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 #include <petsc/private/sfimpl.h> 13 14 #if defined(PETSC_HAVE_HYPRE) 15 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat); 16 #endif 17 18 PETSC_INTERN PetscErrorCode MatProductSymbolic_AB_MPIAIJ_MPIAIJ(Mat C) 19 { 20 PetscErrorCode ierr; 21 Mat_Product *product = C->product; 22 Mat A=product->A,B=product->B; 23 MatProductAlgorithm alg=product->alg; 24 PetscReal fill=product->fill; 25 PetscBool flg; 26 27 PetscFunctionBegin; 28 /* scalable */ 29 ierr = PetscStrcmp(alg,"scalable",&flg);CHKERRQ(ierr); 30 if (flg) { 31 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 32 PetscFunctionReturn(0); 33 } 34 35 /* nonscalable */ 36 ierr = PetscStrcmp(alg,"nonscalable",&flg);CHKERRQ(ierr); 37 if (flg) { 38 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C);CHKERRQ(ierr); 39 PetscFunctionReturn(0); 40 } 41 42 /* seqmpi */ 43 ierr = PetscStrcmp(alg,"seqmpi",&flg);CHKERRQ(ierr); 44 if (flg) { 45 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(A,B,fill,C);CHKERRQ(ierr); 46 PetscFunctionReturn(0); 47 } 48 49 #if defined(PETSC_HAVE_HYPRE) 50 ierr = PetscStrcmp(alg,"hypre",&flg);CHKERRQ(ierr); 51 if (flg) { 52 ierr = MatMatMultSymbolic_AIJ_AIJ_wHYPRE(A,B,fill,C);CHKERRQ(ierr); 53 PetscFunctionReturn(0); 54 } 55 #endif 56 SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_SUP,"Mat Product Algorithm is not supported"); 57 } 58 59 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(void *data) 60 { 61 PetscErrorCode ierr; 62 Mat_APMPI *ptap = (Mat_APMPI*)data; 63 64 PetscFunctionBegin; 65 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 66 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 67 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 68 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 69 ierr = MatDestroy(&ptap->Pt);CHKERRQ(ierr); 70 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 71 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 72 ierr = PetscFree(ptap->apa);CHKERRQ(ierr); 73 ierr = PetscFree(ptap);CHKERRQ(ierr); 74 PetscFunctionReturn(0); 75 } 76 77 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,Mat C) 78 { 79 PetscErrorCode ierr; 80 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 81 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 82 Mat_SeqAIJ *cd =(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 83 PetscScalar *cda=cd->a,*coa=co->a; 84 Mat_SeqAIJ *p_loc,*p_oth; 85 PetscScalar *apa,*ca; 86 PetscInt cm =C->rmap->n; 87 Mat_APMPI *ptap; 88 PetscInt *api,*apj,*apJ,i,k; 89 PetscInt cstart=C->cmap->rstart; 90 PetscInt cdnz,conz,k0,k1; 91 MPI_Comm comm; 92 PetscMPIInt size; 93 94 PetscFunctionBegin; 95 MatCheckProduct(C,3); 96 ptap = (Mat_APMPI*)C->product->data; 97 if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 98 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 99 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 100 101 if (!ptap->P_oth && size>1) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatProductClear()"); 102 103 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 104 /*-----------------------------------------------------*/ 105 /* update numerical values of P_oth and P_loc */ 106 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 107 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 108 109 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 110 /*----------------------------------------------------------*/ 111 /* get data from symbolic products */ 112 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 113 p_oth = NULL; 114 if (size >1) { 115 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 116 } 117 118 /* get apa for storing dense row A[i,:]*P */ 119 apa = ptap->apa; 120 121 api = ptap->api; 122 apj = ptap->apj; 123 for (i=0; i<cm; i++) { 124 /* compute apa = A[i,:]*P */ 125 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 126 127 /* set values in C */ 128 apJ = apj + api[i]; 129 cdnz = cd->i[i+1] - cd->i[i]; 130 conz = co->i[i+1] - co->i[i]; 131 132 /* 1st off-diagonal part