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 #if defined(PETSC_HAVE_HYPRE) 14 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat*); 15 #endif 16 17 PETSC_INTERN PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill, Mat *C) 18 { 19 PetscErrorCode ierr; 20 #if defined(PETSC_HAVE_HYPRE) 21 const char *algTypes[4] = {"scalable","nonscalable","seqmpi","hypre"}; 22 PetscInt nalg = 4; 23 #else 24 const char *algTypes[3] = {"scalable","nonscalable","seqmpi"}; 25 PetscInt nalg = 3; 26 #endif 27 PetscInt alg = 1; /* set nonscalable algorithm as default */ 28 MPI_Comm comm; 29 PetscBool flg; 30 31 PetscFunctionBegin; 32 if (scall == MAT_INITIAL_MATRIX) { 33 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 34 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 35 36 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 37 PetscOptionsObject->alreadyprinted = PETSC_FALSE; /* a hack to ensure the option shows in '-help' */ 38 ierr = PetscOptionsEList("-matmatmult_via","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[1],&alg,&flg);CHKERRQ(ierr); 39 ierr = PetscOptionsEnd();CHKERRQ(ierr); 40 41 if (!flg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 42 MatInfo Ainfo,Binfo; 43 PetscInt nz_local; 44 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 45 46 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 47 ierr = MatGetInfo(B,MAT_LOCAL,&Binfo);CHKERRQ(ierr); 48 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 49 50 if (B->cmap->N > fill*nz_local) alg_scalable_loc = PETSC_TRUE; 51 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 52 53 if (alg_scalable) { 54 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 55 ierr = PetscInfo2(B,"Use scalable algorithm, BN %D, fill*nz_allocated %g\n",B->cmap->N,fill*nz_local);CHKERRQ(ierr); 56 } 57 } 58 59 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 60 switch (alg) { 61 case 1: 62 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C);CHKERRQ(ierr); 63 break; 64 case 2: 65 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(A,B,fill,C);CHKERRQ(ierr); 66 break; 67 #if defined(PETSC_HAVE_HYPRE) 68 case 3: 69 ierr = MatMatMultSymbolic_AIJ_AIJ_wHYPRE(A,B,fill,C);CHKERRQ(ierr); 70 break; 71 #endif 72 default: 73 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 74 break; 75 } 76 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 77 } 78 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 79 ierr = (*(*C)->ops->matmultnumeric)(A,B,*C);CHKERRQ(ierr); 80 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 81 PetscFunctionReturn(0); 82 } 83 84 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat A) 85 { 86 PetscErrorCode ierr; 87 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 88 Mat_PtAPMPI *ptap = a->ptap; 89 90 PetscFunctionBegin; 91 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 92 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 93 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 94 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 95 ierr = MatDestroy(&ptap->Pt);CHKERRQ(ierr); 96 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 97 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 98 ierr = PetscFree(ptap->apa);CHKERRQ(ierr); 99 ierr = ptap->destroy(A);CHKERRQ(ierr); 100 ierr = PetscFree(ptap);CHKERRQ(ierr); 101 PetscFunctionReturn(0); 102 } 103 104 PetscErrorCode MatDuplicate_MPIAIJ_MatMatMult(Mat A, MatDuplicateOption op, Mat *M) 105 { 106 PetscErrorCode ierr; 107 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 108 Mat_PtAPMPI *ptap = a->ptap; 109 110 PetscFunctionBegin; 111 ierr = (*ptap->duplicate)(A,op,M);CHKERRQ(ierr); 112 113 (*M)->ops->destroy = ptap->destroy; /* = MatDestroy_MPIAIJ, *M doesn't duplicate A's special structure! */ 114 (*M)->ops->duplicate = ptap->duplicate; /* = MatDuplicate_MPIAIJ */ 115 PetscFunctionReturn(0); 116 } 117 118 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,Mat C) 119 { 120 PetscErrorCode ierr; 121 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 122 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 123 Mat_SeqAIJ *cd =(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 124 PetscScalar *cda=cd->a,*coa=co->a; 125 Mat_SeqAIJ *p_loc,*p_oth; 126 PetscScalar *apa,*ca; 127 PetscInt cm =C->rmap->n; 128 Mat_PtAPMPI *ptap=c->ptap; 129 PetscInt *api,*apj,*apJ,i,k; 130 PetscInt cstart=C->cmap->rstart; 131 PetscInt cdnz,conz,k0,k1; 132 MPI_Comm comm; 133 PetscMPIInt size; 134 135 PetscFunctionBegin; 136 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 137 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 138 139 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 140 /*-----------------------------------------------------*/ 141 /* update numerical values of P_oth and P_loc */ 142 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 143 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 144 145 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 146 /*----------------------------------------------------------*/ 147 /* get data from symbolic products */ 148 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 149 p_oth = NULL; 150 if (size >1) { 151 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 152 } 153 154 /* get apa for storing dense row A[i,:]*P */ 155 apa = ptap->apa; 156 157 api = ptap->api; 158 apj = ptap->apj; 159 for (i=0; i<cm; i++) { 160 /* compute apa = A[i,:]*P */ 161 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 162 163 /* set values in C */ 164 apJ = apj + api[i]; 165 cdnz = cd->i[i+1] - cd->i[i]; 166 conz = co->i[i+1] - co->i[i]; 167 168 /* 1st off-diagoanl part of C */ 169 ca = coa + co->i[i]; 170 k = 0; 171 for (k0=0; k0<conz; k0++) { 172 if (apJ[k] >= cstart) break; 173 ca[k0] = apa[apJ[k]]; 174 apa[apJ[k++]] = 0.0; 175 } 176 177 /* diagonal part of C */ 178 ca = cda + cd->i[i]; 179 for (k1=0; k1<cdnz; k1++) { 180 ca[k1] = apa[apJ[k]]; 181 apa[apJ[k++]] = 0.0; 182 } 183 184 /* 2nd off-diagoanl part of C */ 185 ca = coa + co->i[i]; 186 for (; k0<conz; k0++) { 187 ca[k0] = apa[apJ[k]]; 188 apa[apJ[k++]] = 0.0; 189 } 190 } 191 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 192 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 193 PetscFunctionReturn(0); 194 } 195 196 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,PetscReal fill,Mat *C) 197 { 198 PetscErrorCode ierr; 199 MPI_Comm comm; 200 PetscMPIInt size; 201 Mat Cmpi; 202 Mat_PtAPMPI *ptap; 203 PetscFreeSpaceList free_space=NULL,current_space=NULL; 204 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c; 205 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 206 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 207 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 208 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 209 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 210 PetscBT lnkbt; 211 PetscScalar *apa; 212 PetscReal afill; 213 214 PetscFunctionBegin; 215 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 216 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 217 218 /* create struct Mat_PtAPMPI and attached it to C later */ 219 ierr = PetscNew(&ptap);CHKERRQ(ierr); 220 221 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 222 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 223 224 /* get P_loc by taking all local rows of P */ 225 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 226 227 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 228 pi_loc = p_loc->i; pj_loc = p_loc->j; 229 if (size > 1) { 230 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 231 pi_oth = p_oth->i; pj_oth = p_oth->j; 232 } else { 233 p_oth = NULL; 234 pi_oth = NULL; pj_oth = NULL; 235 } 236 237 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 238 /*-------------------------------------------------------------------*/ 239 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 240 ptap->api = api; 241 api[0] = 0; 242 243 /* create and initialize a linked list */ 244 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 245 246 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 247 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 248 current_space = free_space; 249 250 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 251 for (i=0; i<am; i++) { 252 /* diagonal portion of A */ 253 nzi = adi[i+1] - adi[i]; 254 for (j=0; j<nzi; j++) { 255 row = *adj++; 256 pnz = pi_loc[row+1] - pi_loc[row]; 257 Jptr = pj_loc + pi_loc[row]; 258 /* add non-zero cols of P into the sorted linked list lnk */ 259 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 260 } 261 /* off-diagonal portion of A */ 262 nzi = aoi[i+1] - aoi[i]; 263 for (j=0; j<nzi; j++) { 264 row = *aoj++; 265 pnz = pi_oth[row+1] - pi_oth[row]; 266 Jptr = pj_oth + pi_oth[row]; 267 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 268 } 269 270 apnz = lnk[0]; 271 api[i+1] = api[i] + apnz; 272 273 /* if free space is not available, double the total space in the list */ 274 if (current_space->local_remaining<apnz) { 275 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 276 nspacedouble++; 277 } 278 279 /* Copy data into free space, then initialize lnk */ 280 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 281 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 282 283 current_space->array += apnz; 284 current_space->local_used += apnz; 285 current_space->local_remaining -= apnz; 286 } 287 288 /* Allocate space for apj, initialize apj, and */ 289 /* destroy list of free space and other temporary array(s) */ 290 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 291 apj = ptap->apj; 292 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 293 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 294 295 /* malloc apa to store dense row A[i,:]*P */ 296 ierr = PetscCalloc1(pN,&apa);CHKERRQ(ierr); 297 298 ptap->apa = apa; 299 300 /* create and assemble symbolic parallel matrix Cmpi */ 301 /*----------------------------------------------------*/ 302 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 303 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 304 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 305 306 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 307 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 308 309 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Cmpi, apj, api);CHKERRQ(ierr); 310 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 311 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 312 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 313 314 ptap->destroy = Cmpi->ops->destroy; 315 ptap->duplicate = Cmpi->ops->duplicate; 316 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 317 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 318 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 319 320 /* attach the supporting struct to Cmpi for reuse */ 321 c = (Mat_MPIAIJ*)Cmpi->data; 322 c->ptap = ptap; 323 324 *C = Cmpi; 325 326 /* set MatInfo */ 327 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 328 if (afill < 1.0) afill = 1.0; 329 Cmpi->info.mallocs = nspacedouble; 330 Cmpi->info.fill_ratio_given = fill; 331 Cmpi->info.fill_ratio_needed = afill; 332 333 #if defined(PETSC_USE_INFO) 334 if (api[am]) { 335 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 336 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 337 } else { 338 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 339 } 340 #endif 341 PetscFunctionReturn(0); 342 } 343 344 PETSC_INTERN PetscErrorCode MatMatMult_MPIAIJ_MPIDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 345 { 346 PetscErrorCode ierr; 347 348 PetscFunctionBegin; 349 if (scall == MAT_INITIAL_MATRIX) { 350 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 351 ierr = MatMatMultSymbolic_MPIAIJ_MPIDense(A,B,fill,C);CHKERRQ(ierr); 352 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 353 } 354 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 355 ierr = MatMatMultNumeric_MPIAIJ_MPIDense(A,B,*C);CHKERRQ(ierr); 356 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 357 PetscFunctionReturn(0); 358 } 359 360 typedef struct { 361 Mat workB; 362 PetscScalar *rvalues,*svalues; 363 MPI_Request *rwaits,*swaits; 364 } MPIAIJ_MPIDense; 365 366 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 367 { 368 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*) ctx; 369 PetscErrorCode ierr; 370 371 PetscFunctionBegin; 372 ierr = MatDestroy(&contents->workB);CHKERRQ(ierr); 373 ierr = PetscFree4(contents->rvalues,contents->svalues,contents->rwaits,contents->swaits);CHKERRQ(ierr); 374 ierr = PetscFree(contents);CHKERRQ(ierr); 375 PetscFunctionReturn(0); 376 } 377 378 #include <petsc/private/vecscatterimpl.h> 379 /* 380 This is a "dummy function" that handles the case where matrix C was created as a dense matrix 381 directly by the user and passed to MatMatMult() with the MAT_REUSE_MATRIX option 382 383 It is the same as MatMatMultSymbolic_MPIAIJ_MPIDense() except does not create C 384 385 Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product 386 for this matrix. This is not desirable.. 387 388 */ 389 PetscErrorCode MatMatMultNumeric_MPIDense(Mat A,Mat B,Mat C) 390 { 391 PetscErrorCode ierr; 392 PetscBool flg; 393 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 394 PetscInt nz = aij->B->cmap->n; 395 PetscContainer container; 396 MPIAIJ_MPIDense *contents; 397 VecScatter ctx = aij->Mvctx; 398 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 399 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 400 401 PetscFunctionBegin; 402 ierr = PetscObjectTypeCompare((PetscObject)B,MATMPIDENSE,&flg);CHKERRQ(ierr); 403 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Second matrix must be mpidense"); 404 405 /* Handle case where where user provided the final C matrix rather than calling MatMatMult() with MAT_INITIAL_MATRIX*/ 406 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 407 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"First matrix must be MPIAIJ"); 408 409 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 410 411 ierr = PetscNew(&contents);CHKERRQ(ierr); 412 /* Create work matrix used to store off processor rows of B needed for local product */ 413 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,NULL,&contents->workB);CHKERRQ(ierr); 414 /* Create work arrays needed */ 415 ierr = PetscMalloc4(B->cmap->N*from->starts[from->n],&contents->rvalues, 416 B->cmap->N*to->starts[to->n],&contents->svalues, 417 from->n,&contents->rwaits, 418 to->n,&contents->swaits);CHKERRQ(ierr); 419 420 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);CHKERRQ(ierr); 421 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 422 ierr = PetscContainerSetUserDestroy(container,MatMPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 423 ierr = PetscObjectCompose((PetscObject)C,"workB",(PetscObject)container);CHKERRQ(ierr); 424 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 425 426 ierr = (*C->ops->matmultnumeric)(A,B,C);CHKERRQ(ierr); 427 PetscFunctionReturn(0); 428 } 429 430 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C) 431 { 432 PetscErrorCode ierr; 433 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 434 PetscInt nz = aij->B->cmap->n; 435 PetscContainer container; 436 MPIAIJ_MPIDense *contents; 437 VecScatter ctx = aij->Mvctx; 438 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 439 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 440 PetscInt m = A->rmap->n,n=B->cmap->n; 441 442 PetscFunctionBegin; 443 ierr = MatCreate(PetscObjectComm((PetscObject)B),C);CHKERRQ(ierr); 444 ierr = MatSetSizes(*C,m,n,A->rmap->N,B->cmap->N);CHKERRQ(ierr); 445 ierr = MatSetBlockSizesFromMats(*C,A,B);CHKERRQ(ierr); 446 ierr = MatSetType(*C,MATMPIDENSE);CHKERRQ(ierr); 447 ierr = MatMPIDenseSetPreallocation(*C,NULL);CHKERRQ(ierr); 448 ierr = MatAssemblyBegin(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 449 ierr = MatAssemblyEnd(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 450 451 (*C)->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 452 453 ierr = PetscNew(&contents);CHKERRQ(ierr); 454 /* Create work matrix used to store off processor rows of B needed for local product */ 455 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,NULL,&contents->workB);CHKERRQ(ierr); 456 /* Create work arrays needed */ 457 ierr = PetscMalloc4(B->cmap->N*from->starts[from->n],&contents->rvalues, 458 B->cmap->N*to->starts[to->n],&contents->svalues, 459 from->n,&contents->rwaits, 460 to->n,&contents->swaits);CHKERRQ(ierr); 461 462 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);CHKERRQ(ierr); 463 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 464 ierr = PetscContainerSetUserDestroy(container,MatMPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 465 ierr = PetscObjectCompose((PetscObject)(*C),"workB",(PetscObject)container);CHKERRQ(ierr); 466 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 467 PetscFunctionReturn(0); 468 } 469 470 /* 471 Performs an efficient scatter on the rows of B needed by this process; this is 472 a modification of the VecScatterBegin_() routines. 473 */ 474 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,Mat C,Mat *outworkB) 475 { 476 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 477 PetscErrorCode ierr; 478 PetscScalar *b,*w,*svalues,*rvalues; 479 VecScatter ctx = aij->Mvctx; 480 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 481 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 482 PetscInt i,j,k; 483 PetscInt *sindices,*sstarts,*rindices,*rstarts; 484 PetscMPIInt *sprocs,*rprocs,nrecvs; 485 MPI_Request *swaits,*rwaits; 486 MPI_Comm comm; 487 PetscMPIInt tag = ((PetscObject)ctx)->tag,ncols = B->cmap->N, nrows = aij->B->cmap->n,imdex,nrowsB = B->rmap->n; 488 MPI_Status status; 489 MPIAIJ_MPIDense *contents; 490 PetscContainer container; 491 Mat workB; 492 493 PetscFunctionBegin; 494 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 495 ierr = PetscObjectQuery((PetscObject)C,"workB",(PetscObject*)&container);CHKERRQ(ierr); 496 if (!container) SETERRQ(comm,PETSC_ERR_PLIB,"Container does not exist"); 497 ierr = PetscContainerGetPointer(container,(void**)&contents);CHKERRQ(ierr); 498 499 workB = *outworkB = contents->workB; 500 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); 501 sindices = to->indices; 502 sstarts = to->starts; 503 sprocs = to->procs; 504 swaits = contents->swaits; 505 svalues = contents->svalues; 506 507 rindices = from->indices; 508 rstarts = from->starts; 509 rprocs = from->procs; 510 rwaits = contents->rwaits; 511 rvalues = contents->rvalues; 512 513 ierr = MatDenseGetArray(B,&b);CHKERRQ(ierr); 514 ierr = MatDenseGetArray(workB,&w);CHKERRQ(ierr); 515 516 for (i=0; i<from->n; i++) { 517 ierr = MPI_Irecv(rvalues+ncols*rstarts[i],ncols*(rstarts[i+1]-rstarts[i]),MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 518 } 519 520 for (i=0; i<to->n; i++) { 521 /* pack a message at a time */ 522 for (j=0; j<sstarts[i+1]-sstarts[i]; j++) { 523 for (k=0; k<ncols; k++) { 524 svalues[ncols*(sstarts[i] + j) + k] = b[sindices[sstarts[i]+j] + nrowsB*k]; 525 } 526 } 527 ierr = MPI_Isend(svalues+ncols*sstarts[i],ncols*(sstarts[i+1]-sstarts[i]),MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 528 } 529 530 nrecvs = from->n; 531 while (nrecvs) { 532 ierr = MPI_Waitany(from->n,rwaits,&imdex,&status);CHKERRQ(ierr); 533 nrecvs--; 534 /* unpack a message at a time */ 535 for (j=0; j<rstarts[imdex+1]-rstarts[imdex]; j++) { 536 for (k=0; k<ncols; k++) { 537 w[rindices[rstarts[imdex]+j] + nrows*k] = rvalues[ncols*(rstarts[imdex] + j) + k]; 538 } 539 } 540 } 541 if (to->n) {ierr = MPI_Waitall(to->n,swaits,to->sstatus);CHKERRQ(ierr);} 542 543 ierr = MatDenseRestoreArray(B,&b);CHKERRQ(ierr); 544 ierr = MatDenseRestoreArray(workB,&w);CHKERRQ(ierr); 545 ierr = MatAssemblyBegin(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 546 ierr = MatAssemblyEnd(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 547 PetscFunctionReturn(0); 548 } 549 extern PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat); 550 551 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 552 { 553 PetscErrorCode ierr; 554 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 555 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 556 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 557 Mat workB; 558 559 PetscFunctionBegin; 560 /* diagonal block of A times all local rows of B*/ 561 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(aij->A,bdense->A,cdense->A);CHKERRQ(ierr); 562 563 /* get off processor parts of B needed to complete the product */ 564 ierr = MatMPIDenseScatter(A,B,C,&workB);CHKERRQ(ierr); 565 566 /* off-diagonal block of A times nonlocal rows of B */ 567 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A);CHKERRQ(ierr); 568 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 569 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 570 PetscFunctionReturn(0); 571 } 572 573 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 574 { 575 PetscErrorCode ierr; 576 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 577 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 578 Mat_SeqAIJ *cd = (Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 579 PetscInt *adi = ad->i,*adj,*aoi=ao->i,*aoj; 580 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 581 Mat_SeqAIJ *p_loc,*p_oth; 582 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 583 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 584 PetscInt cm = C->rmap->n,anz,pnz; 585 Mat_PtAPMPI *ptap = c->ptap; 586 PetscScalar *apa_sparse = ptap->apa; 587 PetscInt *api,*apj,*apJ,i,j,k,row; 588 PetscInt cstart = C->cmap->rstart; 589 PetscInt cdnz,conz,k0,k1,nextp; 590 MPI_Comm comm; 591 PetscMPIInt size; 592 593 PetscFunctionBegin; 594 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 595 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 596 597 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 598 /*-----------------------------------------------------*/ 599 /* update numerical values of P_oth and P_loc */ 600 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 601 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 602 603 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 604 /*----------------------------------------------------------*/ 605 /* get data from symbolic products */ 606 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 607 pi_loc = p_loc->i; pj_loc = p_loc->j; pa_loc = p_loc->a; 608 if (size >1) { 609 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 610 pi_oth = p_oth->i; pj_oth = p_oth->j; pa_oth = p_oth->a; 611 } else { 612 p_oth = NULL; pi_oth = NULL; pj_oth = NULL; pa_oth = NULL; 613 } 614 615 api = ptap->api; 616 apj = ptap->apj; 617 for (i=0; i<cm; i++) { 618 apJ = apj + api[i]; 619 620 /* diagonal portion of A */ 621 anz = adi[i+1] - adi[i]; 622 adj = ad->j + adi[i]; 623 ada = ad->a + adi[i]; 624 for (j=0; j<anz; j++) { 625 row = adj[j]; 626 pnz = pi_loc[row+1] - pi_loc[row]; 627 pj = pj_loc + pi_loc[row]; 628 pa = pa_loc + pi_loc[row]; 629 /* perform sparse axpy */ 630 valtmp = ada[j]; 631 nextp = 0; 632 for (k=0; nextp<pnz; k++) { 633 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 634 apa_sparse[k] += valtmp*pa[nextp++]; 635 } 636 } 637 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 638 } 639 640 /* off-diagonal portion of A */ 641 anz = aoi[i+1] - aoi[i]; 642 aoj = ao->j + aoi[i]; 643 aoa = ao->a + aoi[i]; 644 for (j=0; j<anz; j++) { 645 row = aoj[j]; 646 pnz = pi_oth[row+1] - pi_oth[row]; 647 pj = pj_oth + pi_oth[row]; 648 pa = pa_oth + pi_oth[row]; 649 /* perform sparse axpy */ 650 valtmp = aoa[j]; 651 nextp = 0; 652 for (k=0; nextp<pnz; k++) { 653 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 654 apa_sparse[k] += valtmp*pa[nextp++]; 655 } 656 } 657 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 658 } 659 660 /* set values in C */ 661 cdnz = cd->i[i+1] - cd->i[i]; 662 conz = co->i[i+1] - co->i[i]; 663 664 /* 1st off-diagoanl part of C */ 665 ca = coa + co->i[i]; 666 k = 0; 667 for (k0=0; k0<conz; k0++) { 668 if (apJ[k] >= cstart) break; 669 ca[k0] = apa_sparse[k]; 670 apa_sparse[k] = 0.0; 671 k++; 672 } 673 674 /* diagonal part of C */ 675 ca = cda + cd->i[i]; 676 for (k1=0; k1<cdnz; k1++) { 677 ca[k1] = apa_sparse[k]; 678 apa_sparse[k] = 0.0; 679 k++; 680 } 681 682 /* 2nd off-diagoanl part of C */ 683 ca = coa + co->i[i]; 684 for (; k0<conz; k0++) { 685 ca[k0] = apa_sparse[k]; 686 apa_sparse[k] = 0.0; 687 k++; 688 } 689 } 690 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 691 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 692 PetscFunctionReturn(0); 693 } 694 695 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 696 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 697 { 698 PetscErrorCode ierr; 699 MPI_Comm comm; 700 PetscMPIInt size; 701 Mat Cmpi; 702 Mat_PtAPMPI *ptap; 703 PetscFreeSpaceList free_space = NULL,current_space=NULL; 704 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c; 705 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 706 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 707 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 708 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=0; 709 PetscInt am=A->rmap->n,pn=P->cmap->n,pm=P->rmap->n,lsize=pn+20; 710 PetscReal afill; 711 PetscScalar *apa; 712 713 PetscFunctionBegin; 714 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 715 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 716 717 /* create struct Mat_PtAPMPI and attached it to C later */ 718 ierr = PetscNew(&ptap);CHKERRQ(ierr); 719 720 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 721 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 722 723 /* get P_loc by taking all local rows of P */ 724 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 725 726 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 727 pi_loc = p_loc->i; pj_loc = p_loc->j; 728 if (size > 1) { 729 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 730 pi_oth = p_oth->i; pj_oth = p_oth->j; 731 } else { 732 p_oth = NULL; 733 pi_oth = NULL; pj_oth = NULL; 734 } 735 736 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 737 /*-------------------------------------------------------------------*/ 738 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 739 ptap->api = api; 740 api[0] = 0; 741 742 ierr = PetscLLCondensedCreate_Scalable(lsize,&lnk);CHKERRQ(ierr); 743 744 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 745 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 746 current_space = free_space; 747 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 748 for (i=0; i<am; i++) { 749 /* diagonal portion of A */ 750 nzi = adi[i+1] - adi[i]; 751 for (j=0; j<nzi; j++) { 752 row = *adj++; 753 pnz = pi_loc[row+1] - pi_loc[row]; 754 Jptr = pj_loc + pi_loc[row]; 755 /* Expand list if it is not long enough */ 756 if (pnz+apnz_max > lsize) { 757 lsize = pnz+apnz_max; 758 ierr = PetscLLCondensedExpand_Scalable(lsize, &lnk);CHKERRQ(ierr); 759 } 760 /* add non-zero cols of P into the sorted linked list lnk */ 761 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 762 apnz = *lnk; /* The first element in the list is the number of items in the list */ 763 api[i+1] = api[i] + apnz; 764 if (apnz > apnz_max) apnz_max = apnz; 765 } 766 /* off-diagonal portion of A */ 767 nzi = aoi[i+1] - aoi[i]; 768 for (j=0; j<nzi; j++) { 769 row = *aoj++; 770 pnz = pi_oth[row+1] - pi_oth[row]; 771 Jptr = pj_oth + pi_oth[row]; 772 /* Expand list if it is not long enough */ 773 if (pnz+apnz_max > lsize) { 774 lsize = pnz + apnz_max; 775 ierr = PetscLLCondensedExpand_Scalable(lsize, &lnk);CHKERRQ(ierr); 776 } 777 /* add non-zero cols of P into the sorted linked list lnk */ 778 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 779 apnz = *lnk; /* The