1 2 /* 3 Defines matrix-matrix product routines for pairs of MPIAIJ matrices 4 C = A * B 5 */ 6 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 7 #include <../src/mat/utils/freespace.h> 8 #include <../src/mat/impls/aij/mpi/mpiaij.h> 9 #include <petscbt.h> 10 #include <../src/mat/impls/dense/mpi/mpidense.h> 11 #include <petsc-private/vecimpl.h> 12 13 #undef __FUNCT__ 14 #define __FUNCT__ "MatMatMult_MPIAIJ_MPIAIJ" 15 PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill, Mat *C) 16 { 17 PetscErrorCode ierr; 18 19 PetscFunctionBegin; 20 if (scall == MAT_INITIAL_MATRIX) { 21 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 22 } 23 24 ierr = (*(*C)->ops->matmultnumeric)(A,B,*C);CHKERRQ(ierr); 25 PetscFunctionReturn(0); 26 } 27 28 #undef __FUNCT__ 29 #define __FUNCT__ "MatDestroy_MPIAIJ_MatMatMult" 30 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat A) 31 { 32 PetscErrorCode ierr; 33 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 34 Mat_PtAPMPI *ptap = a->ptap; 35 36 PetscFunctionBegin; 37 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 38 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 39 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 40 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 41 ierr = MatDestroy(&ptap->Pt);CHKERRQ(ierr); 42 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 43 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 44 ierr = PetscFree(ptap->apa);CHKERRQ(ierr); 45 ierr = ptap->destroy(A);CHKERRQ(ierr); 46 ierr = PetscFree(ptap);CHKERRQ(ierr); 47 PetscFunctionReturn(0); 48 } 49 50 #undef __FUNCT__ 51 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatMatMult" 52 PetscErrorCode MatDuplicate_MPIAIJ_MatMatMult(Mat A, MatDuplicateOption op, Mat *M) 53 { 54 PetscErrorCode ierr; 55 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 56 Mat_PtAPMPI *ptap = a->ptap; 57 58 PetscFunctionBegin; 59 ierr = (*ptap->duplicate)(A,op,M);CHKERRQ(ierr); 60 61 (*M)->ops->destroy = ptap->destroy; /* = MatDestroy_MPIAIJ, *M doesn't duplicate A's special structure! */ 62 (*M)->ops->duplicate = ptap->duplicate; /* = MatDuplicate_MPIAIJ */ 63 PetscFunctionReturn(0); 64 } 65 66 #undef __FUNCT__ 67 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIAIJ" 68 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 69 { 70 PetscErrorCode ierr; 71 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 72 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 73 Mat_SeqAIJ *cd =(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 74 PetscInt *adi=ad->i,*adj,*aoi=ao->i,*aoj; 75 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 76 Mat_SeqAIJ *p_loc,*p_oth; 77 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 78 PetscScalar *pa_loc,*pa_oth,*pa,*apa,valtmp,*ca; 79 PetscInt cm =C->rmap->n,anz,pnz; 80 Mat_PtAPMPI *ptap=c->ptap; 81 PetscInt *api,*apj,*apJ,i,j,k,row; 82 PetscInt cstart=C->cmap->rstart; 83 PetscInt cdnz,conz,k0,k1; 84 85 PetscFunctionBegin; 86 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 87 /*-----------------------------------------------------*/ 88 /* update numerical values of P_oth and P_loc */ 89 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 90 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 91 92 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 93 /*----------------------------------------------------------*/ 94 /* get data from symbolic products */ 95 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 96 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 97 pi_loc=p_loc->i; pj_loc=p_loc->j; pa_loc=p_loc->a; 98 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 99 100 /* get apa for storing dense row A[i,:]*P */ 101 apa = ptap->apa; 102 103 api = ptap->api; 104 apj = ptap->apj; 105 for (i=0; i<cm; i++) { 106 /* diagonal portion of A */ 107 anz = adi[i+1] - adi[i]; 108 adj = ad->j + adi[i]; 109 ada = ad->a + adi[i]; 110 for (j=0; j<anz; j++) { 111 row = adj[j]; 112 pnz = pi_loc[row+1] - pi_loc[row]; 113 pj = pj_loc + pi_loc[row]; 114 pa = pa_loc + pi_loc[row]; 115 116 /* perform dense axpy */ 117 valtmp = ada[j]; 118 for (k=0; k<pnz; k++) { 119 apa[pj[k]] += valtmp*pa[k]; 120 } 121 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 122 } 123 124 /* off-diagonal portion of A */ 125 anz = aoi[i+1] - aoi[i]; 126 aoj = ao->j + aoi[i]; 127 aoa = ao->a + aoi[i]; 128 for (j=0; j<anz; j++) { 129 row = aoj[j]; 130 pnz = pi_oth[row+1] - pi_oth[row]; 131 pj = pj_oth + pi_oth[row]; 132 pa = pa_oth + pi_oth[row]; 133 134 /* perform dense axpy */ 135 valtmp = aoa[j]; 136 for (k=0; k<pnz; k++) { 137 apa[pj[k]] += valtmp*pa[k]; 138 } 139 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 140 } 141 142 /* set values in C */ 143 apJ = apj + api[i]; 144 cdnz = cd->i[i+1] - cd->i[i]; 145 conz = co->i[i+1] - co->i[i]; 146 147 /* 1st off-diagoanl part of C */ 148 ca = coa + co->i[i]; 149 k = 0; 150 for (k0=0; k0<conz; k0++) { 151 if (apJ[k] >= cstart) break; 152 ca[k0] = apa[apJ[k]]; 153 apa[apJ[k]] = 0.0; 154 k++; 155 } 156 157 /* diagonal part of C */ 158 ca = cda + cd->i[i]; 159 for (k1=0; k1<cdnz; k1++) { 160 ca[k1] = apa[apJ[k]]; 161 apa[apJ[k]] = 0.0; 162 k++; 163 } 164 165 /* 2nd off-diagoanl part of C */ 166 ca = coa + co->i[i]; 167 for (; k0<conz; k0++) { 168 ca[k0] = apa[apJ[k]]; 169 apa[apJ[k]] = 0.0; 170 k++; 171 } 172 } 173 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 174 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 175 PetscFunctionReturn(0); 176 } 177 178 #undef __FUNCT__ 179 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ" 180 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 181 { 182 PetscErrorCode ierr; 183 MPI_Comm comm; 184 Mat Cmpi; 185 Mat_PtAPMPI *ptap; 186 PetscFreeSpaceList free_space=NULL,current_space=NULL; 187 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c; 188 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 189 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 190 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 191 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 192 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 193 PetscBT lnkbt; 194 PetscScalar *apa; 195 PetscReal afill; 196 PetscBool scalable=PETSC_TRUE; 197 PetscInt nlnk_max,armax,prmax; 198 199 PetscFunctionBegin; 200 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 201 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 202 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 203 } 204 205 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 206 ierr = PetscOptionsBool("-matmatmult_scalable","Use a scalable but slower C=A*B","",scalable,&scalable,NULL);CHKERRQ(ierr); 207 if (scalable) { 208 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(A,P,fill,C);CHKERRQ(ierr); 209 PetscFunctionReturn(0); 210 } 211 ierr = PetscOptionsEnd();CHKERRQ(ierr); 212 213 /* create struct Mat_PtAPMPI and attached it to C later */ 214 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 215 216 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 217 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 218 219 /* get P_loc by taking all local rows of P */ 220 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 221 222 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 223 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 224 pi_loc = p_loc->i; pj_loc = p_loc->j; 225 pi_oth = p_oth->i; pj_oth = p_oth->j; 226 227 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 228 /*-------------------------------------------------------------------*/ 229 ierr = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr); 230 ptap->api = api; 231 api[0] = 0; 232 233 /* create and initialize a linked list */ 234 armax = ad->rmax+ao->rmax; 235 prmax = PetscMax(p_loc->rmax,p_oth->rmax); 236 nlnk_max = armax*prmax; 237 if (!nlnk_max || nlnk_max > pN) nlnk_max = pN; 238 ierr = PetscLLCondensedCreate(nlnk_max,pN,&lnk,&lnkbt);CHKERRQ(ierr); 239 240 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 241 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 242 243 current_space = free_space; 244 245 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 246 for (i=0; i<am; i++) { 247 /* diagonal portion of A */ 248 nzi = adi[i+1] - adi[i]; 249 for (j=0; j<nzi; j++) { 250 row = *adj++; 251 pnz = pi_loc[row+1] - pi_loc[row]; 252 Jptr = pj_loc + pi_loc[row]; 253 /* add non-zero cols of P into the sorted linked list lnk */ 254 