 ierr = MatTranspose_SeqBAIJ(A,MAT_INITIAL_MATRIX,&Btrans);CHKERRQ(ierr); 1410 ierr = MatEqual_SeqBAIJ(B,Btrans,f);CHKERRQ(ierr); 1411 ierr = MatDestroy(&Btrans);CHKERRQ(ierr); 1412 PetscFunctionReturn(0); 1413 } 1414 1415 #undef __FUNCT__ 1416 #define __FUNCT__ "MatView_SeqBAIJ_Binary" 1417 static PetscErrorCode MatView_SeqBAIJ_Binary(Mat A,PetscViewer viewer) 1418 { 1419 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1420 PetscErrorCode ierr; 1421 PetscInt i,*col_lens,bs = A->rmap->bs,count,*jj,j,k,l,bs2=a->bs2; 1422 int fd; 1423 PetscScalar *aa; 1424 FILE *file; 1425 1426 PetscFunctionBegin; 1427 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1428 ierr = PetscMalloc1((4+A->rmap->N),&col_lens);CHKERRQ(ierr); 1429 col_lens[0] = MAT_FILE_CLASSID; 1430 1431 col_lens[1] = A->rmap->N; 1432 col_lens[2] = A->cmap->n; 1433 col_lens[3] = a->nz*bs2; 1434 1435 /* store lengths of each row and write (including header) to file */ 1436 count = 0; 1437 for (i=0; i<a->mbs; i++) { 1438 for (j=0; j<bs; j++) { 1439 col_lens[4+count++] = bs*(a->i[i+1] - a->i[i]); 1440 } 1441 } 1442 ierr = PetscBinaryWrite(fd,col_lens,4+A->rmap->N,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1443 ierr = PetscFree(col_lens);CHKERRQ(ierr); 1444 1445 /* store column indices (zero start index) */ 1446 ierr = PetscMalloc1((a->nz+1)*bs2,&jj);CHKERRQ(ierr); 1447 count = 0; 1448 for (i=0; i<a->mbs; i++) { 1449 for (j=0; j<bs; j++) { 1450 for (k=a->i[i]; k<a->i[i+1]; k++) { 1451 for (l=0; l<bs; l++) { 1452 jj[count++] = bs*a->j[k] + l; 1453 } 1454 } 1455 } 1456 } 1457 ierr = PetscBinaryWrite(fd,jj,bs2*a->nz,PETSC_INT,PETSC_FALSE);CHKERRQ(ierr); 1458 ierr = PetscFree(jj);CHKERRQ(ierr); 1459 1460 /* store nonzero values */ 1461 ierr = PetscMalloc1((a->nz+1)*bs2,&aa);CHKERRQ(ierr); 1462 count = 0; 1463 for (i=0; i<a->mbs; i++) { 1464 for (j=0; j<bs; j++) { 1465 for (k=a->i[i]; k<a->i[i+1]; k++) { 1466 for (l=0; l<bs; l++) { 1467 aa[count++] = a->a[bs2*k + l*bs + j]; 1468 } 1469 } 1470 } 1471 } 1472 ierr = PetscBinaryWrite(fd,aa,bs2*a->nz,PETSC_SCALAR,PETSC_FALSE);CHKERRQ(ierr); 1473 ierr = PetscFree(aa);CHKERRQ(ierr); 1474 1475 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1476 if (file) { 1477 fprintf(file,"-matload_block_size %d\n",(int)A->rmap->bs); 1478 } 1479 PetscFunctionReturn(0); 1480 } 1481 1482 #undef __FUNCT__ 1483 #define __FUNCT__ "MatView_SeqBAIJ_ASCII" 1484 static PetscErrorCode MatView_SeqBAIJ_ASCII(Mat A,PetscViewer viewer) 1485 { 1486 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1487 PetscErrorCode ierr; 1488 PetscInt i,j,bs = A->rmap->bs,k,l,bs2=a->bs2; 1489 PetscViewerFormat format; 1490 1491 PetscFunctionBegin; 1492 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1493 if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1494 ierr = PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);CHKERRQ(ierr); 1495 } else if (format == PETSC_VIEWER_ASCII_MATLAB) { 1496 Mat aij; 1497 ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&aij);CHKERRQ(ierr); 1498 ierr = MatView(aij,viewer);CHKERRQ(ierr); 1499 ierr = MatDestroy(&aij);CHKERRQ(ierr); 1500 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1501 PetscFunctionReturn(0); 1502 } else if (format == PETSC_VIEWER_ASCII_COMMON) { 1503 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 1504 for (i=0; i<a->mbs; i++) { 1505 for (j=0; j<bs; j++) { 1506 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i*bs+j);CHKERRQ(ierr); 1507 for (k=a->i[i]; k<a->i[i+1]; k++) { 1508 for (l=0; l<bs; l++) { 1509 #if defined(PETSC_USE_COMPLEX) 1510 if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) > 0.0 && PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) { 1511 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %gi) ",bs*a->j[k]+l, 1512 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1513 } else if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) < 0.0 && PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) { 1514 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %gi) ",bs*a->j[k]+l, 1515 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),-(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1516 } else if (PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) { 1517 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)PetscRealPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1518 } 1519 #else 1520 if (a->a[bs2*k + l*bs + j] != 0.0) { 1521 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)a->a[bs2*k + l*bs + j]);CHKERRQ(ierr); 1522 } 1523 #endif 1524 } 1525 } 1526 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 1527 } 1528 } 1529 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 1530 } else { 1531 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 1532 for (i=0; i<a->mbs; i++) { 1533 for (j=0; j<bs; j++) { 1534 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i*bs+j);CHKERRQ(ierr); 1535 for (k=a->i[i]; k<a->i[i+1]; k++) { 1536 for (l=0; l<bs; l++) { 1537 #if defined(PETSC_USE_COMPLEX) 1538 if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) > 0.0) { 1539 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i) ",bs*a->j[k]+l, 1540 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1541 } else if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) < 0.0) { 1542 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i) ",bs*a->j[k]+l, 1543 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),-(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1544 } else { 1545 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)PetscRealPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1546 } 1547 #else 1548 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)a->a[bs2*k + l*bs + j]);CHKERRQ(ierr); 1549 #endif 1550 } 1551 } 1552 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 1553 } 1554 } 1555 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 1556 } 1557 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1558 PetscFunctionReturn(0); 1559 } 1560 1561 #include <petscdraw.h> 1562 #undef __FUNCT__ 1563 #define __FUNCT__ "MatView_SeqBAIJ_Draw_Zoom" 1564 static PetscErrorCode MatView_SeqBAIJ_Draw_Zoom(PetscDraw draw,void *Aa) 1565 { 1566 Mat A = (Mat) Aa; 1567 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data; 1568 PetscErrorCode ierr; 1569 PetscInt row,i,j,k,l,mbs=a->mbs,color,bs=A->rmap->bs,bs2=a->bs2; 1570 PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r; 1571 MatScalar *aa; 1572 PetscViewer viewer; 1573 PetscViewerFormat format; 1574 1575 PetscFunctionBegin; 1576 ierr = PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);CHKERRQ(ierr); 1577 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1578 1579 ierr = PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);CHKERRQ(ierr); 1580 1581 /* loop over matrix elements drawing boxes */ 1582 1583 if (format != PETSC_VIEWER_DRAW_CONTOUR) { 1584 color = PETSC_DRAW_BLUE; 1585 for (i=0,row=0; i<mbs; i++,row+=bs) { 1586 for (j=a->i[i]; j<a->i[i+1]; j++) { 1587 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1588 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1589 aa = a->a + j*bs2; 1590 for (k=0; k<bs; k++) { 1591 for (l=0; l<bs; l++) { 1592 if (PetscRealPart(*aa++) >= 0.) continue; 1593 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1594 } 1595 } 1596 } 1597 } 1598 color = PETSC_DRAW_CYAN; 1599 for (i=0,row=0; i<mbs; i++,row+=bs) { 1600 for (j=a->i[i]; j<a->i[i+1]; j++) { 1601 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1602 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1603 aa = a->a + j*bs2; 1604 for (k=0; k<bs; k++) { 1605 for (l=0; l<bs; l++) { 1606 if (PetscRealPart(*aa++) != 0.) continue; 1607 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1608 } 1609 } 1610 } 1611 } 1612 color = PETSC_DRAW_RED; 1613 for (i=0,row=0; i<mbs; i++,row+=bs) { 1614 for (j=a->i[i]; j<a->i[i+1]; j++) { 1615 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1616 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1617 aa = a->a + j*bs2; 1618 for (k=0; k<bs; k++) { 1619 for (l=0; l<bs; l++) { 1620 if (PetscRealPart(*aa++) <= 0.) continue; 1621 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1622 } 1623 } 1624 } 1625 } 1626 } else { 1627 /* use contour shading to indicate magnitude of values */ 1628 /* first determine max of all nonzero values */ 1629 PetscDraw popup; 1630 PetscReal scale,maxv = 0.0; 1631 1632 for (i=0; i<a->nz*a->bs2; i++) { 1633 if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]); 1634 } 1635 scale = (245.0 - PETSC_DRAW_BASIC_COLORS)/maxv; 1636 ierr = PetscDrawGetPopup(draw,&popup);CHKERRQ(ierr); 1637 if (popup) { 1638 ierr = PetscDrawScalePopup(popup,0.0,maxv);CHKERRQ(ierr); 1639 } 1640 for (i=0,row=0; i<mbs; i++,row+=bs) { 1641 for (j=a->i[i]; j<a->i[i+1]; j++) { 1642 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1643 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1644 aa = a->a + j*bs2; 1645 for (k=0; k<bs; k++) { 1646 for (l=0; l<bs; l++) { 1647 color = PETSC_DRAW_BASIC_COLORS + (PetscInt)(scale*PetscAbsScalar(*aa++)); 1648 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1649 } 1650 } 1651 } 1652 } 1653 } 1654 PetscFunctionReturn(0); 1655 } 1656 1657 #undef __FUNCT__ 1658 #define __FUNCT__ "MatView_SeqBAIJ_Draw" 1659 static PetscErrorCode MatView_SeqBAIJ_Draw(Mat A,PetscViewer viewer) 1660 { 1661 PetscErrorCode ierr; 1662 PetscReal xl,yl,xr,yr,w,h; 1663 PetscDraw draw; 1664 PetscBool isnull; 1665 1666 PetscFunctionBegin; 1667 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1668 