of C */ 133 ca = coa + co->i[i]; 134 k = 0; 135 for (k0=0; k0<conz; k0++) { 136 if (apJ[k] >= cstart) break; 137 ca[k0] = apa[apJ[k]]; 138 apa[apJ[k++]] = 0.0; 139 } 140 141 /* diagonal part of C */ 142 ca = cda + cd->i[i]; 143 for (k1=0; k1<cdnz; k1++) { 144 ca[k1] = apa[apJ[k]]; 145 apa[apJ[k++]] = 0.0; 146 } 147 148 /* 2nd off-diagonal part of C */ 149 ca = coa + co->i[i]; 150 for (; k0<conz; k0++) { 151 ca[k0] = apa[apJ[k]]; 152 apa[apJ[k++]] = 0.0; 153 } 154 } 155 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 156 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 157 PetscFunctionReturn(0); 158 } 159 160 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,PetscReal fill,Mat C) 161 { 162 PetscErrorCode ierr; 163 MPI_Comm comm; 164 PetscMPIInt size; 165 Mat_APMPI *ptap; 166 PetscFreeSpaceList free_space=NULL,current_space=NULL; 167 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 168 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 169 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 170 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 171 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 172 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 173 PetscBT lnkbt; 174 PetscReal afill; 175 MatType mtype; 176 177 PetscFunctionBegin; 178 MatCheckProduct(C,4); 179 if (C->product->data) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 180 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 181 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 182 183 /* create struct Mat_APMPI and attached it to C later */ 184 ierr = PetscNew(&ptap);CHKERRQ(ierr); 185 186 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 187 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 188 189 /* get P_loc by taking all local rows of P */ 190 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 191 192 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 193 pi_loc = p_loc->i; pj_loc = p_loc->j; 194 if (size > 1) { 195 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 196 pi_oth = p_oth->i; pj_oth = p_oth->j; 197 } else { 198 p_oth = NULL; 199 pi_oth = NULL; pj_oth = NULL; 200 } 201 202 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 203 /*-------------------------------------------------------------------*/ 204 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 205 ptap->api = api; 206 api[0] = 0; 207 208 /* create and initialize a linked list */ 209 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 210 211 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 212 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 213 current_space = free_space; 214 215 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 216 for (i=0; i<am; i++) { 217 /* diagonal portion of A */ 218 nzi = adi[i+1] - adi[i]; 219 for (j=0; j<nzi; j++) { 220 row = *adj++; 221 pnz = pi_loc[row+1] - pi_loc[row]; 222 Jptr = pj_loc + pi_loc[row]; 223 /* add non-zero cols of P into the sorted linked list lnk */ 224 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 225 } 226 /* off-diagonal portion of A */ 227 nzi = aoi[i+1] - aoi[i]; 228 for (j=0; j<nzi; j++) { 229 row = *aoj++; 230 pnz = pi_oth[row+1] - pi_oth[row]; 231 Jptr = pj_oth + pi_oth[row]; 232 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 233 } 234 /* add possible missing diagonal entry */ 235 if (C->force_diagonals) { 236 j = i + rstart; /* column index */ 237 ierr = PetscLLCondensedAddSorted(1,&j,lnk,lnkbt);CHKERRQ(ierr); 238 } 239 240 apnz = lnk[0]; 241 api[i+1] = api[i] + apnz; 242 243 /* if free space is not available, double the total space in the list */ 244 if (current_space->local_remaining<apnz) { 245 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 246 nspacedouble++; 247 } 248 249 /* Copy data into free space, then initialize lnk */ 250 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 251 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 252 253 current_space->array += apnz; 254 current_space->local_used += apnz; 255 current_space->local_remaining -= apnz; 