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 255 } 256 /* off-diagonal portion of A */ 257 nzi = aoi[i+1] - aoi[i]; 258 for (j=0; j<nzi; j++) { 259 row = *aoj++; 260 pnz = pi_oth[row+1] - pi_oth[row]; 261 Jptr = pj_oth + pi_oth[row]; 262 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 263 } 264 265 apnz = lnk[0]; 266 api[i+1] = api[i] + apnz; 267 268 /* if free space is not available, double the total space in the list */ 269 if (current_space->local_remaining<apnz) { 270 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 271 nspacedouble++; 272 } 273 274 /* Copy data into free space, then initialize lnk */ 275 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 276 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 277 278 current_space->array += apnz; 279 current_space->local_used += apnz; 280 current_space->local_remaining -= apnz; 281 } 282 283 /* Allocate space for apj, initialize apj, and */ 284 /* destroy list of free space and other temporary array(s) */ 285 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ptap->apj);CHKERRQ(ierr); 286 apj = ptap->apj; 287 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 288 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 289 290 /* malloc apa to store dense row A[i,:]*P */ 291 ierr = PetscMalloc(pN*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 292 ierr = PetscMemzero(apa,pN*sizeof(PetscScalar));CHKERRQ(ierr); 293 294 ptap->apa = apa; 295 296 /* create and assemble symbolic parallel matrix Cmpi */ 297 /*----------------------------------------------------*/ 298 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 299 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 300 ierr = MatSetBlockSizes(Cmpi,A->rmap->bs,P->cmap->bs);CHKERRQ(ierr); 301 302 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 303 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 304 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 305 for (i=0; i<am; i++) { 306 row = i + rstart; 307 apnz = api[i+1] - api[i]; 308 ierr = MatSetValues(Cmpi,1,&row,apnz,apj,apa,INSERT_VALUES);CHKERRQ(ierr); 309 apj += apnz; 310 } 311 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 312 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 313 314 ptap->destroy = Cmpi->ops->destroy; 315 ptap->duplicate = Cmpi->ops->duplicate; 316 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 317 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 318 319 /* attach the supporting struct to Cmpi for reuse */ 320 c = (Mat_MPIAIJ*)Cmpi->data; 321 c->ptap = ptap; 322 323 *C = Cmpi; 324 325 /* set MatInfo */ 326 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 327 if (afill < 1.0) afill = 1.0; 328 Cmpi->info.mallocs = nspacedouble; 329 Cmpi->info.fill_ratio_given = fill; 330 Cmpi->info.fill_ratio_needed = afill; 331 332 #if defined(PETSC_USE_INFO) 333 if (api[am]) { 334 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 335 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 336 } else { 337 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 338 } 339 #endif 340 PetscFunctionReturn(0); 341 } 342 343 #undef __FUNCT__ 344 #define __FUNCT__ "MatMatMult_MPIAIJ_MPIDense" 345 PetscErrorCode MatMatMult_MPIAIJ_MPIDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 346 { 347 PetscErrorCode ierr; 348 349 PetscFunctionBegin; 350 if (scall == MAT_INITIAL_MATRIX) { 351 ierr = MatMatMultSymbolic_MPIAIJ_MPIDense(A,B,fill,C);CHKERRQ(ierr); 352 } 353 ierr = MatMatMultNumeric_MPIAIJ_MPIDense(A,B,*C);CHKERRQ(ierr); 354 PetscFunctionReturn(0); 355 } 356 357 typedef struct { 358 Mat workB; 359 PetscScalar *rvalues,*svalues; 360 MPI_Request *rwaits,*swaits; 361 } MPIAIJ_MPIDense; 362 363 #undef __FUNCT__ 364 #define __FUNCT__ "MatMPIAIJ_MPIDenseDestroy" 365 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 366 { 367 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*) ctx; 368 PetscErrorCode ierr; 369 370 PetscFunctionBegin; 371 ierr = MatDestroy(&contents->workB);CHKERRQ(ierr); 372 ierr = PetscFree4(contents->rvalues,contents->svalues,contents->rwaits,contents->swaits);CHKERRQ(ierr); 373 ierr = PetscFree(contents);CHKERRQ(ierr); 374 PetscFunctionReturn(0); 375 } 376 377 #undef __FUNCT__ 378 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIDense" 379 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C) 380 { 381 PetscErrorCode ierr; 382 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 383 PetscInt nz = aij->B->cmap->n; 384 PetscContainer container; 385 MPIAIJ_MPIDense *contents; 386 VecScatter ctx = aij->Mvctx; 387 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 388 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 389 PetscInt m = A->rmap->n,n=B->cmap->n; 390 391 PetscFunctionBegin; 392 ierr = MatCreate(PetscObjectComm((PetscObject)B),C);CHKERRQ(ierr); 393 ierr = MatSetSizes(*C,m,n,A->rmap->N,B->cmap->N);CHKERRQ(ierr); 394 ierr = MatSetBlockSizes(*C,A->rmap->bs,B->cmap->bs);CHKERRQ(ierr); 395 ierr = MatSetType(*C,MATMPIDENSE);CHKERRQ(ierr); 396 ierr = MatMPIDenseSetPreallocation(*C,NULL);CHKERRQ(ierr); 397 ierr = MatAssemblyBegin(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 398 ierr = MatAssemblyEnd(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 399 400 (*C)->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 401 402 ierr = PetscNew(MPIAIJ_MPIDense,&contents);CHKERRQ(ierr); 403 /* Create work matrix used to store off processor rows of B needed for local product */ 404 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,NULL,&contents->workB);CHKERRQ(ierr); 405 /* Create work arrays needed */ 406 ierr = PetscMalloc4(B->cmap->N*from->starts[from->n],PetscScalar,&contents->rvalues, 407 B->cmap->N*to->starts[to->n],PetscScalar,&contents->svalues, 408 from->n,MPI_Request,&contents->rwaits, 409 to->n,MPI_Request,&contents->swaits);CHKERRQ(ierr); 410 411 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);CHKERRQ(ierr); 412 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 413 ierr = PetscContainerSetUserDestroy(container,MatMPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 414 ierr = PetscObjectCompose((PetscObject)(*C),"workB",(PetscObject)container);CHKERRQ(ierr); 415 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 416 PetscFunctionReturn(0); 417 } 418 419 #undef __FUNCT__ 420 #define __FUNCT__ "MatMPIDenseScatter" 421 /* 422 Performs an efficient scatter on the rows of B needed by this process; this is 423 a modification of the VecScatterBegin_() routines. 424 */ 425 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,Mat C,Mat *outworkB) 426 { 427 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 428 PetscErrorCode ierr; 429 PetscScalar *b,*w,*svalues,*rvalues; 430 VecScatter ctx = aij->Mvctx; 431 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 432 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 433 PetscInt i,j,k; 434 PetscInt *sindices,*sstarts,*rindices,*rstarts; 435 PetscMPIInt *sprocs,*rprocs,nrecvs; 436 MPI_Request *swaits,*rwaits; 437 MPI_Comm comm; 438 PetscMPIInt tag = ((PetscObject)ctx)->tag,ncols = B->cmap->N, nrows = aij->B->cmap->n,imdex,nrowsB = B->rmap->n; 439 MPI_Status status; 440 MPIAIJ_MPIDense *contents; 441 PetscContainer container; 442 Mat workB; 443 444 PetscFunctionBegin; 445 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 446 ierr = PetscObjectQuery((PetscObject)C,"workB",(PetscObject*)&container);CHKERRQ(ierr); 447 if (!container) SETERRQ(comm,PETSC_ERR_PLIB,"Container does not exist"); 448 ierr = PetscContainerGetPointer(container,(void**)&contents);CHKERRQ(ierr); 449 450 workB = *outworkB = contents->workB; 451 if (nrows != workB->rmap->n) SETERRQ2(comm,PETSC_ERR_PLIB,"Number of rows of workB %D not equal to columns of aij->B %D",nrows,workB->cmap->n); 452 sindices = to->indices; 453 sstarts = to->starts; 454 sprocs = to->procs; 455 swaits = contents->swaits; 456 svalues = contents->svalues; 457 458 rindices = from->indices; 459 rstarts = from->starts; 460 rprocs = from->procs; 461 rwaits = contents->rwaits; 462 rvalues = contents->rvalues; 463 464 ierr = MatDenseGetArray(B,&b);CHKERRQ(ierr); 465 ierr = MatDenseGetArray(workB,&w);CHKERRQ(ierr); 466 467 for (i=0; i<from->n; i++) { 468 ierr = MPI_Irecv(rvalues+ncols*rstarts[i],ncols*(rstarts[i+1]-rstarts[i]),MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 469 } 470 471 for (i=0; i<to->n; i++) { 472 /* pack a message at a time */ 473 CHKMEMQ; 474 for (j=0; j<sstarts[i+1]-sstarts[i]; j++) { 475 for (k=0; k<ncols; k++) { 476 svalues[ncols*(sstarts[i] + j) + k] = b[sindices[sstarts[i]+j] + nrowsB*k]; 477 } 478 } 479 CHKMEMQ; 480 ierr = MPI_Isend(svalues+ncols*sstarts[i],ncols*(sstarts[i+1]-sstarts[i]),MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 481 } 482 483 nrecvs = from->n; 484 while (nrecvs) { 485 ierr = MPI_Waitany(from->n,rwaits,&imdex,&status);CHKERRQ(ierr); 486 nrecvs--; 487 /* unpack a message at a time */ 488 CHKMEMQ; 489 for (j=0; j<rstarts[imdex+1]-rstarts[imdex]; j++) { 490 for (k=0; k<ncols; k++) { 491 w[rindices[rstarts[imdex]+j] + nrows*k] = rvalues[ncols*(rstarts[imdex] + j) + k]; 492 } 493 } 494 CHKMEMQ; 495 } 496 if (to->n) {ierr = MPI_Waitall(to->n,swaits,to->sstatus);CHKERRQ(ierr);} 497 498 ierr = MatDenseRestoreArray(B,&b);CHKERRQ(ierr); 499 ierr = MatDenseRestoreArray(workB,&w);CHKERRQ(ierr); 500 ierr = MatAssemblyBegin(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 501 ierr = MatAssemblyEnd(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 502 PetscFunctionReturn(0); 503 } 504 extern PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat); 505 506 #undef __FUNCT__ 507 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIDense" 508 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 509 { 510 PetscErrorCode ierr; 511 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 512 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 513 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 514 Mat workB; 515 516 PetscFunctionBegin; 517 /* diagonal block of A times all local rows of B*/ 518 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(aij->A,bdense->A,cdense->A);CHKERRQ(ierr); 519 520 /* get off processor parts of B needed to complete the product */ 521 ierr = MatMPIDenseScatter(A,B,C,&workB);CHKERRQ(ierr); 522 523 /* off-diagonal block of A times nonlocal rows of B */ 524 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A);CHKERRQ(ierr); 525 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 526 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 527 PetscFunctionReturn(0); 528 } 529 530 #undef __FUNCT__ 531 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable" 532 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable(Mat A,Mat P,Mat C) 533 { 534 PetscErrorCode ierr; 535 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 536 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 537 Mat_SeqAIJ *cd = (Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 538 PetscInt *adi = ad->i,*adj,*aoi=ao->i,*aoj; 539 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 540 Mat_SeqAIJ *p_loc,*p_oth; 541 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 542 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 543 PetscInt cm = C->rmap->n,anz,pnz; 544 Mat_PtAPMPI *ptap = c->ptap; 545 PetscScalar *apa_sparse = ptap->apa; 546 PetscInt *api,*apj,*apJ,i,j,k,row; 547 PetscInt cstart = C->cmap->rstart; 548 PetscInt cdnz,conz,k0,k1,nextp; 549 550 PetscFunctionBegin; 551 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 552 /*-----------------------------------------------------*/ 553 /* update numerical values of P_oth and P_loc */ 554 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 555 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 556 557 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 558 /*----------------------------------------------------------*/ 559 /* get data from symbolic products */ 560 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 561 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 562 pi_loc=p_loc->i; pj_loc=p_loc->j; pa_loc=p_loc->a; 563 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 564 565 api = ptap->api; 566 apj = ptap->apj; 567 for (i=0; i<cm; i++) { 568 apJ = apj + api[i]; 569 570 /* diagonal portion of A */ 571 anz = adi[i+1] - adi[i]; 572 adj = ad->j + adi[i]; 573 ada = ad->a + adi[i]; 574 for (j=0; j<anz; j++) { 575 row = adj[j]; 576 pnz = pi_loc[row+1] - pi_loc[row]; 577 pj = pj_loc + pi_loc[row]; 578 pa = pa_loc + pi_loc[row]; 579 /* perform sparse axpy */ 580 valtmp = ada[j]; 581 nextp = 0; 582 for (k=0; nextp<pnz; k++) { 583 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 584 apa_sparse[k] += valtmp*pa[nextp++]; 585 } 586 } 587 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 588 } 589 590 /* off-diagonal portion of A */ 591 anz = aoi[i+1] - aoi[i]; 592 aoj = ao->j + aoi[i]; 593 aoa = ao->a + aoi[i]; 594 for (j=0; j<anz; j++) { 595 row = aoj[j]; 596 pnz = pi_oth[row+1] - pi_oth[row]; 597 pj = pj_oth + pi_oth[row]; 598 pa = pa_oth + pi_oth[row]; 599 /* perform sparse axpy */ 600 valtmp = aoa[j]; 601 nextp = 0; 602 for (k=0; nextp<pnz; k++) { 603 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 604 apa_sparse[k] += valtmp*pa[nextp++]; 605 } 606 } 607 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 608 } 609 610 /* set values in C */ 611 cdnz = cd->i[i+1] - cd->i[i]; 612 conz = co->i[i+1] - co->i[i]; 613 614 /* 1st off-diagoanl part of C */ 615 ca = coa + co->i[i]; 616 k = 0; 617 for (k0=0; k0<conz; k0++) { 618 if (apJ[k] >= cstart) break; 619 ca[k0] = apa_sparse[k]; 620 apa_sparse[k] = 0.0; 621 k++; 622 } 623 624 /* diagonal part of C */ 625 ca = cda + cd->i[i]; 626 for (k1=0; k1<cdnz; k1++) { 627 ca[k1] = apa_sparse[k]; 628 apa_sparse[k] = 0.0; 629 k++; 630 } 631 632 /* 2nd off-diagoanl part of C */ 633 ca = coa + co->i[i]; 634 for (; k0<conz; k0++) { 635 ca[k0] = apa_sparse[k]; 636 apa_sparse[k] = 0.0; 637 k++; 638 } 639 } 640 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 641 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 642 PetscFunctionReturn(0); 643 } 644 645 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ(), except using LLCondensed to avoid O(BN) memory requirement */ 646 #undef __FUNCT__ 647 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable" 648 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(Mat A,Mat P,PetscReal fill,Mat *C) 649 { 650 PetscErrorCode ierr; 651 MPI_Comm comm; 652 Mat Cmpi; 653 Mat_PtAPMPI *ptap; 654 PetscFreeSpaceList free_space = NULL,current_space=NULL; 655 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c; 656 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 657 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 658 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 659 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=0; 660 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 661 PetscInt nlnk_max,armax,prmax; 662 PetscReal afill; 663 PetscScalar *apa; 664 665 PetscFunctionBegin; 666 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 667 /* create struct Mat_PtAPMPI and attached it to C later */ 668 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 669 670 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 671 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 672 673 /* get P_loc by taking all local rows of P */ 674 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 675 676 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 677 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 678 pi_loc = p_loc->i; pj_loc = p_loc->j; 679 pi_oth = p_oth->i; pj_oth = p_oth->j; 680 681 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 682 /*-------------------------------------------------------------------*/ 683 ierr = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr); 684 ptap->api = api; 685 api[0] = 0; 686 687 /* create and initialize a linked list */ 688 armax = ad->rmax+ao->rmax; 689 prmax = PetscMax(p_loc->rmax,p_oth->rmax); 690 nlnk_max = armax*prmax; 691 if (!nlnk_max || nlnk_max > pN) nlnk_max = pN; 692 ierr = PetscLLCondensedCreate_Scalable(nlnk_max,&lnk);CHKERRQ(ierr); 693 694 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 695 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 696 697 current_space = free_space; 698 699 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 700 for (i=0; i<am; i++) { 701 /* diagonal portion of A */ 702 nzi = adi[i+1] - adi[i]; 703 for (j=0; j<nzi; j++) { 704 row = *adj++; 705 pnz = pi_loc[row+1] - pi_loc[row]; 706 Jptr = pj_loc + pi_loc[row]; 707 /* add non-zero cols of P into the sorted linked list lnk */ 708 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 709 } 710 /* off-diagonal portion of A */ 711 nzi = aoi[i+1] - aoi[i]; 712 for (j=0; j<nzi; j++) { 713 row = *aoj++; 714 pnz = pi_oth[row+1] - pi_oth[row]; 