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 = PetscFree2(a->compressedrow.i,a->compressedrow.rindex);CHKERRQ(ierr); 1211 1212 ierr = MatDestroy(&a->sbaijMat);CHKERRQ(ierr); 1213 ierr = MatDestroy(&a->parent);CHKERRQ(ierr); 1214 ierr = PetscFree(A->data);CHKERRQ(ierr); 1215 1216 ierr = PetscObjectChangeTypeName((PetscObject)A,0);CHKERRQ(ierr); 1217 ierr = PetscObjectComposeFunction((PetscObject)A,"MatInvertBlockDiagonal_C",NULL);CHKERRQ(ierr); 1218 ierr = PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1219 ierr = PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1220 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetColumnIndices_C",NULL);CHKERRQ(ierr); 1221 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqaij_C",NULL);CHKERRQ(ierr); 1222 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqsbaij_C",NULL);CHKERRQ(ierr); 1223 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1224 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1225 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqbstrm_C",NULL);CHKERRQ(ierr); 1226 ierr = PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1227 PetscFunctionReturn(0); 1228 } 1229 1230 #undef __FUNCT__ 1231 #define __FUNCT__ "MatSetOption_SeqBAIJ" 1232 PetscErrorCode MatSetOption_SeqBAIJ(Mat A,MatOption op,PetscBool flg) 1233 { 1234 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1235 PetscErrorCode ierr; 1236 1237 PetscFunctionBegin; 1238 switch (op) { 1239 case MAT_ROW_ORIENTED: 1240 a->roworiented = flg; 1241 break; 1242 case MAT_KEEP_NONZERO_PATTERN: 1243 a->keepnonzeropattern = flg; 1244 break; 1245 case MAT_NEW_NONZERO_LOCATIONS: 1246 a->nonew = (flg ? 0 : 1); 1247 break; 1248 case MAT_NEW_NONZERO_LOCATION_ERR: 1249 a->nonew = (flg ? -1 : 0); 1250 break; 1251 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1252 a->nonew = (flg ? -2 : 0); 1253 break; 1254 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1255 a->nounused = (flg ? -1 : 0); 1256 break; 1257 case MAT_NEW_DIAGONALS: 1258 case MAT_IGNORE_OFF_PROC_ENTRIES: 1259 case MAT_USE_HASH_TABLE: 1260 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1261 break; 1262 case MAT_SPD: 1263 case MAT_SYMMETRIC: 1264 case MAT_STRUCTURALLY_SYMMETRIC: 1265 case MAT_HERMITIAN: 1266 case MAT_SYMMETRY_ETERNAL: 1267 /* These options are handled directly by MatSetOption() */ 1268 break; 1269 default: 1270 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1271 } 1272 PetscFunctionReturn(0); 1273 } 1274 1275 /* used for both SeqBAIJ and SeqSBAIJ matrices */ 1276 #undef __FUNCT__ 1277 #define __FUNCT__ "MatGetRow_SeqBAIJ_private" 1278 PetscErrorCode MatGetRow_SeqBAIJ_private(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v,PetscInt *ai,PetscInt *aj,PetscScalar *aa) 1279 { 1280 PetscErrorCode ierr; 1281 PetscInt itmp,i,j,k,M,bn,bp,*idx_i,bs,bs2; 1282 MatScalar *aa_i; 1283 PetscScalar *v_i; 1284 1285 PetscFunctionBegin; 1286 bs = A->rmap->bs; 1287 bs2 = bs*bs; 1288 if (row < 0 || row >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range", row); 1289 1290 bn = row/bs; /* Block number */ 1291 bp = row % bs; /* Block Position */ 1292 M = ai[bn+1] - ai[bn]; 1293 *nz = bs*M; 1294 1295 if (v) { 1296 *v = 0; 1297 if (*nz) { 1298 ierr = PetscMalloc1((*nz),v);CHKERRQ(ierr); 1299 for (i=0; i<M; i++) { /* for each block in the block row */ 1300 v_i = *v + i*bs; 1301 aa_i = aa + bs2*(ai[bn] + i); 1302 for (j=bp,k=0; j<bs2; j+=bs,k++) v_i[k] = aa_i[j]; 1303 } 1304 } 1305 } 1306 1307 if (idx) { 1308 *idx = 0; 1309 if (*nz) { 1310 ierr = PetscMalloc1((*nz),idx);CHKERRQ(ierr); 1311 for (i=0; i<M; i++) { /* for each block in the block row */ 1312 idx_i = *idx + i*bs; 1313 itmp = bs*aj[ai[bn] + i]; 1314 for (j=0; j<bs; j++) idx_i[j] = itmp++; 1315 } 1316 } 1317 } 1318 PetscFunctionReturn(0); 1319 } 1320 1321 #undef __FUNCT__ 1322 #define __FUNCT__ "MatGetRow_SeqBAIJ" 1323 PetscErrorCode MatGetRow_SeqBAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1324 { 1325 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1326 PetscErrorCode ierr; 1327 1328 PetscFunctionBegin; 1329 ierr = MatGetRow_SeqBAIJ_private(A,row,nz,idx,v,a->i,a->j,a->a);CHKERRQ(ierr); 1330 PetscFunctionReturn(0); 1331 } 1332 1333 #undef __FUNCT__ 1334 #define __FUNCT__ "MatRestoreRow_SeqBAIJ" 1335 PetscErrorCode MatRestoreRow_SeqBAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1336 { 1337 PetscErrorCode ierr; 1338 1339 PetscFunctionBegin; 1340 if (idx) {ierr = PetscFree(*idx);CHKERRQ(ierr);} 1341 if (v) {ierr = PetscFree(*v);CHKERRQ(ierr);} 1342 PetscFunctionReturn(0); 1343 } 1344 1345 extern PetscErrorCode MatSetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode); 1346 1347 #undef __FUNCT__ 1348 #define __FUNCT__ "MatTranspose_SeqBAIJ" 1349 PetscErrorCode MatTranspose_SeqBAIJ(Mat A,MatReuse reuse,Mat *B) 1350 { 1351 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data; 1352 Mat C; 1353 PetscErrorCode ierr; 1354 PetscInt i,j,k,*aj=a->j,*ai=a->i,bs=A->rmap->bs,mbs=a->mbs,nbs=a->nbs,len,*col; 1355 PetscInt *rows,*cols,bs2=a->bs2; 1356 MatScalar *array; 1357 1358 PetscFunctionBegin; 1359 if (reuse == MAT_REUSE_MATRIX && A == *B && mbs != nbs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1360 if (reuse == MAT_INITIAL_MATRIX || A == *B) { 1361 ierr = PetscCalloc1((1+nbs),&col);CHKERRQ(ierr); 1362 1363 for (i=0; i<ai[mbs]; i++) col[aj[i]] += 1; 1364 ierr = MatCreate(PetscObjectComm((PetscObject)A),&C);CHKERRQ(ierr); 1365 ierr = MatSetSizes(C,A->cmap->n,A->rmap->N,A->cmap->n,A->rmap->N);CHKERRQ(ierr); 1366 ierr = MatSetType(C,((PetscObject)A)->type_name);CHKERRQ(ierr); 1367 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(C,bs,0,col);CHKERRQ(ierr); 1368 ierr = PetscFree(col);CHKERRQ(ierr); 1369 } else { 1370 C = *B; 1371 } 1372 1373 array = a->a; 1374 ierr = PetscMalloc2(bs,&rows,bs,&cols);CHKERRQ(ierr); 1375 for (i=0; i<mbs; i++) { 1376 cols[0] = i*bs; 1377 for (k=1; k<bs; k++) cols[k] = cols[k-1] + 1; 1378 len = ai[i+1] - ai[i]; 1379 for (j=0; j<len; j++) { 1380 rows[0] = (*aj++)*bs; 1381 for (k=1; k<bs; k++) rows[k] = rows[k-1] + 1; 1382 ierr = MatSetValues_SeqBAIJ(C,bs,rows,bs,cols,array,INSERT_VALUES);CHKERRQ(ierr); 1383 array += bs2; 1384 } 1385 } 1386 ierr = PetscFree2(rows,cols);CHKERRQ(ierr); 1387 1388 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1389 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1390 1391 if (reuse == MAT_INITIAL_MATRIX || *B != A) { 1392 *B = C; 1393 } else { 1394 ierr = MatHeaderMerge(A,C);CHKERRQ(ierr); 1395 } 1396 PetscFunctionReturn(0); 1397 } 1398 1399 #undef __FUNCT__ 1400 #define __FUNCT__ "MatIsTranspose_SeqBAIJ" 1401 PetscErrorCode MatIsTranspose_SeqBAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f) 1402 { 1403 PetscErrorCode ierr; 1404 Mat Btrans; 1405 1406 PetscFunctionBegin; 1407 *f = PETSC_FALSE; 1408 ierr = MatTranspose_SeqBAIJ(A,MAT_INITIAL_MATRIX,&Btrans);CHKERRQ(ierr); 1409 ierr = MatEqual_SeqBAIJ(B,Btrans,f);CHKERRQ(ierr); 1410 ierr = MatDestroy(&Btrans);CHKERRQ(ierr); 1411 PetscFunctionReturn(0); 1412 } 1413 1414 #undef __FUNCT__ 1415 #define __FUNCT__ "MatView_SeqBAIJ_Binary" 1416 static PetscErrorCode MatView_SeqBAIJ_Binary(Mat A,PetscViewer viewer) 1417 { 1418 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1419 PetscErrorCode ierr; 1420 PetscInt i,*col_lens,bs = A->rmap->bs,count,*jj,j,k,l,bs2=a->bs2; 1421 int fd; 1422 PetscScalar *aa; 1423 FILE *file; 1424 1425 PetscFunctionBegin; 1426 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1427 ierr = PetscMalloc1((4+A->rmap->N),&col_lens);CHKERRQ(ierr); 1428 col_lens[0] = MAT_FILE_CLASSID; 1429 1430 col_lens[1] = A->rmap->N; 1431 col_lens[2] = A->cmap->n; 1432 col_lens[3] = a->nz*bs2; 1433 1434 /* store lengths of each row and write (including header) to file */ 1435 count = 0; 1436 for (i=0; i<a->mbs; i++) { 1437 for (j=0; j<bs; j++) { 1438 col_lens[4+count++] = bs*(a->i[i+1] - a->i[i]); 1439 } 1440 } 1441 ierr = PetscBinaryWrite(fd,col_lens,4+A->rmap->N,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1442 ierr = PetscFree(col_lens);CHKERRQ(ierr); 1443 1444 /* store column indices (zero start index) */ 1445 ierr = PetscMalloc1((a->nz+1)*bs2,&jj);CHKERRQ(ierr); 1446 count = 0; 1447 for (i=0; i<a->mbs; i++) { 1448 for (j=0; j<bs; j++) { 1449 for (k=a->i[i]; k<a->i[i+1]; k++) { 1450 for (l=0; l<bs; l++) { 1451 jj[count++] = bs*a->j[k] + l; 1452 } 1453 } 1454 } 1455 } 1456 ierr = PetscBinaryWrite(fd,jj,bs2*a->nz,PETSC_INT,PETSC_FALSE);CHKERRQ(ierr); 1457 ierr = PetscFree(jj);CHKERRQ(ierr); 1458 1459 /* store nonzero values */ 1460 ierr = PetscMalloc1((a->nz+1)*bs2,&aa);CHKERRQ(ierr); 1461 count = 0; 1462 for (i=0; i<a->mbs; i++) { 1463 for (j=0; j<bs; j++) { 1464 for (k=a->i[i]; k<a->i[i+1]; k++) { 1465 for (l=0; l<bs; l++) { 1466 aa[count++] = a->a[bs2*k + l*bs + j]; 1467 } 1468 } 1469 } 1470 } 1471 ierr = PetscBinaryWrite(fd,aa,bs2*a->nz,PETSC_SCALAR,PETSC_FALSE);CHKERRQ(ierr); 1472 ierr = PetscFree(aa);CHKERRQ(ierr); 1473 1474 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1475 if (file) { 1476 fprintf(file,"-matload_block_size %d\n",(int)A->rmap->bs); 1477 } 1478 PetscFunctionReturn(0); 1479 } 1480 1481 #undef __FUNCT__ 1482 #define __FUNCT__ "MatView_SeqBAIJ_ASCII" 1483 static PetscErrorCode MatView_SeqBAIJ_ASCII(Mat A,PetscViewer viewer) 1484 { 1485 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1486 PetscErrorCode ierr; 1487 PetscInt i,j,bs = A->rmap->bs,k,l,bs2=a->bs2; 1488 PetscViewerFormat format; 1489 1490 PetscFunctionBegin; 1491 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1492 if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1493 ierr = PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);CHKERRQ(ierr); 1494 } else if (format == PETSC_VIEWER_ASCII_MATLAB) { 1495 Mat aij; 1496 ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&aij);CHKERRQ(ierr); 1497 ierr = MatView(aij,viewer);CHKERRQ(ierr); 1498 ierr = MatDestroy(&aij);CHKERRQ(ierr); 1499 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1500 PetscFunctionReturn(0); 1501 } else if (format == PETSC_VIEWER_ASCII_COMMON) { 1502 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 1503 for (i=0; i<a->mbs; i++) { 1504 for (j=0; j<bs; j++) { 1505 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i*bs+j);CHKERRQ(ierr); 1506 for (k=a->i[i]; k<a->i[i+1]; k++) { 1507 for (l=0; l<bs; l++) { 1508 #if defined(PETSC_USE_COMPLEX) 1509 if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) > 0.0 && PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) { 1510 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %gi) ",bs*a->j[k]+l, 1511 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1512 } else if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) < 0.0 && PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) { 1513 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %gi) ",bs*a->j[k]+l, 1514 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),-(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1515 } else if (PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) { 1516 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)PetscRealPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1517 } 1518 #else 1519 if (a->a[bs2*k + l*bs + j] != 0.0) { 1520 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)a->a[bs2*k + l*bs + j]);CHKERRQ(ierr); 1521 } 1522 #endif 1523 } 1524 } 1525 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 1526 } 1527 } 1528 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 1529 } else { 1530 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 1531 for (i=0; i<a->mbs; i++) { 1532 for (j=0; j<bs; j++) { 1533 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i*bs+j);CHKERRQ(ierr); 1534 for (k=a->i[i]; k<a->i[i+1]; k++) { 1535 for (l=0; l<bs; l++) { 1536 #if defined(PETSC_USE_COMPLEX) 1537 if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) > 0.0) { 1538 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i) ",bs*a->j[k]+l, 1539 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1540 } else if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) < 0.0) { 1541 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i) ",bs*a->j[k]+l, 1542 (double)PetscRealPart(a->a[bs2*k + l*bs + j]),-(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1543 } else { 1544 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)PetscRealPart(a->a[bs2*k + l*bs + j]));CHKERRQ(ierr); 1545 } 1546 #else 1547 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)a->a[bs2*k + l*bs + j]);CHKERRQ(ierr); 1548 #endif 1549 } 1550 } 1551 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 1552 } 1553 } 1554 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 1555 } 1556 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1557 PetscFunctionReturn(0); 1558 } 1559 1560 #include <petscdraw.h> 1561 #undef __FUNCT__ 1562 #define __FUNCT__ "MatView_SeqBAIJ_Draw_Zoom" 1563 static PetscErrorCode MatView_SeqBAIJ_Draw_Zoom(PetscDraw draw,void *Aa) 1564 { 1565 Mat A = (Mat) Aa; 1566 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data; 1567 PetscErrorCode ierr; 1568 PetscInt row,i,j,k,l,mbs=a->mbs,color,bs=A->rmap->bs,bs2=a->bs2; 1569 PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r; 1570 MatScalar *aa; 1571 PetscViewer viewer; 1572 PetscViewerFormat format; 1573 1574 PetscFunctionBegin; 1575 ierr = PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);CHKERRQ(ierr); 1576 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1577 1578 ierr = PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);CHKERRQ(ierr); 1579 1580 /* loop over matrix elements drawing boxes */ 1581 1582 if (format != PETSC_VIEWER_DRAW_CONTOUR) { 1583 color = PETSC_DRAW_BLUE; 1584 for (i=0,row=0; i<mbs; i++,row+=bs) { 1585 for (j=a->i[i]; j<a->i[i+1]; j++) { 1586 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1587 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1588 aa = a->a + j*bs2; 1589 for (k=0; k<bs; k++) { 1590 for (l=0; l<bs; l++) { 1591 if (PetscRealPart(*aa++) >= 0.) continue; 1592 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1593 } 1594 } 1595 } 1596 } 1597 color = PETSC_DRAW_CYAN; 1598 for (i=0,row=0; i<mbs; i++,row+=bs) { 1599 for (j=a->i[i]; j<a->i[i+1]; j++) { 1600 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1601 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1602 aa = a->a + j*bs2; 1603 for (k=0; k<bs; k++) { 1604 for (l=0; l<bs; l++) { 1605 if (PetscRealPart(*aa++) != 0.) continue; 1606 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1607 } 1608 } 1609 } 1610 } 1611 color = PETSC_DRAW_RED; 1612 for (i=0,row=0; i<mbs; i++,row+=bs) { 1613 for (j=a->i[i]; j<a->i[i+1]; j++) { 1614 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1615 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1616 aa = a->a + j*bs2; 1617 for (k=0; k<bs; k++) { 1618 for (l=0; l<bs; l++) { 1619 if (PetscRealPart(*aa++) <= 0.) continue; 1620 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1621 } 1622 } 1623 } 1624 } 1625 } else { 1626 /* use contour shading to indicate magnitude of values */ 1627 /* first determine max of all nonzero values */ 1628 PetscDraw popup; 1629 PetscReal scale,maxv = 0.0; 1630 1631 for (i=0; i<a->nz*a->bs2; i++) { 1632 if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]); 1633 } 1634 scale = (245.0 - PETSC_DRAW_BASIC_COLORS)/maxv; 1635 ierr = PetscDrawGetPopup(draw,&popup);CHKERRQ(ierr); 1636 if (popup) { 1637 ierr = PetscDrawScalePopup(popup,0.0,maxv);CHKERRQ(ierr); 1638 } 1639 for (i=0,row=0; i<mbs; i++,row+=bs) { 1640 for (j=a->i[i]; j<a->i[i+1]; j++) { 1641 y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0; 1642 x_l = a->j[j]*bs; x_r = x_l + 1.0; 1643 aa = a->a + j*bs2; 1644 for (k=0; k<bs; k++) { 1645 for (l=0; l<bs; l++) { 1646 color = PETSC_DRAW_BASIC_COLORS + (PetscInt)(scale*PetscAbsScalar(*aa++)); 1647 ierr = PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);CHKERRQ(ierr); 1648 } 1649 } 1650 } 1651 } 1652 } 1653 PetscFunctionReturn(0); 1654 } 1655 1656 #undef __FUNCT__ 1657 #define __FUNCT__ "MatView_SeqBAIJ_Draw" 1658 static PetscErrorCode MatView_SeqBAIJ_Draw(Mat A,PetscViewer viewer) 1659 { 1660 PetscErrorCode ierr; 1661 PetscReal xl,yl,xr,yr,w,h; 1662 PetscDraw draw; 1663 PetscBool isnull; 1664 1665 PetscFunctionBegin; 1666 