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1669 1670 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);CHKERRQ(ierr); 1671 xr = A->cmap->n; yr = A->rmap->N; h = yr/10.0; w = xr/10.0; 1672 xr += w; yr += h; xl = -w; yl = -h; 1673 ierr = PetscDrawSetCoordinates(draw,xl,yl,xr,yr);CHKERRQ(ierr); 1674 ierr = PetscDrawZoom(draw,MatView_SeqBAIJ_Draw_Zoom,A);CHKERRQ(ierr); 1675 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);CHKERRQ(ierr); 1676 PetscFunctionReturn(0); 1677 } 1678 1679 #undef __FUNCT__ 1680 #define __FUNCT__ "MatView_SeqBAIJ" 1681 PetscErrorCode MatView_SeqBAIJ(Mat A,PetscViewer viewer) 1682 { 1683 PetscErrorCode ierr; 1684 PetscBool iascii,isbinary,isdraw; 1685 1686 PetscFunctionBegin; 1687 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1688 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1689 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1690 if (iascii) { 1691 ierr = MatView_SeqBAIJ_ASCII(A,viewer);CHKERRQ(ierr); 1692 } else if (isbinary) { 1693 ierr = MatView_SeqBAIJ_Binary(A,viewer);CHKERRQ(ierr); 1694 } else if (isdraw) { 1695 ierr = MatView_SeqBAIJ_Draw(A,viewer);CHKERRQ(ierr); 1696 } else { 1697 Mat B; 1698 ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 1699 ierr = MatView(B,viewer);CHKERRQ(ierr); 1700 ierr = MatDestroy(&B);CHKERRQ(ierr); 1701 } 1702 PetscFunctionReturn(0); 1703 } 1704 1705 1706 #undef __FUNCT__ 1707 #define __FUNCT__ "MatGetValues_SeqBAIJ" 1708 PetscErrorCode MatGetValues_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[]) 1709 { 1710 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1711 PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j; 1712 PetscInt *ai = a->i,*ailen = a->ilen; 1713 PetscInt brow,bcol,ridx,cidx,bs=A->rmap->bs,bs2=a->bs2; 1714 MatScalar *ap,*aa = a->a; 1715 1716 PetscFunctionBegin; 1717 for (k=0; k<m; k++) { /* loop over rows */ 1718 row = im[k]; brow = row/bs; 1719 if (row < 0) {v += n; continue;} /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); */ 1720 if (row >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D too large", row); 1721 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; 1722 nrow = ailen[brow]; 1723 for (l=0; l<n; l++) { /* loop over columns */ 1724 if (in[l] < 0) {v++; continue;} /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column"); */ 1725 if (in[l] >= A->cmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column %D too large", in[l]); 1726 col = in[l]; 1727 bcol = col/bs; 1728 cidx = col%bs; 1729 ridx = row%bs; 1730 high = nrow; 1731 low = 0; /* assume unsorted */ 1732 while (high-low > 5) { 1733 t = (low+high)/2; 1734 if (rp[t] > bcol) high = t; 1735 else low = t; 1736 } 1737 for (i=low; i<high; i++) { 1738 if (rp[i] > bcol) break; 1739 if (rp[i] == bcol) { 1740 *v++ = ap[bs2*i+bs*cidx+ridx]; 1741 goto finished; 1742 } 1743 } 1744 *v++ = 0.0; 1745 finished:; 1746 } 1747 } 1748 PetscFunctionReturn(0); 1749 } 1750 1751 #undef __FUNCT__ 1752 #define __FUNCT__ "MatSetValuesBlocked_SeqBAIJ" 1753 PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 1754 { 1755 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1756 PetscInt *rp,k,low,high,t,ii,jj,row,nrow,i,col,l,rmax,N,lastcol = -1; 1757 PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen; 1758 PetscErrorCode ierr; 1759 PetscInt *aj =a->j,nonew=a->nonew,bs2=a->bs2,bs=A->rmap->bs,stepval; 1760 PetscBool roworiented=a->roworiented; 1761 const PetscScalar *value = v; 1762 MatScalar *ap,*aa = a->a,*bap; 1763 1764 PetscFunctionBegin; 1765 if (roworiented) { 1766 stepval = (n-1)*bs; 1767 } else { 1768 stepval = (m-1)*bs; 1769 } 1770 for (k=0; k<m; k++) { /* loop over added rows */ 1771 row = im[k]; 1772 if (row < 0) continue; 1773 #if defined(PETSC_USE_DEBUG) 1774 if (row >= a->mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,a->mbs-1); 1775 #endif 1776 rp = aj + ai[row]; 1777 ap = aa + bs2*ai[row]; 1778 rmax = imax[row]; 1779 nrow = ailen[row]; 1780 low = 0; 1781 high = nrow; 1782 for (l=0; l<n; l++) { /* loop over added columns */ 1783 if (in[l] < 0) continue; 1784 #if defined(PETSC_USE_DEBUG) 1785 if (in[l] >= a->nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],a->nbs-1); 1786 #endif 1787 col = in[l]; 1788 if (roworiented) { 1789 value = v + (k*(stepval+bs) + l)*bs; 1790 } else { 1791 value = v + (l*(stepval+bs) + k)*bs; 1792 } 1793 if (col <= lastcol) low = 0; 1794 else high = nrow; 1795 lastcol = col; 1796 while (high-low > 7) { 1797 t = (low+high)/2; 1798 if (rp[t] > col) high = t; 1799 else low = t; 1800 } 1801 for (i=low; i<high; i++) { 1802 if (rp[i] > col) break; 1803 if (rp[i] == col) { 1804 bap = ap + bs2*i; 1805 if (roworiented) { 1806 if (is == ADD_VALUES) { 1807 for (ii=0; ii<bs; ii++,value+=stepval) { 1808 for (jj=ii; jj<bs2; jj+=bs) { 1809 bap[jj] += *value++; 1810 } 1811 } 1812 } else { 1813 for (ii=0; ii<bs; ii++,value+=stepval) { 1814 for (jj=ii; jj<bs2; jj+=bs) { 1815 bap[jj] = *value++; 1816 } 1817 } 1818 } 1819 } else { 1820 if (is == ADD_VALUES) { 1821 for (ii=0; ii<bs; ii++,value+=bs+stepval) { 1822 for (jj=0; jj<bs; jj++) { 1823 bap[jj] += value[jj]; 1824 } 1825 bap += bs; 1826 } 1827 } else { 1828 for (ii=0; ii<bs; ii++,value+=bs+stepval) { 1829 for (jj=0; jj<bs; jj++) { 1830 bap[jj] = value[jj]; 1831 } 1832 bap += bs; 1833 } 1834 } 1835 } 1836 goto noinsert2; 1837 } 1838 } 1839 if (nonew == 1) goto noinsert2; 1840 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col); 1841 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 1842 N = nrow++ - 1; high++; 1843 /* shift up all the later entries in this row */ 1844 for (ii=N; ii>=i; ii--) { 1845 rp[ii+1] = rp[ii]; 1846 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); 1847 } 1848 if (N >= i) { 1849 ierr = PetscMemzero(ap+bs2*i,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1850 } 1851 rp[i] = col; 1852 bap = ap + bs2*i; 1853 if (roworiented) { 1854 for (ii=0; ii<bs; ii++,value+=stepval) { 1855 for (jj=ii; jj<bs2; jj+=bs) { 1856 bap[jj] = *value++; 1857 } 1858 } 1859 } else { 1860 for (ii=0; ii<bs; ii++,value+=stepval) { 1861 for (jj=0; jj<bs; jj++) { 1862 *bap++ = *value++; 1863 } 1864 } 1865 } 1866 noinsert2:; 1867 low = i; 1868 } 1869 ailen[row] = nrow; 1870 } 1871 PetscFunctionReturn(0); 1872 } 1873 1874 #undef __FUNCT__ 1875 #define __FUNCT__ "MatAssemblyEnd_SeqBAIJ" 1876 PetscErrorCode MatAssemblyEnd_SeqBAIJ(Mat A,MatAssemblyType mode) 1877 { 1878 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1879 PetscInt fshift = 0,i,j,*ai = a->i,*aj = a->j,*imax = a->imax; 1880 PetscInt m = A->rmap->N,*ip,N,*ailen = a->ilen; 1881 PetscErrorCode ierr; 1882 PetscInt mbs = a->mbs,bs2 = a->bs2,rmax = 0; 1883 MatScalar *aa = a->a,*ap; 1884 PetscReal ratio=0.6; 1885 1886 PetscFunctionBegin; 1887 if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0); 1888 1889 if (m) rmax = ailen[0]; 1890 for (i=1; i<mbs; i++) { 1891 /* move each row back by the amount of empty slots (fshift) before it*/ 1892 fshift += imax[i-1] - ailen[i-1]; 1893 rmax = PetscMax(rmax,ailen[i]); 1894 if (fshift) { 1895 ip = aj + ai[i]; ap = aa + bs2*ai[i]; 1896 N = ailen[i]; 1897 for (j=0; j<N; j++) { 1898 ip[j-fshift] = ip[j]; 1899 1900 ierr = PetscMemcpy(ap+(j-fshift)*bs2,ap+j*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1901 } 1902 } 1903 ai[i] = ai[i-1] + ailen[i-1]; 1904 } 1905 if (mbs) { 1906 fshift += imax[mbs-1] - ailen[mbs-1]; 1907 ai[mbs] = ai[mbs-1] + ailen[mbs-1]; 1908 } 1909 1910 /* reset ilen and imax for each row */ 1911 a->nonzerorowcnt = 0; 1912 for (i=0; i<mbs; i++) { 1913 ailen[i] = imax[i] = ai[i+1] - ai[i]; 1914 a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0); 1915 } 1916 a->nz = ai[mbs]; 1917 1918 /* diagonals may have moved, so kill the diagonal pointers */ 1919 a->idiagvalid = PETSC_FALSE; 1920 if (fshift && a->diag) { 1921 ierr = PetscFree(a->diag);CHKERRQ(ierr); 1922 ierr = PetscLogObjectMemory((PetscObject)A,-(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1923 a->diag = 0; 1924 } 1925 if (fshift && a->nounused == -1) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB, "Unused space detected in matrix: %D X %D block size %D, %D unneeded", m, A->cmap->n, A->rmap->bs, fshift*bs2); 1926 ierr = PetscInfo5(A,"Matrix size: %D X %D, block size %D; storage space: %D unneeded, %D used\n",m,A->cmap->n,A->rmap->bs,fshift*bs2,a->nz*bs2);CHKERRQ(ierr); 1927 ierr = PetscInfo1(A,"Number of mallocs during MatSetValues is %D\n",a->reallocs);CHKERRQ(ierr); 1928 ierr = PetscInfo1(A,"Most nonzeros blocks in any row is %D\n",rmax);CHKERRQ(ierr); 1929 1930 A->info.mallocs += a->reallocs; 1931 a->reallocs = 0; 1932 A->info.nz_unneeded = (PetscReal)fshift*bs2; 1933 1934 ierr = MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,mbs,ratio);CHKERRQ(ierr); 1935 PetscFunctionReturn(0); 1936 } 1937 1938 /* 1939 This function returns an array of flags which indicate the locations of contiguous 1940 blocks that should be zeroed. for eg: if bs = 3 and is = [0,1,2,3,5,6,7,8,9] 1941 then the resulting sizes = [3,1,1,3,1] correspondig to sets [(0,1,2),(3),(5),(6,7,8),(9)] 1942 Assume: sizes should be long enough to hold all the values. 1943 */ 1944 #undef __FUNCT__ 1945 #define __FUNCT__ "MatZeroRows_SeqBAIJ_Check_Blocks" 1946 static PetscErrorCode MatZeroRows_SeqBAIJ_Check_Blocks(PetscInt idx[],PetscInt n,PetscInt bs,PetscInt sizes[], PetscInt *bs_max) 1947 { 1948 PetscInt i,j,k,row; 1949 PetscBool flg; 1950 1951 PetscFunctionBegin; 1952 for (i=0,j=0; i<n; j++) { 1953 row = idx[i]; 1954 if (row%bs!