first element in the list is the number of items in the list */ 780 api[i+1] = api[i] + apnz; 781 if (apnz > apnz_max) apnz_max = apnz; 782 } 783 apnz = *lnk; 784 api[i+1] = api[i] + apnz; 785 if (apnz > apnz_max) apnz_max = apnz; 786 787 /* if free space is not available, double the total space in the list */ 788 if (current_space->local_remaining<apnz) { 789 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 790 nspacedouble++; 791 } 792 793 /* Copy data into free space, then initialize lnk */ 794 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 795 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 796 797 current_space->array += apnz; 798 current_space->local_used += apnz; 799 current_space->local_remaining -= apnz; 800 } 801 802 /* Allocate space for apj, initialize apj, and */ 803 /* destroy list of free space and other temporary array(s) */ 804 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 805 apj = ptap->apj; 806 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 807 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 808 809 /* create and assemble symbolic parallel matrix Cmpi */ 810 /*----------------------------------------------------*/ 811 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 812 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 813 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 814 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 815 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 816 817 /* malloc apa for assembly Cmpi */ 818 ierr = PetscCalloc1(apnz_max,&apa);CHKERRQ(ierr); 819 ptap->apa = apa; 820 821 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Cmpi, apj, api);CHKERRQ(ierr); 822 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 823 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 824 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 825 826 ptap->destroy = Cmpi->ops->destroy; 827 ptap->duplicate = Cmpi->ops->duplicate; 828 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 829 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 830 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 831 832 /* attach the supporting struct to Cmpi for reuse */ 833 c = (Mat_MPIAIJ*)Cmpi->data; 834 c->ptap = ptap; 835 *C = Cmpi; 836 837 /* set MatInfo */ 838 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 839 if (afill < 1.0) afill = 1.0; 840 Cmpi->info.mallocs = nspacedouble; 841 Cmpi->info.fill_ratio_given = fill; 842 Cmpi->info.fill_ratio_needed = afill; 843 844 #if defined(PETSC_USE_INFO) 845 if (api[am]) { 846 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 847 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 848 } else { 849 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 850 } 851 #endif 852 PetscFunctionReturn(0); 853 } 854 855 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 856 /* Three input arrays are merged to one output array. The size of the */ 857 /* output array is also output. Duplicate entries only show up once. */ 858 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, 859 PetscInt size2, PetscInt *in2, 860 PetscInt size3, PetscInt *in3, 861 PetscInt *size4, PetscInt *out) 862 { 863 int i = 0, j = 0, k = 0, l = 0; 864 865 /* Traverse all three arrays */ 866 while (i<size1 && j<size2 && k<size3) { 867 if (in1[i] < in2[j] && in1[i] < in3[k]) { 868 out[l++] = in1[i++]; 869 } 870 else if(in2[j] < in1[i] && in2[j] < in3[k]) { 871 out[l++] = in2[j++]; 872 } 873 else if(in3[k] < in1[i] && in3[k] < in2[j]) { 874 out[l++] = in3[k++]; 875 } 876 else if(in1[i] == in2[j] && in1[i] < in3[k]) { 877 out[l++] = in1[i]; 878 i++, j++; 879 } 880 else if(in1[i] == in3[k] && in1[i] < in2[j]) { 881 out[l++] = in1[i]; 882 i++, k++; 883 } 884 else if(in3[k] == in2[j] && in2[j] < in1[i]) { 885 out[l++] = in2[j]; 886 k++, j++; 887 } 888 else if(in1[i] == in2[j] && in1[i] == in3[k]) { 889 out[l++] = in1[i]; 890 i++, j++, k++; 891 } 892 } 893 894 /* Traverse two remaining arrays */ 895 while (i<size1 && j<size2) { 896 if (in1[i] < in2[j]) { 897 out[l++] = in1[i++]; 898 } 899 else if(in1[i] > in2[j]) { 900 out[l++] = in2[j++]; 901 } 902 else { 903 out[l++] = in1[i]; 904 i++, j++; 905 } 906 } 907 908 while (i<size1 && k<size3) { 909 if (in1[i] < in3[k]) { 910 out[l++] = in1[i++]; 911 } 912 else if(in1[i] > in3[k]) { 913 out[l++] = in3[k++]; 914 } 915 else { 916 out[l++] = in1[i]; 917 i++, k++; 918 } 919 } 920 921 while (k<size3 && j<size2) { 922 if (in3[k] < in2[j]) { 923 out[l++] = in3[k++]; 924 } 925 else if(in3[k] > in2[j]) { 926 out[l++] = in2[j++]; 927 } 928 else { 929 out[l++] = in3[k]; 930 k++, j++; 931 } 932 } 933 934 /* Traverse one remaining array */ 935 while (i<size1) out[l++] = in1[i++]; 936 while (j<size2) out[l++] = in2[j++]; 937 while (k<size3) out[l++] = in3[k++]; 938 939 *size4 = l; 940 } 941 942 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 943 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 944 /* matrix-matrix multiplications. */ 945 #undef __FUNCT__ 946 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI" 947 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat *C) 948 { 949 PetscErrorCode ierr; 950 MPI_Comm comm; 951 PetscMPIInt size; 952 Mat Cmpi; 953 Mat_PtAPMPI *ptap; 954 PetscFreeSpaceList free_space_diag=NULL, current_space=NULL; 955 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 956 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc; 957 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 958 Mat_MPIAIJ *c; 959 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 960 PetscInt adponz, adpdnz; 961 PetscInt *pi_loc,*dnz,*onz; 962 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,rstart=A->rmap->rstart; 963 PetscInt *lnk,i, i1=0,pnz,row,*adpoi,*adpoj, *api, *adpoJ, *aopJ, *apJ,*Jptr, aopnz, nspacedouble=0,j,nzi, 964 *apj,apnz, *adpdi, *adpdj, *adpdJ, *poff_i, *poff_j, *j_temp, *aopothi, *aopothj; 965 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n, p_colstart, p_colend; 966 PetscBT lnkbt; 967 PetscScalar *apa; 968 PetscReal afill; 969 PetscMPIInt rank; 970 Mat adpd, aopoth; 971 972 PetscFunctionBegin; 973 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 974 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 975 ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr); 976 ierr = MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend); CHKERRQ(ierr); 977 978 /* create struct Mat_PtAPMPI and attached it to C later */ 979 ierr = PetscNew(&ptap);CHKERRQ(ierr); 980 981 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 982 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 983 984 /* get P_loc by taking all local rows of P */ 985 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 986 987 988 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 989 pi_loc = p_loc->i; 990 991 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 992 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 993 ierr = PetscMalloc1(am+2,&adpoi);CHKERRQ(ierr); 994 995 adpoi[0] = 0; 996 ptap->api = api; 997 api[0] = 0; 998 999 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1000 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 1001 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 1002 1003 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1004 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->A, p->A, fill, &adpd);CHKERRQ(ierr); 1005 adpd_seq = (Mat_SeqAIJ*)((adpd)->data); 1006 adpdi = adpd_seq->i; adpdj = adpd_seq->j; 1007 p_off = (Mat_SeqAIJ*)((p->B)->data); 1008 poff_i = p_off->i; poff_j = p_off->j; 1009 1010 /* j_temp stores indices of a result row before they are added to the linked list */ 1011 ierr = PetscMalloc1(pN+2,&j_temp);CHKERRQ(ierr); 1012 1013 1014 /* Symbolic calc of the A_diag * p_loc_off */ 1015 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1016 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space_diag);CHKERRQ(ierr); 1017 current_space = free_space_diag; 1018 1019 for (i=0; i<am; i++) { 1020 /* A_diag * P_loc_off */ 1021 nzi = adi[i+1] - adi[i]; 1022 for (j=0; j<nzi; j++) { 1023 row = *adj++; 1024 pnz = poff_i[row+1] - poff_i[row]; 1025 Jptr = poff_j + poff_i[row]; 1026 for(i1 = 0; i1 < pnz; i1++) { 1027 j_temp[i1] = p->garray[Jptr[i1]]; 1028 } 1029 /* add non-zero cols of P into the sorted linked list lnk */ 1030 ierr = PetscLLCondensedAddSorted(pnz,j_temp,lnk,lnkbt);CHKERRQ(ierr); 1031 } 1032 1033 adponz = lnk[0]; 1034 adpoi[i+1] = adpoi[i] + adponz; 1035 1036 /* if free space is not available, double the total space in the list */ 1037 if (current_space->local_remaining<adponz) { 1038 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(adponz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1039 nspacedouble++; 1040 } 1041 1042 /* Copy data into free space, then initialize lnk */ 1043 ierr = PetscLLCondensedClean(pN,adponz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1044 1045 current_space->array += adponz; 1046 current_space->local_used += adponz; 1047 current_space->local_remaining -= adponz; 1048 } 1049 1050 /* Symbolic calc of A_off * P_oth */ 1051 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, &aopoth);CHKERRQ(ierr); 1052 aopoth_seq = (Mat_SeqAIJ*)((aopoth)->data); 1053 aopothi = aopoth_seq->i; aopothj = aopoth_seq->j; 1054 1055 /* Allocate space for apj, adpj, aopj, ... */ 1056 /* destroy lists of free space and other temporary array(s) */ 1057 1058 ierr = PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am]+2, &ptap->apj);CHKERRQ(ierr); 1059 ierr = PetscMalloc1(adpoi[am]+2, &adpoj);CHKERRQ(ierr); 1060 1061 /* Copy from linked list to j-array */ 1062 ierr = PetscFreeSpaceContiguous(&free_space_diag,adpoj);CHKERRQ(ierr); 1063 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1064 1065 adpoJ = adpoj; 1066 adpdJ = adpdj; 1067 aopJ = aopothj; 1068 apj = ptap->apj; 1069 apJ = apj; /* still empty */ 1070 1071 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1072 /* A_diag * P_loc_diag to get A*P */ 1073 for (i = 0; i < am; i++) { 1074 aopnz = aopothi[i+1] - aopothi[i]; 1075 adponz = adpoi[i+1] - adpoi[i]; 1076 adpdnz = adpdi[i+1] - adpdi[i]; 1077 1078 /* Correct indices from A_diag*P_diag */ 1079 for(i1 = 0; i1 < adpdnz; i1++) { 1080 adpdJ[i1] += p_colstart; 1081 } 1082 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1083 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1084 ierr = MatPreallocateSet(i+rstart, apnz, apJ, dnz, onz); CHKERRQ(ierr); 1085 1086 aopJ += aopnz; 1087 adpoJ += adponz; 1088 adpdJ += adpdnz; 1089 apJ += apnz; 1090 api[i+1] = api[i] + apnz; 1091 } 1092 1093 /* malloc apa to store dense row A[i,:]*P */ 1094 ierr = PetscCalloc1(pN+2,&apa);CHKERRQ(ierr); 1095 1096 ptap->apa = apa; 1097 /* create and assemble symbolic parallel matrix Cmpi */ 1098 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1099 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1100 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 1101 1102 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1103 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1104 1105 1106 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Cmpi, apj, api);CHKERRQ(ierr); 1107 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1108 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1109 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1110 1111 1112 ptap->destroy = Cmpi->ops->destroy; 1113 ptap->duplicate = Cmpi->ops->duplicate; 1114 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1115 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 1116 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 1117 1118 /* attach the supporting struct to Cmpi for reuse */ 1119 c = (Mat_MPIAIJ*)Cmpi->data; 1120 c->ptap = ptap; 1121 *C = Cmpi; 1122 1123 /* set MatInfo */ 1124 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 1125 if (afill < 1.0) afill = 1.0; 1126 Cmpi->info.mallocs = nspacedouble; 1127 Cmpi->info.fill_ratio_given = fill; 1128 Cmpi->info.fill_ratio_needed = afill; 1129 1130 #if defined(PETSC_USE_INFO) 1131 if (api[am]) { 1132 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 1133 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 1134 } else { 1135 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1136 } 1137 #endif 1138 1139 ierr = MatDestroy(&aopoth);CHKERRQ(ierr); 1140 ierr = MatDestroy(&adpd);CHKERRQ(ierr); 1141 ierr = PetscFree(j_temp);CHKERRQ(ierr); 1142 ierr = PetscFree(adpoj);CHKERRQ(ierr); 1143 ierr = PetscFree(adpoi);CHKERRQ(ierr); 1144 PetscFunctionReturn(0); 1145 } 1146 1147 1148 /*-------------------------------------------------------------------------*/ 1149 PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat P,Mat A,MatReuse scall,PetscReal fill,Mat *C) 1150 { 1151 PetscErrorCode ierr; 1152 const char *algTypes[3] = {"scalable","nonscalable","matmatmult"}; 1153 PetscInt aN=A->cmap->N,alg=1; /* set default algorithm */ 1154 PetscBool flg; 1155 1156 PetscFunctionBegin; 1157 if (scall == MAT_INITIAL_MATRIX) { 1158 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 1159 PetscOptionsObject->alreadyprinted = PETSC_FALSE; /* a hack to ensure the option shows in '-help' */ 1160 ierr = PetscOptionsEList("-mattransposematmult_via","Algorithmic approach","MatTransposeMatMult",algTypes,3,algTypes[1],&alg,&flg);CHKERRQ(ierr); 1161 ierr = PetscOptionsEnd();CHKERRQ(ierr); 1162 1163 ierr = PetscLogEventBegin(MAT_TransposeMatMultSymbolic,P,A,0,0);CHKERRQ(ierr); 1164 switch (alg) { 1165 case 1: 1166 if (!flg && aN > 100000) { /* may switch to scalable algorithm as default */ 1167 MatInfo Ainfo,Pinfo; 1168 PetscInt nz_local; 1169 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 1170 MPI_Comm comm; 1171 1172 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 1173 ierr = MatGetInfo(P,MAT_LOCAL,&Pinfo);CHKERRQ(ierr); 1174 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); /* estimated local nonzero entries */ 1175 1176 if (aN > fill*nz_local) alg_scalable_loc = PETSC_TRUE; 1177 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1178 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 1179 1180 if (alg_scalable) { 1181 alg = 0; /* scalable algorithm would slower than nonscalable algorithm */ 1182 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 1183 break; 1184 } 1185 } 1186 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(P,A,fill,C);CHKERRQ(ierr); 1187 break; 1188 case 2: 1189 { 1190 Mat Pt; 1191 Mat_PtAPMPI *ptap; 1192 Mat_MPIAIJ *c; 1193 ierr = MatTranspose(P,MAT_INITIAL_MATRIX,&Pt);CHKERRQ(ierr); 1194 ierr = MatMatMult(Pt,A,MAT_INITIAL_MATRIX,fill,C);CHKERRQ(ierr); 1195 c = (Mat_MPIAIJ*)(*C)->data; 1196 ptap = c->ptap; 1197 ptap->Pt = Pt; 1198 (*C)->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 1199 PetscFunctionReturn(0); 1200 } 1201 break; 1202 default: /* scalable algorithm */ 1203 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 1204 break; 1205 } 1206 ierr = PetscLogEventEnd(MAT_TransposeMatMultSymbolic,P,A,0,0);CHKERRQ(ierr); 1207 } 1208 ierr = PetscLogEventBegin(MAT_TransposeMatMultNumeric,P,A,0,0);CHKERRQ(ierr); 1209 ierr = (*(*C)->ops->mattransposemultnumeric)(P,A,*C);CHKERRQ(ierr); 1210 ierr = PetscLogEventEnd(MAT_TransposeMatMultNumeric,P,A,0,0);CHKERRQ(ierr); 1211 PetscFunctionReturn(0); 1212 } 1213 1214 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1215 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P,Mat A,Mat C) 1216 { 1217 PetscErrorCode ierr; 1218 Mat_MPIAIJ *c=(Mat_MPIAIJ*)C->data; 1219 Mat_PtAPMPI *ptap= c->ptap; 1220 Mat Pt=ptap->Pt; 1221 1222 PetscFunctionBegin; 1223 ierr = MatTranspose(P,MAT_REUSE_MATRIX,&Pt);CHKERRQ(ierr); 1224 ierr = MatMatMultNumeric(Pt,A,C);CHKERRQ(ierr); 1225 PetscFunctionReturn(0); 1226 } 1227 1228 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat,MatDuplicateOption,Mat*); 1229 1230 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1231 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,PetscReal fill,Mat *C) 1232 { 1233 PetscErrorCode ierr; 1234 Mat_PtAPMPI *ptap; 1235 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c; 1236 MPI_Comm comm; 1237 PetscMPIInt size,rank; 1238 Mat Cmpi; 1239 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1240 PetscInt pn=P->cmap->n,aN=A->cmap->N,an=A->cmap->n; 1241 PetscInt *lnk,i,k,nsend; 1242 PetscBT lnkbt; 1243 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 1244 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1245 PetscInt len,proc,*dnz,*onz,*owners,nzi; 1246 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1247 MPI_Request *swaits,*rwaits; 1248 MPI_Status *sstatus,rstatus; 1249 PetscLayout rowmap; 1250 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1251 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 1252 PetscInt *Jptr,*prmap=p->garray,con,j,Crmax; 1253 Mat_SeqAIJ *a_loc,*c_loc,*c_oth; 1254 PetscTable ta; 1255 1256 PetscFunctionBegin; 1257 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1258 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1259 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1260 1261 /* create symbolic parallel matrix Cmpi */ 1262 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1263 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1264 1265 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1266 1267 /* create struct Mat_PtAPMPI and attached it to C later */ 1268 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1269 ptap->reuse = MAT_INITIAL_MATRIX; 1270 1271 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1272 /* --------------------------------- */ 1273 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1274 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1275 1276 /* (1) compute symbolic A_loc */ 1277 /* ---------------------------*/ 1278 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1279 1280 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1281 /* ------------------------------------ */ 1282 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->A_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 1283 1284 /* (3) send coj of C_oth to other processors */ 1285 /* ------------------------------------------ */ 1286 /* determine row ownership */ 1287 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 1288 rowmap->n = pn; 1289 rowmap->bs = 1; 1290 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 1291 owners = rowmap->range; 1292 1293 /* determine the number of messages to send, their lengths */ 1294 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 1295 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1296 