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1667 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1668 1669 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);CHKERRQ(ierr); 1670 xr = A->cmap->n; yr = A->rmap->N; h = yr/10.0; w = xr/10.0; 1671 xr += w; yr += h; xl = -w; yl = -h; 1672 ierr = PetscDrawSetCoordinates(draw,xl,yl,xr,yr);CHKERRQ(ierr); 1673 ierr = PetscDrawZoom(draw,MatView_SeqBAIJ_Draw_Zoom,A);CHKERRQ(ierr); 1674 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);CHKERRQ(ierr); 1675 PetscFunctionReturn(0); 1676 } 1677 1678 #undef __FUNCT__ 1679 #define __FUNCT__ "MatView_SeqBAIJ" 1680 PetscErrorCode MatView_SeqBAIJ(Mat A,PetscViewer viewer) 1681 { 1682 PetscErrorCode ierr; 1683 PetscBool iascii,isbinary,isdraw; 1684 1685 PetscFunctionBegin; 1686 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1687 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1688 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1689 if (iascii) { 1690 ierr = MatView_SeqBAIJ_ASCII(A,viewer);CHKERRQ(ierr); 1691 } else if (isbinary) { 1692 ierr = MatView_SeqBAIJ_Binary(A,viewer);CHKERRQ(ierr); 1693 } else if (isdraw) { 1694 ierr = MatView_SeqBAIJ_Draw(A,viewer);CHKERRQ(ierr); 1695 } else { 1696 Mat B; 1697 ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 1698 ierr = MatView(B,viewer);CHKERRQ(ierr); 1699 ierr = MatDestroy(&B);CHKERRQ(ierr); 1700 } 1701 PetscFunctionReturn(0); 1702 } 1703 1704 1705 #undef __FUNCT__ 1706 #define __FUNCT__ "MatGetValues_SeqBAIJ" 1707 PetscErrorCode MatGetValues_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[]) 1708 { 1709 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1710 PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j; 1711 PetscInt *ai = a->i,*ailen = a->ilen; 1712 PetscInt brow,bcol,ridx,cidx,bs=A->rmap->bs,bs2=a->bs2; 1713 MatScalar *ap,*aa = a->a; 1714 1715 PetscFunctionBegin; 1716 for (k=0; k<m; k++) { /* loop over rows */ 1717 row = im[k]; brow = row/bs; 1718 if (row < 0) {v += n; continue;} /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); */ 1719 if (row >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D too large", row); 1720 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; 1721 nrow = ailen[brow]; 1722 for (l=0; l<n; l++) { /* loop over columns */ 1723 if (in[l] < 0) {v++; continue;} /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column"); */ 1724 if (in[l] >= A->cmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column %D too large", in[l]); 1725 col = in[l]; 1726 bcol = col/bs; 1727 cidx = col%bs; 1728 ridx = row%bs; 1729 high = nrow; 1730 low = 0; /* assume unsorted */ 1731 while (high-low > 5) { 1732 t = (low+high)/2; 1733 if (rp[t] > bcol) high = t; 1734 else low = t; 1735 } 1736 for (i=low; i<high; i++) { 1737 if (rp[i] > bcol) break; 1738 if (rp[i] == bcol) { 1739 *v++ = ap[bs2*i+bs*cidx+ridx]; 1740 goto finished; 1741 } 1742 } 1743 *v++ = 0.0; 1744 finished:; 1745 } 1746 } 1747 PetscFunctionReturn(0); 1748 } 1749 1750 #undef __FUNCT__ 1751 #define __FUNCT__ "MatSetValuesBlocked_SeqBAIJ" 1752 PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 1753 { 1754 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1755 PetscInt *rp,k,low,high,t,ii,jj,row,nrow,i,col,l,rmax,N,lastcol = -1; 1756 PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen; 1757 PetscErrorCode ierr; 1758 PetscInt *aj =a->j,nonew=a->nonew,bs2=a->bs2,bs=A->rmap->bs,stepval; 1759 PetscBool roworiented=a->roworiented; 1760 const PetscScalar *value = v; 1761 MatScalar *ap,*aa = a->a,*bap; 1762 1763 PetscFunctionBegin; 1764 if (roworiented) { 1765 stepval = (n-1)*bs; 1766 } else { 1767 stepval = (m-1)*bs; 1768 } 1769 for (k=0; k<m; k++) { /* loop over added rows */ 1770 row = im[k]; 1771 if (row < 0) continue; 1772 #if defined(PETSC_USE_DEBUG) 1773 if (row >= a->mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,a->mbs-1); 1774 #endif 1775 rp = aj + ai[row]; 1776 ap = aa + bs2*ai[row]; 1777 rmax = imax[row]; 1778 nrow = ailen[row]; 1779 low = 0; 1780 high = nrow; 1781 for (l=0; l<n; l++) { /* loop over added columns */ 1782 if (in[l] < 0) continue; 1783 #if defined(PETSC_USE_DEBUG) 1784 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); 1785 #endif 1786 col = in[l]; 1787 if (roworiented) { 1788 value = v + (k*(stepval+bs) + l)*bs; 1789 } else { 1790 value = v + (l*(stepval+bs) + k)*bs; 1791 } 1792 if (col <= lastcol) low = 0; 1793 else high = nrow; 1794 lastcol = col; 1795 while (high-low > 7) { 1796 t = (low+high)/2; 1797 if (rp[t] > col) high = t; 1798 else low = t; 1799 } 1800 for (i=low; i<high; i++) { 1801 if (rp[i] > col) break; 1802 if (rp[i] == col) { 1803 bap = ap + bs2*i; 1804 if (roworiented) { 1805 if (is == ADD_VALUES) { 1806 for (ii=0; ii<bs; ii++,value+=stepval) { 1807 for (jj=ii; jj<bs2; jj+=bs) { 1808 bap[jj] += *value++; 1809 } 1810 } 1811 } else { 1812 for (ii=0; ii<bs; ii++,value+=stepval) { 1813 for (jj=ii; jj<bs2; jj+=bs) { 1814 bap[jj] = *value++; 1815 } 1816 } 1817 } 1818 } else { 1819 if (is == ADD_VALUES) { 1820 for (ii=0; ii<bs; ii++,value+=bs+stepval) { 1821 for (jj=0; jj<bs; jj++) { 1822 bap[jj] += value[jj]; 1823 } 1824 bap += bs; 1825 } 1826 } else { 1827 for (ii=0; ii<bs; ii++,value+=bs+stepval) { 1828 for (jj=0; jj<bs; jj++) { 1829 bap[jj] = value[jj]; 1830 } 1831 bap += bs; 1832 } 1833 } 1834 } 1835 goto noinsert2; 1836 } 1837 } 1838 if (nonew == 1) goto noinsert2; 1839 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col); 1840 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 1841 N = nrow++ - 1; high++; 1842 /* shift up all the later entries in this row */ 1843 for (ii=N; ii>=i; ii--) { 1844 rp[ii+1] = rp[ii]; 1845 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); 1846 } 1847 if (N >= i) { 1848 ierr = PetscMemzero(ap+bs2*i,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1849 } 1850 rp[i] = col; 1851 bap = ap + bs2*i; 1852 if (roworiented) { 1853 for (ii=0; ii<bs; ii++,value+=stepval) { 1854 for (jj=ii; jj<bs2; jj+=bs) { 1855 bap[jj] = *value++; 1856 } 1857 } 1858 } else { 1859 for (ii=0; ii<bs; ii++,value+=stepval) { 1860 for (jj=0; jj<bs; jj++) { 1861 *bap++ = *value++; 1862 } 1863 } 1864 } 1865 noinsert2:; 1866 low = i; 1867 } 1868 ailen[row] = nrow; 1869 } 1870 PetscFunctionReturn(0); 1871 } 1872 1873 #undef __FUNCT__ 1874 #define __FUNCT__ "MatAssemblyEnd_SeqBAIJ" 1875 PetscErrorCode MatAssemblyEnd_SeqBAIJ(Mat A,MatAssemblyType mode) 1876 { 1877 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 1878 PetscInt fshift = 0,i,j,*ai = a->i,*aj = a->j,*imax = a->imax; 1879 PetscInt m = A->rmap->N,*ip,N,*ailen = a->ilen; 1880 PetscErrorCode ierr; 1881 PetscInt mbs = a->mbs,bs2 = a->bs2,rmax = 0; 1882 MatScalar *aa = a->a,*ap; 1883 PetscReal ratio=0.6; 1884 1885 PetscFunctionBegin; 1886 if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0); 1887 1888 if (m) rmax = ailen[0]; 1889 for (i=1; i<mbs; i++) { 1890 /* move each row back by the amount of empty slots (fshift) before it*/ 1891 fshift += imax[i-1] - ailen[i-1]; 1892 rmax = PetscMax(rmax,ailen[i]); 1893 if (fshift) { 1894 ip = aj + ai[i]; ap = aa + bs2*ai[i]; 1895 N = ailen[i]; 1896 for (j=0; j<N; j++) { 1897 ip[j-fshift] = ip[j]; 1898 1899 ierr = PetscMemcpy(ap+(j-fshift)*bs2,ap+j*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1900 } 1901 } 1902 ai[i] = ai[i-1] + ailen[i-1]; 1903 } 1904 if (mbs) { 1905 fshift += imax[mbs-1] - ailen[mbs-1]; 1906 ai[mbs] = ai[mbs-1] + ailen[mbs-1]; 1907 } 1908 1909 /* reset ilen and imax for each row */ 1910 a->nonzerorowcnt = 0; 1911 for (i=0; i<mbs; i++) { 1912 ailen[i] = imax[i] = ai[i+1] - ai[i]; 1913 a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0); 1914 } 1915 a->nz = ai[mbs]; 1916 1917 /* diagonals may have moved, so kill the diagonal pointers */ 1918 a->idiagvalid = PETSC_FALSE; 1919 if (fshift && a->diag) { 1920 ierr = PetscFree(a->diag);CHKERRQ(ierr); 1921 ierr = PetscLogObjectMemory((PetscObject)A,-(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1922 a->diag = 0; 1923 } 1924 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); 1925 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); 1926 ierr = PetscInfo1(A,"Number of mallocs