256 } 257 258 /* Allocate space for apj, initialize apj, and */ 259 /* destroy list of free space and other temporary array(s) */ 260 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 261 apj = ptap->apj; 262 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 263 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 264 265 /* malloc apa to store dense row A[i,:]*P */ 266 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 267 268 /* set and assemble symbolic parallel matrix C */ 269 /*---------------------------------------------*/ 270 ierr = MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 271 ierr = MatSetBlockSizesFromMats(C,A,P);CHKERRQ(ierr); 272 273 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 274 ierr = MatSetType(C,mtype);CHKERRQ(ierr); 275 ierr = MatMPIAIJSetPreallocation(C,0,dnz,0,onz);CHKERRQ(ierr); 276 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 277 278 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api);CHKERRQ(ierr); 279 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 280 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 281 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 282 283 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 284 C->ops->productnumeric = MatProductNumeric_AB; 285 286 /* attach the supporting struct to C for reuse */ 287 C->product->data = ptap; 288 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 289 290 /* set MatInfo */ 291 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 292 if (afill < 1.0) afill = 1.0; 293 C->info.mallocs = nspacedouble; 294 C->info.fill_ratio_given = fill; 295 C->info.fill_ratio_needed = afill; 296 297 #if defined(PETSC_USE_INFO) 298 if (api[am]) { 299 ierr = PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 300 ierr = PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 301 } else { 302 ierr = PetscInfo(C,"Empty matrix product\n");CHKERRQ(ierr); 303 } 304 #endif 305 PetscFunctionReturn(0); 306 } 307 308 /* ------------------------------------------------------- */ 309 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat,Mat,PetscReal,Mat); 310 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat,Mat,Mat); 311 312 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AB(Mat C) 313 { 314 Mat_Product *product = C->product; 315 Mat A = product->A,B=product->B; 316 317 PetscFunctionBegin; 318 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) 319 SETERRQ4(PETSC_COMM_SELF,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); 320 321 C->ops->matmultsymbolic = MatMatMultSymbolic_MPIAIJ_MPIDense; 322 C->ops->productsymbolic = MatProductSymbolic_AB; 323 PetscFunctionReturn(0); 324 } 325 /* -------------------------------------------------------------------- */ 326 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(Mat C) 327 { 328 Mat_Product *product = C->product; 329 Mat A = product->A,B=product->B; 330 331 PetscFunctionBegin; 332 if (A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend) 333 SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->rmap->rstart,A->rmap->rend,B->rmap->rstart,B->rmap->rend); 334 335 C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIAIJ_MPIDense; 336 C->ops->productsymbolic = MatProductSymbolic_AtB; 337 PetscFunctionReturn(0); 338 } 339 340 /* --------------------------------------------------------------------- */ 341 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense(Mat C) 342 { 343 PetscErrorCode ierr; 344 Mat_Product *product = C->product; 345 346 PetscFunctionBegin; 347 switch (product->type) { 348 case MATPRODUCT_AB: 349 ierr = MatProductSetFromOptions_MPIAIJ_MPIDense_AB(C);CHKERRQ(ierr); 350 break; 351 case MATPRODUCT_AtB: 352 ierr = MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(C);CHKERRQ(ierr); 353 break; 354 default: 355 break; 356 } 357 PetscFunctionReturn(0); 358 } 359 /* ------------------------------------------------------- */ 360 361 typedef struct { 362 Mat workB,workB1; 363 MPI_Request *rwaits,*swaits; 364 PetscInt nsends,nrecvs; 365 