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1297 1298 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 1299 coi = c_oth->i; coj = c_oth->j; 1300 con = ptap->C_oth->rmap->n; 1301 proc = 0; 1302 for (i=0; i<con; i++) { 1303 while (prmap[i] >= owners[proc+1]) proc++; 1304 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1305 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1306 } 1307 1308 len = 0; /* max length of buf_si[], see (4) */ 1309 owners_co[0] = 0; 1310 nsend = 0; 1311 for (proc=0; proc<size; proc++) { 1312 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1313 if (len_s[proc]) { 1314 nsend++; 1315 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1316 len += len_si[proc]; 1317 } 1318 } 1319 1320 /* determine the number and length of messages to receive for coi and coj */ 1321 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 1322 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 1323 1324 /* post the Irecv and Isend of coj */ 1325 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1326 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1327 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 1328 for (proc=0, k=0; proc<size; proc++) { 1329 if (!len_s[proc]) continue; 1330 i = owners_co[proc]; 1331 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1332 k++; 1333 } 1334 1335 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1336 /* ---------------------------------------- */ 1337 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->A_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 1338 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 1339 1340 /* receives coj are complete */ 1341 for (i=0; i<nrecv; i++) { 1342 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1343 } 1344 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1345 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 1346 1347 /* add received column indices into ta to update Crmax */ 1348 a_loc = (Mat_SeqAIJ*)(ptap->A_loc)->data; 1349 1350 /* create and initialize a linked list */ 1351 ierr = PetscTableCreate(an,aN,&ta);CHKERRQ(ierr); /* for compute Crmax */ 1352 MatRowMergeMax_SeqAIJ(a_loc,ptap->A_loc->rmap->N,ta); 1353 1354 for (k=0; k<nrecv; k++) {/* k-th received message */ 1355 Jptr = buf_rj[k]; 1356 for (j=0; j<len_r[k]; j++) { 1357 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1358 } 1359 } 1360 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 1361 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 1362 1363 /* (4) send and recv coi */ 1364 /*-----------------------*/ 1365 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1366 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1367 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1368 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1369 for (proc=0,k=0; proc<size; proc++) { 1370 if (!len_s[proc]) continue; 1371 /* form outgoing message for i-structure: 1372 buf_si[0]: nrows to be sent 1373 [1:nrows]: row index (global) 1374 [nrows+1:2*nrows+1]: i-structure index 1375 */ 1376 /*-------------------------------------------*/ 1377 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 1378 buf_si_i = buf_si + nrows+1; 1379 buf_si[0] = nrows; 1380 buf_si_i[0] = 0; 1381 nrows = 0; 1382 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1383 nzi = coi[i+1] - coi[i]; 1384 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1385 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1386 nrows++; 1387 } 1388 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1389 k++; 1390 buf_si += len_si[proc]; 1391 } 1392 for (i=0; i<nrecv; i++) { 1393 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1394 } 1395 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1396 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 1397 1398 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 1399 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1400 ierr = PetscFree(swaits);CHKERRQ(ierr); 1401 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1402 1403 /* (5) compute the local portion of Cmpi */ 1404 /* ------------------------------------------ */ 1405 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of Cmpi */ 1406 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 1407 current_space = free_space; 1408 1409 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 1410 for (k=0; k<nrecv; k++) { 1411 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1412 nrows = *buf_ri_k[k]; 1413 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1414 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1415 } 1416 1417 ierr = MatPreallocateInitialize(comm,pn,an,dnz,onz);CHKERRQ(ierr); 1418 ierr = PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt);CHKERRQ(ierr); 1419 for (i=0; i<pn; i++) { 1420 /* add C_loc into Cmpi */ 1421 nzi = c_loc->i[i+1] - c_loc->i[i]; 1422 Jptr = c_loc->j + c_loc->i[i]; 1423 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1424 1425 /* add received col data into lnk */ 1426 for (k=0; k<nrecv; k++) { /* k-th received message */ 1427 if (i == *nextrow[k]) { /* i-th row */ 1428 nzi = *(nextci[k]+1) - *nextci[k]; 1429 Jptr = buf_rj[k] + *nextci[k]; 1430 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1431 nextrow[k]++; nextci[k]++; 1432 } 1433 } 1434 nzi = lnk[0]; 1435 1436 /* copy data into free space, then initialize lnk */ 1437 ierr = PetscLLCondensedClean(aN,nzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1438 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 1439 } 1440 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1441 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1442 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 1443 1444 /* local sizes and preallocation */ 1445 ierr = MatSetSizes(Cmpi,pn,an,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1446 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 1447 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1448 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1449 1450 /* members in merge */ 1451 ierr = PetscFree(id_r);CHKERRQ(ierr); 1452 ierr = PetscFree(len_r);CHKERRQ(ierr); 1453 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 1454 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 1455 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 1456 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 1457 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 1458 1459 /* attach the supporting struct to Cmpi for reuse */ 1460 c = (Mat_MPIAIJ*)Cmpi->data; 1461 c->ptap = ptap; 1462 ptap->duplicate = Cmpi->ops->duplicate; 1463 ptap->destroy = Cmpi->ops->destroy; 1464 1465 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 1466 Cmpi->assembled = PETSC_FALSE; 1467 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1468 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 1469 *C = Cmpi; 1470 PetscFunctionReturn(0); 1471 } 1472 1473 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,Mat C) 1474 { 1475 PetscErrorCode ierr; 1476 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1477 Mat_SeqAIJ *c_seq; 1478 Mat_PtAPMPI *ptap = c->ptap; 1479 Mat A_loc,C_loc,C_oth; 1480 PetscInt i,rstart,rend,cm,ncols,row; 1481 const PetscInt *cols; 1482 const PetscScalar *vals; 1483 1484 PetscFunctionBegin; 1485 ierr = MatZeroEntries(C);CHKERRQ(ierr); 1486 1487 if (ptap->reuse == MAT_REUSE_MATRIX) { 1488 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1489 /* 1) get R = Pd^T, Ro = Po^T */ 1490 /*----------------------------*/ 1491 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1492 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1493 1494 /* 2) compute numeric A_loc */ 1495 /*--------------------------*/ 1496 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1497 } 1498 1499 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1500 A_loc = ptap->A_loc; 1501 ierr = ((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,A_loc,ptap->C_loc);CHKERRQ(ierr); 1502 ierr = ((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,A_loc,ptap->C_oth);CHKERRQ(ierr); 1503 C_loc = ptap->C_loc; 1504 C_oth = ptap->C_oth; 1505 1506 /* add C_loc and Co to to C */ 1507 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 1508 1509 /* C_loc -> C */ 1510 cm = C_loc->rmap->N; 1511 c_seq = (Mat_SeqAIJ*)C_loc->data; 1512 cols = c_seq->j; 1513 vals = c_seq->a; 1514 for (i=0; i<cm; i++) { 1515 ncols = c_seq->i[i+1] - c_seq->i[i]; 1516 row = rstart + i; 1517 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1518 cols += ncols; vals += ncols; 1519 } 1520 1521 /* Co -> C, off-processor part */ 1522 cm = C_oth->rmap->N; 1523 c_seq = (Mat_SeqAIJ*)C_oth->data; 1524 cols = c_seq->j; 1525 vals = c_seq->a; 1526 for (i=0; i<cm; i++) { 1527 ncols = c_seq->i[i+1] - c_seq->i[i]; 1528 row = p->garray[i]; 1529 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1530 cols += ncols; vals += ncols; 1531 } 1532 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1533 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1534 1535 ptap->reuse = MAT_REUSE_MATRIX; 1536 PetscFunctionReturn(0); 1537 } 1538 1539 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 1540 { 1541 PetscErrorCode ierr; 1542 Mat_Merge_SeqsToMPI *merge; 1543 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1544 