during MatSetValues is %D\n",a->reallocs);CHKERRQ(ierr); 1927 ierr = PetscInfo1(A,"Most nonzeros blocks in any row is %D\n",rmax);CHKERRQ(ierr); 1928 1929 A->info.mallocs += a->reallocs; 1930 a->reallocs = 0; 1931 A->info.nz_unneeded = (PetscReal)fshift*bs2; 1932 a->rmax = rmax; 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 bs=Y->rmap->bs,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 { 2427 Mat B; 2428 PetscInt *nnz; 2429 if (bs != X->rmap->bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrices must have same block size"); 2430 ierr = PetscMalloc1(Y->rmap->N,&nnz);CHKERRQ(ierr); 2431 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2432 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2433 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2434 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2435 ierr = MatSetType(B,(MatType) ((PetscObject)Y)->type_name);CHKERRQ(ierr); 2436 ierr = MatAXPYGetPreallocation_SeqBAIJ(Y,X,nnz);CHKERRQ(ierr); 2437 ierr = MatSeqBAIJSetPreallocation(B,bs,0,nnz);CHKERRQ(ierr); 2438 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2439 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2440 ierr = PetscFree(nnz);CHKERRQ(ierr); 2441 } 2442 PetscFunctionReturn(0); 2443 } 2444 2445 #undef __FUNCT__ 2446 #define __FUNCT__ "MatRealPart_SeqBAIJ" 2447 PetscErrorCode MatRealPart_SeqBAIJ(Mat A) 2448 { 2449 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2450 PetscInt i,nz = a->bs2*a->i[a->mbs]; 2451 MatScalar *aa = a->a; 2452 2453 PetscFunctionBegin; 2454 for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]); 2455 PetscFunctionReturn(0); 2456 } 2457 2458 #undef __FUNCT__ 2459 #define __FUNCT__ "MatImaginaryPart_SeqBAIJ" 2460 PetscErrorCode MatImaginaryPart_SeqBAIJ(Mat A) 2461 { 2462 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2463 PetscInt i,nz = a->bs2*a->i[a->mbs]; 2464 MatScalar *aa = a->a; 2465 2466 PetscFunctionBegin; 2467 for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]); 2468 PetscFunctionReturn(0); 2469 } 2470 2471 #undef __FUNCT__ 2472 #define __FUNCT__ "MatGetColumnIJ_SeqBAIJ" 2473 /* 2474 Code almost idential to MatGetColumnIJ_SeqAIJ() should share common code 2475 */ 2476 PetscErrorCode MatGetColumnIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 2477 { 2478 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2479 PetscErrorCode ierr; 2480 PetscInt bs = A->rmap->bs,i,*collengths,*cia,*cja,n = A->cmap->n/bs,m = A->rmap->n/bs; 2481 PetscInt nz = a->i[m],row,*jj,mr,col; 2482 2483 PetscFunctionBegin; 2484 *nn = n; 2485 if (!ia) PetscFunctionReturn(0); 2486 if (symmetric) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for BAIJ matrices"); 2487 else { 2488 ierr = PetscCalloc1((n+1),&collengths);CHKERRQ(ierr); 2489 ierr = PetscMalloc1((n+1),&cia);CHKERRQ(ierr); 2490 ierr = PetscMalloc1((nz+1),&cja);CHKERRQ(ierr); 2491 jj = a->j; 2492 for (i=0; i<nz; i++) { 2493 collengths[jj[i]]++; 2494 } 2495 cia[0] = oshift; 2496 for (i=0; i<n; i++) { 2497 cia[i+1] = cia[i] + collengths[i]; 2498 } 2499 ierr = PetscMemzero(collengths,n*sizeof(PetscInt));CHKERRQ(ierr); 2500 jj = a->j; 2501 for (row=0; row<m; row++) { 2502 mr = a->i[row+1] - a->i[row]; 2503 for (i=0; i<mr; i++) { 2504 col = *jj++; 2505 2506 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 2507 } 2508 } 2509 ierr = PetscFree(collengths);CHKERRQ(ierr); 2510 *ia = cia; *ja = cja; 2511 } 2512 PetscFunctionReturn(0); 2513 } 2514 2515 #undef __FUNCT__ 2516 #define __FUNCT__ "MatRestoreColumnIJ_SeqBAIJ" 2517 PetscErrorCode MatRestoreColumnIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 2518 { 2519 PetscErrorCode ierr; 2520 2521 PetscFunctionBegin; 2522 if (!ia) PetscFunctionReturn(0); 2523 ierr = PetscFree(*ia);CHKERRQ(ierr); 2524 ierr = PetscFree(*ja);CHKERRQ(ierr); 2525 PetscFunctionReturn(0); 2526 } 2527 2528 /* 2529 MatGetColumnIJ_SeqBAIJ_Color() and MatRestoreColumnIJ_SeqBAIJ_Color() are customized from 2530 MatGetColumnIJ_SeqBAIJ() and MatRestoreColumnIJ_SeqBAIJ() by adding an output 2531 spidx[], index of a->a, to be used in MatTransposeColoringCreate() and MatFDColoringCreate() 2532 */ 2533 #undef __FUNCT__ 2534 #define __FUNCT__ "MatGetColumnIJ_SeqBAIJ_Color" 2535 PetscErrorCode MatGetColumnIJ_SeqBAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 2536 { 2537 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 2538 PetscErrorCode ierr; 2539 PetscInt i,*collengths,*cia,*cja,n=a->nbs,m=a->mbs; 2540 PetscInt nz = a->i[m],row,*jj,mr,col; 2541 PetscInt *cspidx; 2542 2543 PetscFunctionBegin; 2544 *nn = n; 2545 if (!ia) PetscFunctionReturn(0); 2546 2547 ierr = PetscCalloc1((n+1),&collengths);CHKERRQ(ierr); 2548 ierr = PetscMalloc1((n+1),&cia);CHKERRQ(ierr); 2549 ierr = PetscMalloc1((nz+1),&cja);CHKERRQ(ierr); 2550 ierr = PetscMalloc1((nz+1),&cspidx);CHKERRQ(ierr); 2551 jj = a->j; 2552 for (i=0; i<nz; i++) { 2553 collengths[jj[i]]++; 2554 } 2555 cia[0] = oshift; 2556 for (i=0; i<n; i++) { 2557 cia[i+1] = cia[i] + collengths[i]; 2558 } 2559 ierr = PetscMemzero(collengths,n*sizeof(PetscInt));CHKERRQ(ierr); 2560 jj = a->j; 2561 for (row=0; row<m; row++) { 2562 mr = a->i[row+1] - a->i[row]; 2563 for (i=0; i<mr; i++) { 2564 col = *jj++; 2565 cspidx[cia[col] + collengths[col] - oshift] = a->i[row] + i; /* index of a->j */ 2566 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 2567 } 2568 } 2569 ierr = PetscFree(collengths);CHKERRQ(ierr); 2570 *ia = cia; *ja = cja; 2571 *spidx = cspidx; 2572 PetscFunctionReturn(0); 2573 } 2574 2575 #undef __FUNCT__ 2576 #define __FUNCT__ "MatRestoreColumnIJ_SeqBAIJ_Color" 2577 PetscErrorCode MatRestoreColumnIJ_SeqBAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 2578 { 2579 PetscErrorCode ierr; 2580 2581 PetscFunctionBegin; 2582 ierr = MatRestoreColumnIJ_SeqBAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);CHKERRQ(ierr); 2583 ierr = PetscFree(*spidx);CHKERRQ(ierr); 2584 PetscFunctionReturn(0); 2585 } 2586 2587 /* -------------------------------------------------------------------*/ 2588 static struct _MatOps MatOps_Values = {MatSetValues_SeqBAIJ, 2589 MatGetRow_SeqBAIJ, 2590 MatRestoreRow_SeqBAIJ, 2591 MatMult_SeqBAIJ_N, 2592 /* 4*/ MatMultAdd_SeqBAIJ_N, 2593 MatMultTranspose_SeqBAIJ, 2594 MatMultTransposeAdd_SeqBAIJ, 2595 0, 2596 0, 2597 0, 2598 /* 10*/ 0, 2599 MatLUFactor_SeqBAIJ, 2600 0, 2601 0, 2602 MatTranspose_SeqBAIJ, 2603 /* 15*/ MatGetInfo_SeqBAIJ, 2604 MatEqual_SeqBAIJ, 2605 MatGetDiagonal_SeqBAIJ, 2606 MatDiagonalScale_SeqBAIJ, 2607 MatNorm_SeqBAIJ, 2608 /* 20*/ 0, 2609 MatAssemblyEnd_SeqBAIJ, 2610 MatSetOption_SeqBAIJ, 2611 MatZeroEntries_SeqBAIJ, 2612 /* 24*/ MatZeroRows_SeqBAIJ, 2613 0, 2614 0, 2615 0, 2616 0, 2617 /* 29*/ MatSetUp_SeqBAIJ, 2618 0, 2619 0, 2620 0, 2621 0, 2622 /* 34*/ MatDuplicate_SeqBAIJ, 2623 0, 2624 0, 2625 MatILUFactor_SeqBAIJ, 2626 0, 2627 /* 39*/ MatAXPY_SeqBAIJ, 2628 MatGetSubMatrices_SeqBAIJ, 2629 MatIncreaseOverlap_SeqBAIJ, 2630 MatGetValues_SeqBAIJ, 2631 MatCopy_SeqBAIJ, 2632 /* 44*/ 0, 2633 MatScale_SeqBAIJ, 2634 0, 2635 0, 2636 MatZeroRowsColumns_SeqBAIJ, 2637 /* 49*/ 0, 2638 MatGetRowIJ_SeqBAIJ, 2639 MatRestoreRowIJ_SeqBAIJ, 2640 MatGetColumnIJ_SeqBAIJ, 2641 MatRestoreColumnIJ_SeqBAIJ, 2642 /* 54*/ MatFDColoringCreate_SeqXAIJ, 2643 0, 2644 0, 2645 0, 2646 MatSetValuesBlocked_SeqBAIJ, 2647 /* 59*/ MatGetSubMatrix_SeqBAIJ, 2648 MatDestroy_SeqBAIJ, 2649 MatView_SeqBAIJ, 2650 0, 2651 0, 2652 /* 64*/ 0, 2653 0, 2654 0, 2655 0, 2656 0, 2657 /* 69*/ MatGetRowMaxAbs_SeqBAIJ, 2658 0, 2659 MatConvert_Basic, 2660 0, 2661 0, 2662 /* 74*/ 0, 2663 MatFDColoringApply_BAIJ, 2664 0, 2665 0, 2666 0, 2667 /* 79*/ 0, 2668 0, 2669 0, 2670 0, 2671 MatLoad_SeqBAIJ, 2672 /* 84*/ 0, 2673 0, 2674 0, 2675 0, 2676 0, 2677 /* 89*/ 0, 2678 0, 2679 0, 2680 0, 2681 0, 2682 /* 94*/ 0, 2683 0, 2684 0, 2685 0, 2686 0, 2687 /* 99*/ 0, 2688 0, 2689 0, 2690 0, 2691 0, 2692 /*104*/ 0, 2693 MatRealPart_SeqBAIJ, 2694 MatImaginaryPart_SeqBAIJ, 2695 0, 2696 0, 2697 /*109*/ 0, 2698 0, 2699 0, 2700 0, 2701 MatMissingDiagonal_SeqBAIJ, 2702 /*114*/ 0, 2703 0, 2704 0, 2705 0, 2706 0, 2707 /*119*/ 0, 2708 0, 2709 MatMultHermitianTranspose_SeqBAIJ, 2710 MatMultHermitianTransposeAdd_SeqBAIJ, 2711 0, 2712 /*124*/ 0, 2713 0, 2714 MatInvertBlockDiagonal_SeqBAIJ, 2715 0, 2716 0, 2717 /*129*/ 0, 2718 0, 2719 0, 2720 0, 2721 0, 2722 /*134*/ 0, 2723 0, 2724 0, 2725 0, 2726 0, 2727 /*139*/ 0, 2728 0, 2729 0, 2730 MatFDColoringSetUp_SeqXAIJ, 2731 0, 2732 /*144*/MatCreateMPIMatConcatenateSeqMat_SeqBAIJ 2733 }; 2734 2735 #undef __FUNCT__ 2736 #define __FUNCT__ "MatStoreValues_SeqBAIJ" 2737 PetscErrorCode MatStoreValues_SeqBAIJ(Mat mat) 2738 { 2739 Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)mat->data; 2740 PetscInt nz = aij->i[aij->mbs]*aij->bs2; 2741 PetscErrorCode ierr; 2742 2743 PetscFunctionBegin; 2744 if (aij->nonew != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 2745 2746 /* allocate space for values if not already there */ 2747 if (!aij->saved_values) { 2748 ierr = PetscMalloc1((nz+1),&aij->saved_values);CHKERRQ(ierr); 2749 ierr = PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));CHKERRQ(ierr); 2750 } 2751 2752 /* copy values over */ 2753 ierr = PetscMemcpy(aij->saved_values,aij->a,nz*sizeof(PetscScalar));CHKERRQ(ierr); 2754 PetscFunctionReturn(0); 2755 } 2756 2757 #undef __FUNCT__ 2758 #define __FUNCT__ "MatRetrieveValues_SeqBAIJ" 2759 PetscErrorCode MatRetrieveValues_SeqBAIJ(Mat mat) 2760 { 2761 Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)mat->data; 2762 PetscErrorCode ierr; 2763 PetscInt nz = aij->i[aij->mbs]*aij->bs2; 2764 2765 PetscFunctionBegin; 2766 if (aij->nonew != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 2767 if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first"); 2768 2769 /* copy values over */ 2770 ierr = PetscMemcpy(aij->a,aij->saved_values,nz*sizeof(PetscScalar));CHKERRQ(ierr); 2771 PetscFunctionReturn(0); 2772 } 2773 2774 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat, MatType,MatReuse,Mat*); 2775 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqSBAIJ(Mat, MatType,MatReuse,Mat*); 2776 2777 #undef __FUNCT__ 2778 #define __FUNCT__ "MatSeqBAIJSetPreallocation_SeqBAIJ" 2779 PetscErrorCode MatSeqBAIJSetPreallocation_SeqBAIJ(Mat B,PetscInt bs,PetscInt nz,PetscInt *nnz) 2780 { 2781 Mat_SeqBAIJ *b; 2782 PetscErrorCode ierr; 2783 PetscInt i,mbs,nbs,bs2; 2784 PetscBool flg = PETSC_FALSE,skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE; 2785 2786 PetscFunctionBegin; 2787 if (nz >= 0 || nnz) realalloc = PETSC_TRUE; 2788 if (nz == MAT_SKIP_ALLOCATION) { 2789 skipallocation = PETSC_TRUE; 2790 nz = 0; 2791 } 2792 2793 ierr = MatSetBlockSize(B,PetscAbs(bs));CHKERRQ(ierr); 2794 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2795 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2796 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2797 2798 B->preallocated = PETSC_TRUE; 2799 2800 mbs = B->rmap->n/bs; 2801 nbs = B->cmap->n/bs; 2802 bs2 = bs*bs; 2803 2804 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); 2805 2806 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 2807 if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz); 2808 if (nnz) { 2809 for (i=0; i<mbs; i++) { 2810 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]); 2811 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); 2812 } 2813 } 2814 2815 b = (Mat_SeqBAIJ*)B->data; 2816 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Optimize options for SEQBAIJ matrix 2 ","Mat");CHKERRQ(ierr); 2817 ierr = PetscOptionsBool("-mat_no_unroll","Do not optimize for block size (slow)",NULL,flg,&flg,NULL);CHKERRQ(ierr); 2818 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2819 2820 if (!flg) { 2821 switch (bs) { 2822 case 1: 2823 B->ops->mult = MatMult_SeqBAIJ_1; 2824 B->ops->multadd = MatMultAdd_SeqBAIJ_1; 2825 break; 2826 case 2: 2827 B->ops->mult = MatMult_SeqBAIJ_2; 2828 B->ops->multadd = MatMultAdd_SeqBAIJ_2; 2829 break; 2830 case 3: 2831 B->ops->mult = MatMult_SeqBAIJ_3; 2832 B->ops->multadd = MatMultAdd_SeqBAIJ_3; 2833 break; 2834 case 4: 2835 B->ops->mult = MatMult_SeqBAIJ_4; 2836 B->ops->multadd = MatMultAdd_SeqBAIJ_4; 2837 break; 2838 case 5: 2839 B->ops->mult = MatMult_SeqBAIJ_5; 2840 B->ops->multadd = MatMultAdd_SeqBAIJ_5; 2841 break; 2842 case 6: 2843 B->ops->mult = MatMult_SeqBAIJ_6; 2844 B->ops->multadd = MatMultAdd_SeqBAIJ_6; 2845 break; 2846 case 7: 2847 B->ops->mult = MatMult_SeqBAIJ_7; 2848 B->ops->multadd = MatMultAdd_SeqBAIJ_7; 2849 break; 2850 case 15: 2851 B->ops->mult = MatMult_SeqBAIJ_15_ver1; 2852 B->ops->multadd = MatMultAdd_SeqBAIJ_N; 2853 break; 2854 default: 2855 B->ops->mult = MatMult_SeqBAIJ_N; 2856 B->ops->multadd = MatMultAdd_SeqBAIJ_N; 2857 break; 2858 } 2859 } 2860 B->ops->sor = MatSOR_SeqBAIJ; 2861 b->mbs = mbs; 2862 b->nbs = nbs; 2863 if (!skipallocation) { 2864 if (!b->imax) { 2865 ierr = PetscMalloc2(mbs,&b->imax,mbs,&b->ilen);CHKERRQ(ierr); 2866 ierr = PetscLogObjectMemory((PetscObject)B,2*mbs*sizeof(PetscInt));CHKERRQ(ierr); 2867 2868 b->free_imax_ilen = PETSC_TRUE; 2869 } 2870 /* b->ilen will count nonzeros in each block row so far. */ 2871 for (i=0; i<mbs; i++) b->ilen[i] = 0; 2872 if (!nnz) { 2873 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 2874 else if (nz < 0) nz = 1; 2875 for (i=0; i<mbs; i++) b->imax[i] = nz; 2876 nz = nz*mbs; 2877 } else { 2878 nz = 0; 2879 for (i=0; i<mbs; i++) {b->imax[i] = nnz[i]; nz += nnz[i];} 2880 } 2881 2882 /* allocate the matrix space */ 2883 ierr = MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);CHKERRQ(ierr); 2884 ierr = PetscMalloc3(bs2*nz,&b->a,nz,&b->j,B->rmap->N+1,&b->i);CHKERRQ(ierr); 2885 ierr = PetscLogObjectMemory((PetscObject)B,(B->rmap->N+1)*sizeof(PetscInt)+nz*(bs2*sizeof(PetscScalar)+sizeof(PetscInt)));CHKERRQ(ierr); 2886 ierr = PetscMemzero(b->a,nz*bs2*sizeof(MatScalar));CHKERRQ(ierr); 2887 ierr = PetscMemzero(b->j,nz*sizeof(PetscInt));CHKERRQ(ierr); 2888 2889 b->singlemalloc = PETSC_TRUE; 2890 b->i[0] = 0; 2891 for (i=1; i<mbs+1; i++) { 2892 b->i[i] = b->i[i-1] + b->imax[i-1]; 2893 } 2894 b->free_a = PETSC_TRUE; 2895 b->free_ij = PETSC_TRUE; 2896 #if defined(PETSC_THREADCOMM_ACTIVE) 2897 ierr = MatZeroEntries_SeqBAIJ(B);CHKERRQ(ierr); 2898 #endif 2899 } else { 2900 b->free_a = PETSC_FALSE; 2901 b->free_ij = PETSC_FALSE; 2902 } 2903 2904 b->bs2 = bs2; 2905 b->mbs = mbs; 2906 b->nz = 0; 2907 b->maxnz = nz; 2908 B->info.nz_unneeded = (PetscReal)b->maxnz*bs2; 2909 if (realalloc) {ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr);} 2910 PetscFunctionReturn(0); 2911 } 2912 2913 #undef __FUNCT__ 2914 #define __FUNCT__ "MatSeqBAIJSetPreallocationCSR_SeqBAIJ" 2915 PetscErrorCode MatSeqBAIJSetPreallocationCSR_SeqBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[]) 2916 { 2917 PetscInt i,m,nz,nz_max=0,*nnz; 2918 PetscScalar *values=0; 2919 PetscBool roworiented = ((Mat_SeqBAIJ*)B->data)->roworiented; 2920 PetscErrorCode ierr; 2921 2922 PetscFunctionBegin; 2923 if (bs < 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive but it is %D",bs); 2924 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 2925 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 2926 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2927 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2928 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2929 m = B->rmap->n/bs; 2930 2931 if (ii[0] != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "ii[0] must be 0 but it is %D",ii[0]); 2932 ierr = PetscMalloc1((m+1), &nnz);CHKERRQ(ierr); 2933 for (i=0; i<m; i++) { 2934 nz = ii[i+1]- ii[i]; 2935 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D",i,nz); 2936 nz_max = PetscMax(nz_max, nz); 2937 nnz[i] = nz; 2938 } 2939 ierr = MatSeqBAIJSetPreallocation(B,bs,0,nnz);CHKERRQ(ierr); 2940 ierr = PetscFree(nnz);CHKERRQ(ierr); 2941 2942 values = (PetscScalar*)V; 2943 if (!values) { 2944 ierr = PetscCalloc1(bs*bs*(nz_max+1),&values);CHKERRQ(ierr); 2945 } 2946 for (i=0; i<m; i++) { 2947 PetscInt ncols = ii[i+1] - ii[i]; 2948 const PetscInt *icols = jj + ii[i]; 2949 const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0); 2950 if (!roworiented) { 2951 ierr = MatSetValuesBlocked_SeqBAIJ(B,1,&i,ncols,icols,svals,INSERT_VALUES);CHKERRQ(ierr); 2952 } else { 2953 PetscInt j; 2954 for (j=0; j<ncols; j++) { 2955 const PetscScalar *svals = values + (V ? (bs*bs*(ii[i]+j)) : 0); 2956 ierr = MatSetValuesBlocked_SeqBAIJ(B,1,&i,1,&icols[j],svals,INSERT_VALUES);CHKERRQ(ierr); 2957 } 2958 } 2959 } 2960 if (!V) { ierr = PetscFree(values);CHKERRQ(ierr); } 2961 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2962 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2963 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2964 PetscFunctionReturn(0); 2965 } 2966 2967 PETSC_EXTERN PetscErrorCode MatGetFactor_seqbaij_petsc(Mat,MatFactorType,Mat*); 2968 PETSC_EXTERN PetscErrorCode MatGetFactor_seqbaij_bstrm(Mat,MatFactorType,Mat*); 2969 #if defined(PETSC_HAVE_MUMPS) 2970 PETSC_EXTERN PetscErrorCode MatGetFactor_baij_mumps(Mat,MatFactorType,Mat*); 2971 #endif 2972 extern PetscErrorCode MatGetFactorAvailable_seqbaij_petsc(Mat,MatFactorType,PetscBool*); 2973 2974 /*MC 2975 MATSEQBAIJ - MATSEQBAIJ = "seqbaij" - A matrix type to be used for sequential block sparse matrices, based on 2976 block sparse compressed row format. 2977 2978 Options Database Keys: 2979 . -mat_type seqbaij - sets the matrix type to "seqbaij" during a call to MatSetFromOptions() 2980 2981 Level: beginner 2982 2983 .seealso: MatCreateSeqBAIJ() 2984 M*/ 2985 2986 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBSTRM(Mat, MatType,MatReuse,Mat*); 2987 2988 #undef __FUNCT__ 2989 #define __FUNCT__ "MatCreate_SeqBAIJ" 2990 PETSC_EXTERN PetscErrorCode MatCreate_SeqBAIJ(Mat B) 2991 { 2992 PetscErrorCode ierr; 2993 PetscMPIInt size; 2994 Mat_SeqBAIJ *b; 2995 2996 PetscFunctionBegin; 2997 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 2998 if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Comm must be of size 1"); 2999 3000 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 3001 B->data = (void*)b; 3002 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 3003 3004 b->row = 0; 3005 b->col = 0; 3006 b->icol = 0; 3007 b->reallocs = 0; 3008 b->saved_values = 0; 3009 3010 b->roworiented = PETSC_TRUE; 3011 b->nonew = 0; 3012 b->diag = 0; 3013 b->solve_work = 0; 3014 b->mult_work = 0; 3015 B->spptr = 0; 3016 B->info.nz_unneeded = (PetscReal)b->maxnz*b->bs2; 3017 b->keepnonzeropattern = PETSC_FALSE; 3018 3019 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactorAvailable_petsc_C",MatGetFactorAvailable_seqbaij_petsc);CHKERRQ(ierr); 3020 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_petsc_C",MatGetFactor_seqbaij_petsc);CHKERRQ(ierr); 3021 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_bstrm_C",MatGetFactor_seqbaij_bstrm);CHKERRQ(ierr); 3022 #if defined(PETSC_HAVE_MUMPS) 3023 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C", MatGetFactor_baij_mumps);CHKERRQ(ierr); 3024 #endif 3025 ierr = PetscObjectComposeFunction((PetscObject)B,"MatInvertBlockDiagonal_C",MatInvertBlockDiagonal_SeqBAIJ);CHKERRQ(ierr); 3026 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqBAIJ);CHKERRQ(ierr); 3027 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqBAIJ);CHKERRQ(ierr); 3028 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetColumnIndices_C",MatSeqBAIJSetColumnIndices_SeqBAIJ);CHKERRQ(ierr); 3029 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqaij_C",MatConvert_SeqBAIJ_SeqAIJ);CHKERRQ(ierr); 3030 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqsbaij_C",MatConvert_SeqBAIJ_SeqSBAIJ);CHKERRQ(ierr); 3031 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetPreallocation_C",MatSeqBAIJSetPreallocation_SeqBAIJ);CHKERRQ(ierr); 3032 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetPreallocationCSR_C",MatSeqBAIJSetPreallocationCSR_SeqBAIJ);CHKERRQ(ierr); 3033 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqbstrm_C",MatConvert_SeqBAIJ_SeqBSTRM);CHKERRQ(ierr); 3034 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqBAIJ);CHKERRQ(ierr); 3035 ierr = PetscObjectChangeTypeName((PetscObject)B,MATSEQBAIJ);CHKERRQ(ierr); 3036 PetscFunctionReturn(0); 3037 } 3038 3039 #undef __FUNCT__ 3040 #define __FUNCT__ "MatDuplicateNoCreate_SeqBAIJ" 3041 PetscErrorCode MatDuplicateNoCreate_SeqBAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace) 3042 { 3043 Mat_SeqBAIJ *c = (Mat_SeqBAIJ*)C->data,*a = (Mat_SeqBAIJ*)A->data; 3044 PetscErrorCode ierr; 3045 PetscInt i,mbs = a->mbs,nz = a->nz,bs2 = a->bs2; 3046 3047 PetscFunctionBegin; 3048 if (a->i[mbs] != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupt matrix"); 3049 3050 if (cpvalues == MAT_SHARE_NONZERO_PATTERN) { 3051 c->imax = a->imax; 3052 c->ilen = a->ilen; 3053 c->free_imax_ilen = PETSC_FALSE; 3054 } else { 3055 ierr = PetscMalloc2(mbs,&c->imax,mbs,&c->ilen);CHKERRQ(ierr); 3056 ierr = PetscLogObjectMemory((PetscObject)C,2*mbs*sizeof(PetscInt));CHKERRQ(ierr); 3057 for (i=0; i<mbs; i++) { 3058 c->imax[i] = a->imax[i]; 3059 c->ilen[i] = a->ilen[i]; 3060 } 3061 c->free_imax_ilen = PETSC_TRUE; 3062 } 3063 3064 /* allocate the matrix space */ 3065 if (mallocmatspace) { 3066 if (cpvalues == MAT_SHARE_NONZERO_PATTERN) { 3067 ierr = PetscCalloc1(bs2*nz,&c->a);CHKERRQ(ierr); 3068 ierr = PetscLogObjectMemory((PetscObject)C,a->i[mbs]*bs2*sizeof(PetscScalar));CHKERRQ(ierr); 3069 3070 c->i = a->i; 3071 c->j = a->j; 3072 c->singlemalloc = PETSC_FALSE; 3073 c->free_a = PETSC_TRUE; 3074 c->free_ij = PETSC_FALSE; 3075 c->parent = A; 3076 C->preallocated = PETSC_TRUE; 3077 C->assembled = PETSC_TRUE; 3078 3079 ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); 3080 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3081 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3082 } else { 3083 ierr = PetscMalloc3(bs2*nz,&c->a,nz,&c->j,mbs+1,&c->i);CHKERRQ(ierr); 3084 ierr = PetscLogObjectMemory((PetscObject)C,a->i[mbs]*(bs2*sizeof(PetscScalar)+sizeof(PetscInt))+(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 3085 3086 c->singlemalloc = PETSC_TRUE; 3087 c->free_a = PETSC_TRUE; 3088 c->free_ij = PETSC_TRUE; 3089 3090 ierr = PetscMemcpy(c->i,a->i,(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 3091 if (mbs > 0) { 3092 ierr = PetscMemcpy(c->j,a->j,nz*sizeof(PetscInt));CHKERRQ(ierr); 3093 if (cpvalues == MAT_COPY_VALUES) { 3094 ierr = PetscMemcpy(c->a,a->a,bs2*nz*sizeof(MatScalar));CHKERRQ(ierr); 3095 } else { 3096 ierr = PetscMemzero(c->a,bs2*nz*sizeof(MatScalar));CHKERRQ(ierr); 3097 } 3098 } 3099 C->preallocated = PETSC_TRUE; 3100 C->assembled = PETSC_TRUE; 3101 } 3102 } 3103 3104 c->roworiented = a->roworiented; 3105 c->nonew = a->nonew; 3106 3107 ierr = PetscLayoutReference(A->rmap,&C->rmap);CHKERRQ(ierr); 3108 ierr = PetscLayoutReference(A->cmap,&C->cmap);CHKERRQ(ierr); 3109 3110 c->bs2 = a->bs2; 3111 c->mbs = a->mbs; 3112 c->nbs = a->nbs; 3113 3114 if (a->diag) { 3115 if (cpvalues == MAT_SHARE_NONZERO_PATTERN) { 3116 c->diag = a->diag; 3117 c->free_diag = PETSC_FALSE; 3118 } else { 3119 ierr = PetscMalloc1((mbs+1),&c->diag);CHKERRQ(ierr); 3120 ierr = PetscLogObjectMemory((PetscObject)C,(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr); 3121 for (i=0; i<mbs; i++) c->diag[i] = a->diag[i]; 3122 c->free_diag = PETSC_TRUE; 3123 } 3124 } else c->diag = 0; 3125 3126 c->nz = a->nz; 3127 c->maxnz = a->nz; /* Since we allocate exactly the right amount */ 3128 c->solve_work = 0; 3129 c->mult_work = 0; 3130 3131 c->compressedrow.use = a->compressedrow.use; 3132 c->compressedrow.nrows = a->compressedrow.nrows; 3133 if (a->compressedrow.use) { 3134 i = a->compressedrow.nrows; 3135 ierr = PetscMalloc2(i+1,&c->compressedrow.i,i+1,&c->compressedrow.rindex);CHKERRQ(ierr); 3136 ierr = PetscLogObjectMemory((PetscObject)C,(2*i+1)*sizeof(PetscInt));CHKERRQ(ierr); 3137 ierr = PetscMemcpy(c->compressedrow.i,a->compressedrow.i,(i+1)*sizeof(PetscInt));CHKERRQ(ierr); 3138 ierr = PetscMemcpy(c->compressedrow.rindex,a->compressedrow.rindex,i*sizeof(PetscInt));CHKERRQ(ierr); 3139 } else { 3140 c->compressedrow.use = PETSC_FALSE; 3141 c->compressedrow.i = NULL; 3142 c->compressedrow.rindex = NULL; 3143 } 3144 C->nonzerostate = A->nonzerostate; 3145 3146 ierr = PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);CHKERRQ(ierr); 3147 ierr = PetscMemcpy(C->ops,A->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 3148 PetscFunctionReturn(0); 3149 } 3150 3151 #undef __FUNCT__ 3152 #define __FUNCT__ "MatDuplicate_SeqBAIJ" 3153 PetscErrorCode MatDuplicate_SeqBAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B) 3154 { 3155 PetscErrorCode ierr; 3156 3157 PetscFunctionBegin; 3158 ierr = MatCreate(PetscObjectComm((PetscObject)A),B);CHKERRQ(ierr); 3159 ierr = MatSetSizes(*B,A->rmap->N,A->cmap->n,A->rmap->N,A->cmap->n);CHKERRQ(ierr); 3160 ierr = MatSetType(*B,MATSEQBAIJ);CHKERRQ(ierr); 3161 ierr = MatDuplicateNoCreate_SeqBAIJ(*B,A,cpvalues,PETSC_TRUE);CHKERRQ(ierr); 3162 PetscFunctionReturn(0); 3163 } 3164 3165 #undef __FUNCT__ 3166 #define __FUNCT__ "MatLoad_SeqBAIJ" 3167 PetscErrorCode MatLoad_SeqBAIJ(Mat newmat,PetscViewer viewer) 3168 { 3169 Mat_SeqBAIJ *a; 3170 PetscErrorCode ierr; 3171 PetscInt i,nz,header[4],*rowlengths=0,M,N,bs=1; 3172 PetscInt *mask,mbs,*jj,j,rowcount,nzcount,k,*browlengths,maskcount; 3173 PetscInt kmax,jcount,block,idx,point,nzcountb,extra_rows,rows,cols; 3174 PetscInt *masked,nmask,tmp,bs2,ishift; 3175 PetscMPIInt size; 3176 int fd; 3177 PetscScalar *aa; 3178 MPI_Comm comm; 3179 3180 PetscFunctionBegin; 3181 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 3182 ierr = PetscOptionsBegin(comm,NULL,"Options for loading SEQBAIJ matrix","Mat");CHKERRQ(ierr); 3183 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 3184 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3185 bs2 = bs*bs; 3186 3187 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3188 if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"view must have one processor"); 3189 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 3190 ierr = PetscBinaryRead(fd,header,4,PETSC_INT);CHKERRQ(ierr); 3191 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not Mat object"); 3192 M = header[1]; N = header[2]; nz = header[3]; 3193 3194 if (header[3] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as SeqBAIJ"); 3195 if (M != N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Can only do square matrices"); 3196 3197 /* 3198 This code adds extra rows to make sure the number of rows is 3199 divisible by the blocksize 3200 */ 3201 mbs = M/bs; 3202 extra_rows = bs - M + bs*(mbs); 3203 if (extra_rows == bs) extra_rows = 0; 3204 else mbs++; 3205 if (extra_rows) { 3206 ierr = PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");CHKERRQ(ierr); 3207 } 3208 3209 /* Set global sizes if not already set */ 3210 if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) { 3211 ierr = MatSetSizes(newmat,PETSC_DECIDE,PETSC_DECIDE,M+extra_rows,N+extra_rows);CHKERRQ(ierr); 3212 } else { /* Check if the matrix global sizes are correct */ 3213 ierr = MatGetSize(newmat,&rows,&cols);CHKERRQ(ierr); 3214 if (rows < 0 && cols < 0) { /* user might provide local size instead of global size */ 3215 ierr = MatGetLocalSize(newmat,&rows,&cols);CHKERRQ(ierr); 3216 } 3217 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); 3218 } 3219 3220 /* read in row lengths */ 3221 ierr = PetscMalloc1((M+extra_rows),&rowlengths);CHKERRQ(ierr); 3222 ierr = PetscBinaryRead(fd,rowlengths,M,PETSC_INT);CHKERRQ(ierr); 3223 for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1; 3224 3225 /* read in column indices */ 3226 ierr = PetscMalloc1((nz+extra_rows),&jj);CHKERRQ(ierr); 3227 ierr = PetscBinaryRead(fd,jj,nz,PETSC_INT);CHKERRQ(ierr); 3228 for (i=0; i<extra_rows; i++) jj[nz+i] = M+i; 3229 3230 /* loop over row lengths determining block row lengths */ 3231 ierr = PetscCalloc1(mbs,&browlengths);CHKERRQ(ierr); 3232 ierr = PetscMalloc2(mbs,&mask,mbs,&masked);CHKERRQ(ierr); 3233 ierr = PetscMemzero(mask,mbs*sizeof(PetscInt));CHKERRQ(ierr); 3234 rowcount = 0; 3235 nzcount = 0; 3236 for (i=0; i<mbs; i++) { 3237 nmask = 0; 3238 for (j=0; j<bs; j++) { 3239 kmax = rowlengths[rowcount]; 3240 for (k=0; k<kmax; k++) { 3241 tmp = jj[nzcount++]/bs; 3242 if (!mask[tmp]) {masked[nmask++] = tmp; mask[tmp] = 1;} 3243 } 3244 rowcount++; 3245 } 3246 browlengths[i] += nmask; 3247 /* zero out the mask elements we set */ 3248 for (j=0; j<nmask; j++) mask[masked[j]] = 0; 3249 } 3250 3251 /* Do preallocation */ 3252 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(newmat,bs,0,browlengths);CHKERRQ(ierr); 3253 a = (Mat_SeqBAIJ*)newmat->data; 3254 3255 /* set matrix "i" values */ 3256 a->i[0] = 0; 3257 for (i=1; i<= mbs; i++) { 3258 a->i[i] = a->i[i-1] + browlengths[i-1]; 3259 a->ilen[i-1] = browlengths[i-1]; 3260 } 3261 a->nz = 0; 3262 for (i=0; i<mbs; i++) a->nz += browlengths[i]; 3263 3264 /* read in nonzero values */ 3265 ierr = PetscMalloc1((nz+extra_rows),&aa);CHKERRQ(ierr); 3266 ierr = PetscBinaryRead(fd,aa,nz,PETSC_SCALAR);CHKERRQ(ierr); 3267 for (i=0; i<extra_rows; i++) aa[nz+i] = 1.0; 3268 3269 /* set "a" and "j" values into matrix */ 3270 nzcount = 0; jcount = 0; 3271 for (i=0; i<mbs; i++) { 3272 nzcountb = nzcount; 3273 nmask = 0; 3274 for (j=0; j<bs; j++) { 3275 kmax = rowlengths[i*bs+j]; 3276 for (k=0; k<kmax; k++) { 3277 tmp = jj[nzcount++]/bs; 3278 if (!mask[tmp]) { masked[nmask++] = tmp; mask[tmp] = 1;} 3279 } 3280 } 3281 /* sort the masked values */ 3282 ierr = PetscSortInt(nmask,masked);CHKERRQ(ierr); 3283 3284 /* set "j" values into matrix */ 3285 maskcount = 1; 3286 for (j=0; j<nmask; j++) { 3287 a->j[jcount++] = masked[j]; 3288 mask[masked[j]] = maskcount++; 3289 } 3290 /* set "a" values into matrix */ 3291 ishift = bs2*a->i[i]; 3292 for (j=0; j<bs; j++) { 3293 kmax = rowlengths[i*bs+j]; 3294 for (k=0; k<kmax; k++) { 3295 tmp = jj[nzcountb]/bs; 3296 block = mask[tmp] - 1; 3297 point = jj[nzcountb] - bs*tmp; 3298 idx = ishift + bs2*block + j + bs*point; 3299 a->a[idx] = (MatScalar)aa[nzcountb++]; 3300 } 3301 } 3302 /* zero out the mask elements we set */ 3303 for (j=0; j<nmask; j++) mask[masked[j]] = 0; 3304 } 3305 if (jcount != a->nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Bad binary matrix"); 3306 3307 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3308 ierr = PetscFree(browlengths);CHKERRQ(ierr); 3309 ierr = PetscFree(aa);CHKERRQ(ierr); 3310 ierr = PetscFree(jj);CHKERRQ(ierr); 3311 ierr = PetscFree2(mask,masked);CHKERRQ(ierr); 3312 3313 ierr = MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3314 ierr = MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3315 PetscFunctionReturn(0); 3316 } 3317 3318 #undef __FUNCT__ 3319 #define __FUNCT__ "MatCreateSeqBAIJ" 3320 /*@C 3321 MatCreateSeqBAIJ - Creates a sparse matrix in block AIJ (block 3322 compressed row) format. For good matrix assembly performance the 3323 user should preallocate the matrix storage by setting the parameter nz 3324 (or the array nnz). By setting these parameters accurately, performance 3325 during matrix assembly can be increased by more than a factor of 50. 3326 3327 Collective on MPI_Comm 3328 3329 Input Parameters: 3330 + comm - MPI communicator, set to PETSC_COMM_SELF 3331 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 3332 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 3333 . m - number of rows 3334 . n - number of columns 3335 . nz - number of nonzero blocks per block row (same for all rows) 3336 - nnz - array containing the number of nonzero blocks in the various block rows 3337 (possibly different for each block row) or NULL 3338 3339 Output Parameter: 3340 . A - the matrix 3341 3342 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3343 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3344 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3345 3346 Options Database Keys: 3347 . -mat_no_unroll - uses code that does not unroll the loops in the 3348 block calculations (much slower) 3349 . -mat_block_size - size of the blocks to use 3350 3351 Level: intermediate 3352 3353 Notes: 3354 The number of rows and columns must be divisible by blocksize. 3355 3356 If the nnz parameter is given then the nz parameter is ignored 3357 3358 A nonzero block is any block that as 1 or more nonzeros in it 3359 3360 The block AIJ format is fully compatible with standard Fortran 77 3361 storage. That is, the stored row and column indices can begin at 3362 either one (as in Fortran) or zero. See the users' manual for details. 3363 3364 Specify the preallocated storage with either nz or nnz (not both). 