=0) { /* Not the begining of a block */ 1955 sizes[j] = 1; 1956 i++; 1957 } else if (i+bs > n) { /* complete block doesn't exist (at idx end) */ 1958 sizes[j] = 1; /* Also makes sure atleast 'bs' values exist for next else */ 1959 i++; 1960 } else { /* Begining of the block, so check if the complete block exists */ 1961 flg = PETSC_TRUE; 1962 for (k=1; k<bs; k++) { 1963 if (row+k != idx[i+k]) { /* break in the block */ 1964 flg = PETSC_FALSE; 1965 break; 1966 } 1967 } 1968 if (flg) { /* No break in the bs */ 1969 sizes[j] = bs; 1970 i += bs; 1971 } else { 1972 sizes[j] = 1; 1973 i++; 1974 } 1975 } 1976 } 1977 *bs_max = j; 1978 PetscFunctionReturn(0); 1979 } 1980 1981 #undef __FUNCT__ 1982 #define __FUNCT__ "MatZeroRows_SeqBAIJ" 1983 PetscErrorCode MatZeroRows_SeqBAIJ(Mat A,PetscInt is_n,const PetscInt is_idx[],PetscScalar diag,Vec x, Vec b) 1984 { 1985 Mat_SeqBAIJ *baij=(Mat_SeqBAIJ*)A->data; 1986 PetscErrorCode ierr; 1987 PetscInt i,j,k,count,*rows; 1988 PetscInt bs=A->rmap->bs,bs2=baij->bs2,*sizes,row,bs_max; 1989 PetscScalar zero = 0.0; 1990 MatScalar *aa; 1991 const PetscScalar *xx; 1992 PetscScalar *bb; 1993 1994 PetscFunctionBegin; 1995 /* fix right hand side if needed */ 1996 if (x && b) { 1997 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 1998 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 1999 for (i=0; i<is_n; i++) { 2000 bb[is_idx[i]] = diag*xx[is_idx[i]]; 2001 } 2002 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 2003 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 2004 } 2005 2006 /* Make a copy of the IS and sort it */ 2007 /* allocate memory for rows,sizes */ 2008 ierr = PetscMalloc2(is_n,&rows,2*is_n,&sizes);CHKERRQ(ierr); 2009 2010 /* copy IS values to rows, and sort them */ 2011 for (i=0; i<is_n; i++) rows[i] = is_idx[i]; 2012 ierr = PetscSortInt(is_n,rows);CHKERRQ(ierr); 2013 2014 if (baij->keepnonzeropattern) { 2015 for (i=0; i<is_n; i++) sizes[i] = 1; 2016 bs_max = is_n; 2017 } else { 2018 ierr = MatZeroRows_SeqBAIJ_Check_Blocks(rows,is_n,bs,sizes,&bs_max);CHKERRQ(ierr); 2019 A->nonzerostate++; 2020 } 2021 2022 for (i=0,j=0; i<bs_max; j+=sizes[i],i++) { 2023 row = rows[j]; 2024 if (row < 0 || row > A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range",row); 2025 count = (baij->i[row/bs +1] - baij->i[row/bs])*bs; 2026 aa = ((MatScalar*)(baij->a)) + baij->i[row/bs]*bs2 + (row%bs); 2027 if (sizes[i] == bs && !baij->keepnonzeropattern) { 2028 if (diag != (PetscScalar)0.0) { 2029 if (baij->ilen[row/bs] > 0) { 2030 baij->ilen[row/bs] = 1; 2031 baij->j[baij->i[row/bs]] = row/bs; 2032 2033 ierr = PetscMemzero(aa,count*bs*sizeof(MatScalar));CHKERRQ(ierr); 2034 } 2035 /* Now insert all the diagonal values for this bs */ 2036 for (k=0; k<bs; k++) { 2037 ierr = (*A->ops->setvalues)(A,1,rows+j+k,1,rows+j+k,&diag,INSERT_VALUES);CHKERRQ(ierr); 2038 } 2039 } else { /* (diag == 0.0) */ 2040 baij->ilen[row/bs] = 0; 2041 } /* end (diag == 0.0) */ 2042 } else { /* (sizes[i] != bs) */ 2043 #if defined(PETSC_USE_DEBUG) 2044 if (sizes[i] != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal Error. Value should be 1"); 2045 #endif 2046 for (k=0; k<count; k++) { 2047 aa[0] = zero; 2048 aa += bs; 2049 } 2050 if (diag != (PetscScalar)0.0) { 2051 ierr = (*A->ops->setvalues)(A,1,rows+j,1,rows+j,&diag,INSERT_VALUES);CHKERRQ(ierr); 2052 } 2053 } 2054 } 2055 2056 ierr = PetscFree2(rows,sizes);CHKERRQ(ierr); 2057 ierr = MatAssemblyEnd_SeqBAIJ(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2058 PetscFunctionReturn(0); 2059 } 2060 2061 #undef __FUNCT__ 2062 #define __FUNCT__ "MatZeroRowsColumns_SeqBAIJ" 2063 PetscErrorCode MatZeroRowsColumns_SeqBAIJ(Mat A,PetscInt is_n,const PetscInt is_idx[],PetscScalar diag,Vec x, Vec b) 2064 { 2065 Mat_SeqBAIJ *baij=(Mat_SeqBAIJ*)A->data; 2066 PetscErrorCode ierr; 2067 PetscInt i,j,k,count; 2068 PetscInt bs =A->rmap->bs,bs2=baij->bs2,row,col; 2069 PetscScalar zero = 0.0; 2070 MatScalar *aa; 2071 const PetscScalar *xx; 2072 PetscScalar *bb; 2073 PetscBool *zeroed,vecs = PETSC_FALSE; 2074 2075 PetscFunctionBegin; 2076 /* fix right hand side if needed */ 2077 if (x && b) { 2078 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 2079 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 2080 vecs = PETSC_TRUE; 2081 } 2082 2083 /* zero the columns */ 2084 ierr = PetscCalloc1(A->rmap->n,&zeroed);CHKERRQ(ierr); 2085 for (i=0; i<is_n; i++) { 2086 if (is_idx[i] < 0 || is_idx[i] >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range",is_idx[i]); 2087 zeroed[is_idx[i]] = PETSC_TRUE; 2088 } 2089 for (i=0; i<A->rmap->N; i++) { 2090 if (!zeroed[i]) { 2091 row = i/bs; 2092 for (j=baij->i[row]; j<baij->i[row+1]; j++) { 2093 for (k=0; k<bs; k++) { 2094 col = bs*baij->j[j] + k; 2095 if (zeroed[col]) { 2096 aa = ((MatScalar*)(baij->a)) + j*bs2 + (i%bs) + bs*k; 2097 if (vecs) bb[i] -= aa[0]*xx[col]; 2098 aa[0] = 0.0; 2099 } 2100 } 2101 } 2102 } else if (vecs) bb[i] = diag*xx[i]; 2103 } 2104 ierr = PetscFree(zeroed);CHKERRQ(ierr); 2105 if (vecs) { 2106 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 2107 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 2108 } 2109 2110 /* zero the rows */ 2111 for (i=0; i<is_n; i++) { 2112 row = is_idx[i]; 2113 count = (baij->i[row/bs +1] - baij->i[row/bs])*bs; 2114 aa = ((MatScalar*)(baij->a)) + baij->i[row/bs]*bs2 + (row%bs); 2115 for (k=0; k<count; k++) { 2116 aa[0] = zero; 2117 aa += bs; 2118 } 2119 if (diag != (PetscScalar)0.0) { 2120 ierr = (*A->ops->setvalues)(A,1,&row,1,&row,&diag,INSERT_VALUES);CHKERRQ(ierr); 2121 } 2122 } 2123 ierr = MatAssemblyEnd_SeqBAIJ(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2124 PetscFunctionReturn(0); 2125 } 2126 2127 #undef __FUNCT__ 2128 #define __FUNCT__ "MatSetValues_SeqBAIJ" 2129 PetscErrorCode MatSetValues_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 2130 { 2131 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2132 PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N,lastcol = -1; 2133 PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen; 2134 PetscInt *aj =a->j,nonew=a->nonew,bs=A->rmap->bs,brow,bcol; 2135 PetscErrorCode ierr; 2136 PetscInt ridx,cidx,bs2=a->bs2; 2137 PetscBool roworiented=a->roworiented; 2138 MatScalar *ap,value,*aa=a->a,*bap; 2139 2140 PetscFunctionBegin; 2141 for (k=0; k<m; k++) { /* loop over added rows */ 2142 row = im[k]; 2143 brow = row/bs; 2144 if (row < 0) continue; 2145 #if defined(PETSC_USE_DEBUG) 2146 if (row >= A->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->N-1); 2147 #endif 2148 rp = aj + ai[brow]; 2149 ap = aa + bs2*ai[brow]; 2150 rmax = imax[brow]; 2151 nrow = ailen[brow]; 2152 low = 0; 2153 high = nrow; 2154 for (l=0; l<n; l++) { /* loop over added columns */ 2155 if (in[l] < 0) continue; 2156 #if defined(PETSC_USE_DEBUG) 2157 if (in[l] >= A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1); 2158 #endif 2159 col = in[l]; bcol = col/bs; 2160 ridx = row % bs; cidx = col % bs; 2161 if (roworiented) { 2162 value = v[l + k*n]; 2163 } else { 2164 value = v[k + l*m]; 2165 } 2166 if (col <= lastcol) low = 0; else high = nrow; 2167 lastcol = col; 2168 while (high-low > 7) { 2169 t = (low+high)/2; 2170 if (rp[t] > bcol) high = t; 2171 else low = t; 2172 } 2173 for (i=low; i<high; i++) { 2174 if (rp[i] > bcol) break; 2175 if (rp[i] == bcol) { 2176 bap = ap + bs2*i + bs*cidx + ridx; 2177 if (is == ADD_VALUES) *bap += value; 2178 else *bap = value; 2179 goto noinsert1; 2180 } 2181 } 2182 if (nonew == 1) goto noinsert1; 2183 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col); 2184 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 2185 N = nrow++ - 1; high++; 2186 /* shift up all the later entries in this row */ 2187 for (ii=N; ii>=i; ii--) { 2188 rp[ii+1] = rp[ii]; 2189 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); 2190 } 2191 if (N>=i) { 2192 ierr = PetscMemzero(ap+bs2*i,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2193 } 2194 rp[i] = bcol; 2195 ap[bs2*i + bs*cidx + ridx] = value; 2196 a->nz++; 2197 A->nonzerostate++; 2198 noinsert1:; 2199 low = i; 2200 } 2201 ailen[brow] = nrow; 2202 } 2203 PetscFunctionReturn(0); 2204 } 2205 2206 #undef __FUNCT__ 2207 #define __FUNCT__ "MatILUFactor_SeqBAIJ" 2208 PetscErrorCode MatILUFactor_SeqBAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info) 2209 { 2210 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)inA->data; 2211 Mat outA; 2212 PetscErrorCode ierr; 2213 PetscBool row_identity,col_identity; 2214 2215 PetscFunctionBegin; 2216 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for in-place ILU"); 2217 ierr = ISIdentity(row,&row_identity);CHKERRQ(ierr); 2218 ierr = ISIdentity(col,&col_identity);CHKERRQ(ierr); 2219 if (!row_identity || !col_identity) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for in-place ILU"); 2220 2221 outA = inA; 2222 inA->factortype = MAT_FACTOR_LU; 2223 2224 ierr = MatMarkDiagonal_SeqBAIJ(inA);CHKERRQ(ierr); 2225 2226 ierr = PetscObjectReference((PetscObject)row);CHKERRQ(ierr); 2227 ierr = ISDestroy(&a->row);CHKERRQ(ierr); 2228 a->row = row; 2229 ierr = PetscObjectReference((PetscObject)col);CHKERRQ(ierr); 2230 ierr = ISDestroy(&a->col);CHKERRQ(ierr); 2231 a->col = col; 2232 2233 /* Create the invert permutation so that it can be used in MatLUFactorNumeric() */ 2234 ierr = ISDestroy(&a->icol);CHKERRQ(ierr); 2235 ierr = ISInvertPermutation(col,PETSC_DECIDE,&a->icol);CHKERRQ(ierr); 2236 ierr = PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);CHKERRQ(ierr); 2237 2238 ierr = MatSeqBAIJSetNumericFactorization_inplace(inA,(PetscBool)(row_identity && col_identity));CHKERRQ(ierr); 2239 if (!a->solve_work) { 2240 ierr = PetscMalloc1((inA->rmap->N+inA->rmap->bs),&a->solve_work);CHKERRQ(ierr); 2241 ierr = PetscLogObjectMemory((PetscObject)inA,(inA->rmap->N+inA->rmap->bs)*sizeof(PetscScalar));CHKERRQ(ierr); 2242 } 2243 ierr = MatLUFactorNumeric(outA,inA,info);CHKERRQ(ierr); 2244 PetscFunctionReturn(0); 2245 } 2246 2247 #undef __FUNCT__ 2248 #define __FUNCT__ "MatSeqBAIJSetColumnIndices_SeqBAIJ" 2249 PetscErrorCode MatSeqBAIJSetColumnIndices_SeqBAIJ(Mat mat,PetscInt *indices) 2250 { 2251 Mat_SeqBAIJ *baij = (Mat_SeqBAIJ*)mat->data; 2252 PetscInt i,nz,mbs; 2253 2254 PetscFunctionBegin; 2255 nz = baij->maxnz; 2256 mbs = baij->mbs; 2257 for (i=0; i<nz; i++) { 2258 baij->j[i] = indices[i]; 2259 } 2260 baij->nz = nz; 2261 for (i=0; i<mbs; i++) { 2262 baij->ilen[i] = baij->imax[i]; 2263 } 2264 PetscFunctionReturn(0); 2265 } 2266 2267 #undef __FUNCT__ 2268 #define __FUNCT__ "MatSeqBAIJSetColumnIndices" 2269 /*@ 2270 MatSeqBAIJSetColumnIndices - Set the column indices for all the rows 2271 in the matrix. 