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1545 Mat_PtAPMPI *ptap; 1546 PetscInt *adj; 1547 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1548 MatScalar *ada,*ca,valtmp; 1549 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1550 MPI_Comm comm; 1551 PetscMPIInt size,rank,taga,*len_s; 1552 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1553 PetscInt **buf_ri,**buf_rj; 1554 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1555 MPI_Request *s_waits,*r_waits; 1556 MPI_Status *status; 1557 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1558 PetscInt *ai,*aj,*coi,*coj,*poJ,*pdJ; 1559 Mat A_loc; 1560 Mat_SeqAIJ *a_loc; 1561 1562 PetscFunctionBegin; 1563 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1564 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1565 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1566 1567 ptap = c->ptap; 1568 merge = ptap->merge; 1569 1570 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1571 /*------------------------------------------*/ 1572 /* get data from symbolic products */ 1573 coi = merge->coi; coj = merge->coj; 1574 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1575 bi = merge->bi; bj = merge->bj; 1576 owners = merge->rowmap->range; 1577 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); 1578 1579 /* get A_loc by taking all local rows of A */ 1580 A_loc = ptap->A_loc; 1581 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 1582 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1583 ai = a_loc->i; 1584 aj = a_loc->j; 1585 1586 for (i=0; i<am; i++) { 1587 anz = ai[i+1] - ai[i]; 1588 adj = aj + ai[i]; 1589 ada = a_loc->a + ai[i]; 1590 1591 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1592 /*-------------------------------------------------------------*/ 1593 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1594 pnz = po->i[i+1] - po->i[i]; 1595 poJ = po->j + po->i[i]; 1596 pA = po->a + po->i[i]; 1597 for (j=0; j<pnz; j++) { 1598 row = poJ[j]; 1599 cj = coj + coi[row]; 1600 ca = coa + coi[row]; 1601 /* perform sparse axpy */ 1602 nexta = 0; 1603 valtmp = pA[j]; 1604 for (k=0; nexta<anz; k++) { 1605 if (cj[k] == adj[nexta]) { 1606 ca[k] += valtmp*ada[nexta]; 1607 nexta++; 1608 } 1609 } 1610 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1611 } 1612 1613 /* put the value into Cd (diagonal part) */ 1614 pnz = pd->i[i+1] - pd->i[i]; 1615 pdJ = pd->j + pd->i[i]; 1616 pA = pd->a + pd->i[i]; 1617 for (j=0; j<pnz; j++) { 1618 row = pdJ[j]; 1619 cj = bj + bi[row]; 1620 ca = ba + bi[row]; 1621 /* perform sparse axpy */ 1622 nexta = 0; 1623 valtmp = pA[j]; 1624 for (k=0; nexta<anz; k++) { 1625 if (cj[k] == adj[nexta]) { 1626 ca[k] += valtmp*ada[nexta]; 1627 nexta++; 1628 } 1629 } 1630 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1631 } 1632 } 1633 1634 /* 3) send and recv matrix values coa */ 1635 /*------------------------------------*/ 1636 buf_ri = merge->buf_ri; 1637 buf_rj = merge->buf_rj; 1638 len_s = merge->len_s; 1639 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1640 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1641 1642 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1643 for (proc=0,k=0; proc<size; proc++) { 1644 if (!len_s[proc]) continue; 1645 i = merge->owners_co[proc]; 1646 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1647 k++; 1648 } 1649 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1650 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1651 1652 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1653 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1654 ierr = PetscFree(coa);CHKERRQ(ierr); 1655 1656 /* 4) insert local Cseq and received values into Cmpi */ 1657 /*----------------------------------------------------*/ 1658 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1659 for (k=0; k<merge->nrecv; k++) { 1660 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1661 nrows = *(buf_ri_k[k]); 1662 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1663 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1664 } 1665 1666 for (i=0; i<cm; i++) { 1667 row = owners[rank] + i; /* global row index of C_seq */ 1668 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1669 ba_i = ba + bi[i]; 1670 bnz = bi[i+1] - bi[i]; 1671 /* add received vals into ba */ 1672 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1673 /* i-th row */ 1674 if (i == *nextrow[k]) { 1675 cnz = *(nextci[k]+1) - *nextci[k]; 1676 cj = buf_rj[k] + *(nextci[k]); 1677 ca = abuf_r[k] + *(nextci[k]); 1678 nextcj = 0; 1679 for (j=0; nextcj<cnz; j++) { 1680 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1681 ba_i[j] += ca[nextcj++]; 1682 } 1683 } 1684 nextrow[k]++; nextci[k]++; 1685 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1686 } 1687 } 1688 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1689 } 1690 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1691 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1692 1693 ierr = PetscFree(ba);CHKERRQ(ierr); 1694 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1695 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1696 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1697 PetscFunctionReturn(0); 1698 } 1699 1700 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat *C) 1701 { 1702 PetscErrorCode ierr; 1703 Mat Cmpi,A_loc,POt,PDt; 1704 Mat_PtAPMPI *ptap; 1705 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1706 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*a=(Mat_MPIAIJ*)A->data,*c; 1707 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1708 PetscInt nnz; 1709 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1710 PetscInt am =A->rmap->n,pn=P->cmap->n; 1711 MPI_Comm comm; 1712 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1713 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1714 PetscInt len,proc,*dnz,*onz,*owners; 1715 PetscInt nzi,*bi,*bj; 1716 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1717 MPI_Request *swaits,*rwaits; 1718 MPI_Status *sstatus,rstatus; 1719 Mat_Merge_SeqsToMPI *merge; 1720 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1721 PetscReal afill =1.0,afill_tmp; 1722 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Armax; 1723 PetscScalar *vals; 1724 Mat_SeqAIJ *a_loc,*pdt,*pot; 1725 PetscTable ta; 1726 1727 PetscFunctionBegin; 1728 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1729 /* check if matrix local sizes are compatible */ 1730 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) 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); 1731 1732 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1733 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1734 1735 /* create struct Mat_PtAPMPI and attached it to C later */ 1736 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1737 1738 /* get A_loc by taking all local rows of A */ 1739 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1740 1741 ptap->A_loc = A_loc; 1742 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1743 ai = a_loc->i; 1744 aj = a_loc->j; 1745 1746 /* determine symbolic Co=(p->B)^T*A - send to others */ 1747 /*----------------------------------------------------*/ 1748 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1749 pdt = (Mat_SeqAIJ*)PDt->data; 1750 pdti = pdt->i; pdtj = pdt->j; 1751 1752 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1753 pot = (Mat_SeqAIJ*)POt->data; 1754 poti = pot->i; potj = pot->j; 1755 1756 /* then, compute symbolic Co = (p->B)^T*A */ 1757 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1758 >= (num of nonzero rows of C_seq) - pn */ 1759 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 1760 coi[0] = 0; 1761 1762 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1763 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(poti[pon],ai[am])); 1764 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1765 current_space = free_space; 1766 1767 /* create and initialize a linked list */ 1768 ierr = PetscTableCreate(A->cmap->n + a->B->cmap->N,aN,&ta);CHKERRQ(ierr); 1769 MatRowMergeMax_SeqAIJ(a_loc,am,ta); 1770 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1771 1772 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1773 1774 for (i=0; i<pon; i++) { 1775 pnz = poti[i+1] - poti[i]; 1776 ptJ = potj + poti[i]; 1777 for (j=0; j<pnz; j++) { 1778 row = ptJ[j]; /* row of A_loc == col of Pot */ 1779 anz = ai[row+1] - ai[row]; 1780 Jptr = aj + ai[row]; 1781 /* add non-zero cols of AP into the sorted linked list lnk */ 1782 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1783 } 1784 nnz = lnk[0]; 1785 1786 /* If free space is not available, double the total space in the list */ 1787 if (current_space->local_remaining<nnz) { 1788 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1789 nspacedouble++; 1790 } 1791 1792 /* Copy data into free space, and zero out denserows */ 1793 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1794 1795 current_space->array += nnz; 1796 current_space->local_used += nnz; 1797 current_space->local_remaining -= nnz; 1798 1799 coi[i+1] = coi[i] + nnz; 1800 } 1801 1802 ierr = PetscMalloc1(coi[pon]+1,&coj);CHKERRQ(ierr); 1803 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1804 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); /* must destroy to