715 Jptr = pj_oth + pi_oth[row]; 716 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 717 } 718 719 apnz = *lnk; 720 api[i+1] = api[i] + apnz; 721 if (apnz > apnz_max) apnz_max = apnz; 722 723 /* if free space is not available, double the total space in the list */ 724 if (current_space->local_remaining<apnz) { 725 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 726 nspacedouble++; 727 } 728 729 /* Copy data into free space, then initialize lnk */ 730 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 731 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 732 733 current_space->array += apnz; 734 current_space->local_used += apnz; 735 current_space->local_remaining -= apnz; 736 } 737 738 /* Allocate space for apj, initialize apj, and */ 739 /* destroy list of free space and other temporary array(s) */ 740 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ptap->apj);CHKERRQ(ierr); 741 apj = ptap->apj; 742 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 743 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 744 745 /* create and assemble symbolic parallel matrix Cmpi */ 746 /*----------------------------------------------------*/ 747 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 748 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 749 ierr = MatSetBlockSizes(Cmpi,A->rmap->bs,P->cmap->bs);CHKERRQ(ierr); 750 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 751 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 752 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 753 754 /* malloc apa for assembly Cmpi */ 755 ierr = PetscMalloc(apnz_max*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 756 ierr = PetscMemzero(apa,apnz_max*sizeof(PetscScalar));CHKERRQ(ierr); 757 758 ptap->apa = apa; 759 for (i=0; i<am; i++) { 760 row = i + rstart; 761 apnz = api[i+1] - api[i]; 762 ierr = MatSetValues(Cmpi,1,&row,apnz,apj,apa,INSERT_VALUES);CHKERRQ(ierr); 763 apj += apnz; 764 } 765 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 766 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 767 768 ptap->destroy = Cmpi->ops->destroy; 769 ptap->duplicate = Cmpi->ops->duplicate; 770 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable; 771 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 772 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 773 774 /* attach the supporting struct to Cmpi for reuse */ 775 c = (Mat_MPIAIJ*)Cmpi->data; 776 c->ptap = ptap; 777 778 *C = Cmpi; 779 780 /* set MatInfo */ 781 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 782 if (afill < 1.0) afill = 1.0; 783 Cmpi->info.mallocs = nspacedouble; 784 Cmpi->info.fill_ratio_given = fill; 785 Cmpi->info.fill_ratio_needed = afill; 786 787 #if defined(PETSC_USE_INFO) 788 if (api[am]) { 789 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 790 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 791 } else { 792 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 793 } 794 #endif 795 PetscFunctionReturn(0); 796 } 797 798 /*-------------------------------------------------------------------------*/ 799 #undef __FUNCT__ 800 #define __FUNCT__ "MatTransposeMatMult_MPIAIJ_MPIAIJ" 801 PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat P,Mat A,MatReuse scall,PetscReal fill,Mat *C) 802 { 803 PetscErrorCode ierr; 804 PetscBool scalable=PETSC_TRUE,viamatmatmult=PETSC_FALSE; 805 806 PetscFunctionBegin; 807 ierr = PetscOptionsBool("-mattransposematmult_viamatmatmult","Use R=Pt and C=R*A","",viamatmatmult,&viamatmatmult,NULL);CHKERRQ(ierr); 808 if (viamatmatmult) { 809 Mat Pt; 810 Mat_PtAPMPI *ptap; 811 Mat_MPIAIJ *c; 812 if (scall == MAT_INITIAL_MATRIX) { 813 ierr = MatTranspose(P,MAT_INITIAL_MATRIX,&Pt);CHKERRQ(ierr); 814 ierr = MatMatMult(Pt,A,MAT_INITIAL_MATRIX,fill,C);CHKERRQ(ierr); 815 816 c = (Mat_MPIAIJ*)(*C)->data; 817 ptap = c->ptap; 818 ptap->Pt = Pt; 819 } else if (scall == MAT_REUSE_MATRIX) { 820 c = (Mat_MPIAIJ*)(*C)->data; 821 ptap = c->ptap; 822 Pt = ptap->Pt; 823 ierr = MatTranspose(P,scall,&Pt);CHKERRQ(ierr); 824 ierr = MatMatMult(Pt,A,scall,fill,C);CHKERRQ(ierr); 825 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONGSTATE,"Not supported"); 826 PetscFunctionReturn(0); 827 } 828 829 if (scall == MAT_INITIAL_MATRIX) { 830 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 831 ierr = PetscOptionsBool("-mattransposematmult_scalable","Use a scalable but slower C=Pt*A","",scalable,&scalable,NULL);CHKERRQ(ierr); 832 if (scalable) { 833 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(P,A,fill,C);CHKERRQ(ierr); 834 } else { 835 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 836 } 837 ierr = PetscOptionsEnd();CHKERRQ(ierr); 838 } 839 ierr = (*(*C)->ops->mattransposemultnumeric)(P,A,*C);CHKERRQ(ierr); 840 PetscFunctionReturn(0); 841 } 842 843 #undef __FUNCT__ 844 #define __FUNCT__ "MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ" 845 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 846 { 847 PetscErrorCode ierr; 848 Mat_Merge_SeqsToMPI *merge; 849 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 850 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 851 Mat_PtAPMPI *ptap; 852 PetscInt *adj,*aJ; 853 PetscInt i,j,k,anz,pnz,row,*cj; 854 MatScalar *ada,*aval,*ca,valtmp; 855 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 856 MPI_Comm comm; 857 PetscMPIInt size,rank,taga,*len_s; 858 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 859 PetscInt **buf_ri,**buf_rj; 860 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 861 MPI_Request *s_waits,*r_waits; 862 MPI_Status *status; 863 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 864 PetscInt *ai,*aj,*coi,*coj; 865 PetscInt *poJ,*pdJ; 866 Mat A_loc; 867 Mat_SeqAIJ *a_loc; 868 869 PetscFunctionBegin; 870 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 871 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 872 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 873 874 ptap = c->ptap; 875 merge = ptap->merge; 876 877 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 878 /*--------------------------------------------------------------*/ 879 /* get data from symbolic products */ 880 coi = merge->coi; coj = merge->coj; 881 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 882 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 883 884 bi = merge->bi; bj = merge->bj; 885 owners = merge->rowmap->range; 886 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 887 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 888 889 /* get A_loc by taking all local rows of A */ 890 A_loc = ptap->A_loc; 891 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 892 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 893 ai = a_loc->i; 894 aj = a_loc->j; 895 896 ierr = PetscMalloc((A->cmap->N)*sizeof(PetscScalar),&aval);CHKERRQ(ierr); /* non-scalable!!! */ 897 ierr = PetscMemzero(aval,A->cmap->N*sizeof(PetscScalar));CHKERRQ(ierr); 898 899 for (i=0; i<am; i++) { 900 /* 2-a) put A[i,:] to dense array aval */ 901 anz = ai[i+1] - ai[i]; 902 adj = aj + ai[i]; 903 ada = a_loc->a + ai[i]; 904 for (j=0; j<anz; j++) { 905 aval[adj[j]] = ada[j]; 906 } 907 908 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 909 /*--------------------------------------------------------------*/ 910 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 911 pnz = po->i[i+1] - po->i[i]; 912 poJ = po->j + po->i[i]; 913 pA = po->a + po->i[i]; 914 for (j=0; j<pnz; j++) { 915 row = poJ[j]; 916 cnz = coi[row+1] - coi[row]; 917 cj = coj + coi[row]; 918 ca = coa + coi[row]; 919 /* perform dense axpy */ 920 valtmp = pA[j]; 921 for (k=0; k<cnz; k++) { 922 ca[k] += valtmp*aval[cj[k]]; 923 } 924 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 925 } 926 927 /* put the value into Cd (diagonal part) */ 928 pnz = pd->i[i+1] - pd->i[i]; 929 pdJ = pd->j + pd->i[i]; 930 pA = pd->a + pd->i[i]; 931 for (j=0; j<pnz; j++) { 932 row = pdJ[j]; 933 cnz = bi[row+1] - bi[row]; 934 cj = bj + bi[row]; 935 ca = ba + bi[row]; 936 /* perform dense axpy */ 937 valtmp = pA[j]; 938 for (k=0; k<cnz; k++) { 939 ca[k] += valtmp*aval[cj[k]]; 940 } 941 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 942 } 943 944 /* zero the current row of Pt*A */ 945 aJ = aj + ai[i]; 946 for (k=0; k<anz; k++) aval[aJ[k]] = 0.0; 947 } 948 949 /* 3) send and recv matrix values coa */ 950 /*------------------------------------*/ 951 buf_ri = merge->buf_ri; 952 buf_rj = merge->buf_rj; 