MPI_Datatype *stype,*rtype; 366 PetscInt blda; 367 } MPIAIJ_MPIDense; 368 369 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 370 { 371 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*)ctx; 372 PetscErrorCode ierr; 373 PetscInt i; 374 375 PetscFunctionBegin; 376 ierr = MatDestroy(&contents->workB);CHKERRQ(ierr); 377 ierr = MatDestroy(&contents->workB1);CHKERRQ(ierr); 378 for (i=0; i<contents->nsends; i++) { 379 ierr = MPI_Type_free(&contents->stype[i]);CHKERRMPI(ierr); 380 } 381 for (i=0; i<contents->nrecvs; i++) { 382 ierr = MPI_Type_free(&contents->rtype[i]);CHKERRMPI(ierr); 383 } 384 ierr = PetscFree4(contents->stype,contents->rtype,contents->rwaits,contents->swaits);CHKERRQ(ierr); 385 ierr = PetscFree(contents);CHKERRQ(ierr); 386 PetscFunctionReturn(0); 387 } 388 389 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat C) 390 { 391 PetscErrorCode ierr; 392 Mat_MPIAIJ *aij=(Mat_MPIAIJ*)A->data; 393 PetscInt nz=aij->B->cmap->n,nsends,nrecvs,i,nrows_to,j,blda,clda; 394 MPIAIJ_MPIDense *contents; 395 VecScatter ctx=aij->Mvctx; 396 PetscInt Am=A->rmap->n,Bm=B->rmap->n,BN=B->cmap->N,Bbn,Bbn1,bs,nrows_from,numBb; 397 MPI_Comm comm; 398 MPI_Datatype type1,*stype,*rtype; 399 const PetscInt *sindices,*sstarts,*rstarts; 400 PetscMPIInt *disp; 401 PetscBool cisdense; 402 403 PetscFunctionBegin; 404 MatCheckProduct(C,4); 405 if (C->product->data) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data not empty"); 406 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 407 ierr = PetscObjectBaseTypeCompare((PetscObject)C,MATMPIDENSE,&cisdense);CHKERRQ(ierr); 408 if (!cisdense) { 409 ierr = MatSetType(C,((PetscObject)B)->type_name);CHKERRQ(ierr); 410 } 411 ierr = MatSetSizes(C,Am,B->cmap->n,A->rmap->N,BN);CHKERRQ(ierr); 412 ierr = MatSetBlockSizesFromMats(C,A,B);CHKERRQ(ierr); 413 ierr = MatSetUp(C);CHKERRQ(ierr); 414 ierr = MatDenseGetLDA(B,&blda);CHKERRQ(ierr); 415 ierr = MatDenseGetLDA(C,&clda);CHKERRQ(ierr); 416 ierr = PetscNew(&contents);CHKERRQ(ierr); 417 418 ierr = VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,NULL,NULL);CHKERRQ(ierr); 419 ierr = VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,NULL,NULL);CHKERRQ(ierr); 420 421 /* Create column block of B and C for memory scalability when BN is too large */ 422 /* Estimate Bbn, column size of Bb */ 423 if (nz) { 424 Bbn1 = 2*Am*BN/nz; 425 } else Bbn1 = BN; 426 427 bs = PetscAbs(B->cmap->bs); 428 Bbn1 = Bbn1/bs *bs; /* Bbn1 is a multiple of bs */ 429 if (Bbn1 > BN) Bbn1 = BN; 430 ierr = MPI_Allreduce(&Bbn1,&Bbn,1,MPIU_INT,MPI_MAX,comm);CHKERRMPI(ierr); 431 432 /* Enable runtime option for Bbn */ 433 ierr = PetscOptionsBegin(comm,((PetscObject)C)->prefix,"MatMatMult","Mat");CHKERRQ(ierr); 434 ierr = PetscOptionsInt("-matmatmult_Bbn","Number of columns in Bb","MatMatMult",Bbn,&Bbn,NULL);CHKERRQ(ierr); 435 ierr = PetscOptionsEnd();CHKERRQ(ierr); 436 Bbn = PetscMin(Bbn,BN); 437 438 if (Bbn > 0 && Bbn < BN) { 439 numBb = BN/Bbn; 440 Bbn1 = BN - numBb*Bbn; 441 } else numBb = 0; 442 443 if (numBb) { 444 ierr = PetscInfo3(C,"use Bb, BN=%D, Bbn=%D; numBb=%D\n",BN,Bbn,numBb);CHKERRQ(ierr); 445 if (Bbn1) { /* Create workB1 for the remaining columns */ 446 ierr = PetscInfo2(C,"use Bb1, BN=%D, Bbn1=%D\n",BN,Bbn1);CHKERRQ(ierr); 447 /* Create work matrix used to store off processor rows of B needed for local product */ 448 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,Bbn1,NULL,&contents->workB1);CHKERRQ(ierr); 449 } else contents->workB1 = NULL; 450 } 451 452 /* Create work matrix used to store off processor rows of B needed for local product */ 453 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,Bbn,NULL,&contents->workB);CHKERRQ(ierr); 454 455 /* Use MPI derived data type to reduce memory required by the send/recv buffers */ 456 ierr = PetscMalloc4(nsends,&stype,nrecvs,&rtype,nrecvs,&contents->rwaits,nsends,&contents->swaits);CHKERRQ(ierr); 457 contents->stype = stype; 458 contents->nsends = nsends; 459 460 contents->rtype = rtype; 461 contents->nrecvs = nrecvs; 462 contents->blda = blda; 463 464 ierr = PetscMalloc1(Bm+1,&disp);CHKERRQ(ierr); 465 for (i=0; i<nsends; i++) { 466 nrows_to = sstarts[i+1]-sstarts[i]; 467 for (j=0; j<nrows_to; j++){ 468 disp[j] = sindices[sstarts[i]+j]; /* rowB to be sent */ 469 } 470 ierr = MPI_Type_create_indexed_block(nrows_to,1,(const PetscMPIInt *)disp,MPIU_SCALAR,&type1);CHKERRMPI(ierr); 471 472 ierr = MPI_Type_create_resized(type1,0,blda*sizeof(PetscScalar),&stype[i]);CHKERRMPI(ierr); 473 ierr = MPI_Type_commit(&stype[i]);CHKERRMPI(ierr); 474 ierr = MPI_Type_free(&type1);CHKERRMPI(ierr); 475 } 476 477 for (i=0; i<nrecvs; i++) { 478 /* received values from a process form a (nrows_from x Bbn) row block in workB (column-wise) */ 479 nrows_from = rstarts[i+1]-rstarts[i]; 480 disp[0] = 0; 481 ierr = MPI_Type_create_indexed_block(1, nrows_from, (const PetscMPIInt *)disp, MPIU_SCALAR, &type1);CHKERRMPI(ierr); 482 ierr = MPI_Type_create_resized(type1, 0, nz*sizeof(PetscScalar), &rtype[i]);CHKERRMPI(ierr); 483 ierr = MPI_Type_commit(&rtype[i]);CHKERRMPI(ierr); 484 ierr = MPI_Type_free(&type1);CHKERRMPI(ierr); 485 } 486 487 ierr = PetscFree(disp);CHKERRQ(ierr); 488 ierr = VecScatterRestoreRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,NULL,NULL);CHKERRQ(ierr); 489 ierr = VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,NULL,NULL);CHKERRQ(ierr); 490 ierr = MatSetOption(C,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 491 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 492 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 493 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 494 495 C->product->data = contents; 496 C->product->destroy = MatMPIAIJ_MPIDenseDestroy; 497 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 498 PetscFunctionReturn(0); 499 } 500 501 PETSC_INTERN PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat,const PetscBool); 502 /* 503 Performs an efficient scatter on the rows of B needed by this process; this is 504 a modification of the VecScatterBegin_() routines. 505 506 Input: Bbidx = 0: B = Bb 507 = 1: B = Bb1, see MatMatMultSymbolic_MPIAIJ_MPIDense() 508 */ 509 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,PetscInt Bbidx,Mat C,Mat *outworkB) 510 { 511 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 512 PetscErrorCode ierr; 513 const PetscScalar *b; 514 PetscScalar *rvalues; 515 VecScatter ctx = aij->Mvctx; 516 const PetscInt *sindices,*sstarts,*rstarts; 517 const PetscMPIInt *sprocs,*rprocs; 518 PetscInt i,nsends,nrecvs; 519 MPI_Request *swaits,*rwaits; 520 MPI_Comm comm; 521 PetscMPIInt tag=((PetscObject)ctx)->tag,ncols=B->cmap->N,nrows=aij->B->cmap->n,nsends_mpi,nrecvs_mpi; 522 MPIAIJ_MPIDense *contents; 523 Mat workB; 524 MPI_Datatype *stype,*rtype; 525 PetscInt blda; 526 527 PetscFunctionBegin; 528 MatCheckProduct(C,4); 529 if (!C->product->data) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data empty"); 530 contents = (MPIAIJ_MPIDense*)C->product->data; 531 ierr = VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL/*bs*/);CHKERRQ(ierr); 532 ierr = VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,&rprocs,NULL/*bs*/);CHKERRQ(ierr); 533 ierr = PetscMPIIntCast(nsends,&nsends_mpi);CHKERRQ(ierr); 534 ierr = PetscMPIIntCast(nrecvs,&nrecvs_mpi);CHKERRQ(ierr); 535 if (Bbidx == 0) { 536 workB = *outworkB = contents->workB; 537 } else { 538 workB = *outworkB = contents->workB1; 539 } 540 if (nrows != workB->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Number of rows of workB %D not equal to columns of aij->B %D",workB->cmap->n,nrows); 541 swaits = contents->swaits; 542 rwaits = contents->rwaits; 543 544 ierr = MatDenseGetArrayRead(B,&b);CHKERRQ(ierr); 545 ierr = MatDenseGetLDA(B,&blda);CHKERRQ(ierr); 546 if (blda != contents->blda) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot reuse an input matrix with lda %D != %D",blda,contents->blda); 547 ierr = MatDenseGetArray(workB,&rvalues);CHKERRQ(ierr); 548 549 /* Post recv, use MPI derived data type to save memory */ 550 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 