get a new one for C */ 1805 1806 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1807 if (afill_tmp > afill) afill = afill_tmp; 1808 1809 /* send j-array (coj) of Co to other processors */ 1810 /*----------------------------------------------*/ 1811 /* determine row ownership */ 1812 ierr = PetscNew(&merge);CHKERRQ(ierr); 1813 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1814 1815 merge->rowmap->n = pn; 1816 merge->rowmap->bs = 1; 1817 1818 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1819 owners = merge->rowmap->range; 1820 1821 /* determine the number of messages to send, their lengths */ 1822 ierr = PetscCalloc1(size,&len_si);CHKERRQ(ierr); 1823 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 1824 1825 len_s = merge->len_s; 1826 merge->nsend = 0; 1827 1828 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 1829 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1830 1831 proc = 0; 1832 for (i=0; i<pon; i++) { 1833 while (prmap[i] >= owners[proc+1]) proc++; 1834 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1835 len_s[proc] += coi[i+1] - coi[i]; 1836 } 1837 1838 len = 0; /* max length of buf_si[] */ 1839 owners_co[0] = 0; 1840 for (proc=0; proc<size; proc++) { 1841 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1842 if (len_si[proc]) { 1843 merge->nsend++; 1844 len_si[proc] = 2*(len_si[proc] + 1); 1845 len += len_si[proc]; 1846 } 1847 } 1848 1849 /* determine the number and length of messages to receive for coi and coj */ 1850 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1851 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1852 1853 /* post the Irecv and Isend of coj */ 1854 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1855 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1856 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 1857 for (proc=0, k=0; proc<size; proc++) { 1858 if (!len_s[proc]) continue; 1859 i = owners_co[proc]; 1860 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1861 k++; 1862 } 1863 1864 /* receives and sends of coj are complete */ 1865 ierr = PetscMalloc1(size,&sstatus);CHKERRQ(ierr); 1866 for (i=0; i<merge->nrecv; i++) { 1867 PetscMPIInt icompleted; 1868 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1869 } 1870 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1871 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1872 1873 /* add received column indices into table to update Armax */ 1874 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/examples/tutorials/ex56.c! */ 1875 for (k=0; k<merge->nrecv; k++) {/* k-th received message */ 1876 Jptr = buf_rj[k]; 1877 for (j=0; j<merge->len_r[k]; j++) { 1878 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1879 } 1880 } 1881 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1882 /* printf("Armax %d, an %d + Bn %d = %d, aN %d\n",Armax,A->cmap->n,a->B->cmap->N,A->cmap->n+a->B->cmap->N,aN); */ 1883 1884 /* send and recv coi */ 1885 /*-------------------*/ 1886 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1887 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1888 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1889 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1890 for (proc=0,k=0; proc<size; proc++) { 1891 if (!len_s[proc]) continue; 1892 /* form outgoing message for i-structure: 1893 buf_si[0]: nrows to be sent 1894 [1:nrows]: row index (global) 1895 [nrows+1:2*nrows+1]: i-structure index 1896 */ 1897 /*-------------------------------------------*/ 1898 nrows = len_si[proc]/2 - 1; 1899 buf_si_i = buf_si + nrows+1; 1900 buf_si[0] = nrows; 1901 buf_si_i[0] = 0; 1902 nrows = 0; 1903 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1904 nzi = coi[i+1] - coi[i]; 1905 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1906 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1907 nrows++; 1908 } 1909 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1910 k++; 1911 buf_si += len_si[proc]; 1912 } 1913 i = merge->nrecv; 1914 while (i--) { 1915 PetscMPIInt icompleted; 1916 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1917 } 1918 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1919 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1920 ierr = PetscFree(len_si);CHKERRQ(ierr); 1921 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1922 ierr = PetscFree(swaits);CHKERRQ(ierr); 1923 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1924 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1925 1926 /* compute the local portion of C (mpi mat) */ 1927 /*------------------------------------------*/ 1928 /* allocate bi array and free space for accumulating nonzero column info */ 1929 ierr = PetscMalloc1(pn+1,&bi);CHKERRQ(ierr); 1930 bi[0] = 0; 1931 1932 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1933 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(pdti[pn],PetscIntSumTruncate(poti[pon],ai[am]))); 1934 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1935 current_space = free_space; 1936 1937 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1938 for (k=0; k<merge->nrecv; k++) { 1939 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1940 nrows = *buf_ri_k[k]; 1941 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1942 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recieved i-structure */ 1943 } 1944 1945 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1946 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1947 rmax = 0; 1948 for (i=0; i<pn; i++) { 1949 /* add pdt[i,:]*AP into lnk */ 1950 pnz = pdti[i+1] - pdti[i]; 1951 ptJ = pdtj + pdti[i]; 1952 for (j=0; j<pnz; j++) { 1953 row = ptJ[j]; /* row of AP == col of Pt */ 1954 anz = ai[row+1] - ai[row]; 1955 Jptr = aj + ai[row]; 1956 /* add non-zero cols of AP into the sorted linked list lnk */ 1957 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1958 } 1959 1960 /* add received col data into lnk */ 1961 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1962 if (i == *nextrow[k]) { /* i-th row */ 1963 nzi = *(nextci[k]+1) - *nextci[k]; 1964 Jptr = buf_rj[k] + *nextci[k]; 1965 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 1966 nextrow[k]++; nextci[k]++; 1967 } 1968 } 1969 nnz = lnk[0]; 1970 1971 /* if free space is not available, make more free space */ 1972 if (current_space->local_remaining<nnz) { 1973 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1974 nspacedouble++; 1975 } 1976 /* copy data into free space, then initialize lnk */ 1977 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1978 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 1979 1980 current_space->array += nnz; 1981 current_space->local_used += nnz; 1982 current_space->local_remaining -= nnz; 1983 1984 bi[i+1] = bi[i] + nnz; 1985 if (nnz > rmax) rmax = nnz; 1986 } 1987 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1988 1989 ierr = PetscMalloc1(bi[pn]+1,&bj);CHKERRQ(ierr); 1990 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 1991 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 1992 if (afill_tmp > afill) afill = afill_tmp; 1993 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 1994 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 1995 1996 ierr = MatDestroy(&POt);CHKERRQ(ierr); 1997 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 1998 1999 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 2000 /*----------------------------------------------------------------------------------*/ 2001 ierr = PetscCalloc1(rmax+1,&vals);CHKERRQ(ierr); 2002 2003 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 2004 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2005 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(A->cmap->bs));CHKERRQ(ierr); 2006 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 2007 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 2008 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 2009 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 2010 for (i=0; i<pn; i++) { 2011 row = i + rstart; 2012 nnz = bi[i+1] - bi[i]; 2013 Jptr = bj + bi[i]; 2014 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 2015 } 2016 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2017 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2018 ierr = PetscFree(vals);CHKERRQ(ierr); 2019 2020 merge->bi = bi; 2021 merge->bj = bj; 2022 merge->coi = coi; 2023 merge->coj = coj; 2024 merge->buf_ri = buf_ri; 2025 merge->buf_rj = buf_rj; 2026 merge->owners_co = owners_co; 2027 2028 /* attach the supporting struct to Cmpi for reuse */ 2029 c = (Mat_MPIAIJ*)Cmpi->data; 2030 2031 c->ptap = ptap; 2032 ptap->api = NULL; 2033 ptap->apj = NULL; 2034 ptap->merge = merge; 2035 ptap->apa = NULL; 2036 ptap->destroy = Cmpi->ops->destroy; 2037 ptap->duplicate = Cmpi->ops->duplicate; 2038 2039 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2040 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 2041 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 2042 2043 *C = Cmpi; 2044 #if defined(PETSC_USE_INFO) 2045 if (bi[pn] != 0) { 2046 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 2047 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 2048 } else { 2049 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 2050 } 2051 #endif 2052 PetscFunctionReturn(0); 2053 } 2054