953 len_s = merge->len_s; 954 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 955 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 956 957 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 958 for (proc=0,k=0; proc<size; proc++) { 959 if (!len_s[proc]) continue; 960 i = merge->owners_co[proc]; 961 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 962 k++; 963 } 964 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 965 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 966 967 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 968 ierr = PetscFree(r_waits);CHKERRQ(ierr); 969 ierr = PetscFree(coa);CHKERRQ(ierr); 970 971 /* 4) insert local Cseq and received values into Cmpi */ 972 /*----------------------------------------------------*/ 973 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 974 for (k=0; k<merge->nrecv; k++) { 975 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 976 nrows = *(buf_ri_k[k]); 977 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 978 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 979 } 980 981 for (i=0; i<cm; i++) { 982 row = owners[rank] + i; /* global row index of C_seq */ 983 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 984 ba_i = ba + bi[i]; 985 bnz = bi[i+1] - bi[i]; 986 /* add received vals into ba */ 987 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 988 /* i-th row */ 989 if (i == *nextrow[k]) { 990 cnz = *(nextci[k]+1) - *nextci[k]; 991 cj = buf_rj[k] + *(nextci[k]); 992 ca = abuf_r[k] + *(nextci[k]); 993 nextcj = 0; 994 for (j=0; nextcj<cnz; j++) { 995 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 996 ba_i[j] += ca[nextcj++]; 997 } 998 } 999 nextrow[k]++; nextci[k]++; 1000 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1001 } 1002 } 1003 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1004 } 1005 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1006 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1007 1008 ierr = PetscFree(ba);CHKERRQ(ierr); 1009 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1010 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1011 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1012 ierr = PetscFree(aval);CHKERRQ(ierr); 1013 PetscFunctionReturn(0); 1014 } 1015 1016 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1017 #undef __FUNCT__ 1018 #define __FUNCT__ "MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ" 1019 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat *C) 1020 { 1021 PetscErrorCode ierr; 1022 Mat Cmpi,A_loc,POt,PDt; 1023 Mat_PtAPMPI *ptap; 1024 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1025 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data,*c; 1026 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1027 PetscInt nnz; 1028 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1029 PetscInt am=A->rmap->n,pn=P->cmap->n; 1030 PetscBT lnkbt; 1031 MPI_Comm comm; 1032 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1033 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1034 PetscInt len,proc,*dnz,*onz,*owners; 1035 PetscInt nzi,*bi,*bj; 1036 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1037 MPI_Request *swaits,*rwaits; 1038 MPI_Status *sstatus,rstatus; 1039 Mat_Merge_SeqsToMPI *merge; 1040 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1041 PetscReal afill =1.0,afill_tmp; 1042 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Crmax; 1043 PetscScalar *vals; 1044 Mat_SeqAIJ *a_loc, *pdt,*pot; 1045 1046 PetscFunctionBegin; 1047 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1048 /* check if matrix local sizes are compatible */ 1049 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 1050 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A (%D, %D) != P (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 1051 } 1052 1053 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1054 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1055 1056 /* create struct Mat_PtAPMPI and attached it to C later */ 1057 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 1058 1059 /* get A_loc by taking all local rows of A */ 1060 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1061 1062 ptap->A_loc = A_loc; 1063 1064 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1065 ai = a_loc->i; 1066 aj = a_loc->j; 1067 1068 /* determine symbolic Co=(p->B)^T*A - send to others */ 1069 /*----------------------------------------------------*/ 1070 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1071 pdt = (Mat_SeqAIJ*)PDt->data; 1072 pdti = pdt->i; pdtj = pdt->j; 1073 1074 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1075 pot = (Mat_SeqAIJ*)POt->data; 1076 poti = pot->i; potj = pot->j; 1077 1078 /* then, compute symbolic Co = (p->B)^T*A */ 1079 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors >= (num of nonzero rows of C_seq) - pn */ 1080 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 1081 coi[0] = 0; 1082 1083 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1084 nnz = fill*(poti[pon] + ai[am]); 1085 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1086 current_space = free_space; 1087 1088 /* create and initialize a linked list */ 1089 i = PetscMax(pdt->rmax,pot->rmax); 1090 Crmax = i*a_loc->rmax*size; 1091 if (!Crmax || Crmax > aN) Crmax = aN; 1092 ierr = PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt);CHKERRQ(ierr); 1093 1094 for (i=0; i<pon; i++) { 1095 pnz = poti[i+1] - poti[i]; 1096 ptJ = potj + poti[i]; 1097 for (j=0; j<pnz; j++) { 1098 row = ptJ[j]; /* row of A_loc == col of Pot */ 1099 anz = ai[row+1] - ai[row]; 1100 Jptr = aj + ai[row]; 1101 /* add non-zero cols of AP into the sorted linked list lnk */ 1102 ierr = PetscLLCondensedAddSorted(anz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1103 } 1104 nnz = lnk[0]; 1105 1106 /* If free space is not available, double the total space in the list */ 1107 if (current_space->local_remaining<nnz) { 1108 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1109 nspacedouble++; 1110 } 1111 1112 /* Copy data into free space, and zero out denserows */ 1113 ierr = PetscLLCondensedClean(aN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1114 1115 current_space->array += nnz; 1116 current_space->local_used += nnz; 1117 current_space->local_remaining -= nnz; 1118 1119 coi[i+1] = coi[i] + nnz; 1120 } 1121 1122 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 1123 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1124 1125 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1126 if (afill_tmp > afill) afill = afill_tmp; 1127 1128 /* send j-array (coj) of Co to other processors */ 1129 /*----------------------------------------------*/ 1130 /* determine row ownership */ 1131 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 1132 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1133 1134 merge->rowmap->n = pn; 1135 merge->rowmap->bs = 1; 1136 1137 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1138 owners = merge->rowmap->range; 1139 1140 /* determine the number of messages to send, their lengths */ 1141 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 1142 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1143 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 1144 1145 len_s = merge->len_s; 1146 merge->nsend = 0; 1147 1148 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 1149 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1150 1151 proc = 0; 1152 for (i=0; i<pon; i++) { 1153 while (prmap[i] >= owners[proc+1]) proc++; 1154 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1155 len_s[proc] += coi[i+1] - coi[i]; 1156 } 1157 1158 len = 0; /* max length of buf_si[] */ 1159 owners_co[0] = 0; 1160 for (proc=0; proc<size; proc++) { 1161 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1162 if (len_si[proc]) { 1163 merge->nsend++; 1164 len_si[proc] = 2*(len_si[proc] + 1); 1165 len += len_si[proc]; 1166 } 1167 } 1168 1169 /* determine the number and length of messages to receive for coi and coj */ 1170 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1171 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1172 1173 /* post the Irecv and Isend of coj */ 1174 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1175 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1176 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 1177 for (proc=0, k=0; proc<size; proc++) { 1178 if (!len_s[proc]) continue; 