3365 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3366 allocation. See Users-Manual: ch_mat for details. 3367 matrices. 3368 3369 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateBAIJ() 3370 @*/ 3371 PetscErrorCode MatCreateSeqBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A) 3372 { 3373 PetscErrorCode ierr; 3374 3375 PetscFunctionBegin; 3376 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3377 ierr = MatSetSizes(*A,m,n,m,n);CHKERRQ(ierr); 3378 ierr = MatSetType(*A,MATSEQBAIJ);CHKERRQ(ierr); 3379 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(*A,bs,nz,(PetscInt*)nnz);CHKERRQ(ierr); 3380 PetscFunctionReturn(0); 3381 } 3382 3383 #undef __FUNCT__ 3384 #define __FUNCT__ "MatSeqBAIJSetPreallocation" 3385 /*@C 3386 MatSeqBAIJSetPreallocation - Sets the block size and expected nonzeros 3387 per row in the matrix. For good matrix assembly performance the 3388 user should preallocate the matrix storage by setting the parameter nz 3389 (or the array nnz). By setting these parameters accurately, performance 3390 during matrix assembly can be increased by more than a factor of 50. 3391 3392 Collective on MPI_Comm 3393 3394 Input Parameters: 3395 + B - the matrix 3396 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 3397 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 3398 . nz - number of block nonzeros per block row (same for all rows) 3399 - nnz - array containing the number of block nonzeros in the various block rows 3400 (possibly different for each block row) or NULL 3401 3402 Options Database Keys: 3403 . -mat_no_unroll - uses code that does not unroll the loops in the 3404 block calculations (much slower) 3405 . -mat_block_size - size of the blocks to use 3406 3407 Level: intermediate 3408 3409 Notes: 3410 If the nnz parameter is given then the nz parameter is ignored 3411 3412 You can call MatGetInfo() to get information on how effective the preallocation was; 3413 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3414 You can also run with the option -info and look for messages with the string 3415 malloc in them to see if additional memory allocation was needed. 3416 3417 The block AIJ format is fully compatible with standard Fortran 77 3418 storage. That is, the stored row and column indices can begin at 3419 either one (as in Fortran) or zero. See the users' manual for details. 3420 3421 Specify the preallocated storage with either nz or nnz (not both). 3422 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3423 allocation. See Users-Manual: ch_mat for details. 3424 3425 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateBAIJ(), MatGetInfo() 3426 @*/ 3427 PetscErrorCode MatSeqBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt nz,const PetscInt nnz[]) 3428 { 3429 PetscErrorCode ierr; 3430 3431 PetscFunctionBegin; 3432 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3433 PetscValidType(B,1); 3434 PetscValidLogicalCollectiveInt(B,bs,2); 3435 ierr = PetscTryMethod(B,"MatSeqBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[]),(B,bs,nz,nnz));CHKERRQ(ierr); 3436 PetscFunctionReturn(0); 3437 } 3438 3439 #undef __FUNCT__ 3440 #define __FUNCT__ "MatSeqBAIJSetPreallocationCSR" 3441 /*@C 3442 MatSeqBAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format 3443 (the default sequential PETSc format). 3444 3445 Collective on MPI_Comm 3446 3447 Input Parameters: 3448 + B - the matrix 3449 . i - the indices into j for the start of each local row (starts with zero) 3450 . j - the column indices for each local row (starts with zero) these must be sorted for each row 3451 - v - optional values in the matrix 3452 3453 Level: developer 3454 3455 Notes: 3456 The order of the entries in values is specified by the MatOption MAT_ROW_ORIENTED. For example, C programs 3457 may want to use the default MAT_ROW_ORIENTED=PETSC_TRUE and use an array v[nnz][bs][bs] where the second index is 3458 over rows within a block and the last index is over columns within a block row. Fortran programs will likely set 3459 MAT_ROW_ORIENTED=PETSC_FALSE and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a 3460 block column and the second index is over columns within a block. 3461 3462 .keywords: matrix, aij, compressed row, sparse 3463 3464 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatSeqBAIJSetPreallocation(), MATSEQBAIJ 3465 @*/ 3466 PetscErrorCode MatSeqBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3467 { 3468 PetscErrorCode ierr; 3469 3470 PetscFunctionBegin; 3471 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3472 PetscValidType(B,1); 3473 PetscValidLogicalCollectiveInt(B,bs,2); 3474 ierr = PetscTryMethod(B,"MatSeqBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));CHKERRQ(ierr); 3475 PetscFunctionReturn(0); 3476 } 3477 3478 3479 #undef __FUNCT__ 3480 #define __FUNCT__ "MatCreateSeqBAIJWithArrays" 3481 /*@ 3482 MatCreateSeqBAIJWithArrays - Creates an sequential BAIJ matrix using matrix elements provided by the user. 3483 3484 Collective on MPI_Comm 3485 3486 Input Parameters: 3487 + comm - must be an MPI communicator of size 1 3488 . bs - size of block 3489 . m - number of rows 3490 . n - number of columns 3491 . i - row indices 3492 . j - column indices 3493 - a - matrix values 3494 3495 Output Parameter: 3496 . mat - the matrix 3497 3498 Level: advanced 3499 3500 Notes: 3501 The i, j, and a arrays are not copied by this routine, the user must free these arrays 3502 once the matrix is destroyed 3503 3504 You cannot set new nonzero locations into this matrix, that will generate an error. 3505 3506 The i and j indices are 0 based 3507 3508 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). 3509 3510 The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is 3511 the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first 3512 block, followed by the second column of the first block etc etc. That is, the blocks are contiguous in memory 3513 with column-major ordering within blocks. 3514 3515 .seealso: MatCreate(), MatCreateBAIJ(), MatCreateSeqBAIJ() 3516 3517 @*/ 3518 PetscErrorCode MatCreateSeqBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt *i,PetscInt *j,PetscScalar *a,Mat *mat) 3519 { 3520 PetscErrorCode ierr; 3521 PetscInt ii; 3522 Mat_SeqBAIJ *baij; 3523 3524 PetscFunctionBegin; 3525 if (bs != 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"block size %D > 1 is not supported yet",bs); 3526 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3527 3528 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3529 ierr = MatSetSizes(*mat,m,n,m,n);CHKERRQ(ierr); 3530 ierr = MatSetType(*mat,MATSEQBAIJ);CHKERRQ(ierr); 3531 ierr = MatSeqBAIJSetPreallocation_SeqBAIJ(*mat,bs,MAT_SKIP_ALLOCATION,0);CHKERRQ(ierr); 3532 baij = (Mat_SeqBAIJ*)(*mat)->data; 3533 ierr = PetscMalloc2(m,&baij->imax,m,&baij->ilen);CHKERRQ(ierr); 3534 ierr = PetscLogObjectMemory((PetscObject)*mat,2*m*sizeof(PetscInt));CHKERRQ(ierr); 3535 3536 baij->i = i; 3537 baij->j = j; 3538 baij->a = a; 3539 3540 baij->singlemalloc = PETSC_FALSE; 3541 baij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/ 3542 baij->free_a = PETSC_FALSE; 3543 baij->free_ij = PETSC_FALSE; 3544 3545 for (ii=0; ii<m; ii++) { 3546 baij->ilen[ii] = baij->imax[ii] = i[ii+1] - i[ii]; 3547 #if defined(PETSC_USE_DEBUG) 3548 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]); 3549 #endif 3550 } 3551 #if defined(PETSC_USE_DEBUG) 3552 for (ii=0; ii<baij->i[m]; ii++) { 3553 if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %d index = %d",ii,j[ii]); 3554 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]); 3555 } 3556 #endif 3557 3558 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3559 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3560 PetscFunctionReturn(0); 3561 } 3562 3563 #undef __FUNCT__ 3564 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_SeqBAIJ" 3565 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqBAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 3566 { 3567 PetscErrorCode ierr; 3568 3569 PetscFunctionBegin; 3570 ierr = MatCreateMPIMatConcatenateSeqMat_MPIBAIJ(comm,inmat,n,scall,outmat);CHKERRQ(ierr); 3571 PetscFunctionReturn(0); 3572 } 3573