2272 2273 Input Parameters: 2274 + mat - the SeqBAIJ matrix 2275 - indices - the column indices 2276 2277 Level: advanced 2278 2279 Notes: 2280 This can be called if you have precomputed the nonzero structure of the 2281 matrix and want to provide it to the matrix object to improve the performance 2282 of the MatSetValues() operation. 2283 2284 You MUST have set the correct numbers of nonzeros per row in the call to 2285 MatCreateSeqBAIJ(), and the columns indices MUST be sorted. 2286 2287 MUST be called before any calls to MatSetValues(); 2288 2289 @*/ 2290 PetscErrorCode MatSeqBAIJSetColumnIndices(Mat mat,PetscInt *indices) 2291 { 2292 PetscErrorCode ierr; 2293 2294 PetscFunctionBegin; 2295 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 2296 PetscValidPointer(indices,2); 2297 ierr = PetscUseMethod(mat,"MatSeqBAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));CHKERRQ(ierr); 2298 PetscFunctionReturn(0); 2299 } 2300 2301 #undef __FUNCT__ 2302 #define __FUNCT__ "MatGetRowMaxAbs_SeqBAIJ" 2303 PetscErrorCode MatGetRowMaxAbs_SeqBAIJ(Mat A,Vec v,PetscInt idx[]) 2304 { 2305 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2306 PetscErrorCode ierr; 2307 PetscInt i,j,n,row,bs,*ai,*aj,mbs; 2308 PetscReal atmp; 2309 PetscScalar *x,zero = 0.0; 2310 MatScalar *aa; 2311 PetscInt ncols,brow,krow,kcol; 2312 2313 PetscFunctionBegin; 2314 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 2315 bs = A->rmap->bs; 2316 aa = a->a; 2317 ai = a->i; 2318 aj = a->j; 2319 mbs = a->mbs; 2320 2321 ierr = VecSet(v,zero);CHKERRQ(ierr); 2322 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 2323 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 2324 if (n != A->rmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 2325 for (i=0; i<mbs; i++) { 2326 ncols = ai[1] - ai[0]; ai++; 2327 brow = bs*i; 2328 for (j=0; j<ncols; j++) { 2329 for (kcol=0; kcol<bs; kcol++) { 2330 for (krow=0; krow<bs; krow++) { 2331 atmp = PetscAbsScalar(*aa);aa++; 2332 row = brow + krow; /* row index */ 2333 if (PetscAbsScalar(x[row]) < atmp) {x[row] = atmp; if (idx) idx[row] = bs*(*aj) + kcol;} 2334 } 2335 } 2336 aj++; 2337 } 2338 } 2339 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 2340 PetscFunctionReturn(0); 2341 } 2342 2343 #undef __FUNCT__ 2344 #define __FUNCT__ "MatCopy_SeqBAIJ" 2345 PetscErrorCode MatCopy_SeqBAIJ(Mat A,Mat B,MatStructure str) 2346 { 2347 PetscErrorCode ierr; 2348 2349 PetscFunctionBegin; 2350 /* If the two matrices have the same copy implementation, use fast copy. */ 2351 if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) { 2352 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2353 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)B->data; 2354 PetscInt ambs=a->mbs,bmbs=b->mbs,abs=A->rmap->bs,bbs=B->rmap->bs,bs2=abs*abs; 2355 2356 if (a->i[ambs] != b->i[bmbs]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number of nonzero blocks in matrices A %D and B %D are different",a->i[ambs],b->i[bmbs]); 2357 if (abs != bbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Block size A %D and B %D are different",abs,bbs); 2358 ierr = PetscMemcpy(b->a,a->a,(bs2*a->i[ambs])*sizeof(PetscScalar));CHKERRQ(ierr); 2359 } else { 2360 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2361 } 2362 PetscFunctionReturn(0); 2363 } 2364 2365 #undef __FUNCT__ 2366 #define __FUNCT__ "MatSetUp_SeqBAIJ" 2367 PetscErrorCode MatSetUp_SeqBAIJ(Mat A) 2368 { 2369 PetscErrorCode ierr; 2370 2371 PetscFunctionBegin; 2372 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(A,A->rmap->bs,PETSC_DEFAULT,0);CHKERRQ(ierr); 2373 PetscFunctionReturn(0); 2374 } 2375 2376 #undef __FUNCT__ 2377 #define __FUNCT__ "MatSeqBAIJGetArray_SeqBAIJ" 2378 PetscErrorCode MatSeqBAIJGetArray_SeqBAIJ(Mat A,PetscScalar *array[]) 2379 { 2380 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2381 2382 PetscFunctionBegin; 2383 *array = a->a; 2384 PetscFunctionReturn(0); 2385 } 2386 2387 #undef __FUNCT__ 2388 #define __FUNCT__ "MatSeqBAIJRestoreArray_SeqBAIJ" 2389 PetscErrorCode MatSeqBAIJRestoreArray_SeqBAIJ(Mat A,PetscScalar *array[]) 2390 { 2391 PetscFunctionBegin; 2392 PetscFunctionReturn(0); 2393 } 2394 2395 #undef __FUNCT__ 2396 #define __FUNCT__ "MatAXPYGetPreallocation_SeqBAIJ" 2397 PetscErrorCode MatAXPYGetPreallocation_SeqBAIJ(Mat Y,Mat X,PetscInt *nnz) 2398 { 2399 PetscInt bs = Y->rmap->bs,mbs = Y->rmap->N/bs; 2400 Mat_SeqBAIJ *x = (Mat_SeqBAIJ*)X->data; 2401 Mat_SeqBAIJ *y = (Mat_SeqBAIJ*)Y->data; 2402 PetscErrorCode ierr; 2403 2404 PetscFunctionBegin; 2405 /* Set the number of nonzeros in the new matrix */ 2406 ierr = MatAXPYGetPreallocation_SeqX_private(mbs,x->i,x->j,y->i,y->j,nnz);CHKERRQ(ierr); 2407 PetscFunctionReturn(0); 2408 } 2409 2410 #undef __FUNCT__ 2411 #define __FUNCT__ "MatAXPY_SeqBAIJ" 2412 PetscErrorCode MatAXPY_SeqBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2413 { 2414 Mat_SeqBAIJ *x = (Mat_SeqBAIJ*)X->data,*y = (Mat_SeqBAIJ*)Y->data; 2415 PetscErrorCode ierr; 2416 PetscInt i,bs=Y->rmap->bs,j,bs2=bs*bs; 2417 PetscBLASInt one=1; 2418 2419 PetscFunctionBegin; 2420 if (str == SAME_NONZERO_PATTERN) { 2421 PetscScalar alpha = a; 2422 PetscBLASInt bnz; 2423 ierr = PetscBLASIntCast(x->nz*bs2,&bnz);CHKERRQ(ierr); 2424 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2425 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2426 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2427 if (y->xtoy && y->XtoY != X) { 2428 ierr = PetscFree(y->xtoy);CHKERRQ(ierr); 2429 ierr = MatDestroy(&y->XtoY);CHKERRQ(ierr); 2430 } 2431 if (!y->xtoy) { /* get xtoy */ 2432 ierr = MatAXPYGetxtoy_Private(x->mbs,x->i,x->j,NULL, y->i,y->j,NULL, &y->xtoy);CHKERRQ(ierr); 2433 y->XtoY = X; 2434 ierr = PetscObjectReference((PetscObject)X);CHKERRQ(ierr); 2435 } 2436 for (i=0; i<x->nz; i++) { 2437 j = 0; 2438 while (j < bs2) { 2439 y->a[bs2*y->xtoy[i]+j] += a*(x->a[bs2*i+j]); 2440 j++; 2441 } 2442 } 2443 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2444 ierr = PetscInfo3(Y,"ratio of nnz(X)/nnz(Y): %D/%D = %g\n",bs2*x->nz,bs2*y->nz,(double)((PetscReal)(bs2*x->nz)/(bs2*y->nz)));CHKERRQ(ierr); 2445 } else { 2446 Mat B; 2447 PetscInt *nnz; 2448 if (bs != X->rmap->bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrices must have same block size"); 2449 ierr = PetscMalloc1(Y->rmap->N,&nnz);CHKERRQ(ierr); 2450 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2451 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2452 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2453 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2454 ierr = MatSetType(B,(MatType) ((PetscObject)Y)->type_name);CHKERRQ(ierr); 2455 ierr = MatAXPYGetPreallocation_SeqBAIJ(Y,X,nnz);CHKERRQ(ierr); 2456 ierr = MatSeqBAIJSetPreallocation(B,bs,0,nnz);CHKERRQ(ierr); 2457 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2458 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2459 ierr = PetscFree(nnz);CHKERRQ(ierr); 2460 } 2461 PetscFunctionReturn(0); 2462 } 2463 2464 #undef __FUNCT__ 2465 #define __FUNCT__ "MatRealPart_SeqBAIJ" 2466 PetscErrorCode MatRealPart_SeqBAIJ(Mat A) 2467 { 2468 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2469 PetscInt i,nz = a->bs2*a->i[a->mbs]; 2470 MatScalar *aa = a->a; 2471 2472 PetscFunctionBegin; 2473 for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]); 2474 PetscFunctionReturn(0); 2475 } 2476 2477 #undef __FUNCT__ 2478 #define __FUNCT__ "MatImaginaryPart_SeqBAIJ" 2479 PetscErrorCode MatImaginaryPart_SeqBAIJ(Mat A) 2480 { 2481 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2482 PetscInt i,nz = a->bs2*a->i[a->mbs]; 2483 MatScalar *aa = a->a; 2484 2485 PetscFunctionBegin; 2486 for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]); 2487 PetscFunctionReturn(0); 2488 } 2489 2490 #undef __FUNCT__ 2491 #define __FUNCT__ "MatGetColumnIJ_SeqBAIJ" 2492 /* 2493 Code almost idential to MatGetColumnIJ_SeqAIJ() should share common code 2494 */ 2495 PetscErrorCode MatGetColumnIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 2496 { 2497 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2498 PetscErrorCode ierr; 2499 PetscInt bs = A->rmap->bs,i,*collengths,*cia,*cja,n = A->cmap->n/bs,m = A->rmap->n/bs; 2500 PetscInt nz = a->i[m],row,*jj,mr,col; 2501 2502 PetscFunctionBegin; 2503 *nn = n; 2504 if (!ia) PetscFunctionReturn(0); 2505 if (symmetric) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for BAIJ matrices"); 2506 else { 2507 ierr = PetscCalloc1((n+1),&collengths);CHKERRQ(ierr); 2508 ierr = PetscMalloc1((n+1),&cia);CHKERRQ(ierr); 2509 ierr = PetscMalloc1((nz+1),&cja);CHKERRQ(ierr); 2510 jj = a->j; 2511 for (i=0; i<nz; i++) { 2512 collengths[jj[i]]++; 2513 } 2514 cia[0] = oshift; 2515 for (i=0; i<n; i++) { 2516 cia[i+1] = cia[i] + collengths[i]; 2517 } 2518 ierr = PetscMemzero(collengths,n*sizeof(PetscInt));CHKERRQ(ierr); 2519 jj = a->j; 2520 for (row=0; row<m; row++) { 2521 mr = a->i[row+1] - a->i[row]; 2522 for (i=0; i<mr; i++) { 2523 col = *jj++; 2524 2525 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 2526 } 2527 } 2528 ierr = PetscFree(collengths);CHKERRQ(ierr); 2529 *ia = cia; *ja = cja; 2530 } 2531 PetscFunctionReturn(0); 2532 } 2533 2534 #undef __FUNCT__ 2535 #define __FUNCT__ "MatRestoreColumnIJ_SeqBAIJ" 2536 PetscErrorCode MatRestoreColumnIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 2537 { 2538 PetscErrorCode ierr; 2539 2540 PetscFunctionBegin; 2541 if (!ia) PetscFunctionReturn(0); 2542 ierr = PetscFree(*ia);CHKERRQ(ierr); 2543 ierr = PetscFree(*ja);CHKERRQ(ierr); 2544 PetscFunctionReturn(0); 2545 } 2546 2547 /* 2548 MatGetColumnIJ_SeqBAIJ_Color() and MatRestoreColumnIJ_SeqBAIJ_Color() are customized from 2549 MatGetColumnIJ_SeqBAIJ() and MatRestoreColumnIJ_SeqBAIJ() by adding an output 2550 spidx[], index of a->a, to be used in MatTransposeColoringCreate() and MatFDColoringCreate() 2551 */ 2552 #undef __FUNCT__ 2553 #define __FUNCT__ "MatGetColumnIJ_SeqBAIJ_Color" 2554 PetscErrorCode MatGetColumnIJ_SeqBAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 2555 { 2556 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2557 PetscErrorCode ierr; 2558 PetscInt i,*collengths,*cia,*cja,n=a->nbs,m=a->mbs; 2559 PetscInt nz = a->i[m],row,*jj,mr,col; 2560 PetscInt *cspidx; 2561 2562 PetscFunctionBegin; 2563 *nn = n; 2564 if (!ia) PetscFunctionReturn(0); 2565 2566 ierr = PetscCalloc1((n+1),&collengths);CHKERRQ(ierr); 2567 ierr = PetscMalloc1((n+1),&cia);CHKERRQ(ierr); 2568 ierr = PetscMalloc1((nz+1),&cja);CHKERRQ(ierr); 2569 ierr = PetscMalloc1((nz+1),&cspidx);CHKERRQ(ierr); 2570 jj = a->j; 2571 for (i=0; i<nz; i++) { 2572 collengths[jj[i]]++; 2573 } 2574 cia[0] = oshift; 2575 for (i=0; i<n; i++) { 2576 cia[i+1] = cia[i] + collengths[i]; 2577 } 2578 ierr = PetscMemzero(collengths,n*sizeof(PetscInt));CHKERRQ(ierr); 2579 jj = a->j; 2580 for (row=0; row<m; row++) { 2581 mr = a->i[row+1] - a->i[row]; 2582 for (i=0; i<mr; i++) { 2583 col = *jj++; 2584 cspidx[cia[col] + collengths[col] - oshift] = a->i[row] + i; /* index of a->j */ 2585 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 2586 } 2587 } 2588 ierr = PetscFree(collengths);CHKERRQ(ierr); 2589 *ia = cia; *ja = cja; 2590 *spidx = cspidx; 2591 PetscFunctionReturn(0); 2592 } 2593 2594 #undef __FUNCT__ 2595 #define __FUNCT__ "MatRestoreColumnIJ_SeqBAIJ_Color" 2596 PetscErrorCode MatRestoreColumnIJ_SeqBAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 2597 { 2598 PetscErrorCode ierr; 2599 2600 PetscFunctionBegin; 2601 ierr = MatRestoreColumnIJ_SeqBAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);CHKERRQ(ierr); 2602 ierr = PetscFree(*spidx);CHKERRQ(ierr); 2603 PetscFunctionReturn(0); 2604 } 2605 2606 /* -------------------------------------------------------------------*/ 2607 static struct _MatOps MatOps_Values = {MatSetValues_SeqBAIJ, 2608 MatGetRow_SeqBAIJ, 2609 MatRestoreRow_SeqBAIJ, 2610 MatMult_SeqBAIJ_N, 2611 /* 4*/ MatMultAdd_SeqBAIJ_N, 2612 MatMultTranspose_SeqBAIJ, 2613 MatMultTransposeAdd_SeqBAIJ, 2614 0, 2615 0, 2616 0, 2617 /* 10*/ 0, 2618 MatLUFactor_SeqBAIJ, 2619 0, 2620 0, 2621 MatTranspose_SeqBAIJ, 2622 /* 15*/ MatGetInfo_SeqBAIJ, 2623 MatEqual_SeqBAIJ, 2624 MatGetDiagonal_SeqBAIJ, 2625 MatDiagonalScale_SeqBAIJ, 2626 MatNorm_SeqBAIJ, 2627 /* 20*/ 0, 2628 MatAssemblyEnd_SeqBAIJ, 2629 MatSetOption_SeqBAIJ, 2630 MatZeroEntries_SeqBAIJ, 2631 /* 24*/ MatZeroRows_SeqBAIJ, 2632 0, 2633 0, 2634 0, 2635 0, 2636 /* 29*/ MatSetUp_SeqBAIJ, 2637 0, 2638 0, 2639 0, 2640 0, 2641 /* 34*/ MatDuplicate_SeqBAIJ, 2642 0, 2643 0, 2644 MatILUFactor_SeqBAIJ, 2645 0, 2646 /* 39*/ MatAXPY_SeqBAIJ, 2647 MatGetSubMatrices_SeqBAIJ, 2648 MatIncreaseOverlap_SeqBAIJ, 2649 MatGetValues_SeqBAIJ, 2650 MatCopy_SeqBAIJ, 2651 /* 44*/ 0, 2652 MatScale_SeqBAIJ, 2653 0, 2654 0, 2655 MatZeroRowsColumns_SeqBAIJ, 2656 /* 49*/ 0, 2657 MatGetRowIJ_SeqBAIJ, 2658 MatRestoreRowIJ_SeqBAIJ, 2659 MatGetColumnIJ_SeqBAIJ, 2660 MatRestoreColumnIJ_SeqBAIJ, 2661 /* 54*/ MatFDColoringCreate_SeqXAIJ, 2662 0, 2663 0, 2664 0, 2665 MatSetValuesBlocked_SeqBAIJ, 2666 /* 59*/ MatGetSubMatrix_SeqBAIJ, 2667 MatDestroy_SeqBAIJ, 2668 MatView_SeqBAIJ, 2669 0, 2670 0, 2671 /* 64*/ 0, 2672 0, 2673 0, 2674 0, 2675 0, 2676 /* 69*/ MatGetRowMaxAbs_SeqBAIJ, 2677 0, 2678 MatConvert_Basic, 2679 0, 2680 0, 2681 /* 74*/ 0, 2682 MatFDColoringApply_BAIJ, 2683 0, 2684 0, 2685 0, 2686 /* 79*/ 0, 2687 0, 2688 0, 2689 0, 2690 MatLoad_SeqBAIJ, 2691 /* 84*/ 0, 2692 0, 2693 0, 2694 0, 2695 0, 2696 /* 89*/ 0, 2697 0, 2698 0, 2699 0, 2700 0, 2701 /* 94*/ 0, 2702 0, 2703 0, 2704 0, 2705 0, 2706 /* 99*/ 0, 2707 0, 2708 0, 2709 0, 2710 0, 2711 /*104*/ 0, 2712 MatRealPart_SeqBAIJ, 2713 MatImaginaryPart_SeqBAIJ, 2714 0, 2715 0, 2716 /*109*/ 0, 2717 0, 2718 0, 2719 0, 2720 MatMissingDiagonal_SeqBAIJ, 2721 /*114*/ 0, 2722 0, 2723 0, 2724 0, 2725 0, 2726 /*119*/ 0, 2727 0, 2728 MatMultHermitianTranspose_SeqBAIJ, 2729 MatMultHermitianTransposeAdd_SeqBAIJ, 2730 0, 2731 /*124*/ 0, 2732 0, 2733 MatInvertBlockDiagonal_SeqBAIJ, 2734 0, 2735 0, 2736 /*129*/ 0, 2737 0, 2738 0, 2739 0, 2740 0, 2741 /*134*/ 0, 2742 0, 2743 0, 2744 0, 2745 0, 2746 /*139*/ 0, 2747 0, 2748 0, 2749 MatFDColoringSetUp_SeqXAIJ 2750 }; 2751 2752 #undef __FUNCT__ 2753 #define __FUNCT__ "MatStoreValues_SeqBAIJ" 2754 PetscErrorCode MatStoreValues_SeqBAIJ(Mat mat) 2755 { 2756 Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)mat->data; 2757 PetscInt nz = aij->i[aij->mbs]*aij->bs2; 2758 PetscErrorCode ierr; 2759 2760 PetscFunctionBegin; 2761 if (aij->nonew != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 2762 2763 /* allocate space for values if not already there */ 2764 if (!aij->saved_values) { 2765 ierr = PetscMalloc1((nz+1),&aij->saved_values);CHKERRQ(ierr); 2766 ierr = PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));CHKERRQ(ierr); 2767 } 2768 2769 /* copy values over */ 2770 ierr = PetscMemcpy(aij->saved_values,aij->a,nz*sizeof(PetscScalar));CHKERRQ(ierr); 2771 PetscFunctionReturn(0); 2772 } 2773 2774 #undef __FUNCT__ 2775 #define __FUNCT__ "MatRetrieveValues_SeqBAIJ" 2776 PetscErrorCode MatRetrieveValues_SeqBAIJ(Mat mat) 2777 { 2778 Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)mat->data; 2779 PetscErrorCode ierr; 2780 PetscInt nz = aij->i[aij->mbs]*aij->bs2; 2781 2782 PetscFunctionBegin; 2783 if (aij->nonew != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 2784 if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first"); 2785 2786 /* copy values over */ 2787 ierr = PetscMemcpy(aij->a,aij->saved_values,nz*sizeof(PetscScalar));CHKERRQ(ierr); 2788 PetscFunctionReturn(0); 2789 } 2790 2791 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat, MatType,MatReuse,Mat*); 2792 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqSBAIJ(Mat, MatType,MatReuse,Mat*); 2793 2794 #undef __FUNCT__ 2795 #define __FUNCT__ "MatSeqBAIJSetPreallocation_SeqBAIJ" 2796 PetscErrorCode MatSeqBAIJSetPreallocation_SeqBAIJ(Mat B,PetscInt bs,PetscInt nz,PetscInt *nnz) 2797 { 2798 Mat_SeqBAIJ *b; 2799 PetscErrorCode ierr; 2800 PetscInt i,mbs,nbs,bs2; 2801 PetscBool flg,skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE; 2802 2803 PetscFunctionBegin; 2804 if (nz >= 0 || nnz) realalloc = PETSC_TRUE; 2805 if (nz == MAT_SKIP_ALLOCATION) { 2806 skipallocation = PETSC_TRUE; 2807 nz = 0; 2808 } 2809 2810 ierr = MatSetBlockSize(B,PetscAbs(bs));CHKERRQ(ierr); 2811 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2812 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2813 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2814 2815 B->preallocated = PETSC_TRUE; 2816 2817 mbs = B->rmap->n/bs; 2818 nbs = B->cmap->n/bs; 2819 bs2 = bs*bs; 2820 2821 if (mbs*bs!=B->rmap->n || nbs*bs!