1179 i = owners_co[proc]; 1180 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1181 k++; 1182 } 1183 1184 /* receives and sends of coj are complete */ 1185 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 1186 for (i=0; i<merge->nrecv; i++) { 1187 PetscMPIInt icompleted; 1188 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1189 } 1190 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1191 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1192 1193 /* send and recv coi */ 1194 /*-------------------*/ 1195 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1196 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1197 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 1198 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1199 for (proc=0,k=0; proc<size; proc++) { 1200 if (!len_s[proc]) continue; 1201 /* form outgoing message for i-structure: 1202 buf_si[0]: nrows to be sent 1203 [1:nrows]: row index (global) 1204 [nrows+1:2*nrows+1]: i-structure index 1205 */ 1206 /*-------------------------------------------*/ 1207 nrows = len_si[proc]/2 - 1; 1208 buf_si_i = buf_si + nrows+1; 1209 buf_si[0] = nrows; 1210 buf_si_i[0] = 0; 1211 nrows = 0; 1212 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1213 nzi = coi[i+1] - coi[i]; 1214 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1215 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1216 nrows++; 1217 } 1218 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1219 k++; 1220 buf_si += len_si[proc]; 1221 } 1222 i = merge->nrecv; 1223 while (i--) { 1224 PetscMPIInt icompleted; 1225 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1226 } 1227 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1228 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1229 ierr = PetscFree(len_si);CHKERRQ(ierr); 1230 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1231 ierr = PetscFree(swaits);CHKERRQ(ierr); 1232 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1233 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1234 1235 /* compute the local portion of C (mpi mat) */ 1236 /*------------------------------------------*/ 1237 /* allocate bi array and free space for accumulating nonzero column info */ 1238 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1239 bi[0] = 0; 1240 1241 /* set initial free space to be fill*(nnz(P) + nnz(A)) */ 1242 nnz = fill*(pdti[pn] + poti[pon] + ai[am]); 1243 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1244 current_space = free_space; 1245 1246 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1247 for (k=0; k<merge->nrecv; k++) { 1248 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1249 nrows = *buf_ri_k[k]; 1250 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1251 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1252 } 1253 1254 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1255 rmax = 0; 1256 for (i=0; i<pn; i++) { 1257 /* add pdt[i,:]*AP into lnk */ 1258 pnz = pdti[i+1] - pdti[i]; 1259 ptJ = pdtj + pdti[i]; 1260 for (j=0; j<pnz; j++) { 1261 row = ptJ[j]; /* row of AP == col of Pt */ 1262 anz = ai[row+1] - ai[row]; 1263 Jptr = aj + ai[row]; 1264 /* add non-zero cols of AP into the sorted linked list lnk */ 1265 ierr = PetscLLCondensedAddSorted(anz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1266 } 1267 1268 /* add received col data into lnk */ 1269 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1270 if (i == *nextrow[k]) { /* i-th row */ 1271 nzi = *(nextci[k]+1) - *nextci[k]; 1272 Jptr = buf_rj[k] + *nextci[k]; 1273 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1274 nextrow[k]++; nextci[k]++; 1275 } 1276 } 1277 nnz = lnk[0]; 1278 1279 /* if free space is not available, make more free space */ 1280 if (current_space->local_remaining<nnz) { 1281 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1282 nspacedouble++; 1283 } 1284 /* copy data into free space, then initialize lnk */ 1285 ierr = PetscLLCondensedClean(aN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1286 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 1287 1288 current_space->array += nnz; 1289 current_space->local_used += nnz; 1290 current_space->local_remaining -= nnz; 1291 1292 bi[i+1] = bi[i] + nnz; 1293 if (nnz > rmax) rmax = nnz; 1294 } 1295 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1296 1297 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1298 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 1299 1300 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 1301 if (afill_tmp > afill) afill = afill_tmp; 1302 ierr = PetscLLCondensedDestroy(lnk,lnkbt);CHKERRQ(ierr); 1303 ierr = MatDestroy(&POt);CHKERRQ(ierr); 1304 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 1305 1306 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 1307 /*----------------------------------------------------------------------------------*/ 1308 ierr = PetscMalloc((rmax+1)*sizeof(PetscScalar),&vals);CHKERRQ(ierr); 1309 ierr = PetscMemzero(vals,rmax*sizeof(PetscScalar));CHKERRQ(ierr); 1310 1311 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1312 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1313 ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,A->cmap->bs);CHKERRQ(ierr); 1314 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1315 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1316 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1317 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 1318 for (i=0; i<pn; i++) { 1319 row = i + rstart; 1320 nnz = bi[i+1] - bi[i]; 1321 Jptr = bj + bi[i]; 1322 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 1323 } 1324 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1325 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1326 ierr = PetscFree(vals);CHKERRQ(ierr); 1327 1328 merge->bi = bi; 1329 merge->bj = bj; 1330 merge->coi = coi; 1331 merge->coj = coj; 1332 merge->buf_ri = buf_ri; 1333 merge->buf_rj = buf_rj; 1334 merge->owners_co = owners_co; 1335 merge->destroy = Cmpi->ops->destroy; 1336 merge->duplicate = Cmpi->ops->duplicate; 1337 1338 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 1339 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1340 1341 /* attach the supporting struct to Cmpi for reuse */ 1342 c = (Mat_MPIAIJ*)Cmpi->data; 1343 c->ptap = ptap; 1344 ptap->api = NULL; 1345 ptap->apj = NULL; 1346 ptap->merge = merge; 1347 ptap->rmax = rmax; 1348 1349 *C = Cmpi; 1350 #if defined(PETSC_USE_INFO) 1351 if (bi[pn] != 0) { 1352 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 1353 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 1354 } else { 1355 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1356 } 1357 #endif 1358 PetscFunctionReturn(0); 1359 } 1360 1361 #undef __FUNCT__ 1362 #define __FUNCT__ "MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_Scalable" 1363 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_Scalable(Mat P,Mat A,Mat C) 1364 { 1365 PetscErrorCode ierr; 1366 Mat_Merge_SeqsToMPI *merge; 1367 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1368 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1369 Mat_PtAPMPI *ptap; 1370 PetscInt *adj; 1371 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1372 MatScalar *ada,*ca,valtmp; 1373 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1374 MPI_Comm comm; 1375 PetscMPIInt size,rank,taga,*len_s; 1376 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1377 PetscInt **buf_ri,**buf_rj; 1378 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1379 MPI_Request *s_waits,*r_waits; 1380 MPI_Status *status; 1381 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1382 PetscInt *ai,*aj,*coi,*coj; 1383 PetscInt *poJ,*pdJ; 1384 Mat A_loc; 1385 Mat_SeqAIJ *a_loc; 1386 1387 PetscFunctionBegin; 1388 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1389 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1390 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1391 1392 ptap = c->ptap; 1393 merge = ptap->merge; 1394 1395 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1396 /*------------------------------------------*/ 1397 /* get data from symbolic products */ 1398 coi = merge->coi; coj = merge->coj; 1399 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 1400 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 1401 bi = merge->bi; bj = merge->bj; 1402 owners = merge->rowmap->range; 1403 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 