=B->cmap->n) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number rows %D, cols %D must be divisible by blocksize %D",B->rmap->N,B->cmap->n,bs); 2822 2823 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 2824 if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz); 2825 if (nnz) { 2826 for (i=0; i<mbs; i++) { 2827 if (nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be less than 0: local row %D value %D",i,nnz[i]); 2828 if (nnz[i] > nbs) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be greater than block row length: local row %D value %D rowlength %D",i,nnz[i],nbs); 2829 } 2830 } 2831 2832 b = (Mat_SeqBAIJ*)B->data; 2833 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Optimize options for SEQBAIJ matrix 2 ","Mat");CHKERRQ(ierr); 2834 ierr = PetscOptionsBool("-mat_no_unroll","Do not optimize for block size (slow)",NULL,PETSC_FALSE,&flg,NULL);CHKERRQ(ierr); 2835 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2836 2837 if (!flg) { 2838 switch (bs) { 2839 case 1: 2840 B->ops->mult = MatMult_SeqBAIJ_1; 2841 B->ops->multadd = MatMultAdd_SeqBAIJ_1; 2842 break; 2843 case 2: 2844 B->ops->mult = MatMult_SeqBAIJ_2; 2845 B->ops->multadd = MatMultAdd_SeqBAIJ_2; 2846 break; 2847 case 3: 2848 B->ops->mult = MatMult_SeqBAIJ_3; 2849 B->ops->multadd = MatMultAdd_SeqBAIJ_3; 2850 break; 2851 case 4: 2852 B->ops->mult = MatMult_SeqBAIJ_4; 2853 B->ops->multadd = MatMultAdd_SeqBAIJ_4; 2854 break; 2855 case 5: 2856 B->ops->mult = MatMult_SeqBAIJ_5; 2857 B->ops->multadd = MatMultAdd_SeqBAIJ_5; 2858 break; 2859 case 6: 2860 B->ops->mult = MatMult_SeqBAIJ_6; 2861 B->ops->multadd = MatMultAdd_SeqBAIJ_6; 2862 break; 2863 case 7: 2864 B->ops->mult = MatMult_SeqBAIJ_7; 2865 B->ops->multadd = MatMultAdd_SeqBAIJ_7; 2866 break; 2867 case 15: 2868 B->ops->mult = MatMult_SeqBAIJ_15_ver1; 2869 B->ops->multadd = MatMultAdd_SeqBAIJ_N; 2870 break; 2871 default: 2872 B->ops->mult = MatMult_SeqBAIJ_N; 2873 B->ops->multadd = MatMultAdd_SeqBAIJ_N; 2874 break; 2875 } 2876 } 2877 B->ops->sor = MatSOR_SeqBAIJ; 2878 b->mbs = mbs; 2879 b->nbs = nbs; 2880 if (!skipallocation) { 2881 if (!b->imax) { 2882 ierr = PetscMalloc2(mbs,&b->imax,mbs,&b->ilen);CHKERRQ(ierr); 2883 ierr = PetscLogObjectMemory((PetscObject)B,2*mbs*sizeof(PetscInt));CHKERRQ(ierr); 2884 2885 b->free_imax_ilen = PETSC_TRUE; 2886 } 2887 /* b->ilen will count nonzeros in each block row so far. */ 2888 for (i=0; i<mbs; i++) b->ilen[i] = 0; 2889 if (!nnz) { 2890 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 2891 else if (nz < 0) nz = 1; 2892 for (i=0; i<mbs; i++) b->imax[i] = nz; 2893 nz = nz*mbs; 2894 } else { 2895 nz = 0; 2896 for (i=0; i<mbs; i++) {b->imax[i] = nnz[i]; nz += nnz[i];} 2897 } 2898 2899 /* allocate the matrix space */ 2900 ierr = MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);CHKERRQ(ierr); 2901 ierr = PetscMalloc3(bs2*nz,&b->a,nz,&b->j,B->rmap->N+1,&b->i);CHKERRQ(ierr); 2902 ierr = PetscLogObjectMemory((PetscObject)B,(B->rmap->N+1)*sizeof(PetscInt)+nz*(bs2*sizeof(PetscScalar)+sizeof(PetscInt)));CHKERRQ(ierr); 2903 ierr = PetscMemzero(b->a,nz*bs2*sizeof(MatScalar));CHKERRQ(ierr); 2904 ierr = PetscMemzero(b->j,nz*sizeof(PetscInt));CHKERRQ(ierr); 2905 2906 b->singlemalloc = PETSC_TRUE; 2907 b->i[0] = 0; 2908 for (i=1; i<mbs+1; i++) { 2909 b->i[i] = b->i[i-1] + b->imax[i-1]; 2910 } 2911 b->free_a = PETSC_TRUE; 2912 b->free_ij = PETSC_TRUE; 2913 #if defined(PETSC_THREADCOMM_ACTIVE) 2914 ierr = MatZeroEntries_SeqBAIJ(B);CHKERRQ(ierr); 2915 #endif 2916 } else { 2917 b->free_a = PETSC_FALSE; 2918 b->free_ij = PETSC_FALSE; 2919 } 2920 2921 b->bs2 = bs2; 2922 b->mbs = mbs; 2923 b->nz = 0; 2924 b->maxnz = nz; 2925 B->info.nz_unneeded = (PetscReal)b->maxnz*bs2; 2926 if (realalloc) {ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr);} 2927 PetscFunctionReturn(0); 2928 } 2929 2930 #undef __FUNCT__ 2931 #define __FUNCT__ "MatSeqBAIJSetPreallocationCSR_SeqBAIJ" 2932 PetscErrorCode MatSeqBAIJSetPreallocationCSR_SeqBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[]) 2933 { 2934 PetscInt i,m,nz,nz_max=0,*nnz; 2935 PetscScalar *values=0; 2936 PetscBool roworiented = ((Mat_SeqBAIJ*)B->data)->roworiented; 2937 PetscErrorCode ierr; 2938 2939 PetscFunctionBegin; 2940 if (bs < 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive but it is %D",bs); 2941 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 2942 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 2943 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2944 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2945 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2946 m = B->rmap->n/bs; 2947 2948 if (ii[0] != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "ii[0] must be 0 but it is %D",ii[0]); 2949 ierr = PetscMalloc1((m+1), &nnz);CHKERRQ(ierr); 2950 for (i=0; i<m; i++) { 2951 nz = ii[i+1]- ii[i]; 2952 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D",i,nz); 2953 nz_max = PetscMax(nz_max, nz); 2954 nnz[i] = nz; 2955 } 2956 ierr = MatSeqBAIJSetPreallocation(B,bs,0,nnz);CHKERRQ(ierr); 2957 ierr = PetscFree(nnz);CHKERRQ(ierr); 2958 2959 values = (PetscScalar*)V; 2960 if (!values) { 2961 ierr = PetscCalloc1(bs*bs*(nz_max+1),&values);CHKERRQ(ierr); 2962 } 2963 for (i=0; i<m; i++) { 2964 PetscInt ncols = ii[i+1] - ii[i]; 2965 const PetscInt *icols = jj + ii[i]; 2966 const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0); 2967 if (!roworiented) { 2968 ierr = MatSetValuesBlocked_SeqBAIJ(B,1,&i,ncols,icols,svals,INSERT_VALUES);CHKERRQ(ierr); 2969 } else { 2970 PetscInt j; 2971 for (j=0; j<ncols; j++) { 2972 const PetscScalar *svals = values + (V ? (bs*bs*(ii[i]+j)) : 0); 2973 ierr = MatSetValuesBlocked_SeqBAIJ(B,1,&i,1,&icols[j],svals,INSERT_VALUES);CHKERRQ(ierr); 2974 } 2975 } 2976 } 2977 if (!V) { ierr = PetscFree(values);CHKERRQ(ierr); } 2978 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2979 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2980 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2981 PetscFunctionReturn(0); 2982 } 2983 2984 PETSC_EXTERN PetscErrorCode MatGetFactor_seqbaij_petsc(Mat,MatFactorType,Mat*); 2985 PETSC_EXTERN PetscErrorCode MatGetFactor_seqbaij_bstrm(Mat,MatFactorType,Mat*); 2986 #if defined(PETSC_HAVE_MUMPS) 2987 PETSC_EXTERN PetscErrorCode MatGetFactor_baij_mumps(Mat,MatFactorType,Mat*); 2988 #endif 2989 extern PetscErrorCode MatGetFactorAvailable_seqbaij_petsc(Mat,MatFactorType,PetscBool*); 2990 2991 /*MC 2992 MATSEQBAIJ - MATSEQBAIJ = "seqbaij" - A matrix type to be used for sequential block sparse matrices, based on 2993 block sparse compressed row format. 2994 2995 Options Database Keys: 2996 . -mat_type seqbaij - sets the matrix type to "seqbaij" during a call to MatSetFromOptions() 2997 2998 Level: beginner 2999 3000 .seealso: MatCreateSeqBAIJ() 3001 M*/ 3002 3003 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBSTRM(Mat, MatType,MatReuse,Mat*); 3004 3005 #undef __FUNCT__ 3006 #define __FUNCT__ "MatCreate_SeqBAIJ" 3007 PETSC_EXTERN PetscErrorCode MatCreate_SeqBAIJ(Mat B) 3008 { 3009 PetscErrorCode ierr; 3010 PetscMPIInt size; 3011 Mat_SeqBAIJ *b; 3012 3013 PetscFunctionBegin; 3014 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 3015 if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Comm must be of size 1"); 3016 3017 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 3018 B->data = (void*)b; 3019 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 3020 3021 b->row = 0; 3022 b->col = 0; 3023 b->icol = 0; 3024 b->reallocs = 0; 3025 b->saved_values = 0; 3026 3027 b->roworiented = PETSC_TRUE; 3028 b->nonew = 0; 3029 b->diag = 0; 3030 b->solve_work = 0; 3031 b->mult_work = 0; 3032 B->spptr = 0; 3033 B->info.nz_unneeded = (PetscReal)b->maxnz*b->bs2; 3034 b->keepnonzeropattern = PETSC_FALSE; 3035 b->xtoy = 0; 3036 b->XtoY = 0; 3037 3038 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactorAvailable_petsc_C",MatGetFactorAvailable_seqbaij_petsc);CHKERRQ(ierr); 3039 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_petsc_C",MatGetFactor_seqbaij_petsc);CHKERRQ(ierr); 3040 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_bstrm_C",MatGetFactor_seqbaij_bstrm);CHKERRQ(ierr); 3041 #if defined(PETSC_HAVE_MUMPS) 3042 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C", MatGetFactor_baij_mumps);CHKERRQ(ierr); 3043 #endif 3044 ierr = PetscObjectComposeFunction((PetscObject)B,"MatInvertBlockDiagonal_C",MatInvertBlockDiagonal_SeqBAIJ);CHKERRQ(ierr); 3045 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqBAIJ);CHKERRQ(ierr); 3046 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqBAIJ);CHKERRQ(ierr); 3047 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetColumnIndices_C",MatSeqBAIJSetColumnIndices_SeqBAIJ);CHKERRQ(ierr); 3048 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqaij_C",MatConvert_SeqBAIJ_SeqAIJ);CHKERRQ(ierr); 3049 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqsbaij_C",MatConvert_SeqBAIJ_SeqSBAIJ);CHKERRQ(ierr); 3050 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetPreallocation_C",MatSeqBAIJSetPreallocation_SeqBAIJ);CHKERRQ(ierr); 3051 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetPreallocationCSR_C",MatSeqBAIJSetPreallocationCSR_SeqBAIJ);CHKERRQ(ierr); 3052 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqbstrm_C",MatConvert_SeqBAIJ_SeqBSTRM);CHKERRQ(ierr); 3053 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqBAIJ);CHKERRQ(ierr); 3054 ierr = PetscObjectChangeTypeName((PetscObject)B,MATSEQBAIJ);CHKERRQ(ierr); 3055 PetscFunctionReturn(0); 3056 } 3057 3058 #undef __FUNCT__ 3059 #define __FUNCT__ "MatDuplicateNoCreate_SeqBAIJ" 3060 PetscErrorCode MatDuplicateNoCreate_SeqBAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace) 3061 { 3062 Mat_SeqBAIJ *c = (Mat_SeqBAIJ*)C->data,*a = (Mat_SeqBAIJ*)A->data; 3063 PetscErrorCode ierr; 3064 PetscInt i,mbs = a->mbs,nz = a->nz,bs2 = a->bs2; 3065 3066 PetscFunctionBegin; 3067 if (a->i[mbs] != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupt matrix"); 3068 3069 if (cpvalues == MAT_SHARE_NONZERO_PATTERN) { 3070 c->imax = a->imax; 3071 c->ilen = a->ilen; 3072 c->free_imax_ilen = PETSC_FALSE; 3073 } else { 3074 ierr = PetscMalloc2(mbs,&c->imax,mbs,&c->ilen);CHKERRQ(ierr); 3075 ierr = PetscLogObjectMemory((PetscObject)C,2*mbs*sizeof(PetscInt));CHKERRQ(ierr); 3076 for (i=0; i<mbs; i++) { 3077 c->imax[i] = a->imax[i]; 3078 c->ilen[i] = a->ilen[i]; 3079 } 3080 c->free_imax_ilen = PETSC_TRUE; 3081 } 3082 3083 /* allocate the matrix space */ 3084 if (mallocmatspace) { 3085 if (cpvalues == MAT_SHARE_NONZERO_PATTERN) { 3086 ierr = PetscCalloc1(bs2*nz,&c->a);CHKERRQ(ierr); 3087 ierr = PetscLogObjectMemory((PetscObject)C,a->i[mbs]*bs2*sizeof(PetscScalar));CHKERRQ(ierr); 3088 3089 c->i = a->i; 3090 c->j = a->j; 3091 c->singlemalloc = PETSC_FALSE; 3092 c->free_a = PETSC_TRUE; 3093 c->free_ij = PETSC_FALSE; 3094 c->parent = A; 3095 C->preallocated = PETSC_TRUE; 3096 C->assembled = PETSC_TRUE; 3097 3098 ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); 3099 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3100 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3101 } else { 3102 ierr = PetscMalloc3(bs2*nz,&c->a,nz,&c->j,mbs+1,&c->i);CHKERRQ(ierr); 3103 ierr = PetscLogObjectMemory((PetscObject)C,a->i[mbs]*(bs2*sizeof(PetscScalar)+sizeof(PetscInt))+(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 3104 3105 c->singlemalloc = PETSC_TRUE; 3106 c->free_a = PETSC_TRUE; 3107 c->free_ij = PETSC_TRUE; 3108 3109 ierr = PetscMemcpy(c->i,a->i,(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 3110 if (mbs > 0) { 3111 ierr = PetscMemcpy(c->j,a->j,nz*sizeof(PetscInt));CHKERRQ(ierr); 3112 if (cpvalues == MAT_COPY_VALUES) { 3113 ierr = PetscMemcpy(c->a,a->a,bs2*nz*sizeof(MatScalar));CHKERRQ(ierr); 3114 } else { 3115 ierr = PetscMemzero(c->a,bs2*nz*sizeof(MatScalar));CHKERRQ(ierr); 3116 } 3117 } 3118 C->preallocated = PETSC_TRUE; 3119 C->assembled = PETSC_TRUE; 3120 } 3121 } 3122 3123 c->roworiented = a->roworiented; 3124 c->nonew = a->nonew; 3125 3126 ierr = PetscLayoutReference(A->rmap,&C->rmap);CHKERRQ(ierr); 3127 ierr = PetscLayoutReference(A->cmap,&C->cmap);CHKERRQ(ierr); 3128 3129 c->bs2 = a->bs2; 3130 c->mbs = a->mbs; 3131 c->nbs = a->nbs; 3132 3133 if (a->diag) { 3134 if (cpvalues == MAT_SHARE_NONZERO_PATTERN) { 3135 c->diag = a->diag; 3136 c->free_diag = PETSC_FALSE; 3137 } else { 3138 ierr = PetscMalloc1((mbs+1),&c->diag);CHKERRQ(ierr); 3139 ierr = PetscLogObjectMemory((PetscObject)C,(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 3140 for (i=0; i<mbs; i++) c->diag[i] = a->diag[i]; 3141 c->free_diag = PETSC_TRUE; 3142 } 3143 } else c->diag = 0; 3144 3145 c->nz = a->nz; 3146 c->maxnz = a->nz; /* Since we allocate exactly the right amount */ 3147 c->solve_work = 0; 3148 c->mult_work = 0; 3149 3150 c->compressedrow.use = a->compressedrow.use; 3151 c->compressedrow.nrows = a->compressedrow.nrows; 3152 if (a->compressedrow.use) { 3153 i = a->compressedrow.nrows; 3154 ierr = PetscMalloc2(i+1,&c->compressedrow.i,i+1,&c->compressedrow.rindex);CHKERRQ(ierr); 3155 ierr = PetscLogObjectMemory((PetscObject)C,(2*i+1)*sizeof(PetscInt));CHKERRQ(ierr); 3156 ierr = PetscMemcpy(c->compressedrow.i,a->compressedrow.i,(i+1)*sizeof(PetscInt));CHKERRQ(ierr); 3157 ierr = PetscMemcpy(c->compressedrow.rindex,a->compressedrow.rindex,i*sizeof(PetscInt));CHKERRQ(ierr); 3158 } else { 3159 c->compressedrow.use = PETSC_FALSE; 3160 c->compressedrow.i = NULL; 3161 c->compressedrow.rindex = NULL; 3162 } 3163 C->nonzerostate = A->nonzerostate; 3164 3165 ierr = PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);CHKERRQ(ierr); 3166 ierr = PetscMemcpy(C->ops,A->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 3167 PetscFunctionReturn(0); 3168 } 3169 3170 #undef __FUNCT__ 3171 #define __FUNCT__ "MatDuplicate_SeqBAIJ" 3172 PetscErrorCode MatDuplicate_SeqBAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B) 3173 { 3174 PetscErrorCode ierr; 3175 3176 PetscFunctionBegin; 3177 ierr = MatCreate(PetscObjectComm((PetscObject)A),B);CHKERRQ(ierr); 3178 ierr = MatSetSizes(*B,A->rmap->N,A->cmap->n,A->rmap->N,A->cmap->n);CHKERRQ(ierr); 3179 ierr = MatSetType(*B,MATSEQBAIJ);CHKERRQ(ierr); 3180 ierr = MatDuplicateNoCreate_SeqBAIJ(*B,A,cpvalues,PETSC_TRUE);CHKERRQ(ierr); 3181 PetscFunctionReturn(0); 3182 } 3183 3184 #undef __FUNCT__ 3185 #define __FUNCT__ "MatLoad_SeqBAIJ" 3186 PetscErrorCode MatLoad_SeqBAIJ(Mat newmat,PetscViewer viewer) 3187 { 3188 Mat_SeqBAIJ *a; 3189 PetscErrorCode ierr; 3190 PetscInt i,nz,header[4],*rowlengths=0,M,N,bs=1; 3191 PetscInt *mask,mbs,*jj,j,rowcount,nzcount,k,*browlengths,maskcount; 3192 PetscInt kmax,jcount,block,idx,point,nzcountb,extra_rows,rows,cols; 3193 PetscInt *masked,nmask,tmp,bs2,ishift; 3194 PetscMPIInt size; 3195 int fd; 3196 PetscScalar *aa; 3197 MPI_Comm comm; 3198 3199 PetscFunctionBegin; 3200 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 3201 ierr = PetscOptionsBegin(comm,NULL,"Options for loading SEQBAIJ matrix","Mat");CHKERRQ(ierr); 3202 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 3203 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3204 bs2 = bs*bs; 3205 3206 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3207 if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"view must have one processor"); 3208 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 3209 ierr = PetscBinaryRead(fd,header,4,PETSC_INT);CHKERRQ(ierr); 3210 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not Mat object"); 3211 M = header[1]; N = header[2]; nz = header[3]; 3212 3213 if (header[3] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as SeqBAIJ"); 3214 if (M != N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Can only do square matrices"); 3215 3216 /* 3217 This code adds extra rows to make sure the number of rows is 3218 divisible by the blocksize 3219 */ 3220 mbs = M/bs; 3221 extra_rows = bs - M + bs*(mbs); 3222 if (extra_rows == bs) extra_rows = 0; 3223 else mbs++; 3224 if (extra_rows) { 3225 ierr = PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");CHKERRQ(ierr); 3226 } 3227 3228 /* Set global sizes if not already set */ 3229 if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) { 3230 ierr = MatSetSizes(newmat,PETSC_DECIDE,PETSC_DECIDE,M+extra_rows,N+extra_rows);CHKERRQ(ierr); 3231 } else { /* Check if the matrix global sizes are correct */ 3232 ierr = MatGetSize(newmat,&rows,&cols);CHKERRQ(ierr); 3233 if (rows < 0 && cols < 0) { /* user might provide local size instead of global size */ 3234 ierr = MatGetLocalSize(newmat,&rows,&cols);CHKERRQ(ierr); 3235 } 3236 if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix in file of different length (%d, %d) than the input matrix (%d, %d)",M,N,rows,cols); 3237 } 3238 3239 /* read in row lengths */ 3240 ierr = PetscMalloc1((M+extra_rows),&rowlengths);CHKERRQ(ierr); 3241 ierr = PetscBinaryRead(fd,rowlengths,M,PETSC_INT);CHKERRQ(ierr); 3242 for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1; 3243 3244 /* read in column indices */ 3245 ierr = PetscMalloc1((nz+extra_rows),&jj);CHKERRQ(ierr); 3246 ierr = PetscBinaryRead(fd,jj,nz,PETSC_INT);CHKERRQ(ierr); 3247 for (i=0; i<extra_rows; i++) jj[nz+i] = M+i; 3248 3249 /* loop over row lengths determining block row lengths */ 3250 ierr = PetscCalloc1(mbs,&browlengths);CHKERRQ(ierr); 3251 ierr = PetscMalloc2(mbs,&mask,mbs,&masked);CHKERRQ(ierr); 3252 ierr = PetscMemzero(mask,mbs*sizeof(PetscInt));CHKERRQ(ierr); 3253 rowcount = 0; 3254 nzcount = 0; 3255 for (i=0; i<mbs; i++) { 3256 nmask = 0; 3257 for (j=0; j<bs; j++) { 3258 kmax = rowlengths[rowcount]; 3259 for (k=0; k<kmax; k++) { 3260 tmp = jj[nzcount++]/bs; 3261 if (!mask[tmp]) {masked[nmask++] = tmp; mask[tmp] = 1;} 3262 } 3263 rowcount++; 3264 } 3265 browlengths[i] += nmask; 3266 /* zero out the mask elements we set */ 3267 for (j=0; j<nmask; j++) mask[masked[j]] = 0; 3268 } 3269 3270 /* Do preallocation */ 3271 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(newmat,bs,0,browlengths);CHKERRQ(ierr); 3272 a = (Mat_SeqBAIJ*)newmat->data; 3273 3274 /* set matrix "i" values */ 3275 a->i[0] = 0; 3276 for (i=1; i<= mbs; i++) { 3277 a->i[i] = a->i[i-1] + browlengths[i-1]; 3278 a->ilen[i-1] = browlengths[i-1]; 3279 } 3280 a->nz = 0; 3281 for (i=0; i<mbs; i++) a->nz += browlengths[i]; 3282 3283 /* read in nonzero values */ 3284 ierr = PetscMalloc1((nz+extra_rows),&aa);CHKERRQ(ierr); 3285 ierr = PetscBinaryRead(fd,aa,nz,PETSC_SCALAR);CHKERRQ(ierr); 3286 for (i=0; i<extra_rows; i++) aa[nz+i] = 1.0; 3287 3288 /* set "a" and "j" values into matrix */ 3289 nzcount = 0; jcount = 0; 3290 for (i=0; i<mbs; i++) { 3291 nzcountb = nzcount; 3292 nmask = 0; 3293 for (j=0; j<bs; j++) { 3294 kmax = rowlengths[i*bs+j]; 3295 for (k=0; k<kmax; k++) { 3296 tmp = jj[nzcount++]/bs; 3297 if (!mask[tmp]) { masked[nmask++] = tmp; mask[tmp] = 1;} 3298 } 3299 } 3300 /* sort the masked values */ 3301 ierr = PetscSortInt(nmask,masked);CHKERRQ(ierr); 3302 3303 /* set "j" values into matrix */ 3304 maskcount = 1; 3305 for (j=0; j<nmask; j++) { 3306 a->j[jcount++] = masked[j]; 3307 mask[masked[j]] = maskcount++; 3308 } 3309 /* set "a" values into matrix */ 3310 ishift = bs2*a->i[i]; 3311 for (j=0; j<bs; j++) { 3312 kmax = rowlengths[i*bs+j]; 3313 for (k=0; k<kmax; k++) { 3314 tmp = jj[nzcountb]/bs; 3315 block = mask[tmp] - 1; 3316 point = jj[nzcountb] - bs*tmp; 3317 idx = ishift + bs2*block + j + bs*point; 3318 a->a[idx] = (MatScalar)aa[nzcountb++]; 3319 } 3320 } 3321 /* zero out the mask elements we set */ 3322 for (j=0; j<nmask; j++) mask[masked[j]] = 0; 3323 } 3324 if (jcount != a->nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Bad binary matrix"); 3325 3326 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3327 ierr = PetscFree(browlengths);CHKERRQ(ierr); 3328 ierr = PetscFree(aa);CHKERRQ(ierr); 3329 ierr = PetscFree(jj);CHKERRQ(ierr); 3330 ierr = PetscFree2(mask,masked);CHKERRQ(ierr); 3331 3332 ierr = MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3333 ierr = MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3334 PetscFunctionReturn(0); 3335 } 3336 3337 #undef __FUNCT__ 3338 #define __FUNCT__ "MatCreateSeqBAIJ" 3339 /*@C 3340 MatCreateSeqBAIJ - Creates a sparse matrix in block AIJ (block 3341 compressed row) format. For good matrix assembly performance the 3342 user should preallocate the matrix storage by setting the parameter nz 3343 (or the array nnz). By setting these parameters accurately, performance 3344 during matrix assembly can be increased by more than a factor of 50. 3345 3346 Collective on MPI_Comm 3347 3348 Input Parameters: 3349 + comm - MPI communicator, set to PETSC_COMM_SELF 3350 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 3351 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 3352 . m - number of rows 3353 . n - number of columns 3354 . nz - number of nonzero blocks per block row (same for all rows) 3355 - nnz - array containing the number of nonzero blocks in the various block rows 3356 (possibly different for each block row) or NULL 3357 3358 Output Parameter: 3359 . A - the matrix 3360 3361 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3362 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3363 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3364 3365 Options Database Keys: 3366 . -mat_no_unroll - uses code that does not unroll the loops in the 3367 block calculations (much slower) 3368 . -mat_block_size - size of the blocks to use 3369 3370 Level: intermediate 3371 3372 Notes: 3373 The number of rows and columns must be divisible by blocksize. 3374 3375 If the nnz parameter is given then the nz parameter is ignored 3376 3377 A nonzero block is any block that as 1 or more nonzeros in it 3378 3379 The block AIJ format is fully compatible with standard Fortran 77 3380 storage. That is, the stored row and column indices can begin at 3381 either one (as in Fortran) or zero. See the users' manual for details. 3382 3383 Specify the preallocated storage with either nz or nnz (not both). 3384 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3385 allocation. See Users-Manual: ch_mat for details. 3386 matrices. 3387 3388 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateBAIJ() 3389 @*/ 3390 PetscErrorCode MatCreateSeqBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A) 3391 { 3392 PetscErrorCode ierr; 3393 3394 PetscFunctionBegin; 3395 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3396 ierr = MatSetSizes(*A,m,n,m,n);CHKERRQ(ierr); 3397 ierr = MatSetType(*A,MATSEQBAIJ);CHKERRQ(ierr); 3398 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(*A,bs,nz,(PetscInt*)nnz);CHKERRQ(ierr); 3399 PetscFunctionReturn(0); 3400 } 3401 3402 #undef __FUNCT__ 3403 #define __FUNCT__ "MatSeqBAIJSetPreallocation" 3404 /*@C 3405 MatSeqBAIJSetPreallocation - Sets the block size and expected nonzeros 3406 per row in the matrix. For good matrix assembly performance the 3407 user should preallocate the matrix storage by setting the parameter nz 3408 (or the array nnz). By setting these parameters accurately, performance 3409 during matrix assembly can be increased by more than a factor of 50. 3410 3411 Collective on MPI_Comm 3412 3413 Input Parameters: 3414 + B - the matrix 3415 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 3416 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 3417 . nz - number of block nonzeros per block row (same for all rows) 3418 - nnz - array containing the number of block nonzeros in the various block rows 3419 (possibly different for each block row) or NULL 3420 3421 Options Database Keys: 3422 . -mat_no_unroll - uses code that does not unroll the loops in the 3423 block calculations (much slower) 3424 . -mat_block_size - size of the blocks to use 3425 3426 Level: intermediate 3427 3428 Notes: 3429 If the nnz parameter is given then the nz parameter is ignored 3430 3431 You can call MatGetInfo() to get information on how effective the preallocation was; 3432 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3433 You can also run with the option -info and look for messages with the string 3434 malloc in them to see if additional memory allocation was needed. 3435 3436 The block AIJ format is fully compatible with standard Fortran 77 3437 storage. That is, the stored row and column indices can begin at 3438 either one (as in Fortran) or zero. See the users' manual for details. 3439 3440 Specify the preallocated storage with either nz or nnz (not both). 3441 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3442 allocation. See Users-Manual: ch_mat for details. 3443 3444 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateBAIJ(), MatGetInfo() 3445 @*/ 3446 PetscErrorCode MatSeqBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt nz,const PetscInt nnz[]) 3447 { 3448 PetscErrorCode ierr; 3449 3450 PetscFunctionBegin; 3451 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3452 PetscValidType(B,1); 3453 PetscValidLogicalCollectiveInt(B,bs,2); 3454 ierr = PetscTryMethod(B,"MatSeqBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[]),(B,bs,nz,nnz));CHKERRQ(ierr); 3455 PetscFunctionReturn(0); 3456 } 3457 3458 #undef __FUNCT__ 3459 #define __FUNCT__ "MatSeqBAIJSetPreallocationCSR" 3460 /*@C 3461 MatSeqBAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format 3462 (the default sequential PETSc format). 3463 3464 Collective on MPI_Comm 3465 3466 Input Parameters: 3467 + B - the matrix 3468 . i - the indices into j for the start of each local row (starts with zero) 3469 . j - the column indices for each local row (starts with zero) these must be sorted for each row 3470 - v - optional values in the matrix 3471 3472 Level: developer 3473 3474 Notes: 3475 The order of the entries in values is specified by the MatOption MAT_ROW_ORIENTED. For example, C programs 3476 may want to use the default MAT_ROW_ORIENTED=PETSC_TRUE and use an array v[nnz][bs][bs] where the second index is 3477 over rows within a block and the last index is over columns within a block row. Fortran programs will likely set 3478 MAT_ROW_ORIENTED=PETSC_FALSE and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a 3479 block column and the second index is over columns within a block. 3480 3481 .keywords: matrix, aij, compressed row, sparse 3482 3483 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatSeqBAIJSetPreallocation(), MATSEQBAIJ 3484 @*/ 3485 PetscErrorCode MatSeqBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3486 { 3487 PetscErrorCode ierr; 3488 3489 PetscFunctionBegin; 3490 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3491 PetscValidType(B,1); 3492 PetscValidLogicalCollectiveInt(B,bs,2); 3493 ierr = PetscTryMethod(B,"MatSeqBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));CHKERRQ(ierr); 3494 PetscFunctionReturn(0); 3495 } 3496 3497 3498 #undef __FUNCT__ 3499 #define __FUNCT__ "MatCreateSeqBAIJWithArrays" 3500 /*@ 3501 MatCreateSeqBAIJWithArrays - Creates an sequential BAIJ matrix using matrix elements provided by the user. 3502 3503 Collective on MPI_Comm 3504 3505 Input Parameters: 3506 + comm - must be an MPI communicator of size 1 3507 . bs - size of block 3508 . m - number of rows 3509 . n - number of columns 3510 . i - row indices 3511 . j - column indices 3512 - a - matrix values 3513 3514 Output Parameter: 3515 . mat - the matrix 3516 3517 Level: advanced 3518 3519 Notes: 3520 The i, j, and a arrays are not copied by this routine, the user must free these arrays 3521 once the matrix is destroyed 3522 3523 You cannot set new nonzero locations into this matrix, that will generate an error. 3524 3525 The i and j indices are 0 based 3526 3527 When block size is greater than 1 the matrix values must be stored using the BAIJ storage format (see the BAIJ code to determine this). 3528 3529 The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is 3530 the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first 3531 block, followed by the second column of the first block etc etc. That is, the blocks are contiguous in memory 3532 with column-major ordering within blocks. 3533 3534 .seealso: MatCreate(), MatCreateBAIJ(), MatCreateSeqBAIJ() 3535 3536 @*/ 3537 PetscErrorCode MatCreateSeqBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt *i,PetscInt *j,PetscScalar *a,Mat *mat) 3538 { 3539 PetscErrorCode ierr; 3540 PetscInt ii; 3541 Mat_SeqBAIJ *baij; 3542 3543 PetscFunctionBegin; 3544 if (bs != 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"block size %D > 1 is not supported yet",bs); 3545 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3546 3547 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3548 ierr = MatSetSizes(*mat,m,n,m,n);CHKERRQ(ierr); 3549 ierr = MatSetType(*mat,MATSEQBAIJ);CHKERRQ(ierr); 3550 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(*mat,bs,MAT_SKIP_ALLOCATION,0);CHKERRQ(ierr); 3551 baij = (Mat_SeqBAIJ*)(*mat)->data; 3552 ierr = PetscMalloc2(m,&baij->imax,m,&baij->ilen);CHKERRQ(ierr); 3553 ierr = PetscLogObjectMemory((PetscObject)*mat,2*m*sizeof(PetscInt));CHKERRQ(ierr); 3554 3555 baij->i = i; 3556 baij->j = j; 3557 baij->a = a; 3558 3559 baij->singlemalloc = PETSC_FALSE; 3560 baij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/ 3561 baij->free_a = PETSC_FALSE; 3562 baij->free_ij = PETSC_FALSE; 3563 3564 for (ii=0; ii<m; ii++) { 3565 baij->ilen[ii] = baij->imax[ii] = i[ii+1] - i[ii]; 3566 #if defined(PETSC_USE_DEBUG) 3567 if (i[ii+1] - i[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row length in i (row indices) row = %d length = %d",ii,i[ii+1] - i[ii]); 3568 #endif 3569 } 3570 #if defined(PETSC_USE_DEBUG) 3571 for (ii=0; ii<baij->i[m]; ii++) { 3572 if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %d index = %d",ii,j[ii]); 3573 if (j[ii] > n - 1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column index to large at location = %d index = %d",ii,j[ii]); 3574 } 3575 #endif 3576 3577 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3578 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3579 PetscFunctionReturn(0); 3580 } 3581