1404 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 1405 1406 /* get A_loc by taking all local rows of A */ 1407 A_loc = ptap->A_loc; 1408 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 1409 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1410 ai = a_loc->i; 1411 aj = a_loc->j; 1412 1413 for (i=0; i<am; i++) { 1414 anz = ai[i+1] - ai[i]; 1415 adj = aj + ai[i]; 1416 ada = a_loc->a + ai[i]; 1417 1418 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1419 /*-------------------------------------------------------------*/ 1420 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1421 pnz = po->i[i+1] - po->i[i]; 1422 poJ = po->j + po->i[i]; 1423 pA = po->a + po->i[i]; 1424 for (j=0; j<pnz; j++) { 1425 row = poJ[j]; 1426 cj = coj + coi[row]; 1427 ca = coa + coi[row]; 1428 /* perform sparse axpy */ 1429 nexta = 0; 1430 valtmp = pA[j]; 1431 for (k=0; nexta<anz; k++) { 1432 if (cj[k] == adj[nexta]) { 1433 ca[k] += valtmp*ada[nexta]; 1434 nexta++; 1435 } 1436 } 1437 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1438 } 1439 1440 /* put the value into Cd (diagonal part) */ 1441 pnz = pd->i[i+1] - pd->i[i]; 1442 pdJ = pd->j + pd->i[i]; 1443 pA = pd->a + pd->i[i]; 1444 for (j=0; j<pnz; j++) { 1445 row = pdJ[j]; 1446 cj = bj + bi[row]; 1447 ca = ba + bi[row]; 1448 /* perform sparse axpy */ 1449 nexta = 0; 1450 valtmp = pA[j]; 1451 for (k=0; nexta<anz; k++) { 1452 if (cj[k] == adj[nexta]) { 1453 ca[k] += valtmp*ada[nexta]; 1454 nexta++; 1455 } 1456 } 1457 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1458 } 1459 } 1460 1461 /* 3) send and recv matrix values coa */ 1462 /*------------------------------------*/ 1463 buf_ri = merge->buf_ri; 1464 buf_rj = merge->buf_rj; 1465 len_s = merge->len_s; 1466 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1467 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1468 1469 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 1470 for (proc=0,k=0; proc<size; proc++) { 1471 if (!len_s[proc]) continue; 1472 i = merge->owners_co[proc]; 1473 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1474 k++; 1475 } 1476 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1477 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1478 1479 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1480 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1481 ierr = PetscFree(coa);CHKERRQ(ierr); 1482 1483 /* 4) insert local Cseq and received values into Cmpi */ 1484 /*----------------------------------------------------*/ 1485 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1486 for (k=0; k<merge->nrecv; k++) { 1487 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1488 nrows = *(buf_ri_k[k]); 1489 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1490 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1491 } 1492 1493 for (i=0; i<cm; i++) { 1494 row = owners[rank] + i; /* global row index of C_seq */ 1495 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1496 ba_i = ba + bi[i]; 1497 bnz = bi[i+1] - bi[i]; 1498 /* add received vals into ba */ 1499 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1500 /* i-th row */ 1501 if (i == *nextrow[k]) { 1502 cnz = *(nextci[k]+1) - *nextci[k]; 1503 cj = buf_rj[k] + *(nextci[k]); 1504 ca = abuf_r[k] + *(nextci[k]); 1505 nextcj = 0; 1506 for (j=0; nextcj<cnz; j++) { 1507 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1508 ba_i[j] += ca[nextcj++]; 1509 } 1510 } 1511 nextrow[k]++; nextci[k]++; 1512 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1513 } 1514 } 1515 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1516 } 1517 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1518 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1519 1520 ierr = PetscFree(ba);CHKERRQ(ierr); 1521 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1522 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1523 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1524 PetscFunctionReturn(0); 1525 } 1526 1527 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1528 #undef __FUNCT__ 1529 #define __FUNCT__ "MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable" 1530 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(Mat P,Mat A,PetscReal fill,Mat *C) 1531 { 1532 PetscErrorCode ierr; 1533 Mat Cmpi,A_loc,POt,PDt; 1534 Mat_PtAPMPI *ptap; 1535 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1536 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data,*c; 1537 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1538 PetscInt nnz; 1539 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1540 PetscInt am =A->rmap->n,pn=P->cmap->n; 1541 MPI_Comm comm; 1542 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1543 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1544 PetscInt len,proc,*dnz,*onz,*owners; 1545 PetscInt nzi,*bi,*bj; 1546 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1547 MPI_Request *swaits,*rwaits; 1548 MPI_Status *sstatus,rstatus; 1549 Mat_Merge_SeqsToMPI *merge; 1550 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1551 PetscReal afill =1.0,afill_tmp; 1552 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Crmax; 1553 PetscScalar *vals; 1554 Mat_SeqAIJ *a_loc, *pdt,*pot; 1555 1556 PetscFunctionBegin; 1557 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1558 /* check if matrix local sizes are compatible */ 1559 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 1560 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A (%D, %D) != P (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 1561 } 1562 1563 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1564 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1565 1566 /* create struct Mat_PtAPMPI and attached it to C later */ 1567 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 1568 1569 /* get A_loc by taking all local rows of A */ 1570 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1571 1572 ptap->A_loc = A_loc; 1573 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1574 ai = a_loc->i; 1575 aj = a_loc->j; 1576 1577 /* determine symbolic Co=(p->B)^T*A - send to others */ 1578 /*----------------------------------------------------*/ 1579 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1580 pdt = (Mat_SeqAIJ*)PDt->data; 1581 pdti = pdt->i; pdtj = pdt->j; 1582 1583 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1584 pot = (Mat_SeqAIJ*)POt->data; 1585 poti = pot->i; potj = pot->j; 1586 1587 /* then, compute symbolic Co = (p->B)^T*A */ 1588 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1589 >= (num of nonzero rows of C_seq) - pn */ 1590 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 1591 coi[0] = 0; 1592 1593 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1594 nnz = fill*(poti[pon] + ai[am]); 1595 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1596 current_space = free_space; 1597 1598 /* create and initialize a linked list */ 1599 i = PetscMax(pdt->rmax,pot->rmax); 1600 Crmax = i*a_loc->rmax*size; /* non-scalable! */ 1601 if (!Crmax || Crmax > aN) Crmax = aN; 1602 ierr = PetscLLCondensedCreate_Scalable(Crmax,&lnk);CHKERRQ(ierr); 1603 1604 for (i=0; i<pon; i++) { 1605 pnz = poti[i+1] - poti[i]; 1606 ptJ = potj + poti[i]; 1607 for (j=0; j<pnz; j++) { 1608 row = ptJ[j]; /* row of A_loc == col of Pot */ 1609 anz = ai[row+1] - ai[row]; 1610 Jptr = aj + ai[row]; 1611 /* add non-zero cols of AP into the sorted linked list lnk */ 1612 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1613 } 1614 nnz = lnk[0]; 1615 1616 /* If free space is not available, double the total space in the list */ 1617 if (current_space->local_remaining<nnz) { 1618 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1619 nspacedouble++; 1620 } 1621 1622 /* Copy data into free space, and zero out denserows */ 1623 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1624 1625 current_space->array += nnz; 1626 current_space->local_used += nnz; 1627 current_space->local_remaining -= nnz; 1628 1629 coi[i+1] = coi[i] + nnz; 1630 } 1631 1632 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 1633 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1634 1635 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1636 if (afill_tmp > afill) afill = afill_tmp; 1637 1638 /* send j-array (coj) of Co to other processors */ 1639 /*----------------------------------------------*/ 1640 /* determine row ownership */ 1641 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 1642 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1643 1644 merge->rowmap->n = pn; 1645 merge->rowmap->bs = 1; 1646 1647 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1648 owners = merge->rowmap->range; 1649 1650 /* determine the number of messages to send, their lengths */ 1651 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 1652 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1653 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 1654 1655 len_s = merge->len_s; 1656 merge->nsend = 0; 1657 1658 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 1659 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1660 1661 proc = 0; 1662 for (i=0; i<pon; i++) { 1663 while (prmap[i] >= owners[proc+1]) proc++; 1664 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1665 len_s[proc] += coi[i+1] - coi[i]; 1666 } 1667 1668 len = 0; /* max length of buf_si[] */ 1669 owners_co[0] = 0; 1670 for (proc=0; proc<size; proc++) { 1671 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1672 if (len_si[proc]) { 1673 merge->nsend++; 1674 len_si[proc] = 2*(len_si[proc] + 1); 1675 len += len_si[proc]; 1676 } 1677 } 1678 1679 /* determine the number and length of messages to receive for coi and coj */ 1680 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1681 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1682 1683 /* post the Irecv and Isend of coj */ 1684 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1685 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1686 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 1687 for (proc=0, k=0; proc<size; proc++) { 1688 if (!len_s[proc]) continue; 1689 i = owners_co[proc]; 1690 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1691 k++; 1692 } 1693 1694 /* receives and sends of coj are complete */ 1695 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 1696 for (i=0; i<merge->nrecv; i++) { 1697 PetscMPIInt icompleted; 1698 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1699 } 1700 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1701 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1702 1703 /* send and recv coi */ 1704 /*-------------------*/ 1705 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1706 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1707 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 1708 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1709 for (proc=0,k=0; proc<size; proc++) { 1710 if (!len_s[proc]) continue; 1711 /* form outgoing message for i-structure: 1712 buf_si[0]: nrows to be sent 1713 [1:nrows]: row index (global) 1714 [nrows+1:2*nrows+1]: i-structure index 1715 */ 1716 /*-------------------------------------------*/ 1717 nrows = len_si[proc]/2 - 1; 1718 buf_si_i = buf_si + nrows+1; 1719 buf_si[0] = nrows; 1720 buf_si_i[0] = 0; 1721 nrows = 0; 1722 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1723 nzi = coi[i+1] - coi[i]; 1724 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1725 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1726 nrows++; 1727 } 1728 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1729 k++; 1730 buf_si += len_si[proc]; 1731 } 1732 i = merge->nrecv; 1733 while (i--) { 1734 PetscMPIInt icompleted; 1735 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1736 } 1737 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1738 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1739 ierr = PetscFree(len_si);CHKERRQ(ierr); 1740 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1741 ierr = PetscFree(swaits);CHKERRQ(ierr); 1742 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1743 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1744 1745 /* compute the local portion of C (mpi mat) */ 1746 /*------------------------------------------*/ 1747 /* allocate bi array and free space for accumulating nonzero column info */ 1748 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1749 bi[0] = 0; 1750 1751 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1752 nnz = fill*(pdti[pn] + poti[pon] + ai[am]); 1753 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1754 current_space = free_space; 1755 1756 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1757 for (k=0; k<merge->nrecv; k++) { 1758 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1759 nrows = *buf_ri_k[k]; 1760 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1761 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recieved i-structure */ 1762 } 1763 1764 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1765 rmax = 0; 1766 for (i=0; i<pn; i++) { 1767 /* add pdt[i,:]*AP into lnk */ 1768 pnz = pdti[i+1] - pdti[i]; 1769 ptJ = pdtj + pdti[i]; 1770 for (j=0; j<pnz; j++) { 1771 row = ptJ[j]; /* row of AP == col of Pt */ 1772 anz = ai[row+1] - ai[row]; 1773 Jptr = aj + ai[row]; 1774 /* add non-zero cols of AP into the sorted linked list lnk */ 1775 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1776 } 1777 1778 /* add received col data into lnk */ 1779 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1780 if (i == *nextrow[k]) { /* i-th row */ 1781 nzi = *(nextci[k]+1) - *nextci[k]; 1782 Jptr = buf_rj[k] + *nextci[k]; 1783 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 1784 nextrow[k]++; nextci[k]++; 1785 } 1786 } 1787 nnz = lnk[0]; 1788 1789 /* if free space is not available, make more free space */ 1790 if (current_space->local_remaining<nnz) { 1791 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1792 nspacedouble++; 1793 } 1794 /* copy data into free space, then initialize lnk */ 1795 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1796 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 1797 1798 current_space->array += nnz; 1799 current_space->local_used += nnz; 1800 current_space->local_remaining -= nnz; 1801 1802 bi[i+1] = bi[i] + nnz; 1803 if (nnz > rmax) rmax = nnz; 1804 } 1805 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1806 1807 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1808 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 1809 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 1810 if (afill_tmp > afill) afill = afill_tmp; 1811 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 1812 ierr = MatDestroy(&POt);CHKERRQ(ierr); 1813 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 1814 1815 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 1816 /*----------------------------------------------------------------------------------*/ 1817 ierr = PetscMalloc((rmax+1)*sizeof(PetscScalar),&vals);CHKERRQ(ierr); 1818 ierr = PetscMemzero(vals,rmax*sizeof(PetscScalar));CHKERRQ(ierr); 1819 1820 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1821 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1822 ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,A->cmap->bs);CHKERRQ(ierr); 1823 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1824 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1825 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1826 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 1827 for (i=0; i<pn; i++) { 1828 row = i + rstart; 1829 nnz = bi[i+1] - bi[i]; 1830 Jptr = bj + bi[i]; 1831 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 1832 } 1833 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1834 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1835 ierr = PetscFree(vals);CHKERRQ(ierr); 1836 1837 merge->bi = bi; 1838 merge->bj = bj; 1839 merge->coi = coi; 1840 merge->coj = coj; 1841 merge->buf_ri = buf_ri; 1842 merge->buf_rj = buf_rj; 1843 merge->owners_co = owners_co; 1844 merge->destroy = Cmpi->ops->destroy; 1845 merge->duplicate = Cmpi->ops->duplicate; 1846 1847 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_Scalable; 1848 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1849 1850 /* attach the supporting struct to Cmpi for reuse */ 1851 c = (Mat_MPIAIJ*)Cmpi->data; 1852 1853 c->ptap = ptap; 1854 ptap->api = NULL; 1855 ptap->apj = NULL; 1856 ptap->merge = merge; 1857 ptap->rmax = rmax; 1858 ptap->apa = NULL; 1859 1860 *C = Cmpi; 1861 #if defined(PETSC_USE_INFO) 1862 if (bi[pn] != 0) { 1863 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 1864 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 1865 } else { 1866 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1867 } 1868 #endif 1869 PetscFunctionReturn(0); 1870 } 1871