551 rtype = contents->rtype; 552 for (i=0; i<nrecvs; i++) { 553 ierr = MPI_Irecv(rvalues+(rstarts[i]-rstarts[0]),ncols,rtype[i],rprocs[i],tag,comm,rwaits+i);CHKERRMPI(ierr); 554 } 555 556 stype = contents->stype; 557 for (i=0; i<nsends; i++) { 558 ierr = MPI_Isend(b,ncols,stype[i],sprocs[i],tag,comm,swaits+i);CHKERRMPI(ierr); 559 } 560 561 if (nrecvs) {ierr = MPI_Waitall(nrecvs_mpi,rwaits,MPI_STATUSES_IGNORE);CHKERRMPI(ierr);} 562 if (nsends) {ierr = MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);CHKERRMPI(ierr);} 563 564 ierr = VecScatterRestoreRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL);CHKERRQ(ierr); 565 ierr = VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,&rprocs,NULL);CHKERRQ(ierr); 566 ierr = MatDenseRestoreArrayRead(B,&b);CHKERRQ(ierr); 567 ierr = MatDenseRestoreArray(workB,&rvalues);CHKERRQ(ierr); 568 PetscFunctionReturn(0); 569 } 570 571 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 572 { 573 PetscErrorCode ierr; 574 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 575 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 576 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 577 Mat workB; 578 MPIAIJ_MPIDense *contents; 579 580 PetscFunctionBegin; 581 MatCheckProduct(C,3); 582 if (!C->product->data) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data empty"); 583 contents = (MPIAIJ_MPIDense*)C->product->data; 584 /* diagonal block of A times all local rows of B */ 585 /* TODO: this calls a symbolic multiplication every time, which could be avoided */ 586 ierr = MatMatMult(aij->A,bdense->A,MAT_REUSE_MATRIX,PETSC_DEFAULT,&cdense->A);CHKERRQ(ierr); 587 if (contents->workB->cmap->n == B->cmap->N) { 588 /* get off processor parts of B needed to complete C=A*B */ 589 ierr = MatMPIDenseScatter(A,B,0,C,&workB);CHKERRQ(ierr); 590 591 /* off-diagonal block of A times nonlocal rows of B */ 592 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A,PETSC_TRUE);CHKERRQ(ierr); 593 } else { 594 Mat Bb,Cb; 595 PetscInt BN=B->cmap->N,n=contents->workB->cmap->n,i; 596 597 for (i=0; i<BN; i+=n) { 598 ierr = MatDenseGetSubMatrix(B,i,PetscMin(i+n,BN),&Bb);CHKERRQ(ierr); 599 ierr = MatDenseGetSubMatrix(C,i,PetscMin(i+n,BN),&Cb);CHKERRQ(ierr); 600 601 /* get off processor parts of B needed to complete C=A*B */ 602 ierr = MatMPIDenseScatter(A,Bb,i+n>BN,C,&workB);CHKERRQ(ierr); 603 604 /* off-diagonal block of A times nonlocal rows of B */ 605 cdense = (Mat_MPIDense*)Cb->data; 606 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A,PETSC_TRUE);CHKERRQ(ierr); 607 608 ierr = MatDenseRestoreSubMatrix(B,&Bb);CHKERRQ(ierr); 609 ierr = MatDenseRestoreSubMatrix(C,&Cb);CHKERRQ(ierr); 610 } 611 } 612 PetscFunctionReturn(0); 613 } 614 615 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 616 { 617 PetscErrorCode ierr; 618 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 619 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 620 Mat_SeqAIJ *cd = (Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 621 PetscInt *adi = ad->i,*adj,*aoi=ao->i,*aoj; 622 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 623 Mat_SeqAIJ *p_loc,*p_oth; 624 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 625 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 626 PetscInt cm = C->rmap->n,anz,pnz; 627 Mat_APMPI *ptap; 628 PetscScalar *apa_sparse; 629 PetscInt *api,*apj,*apJ,i,j,k,row; 630 PetscInt cstart = C->cmap->rstart; 631 PetscInt cdnz,conz,k0,k1,nextp; 632 MPI_Comm comm; 633 PetscMPIInt size; 634 635 PetscFunctionBegin; 636 MatCheckProduct(C,3); 637 ptap = (Mat_APMPI*)C->product->data; 638 if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 639 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 640 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 641 if (!ptap->P_oth && size>1) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatProductClear()"); 642 643 apa_sparse = ptap->apa; 644 645 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 646 /*-----------------------------------------------------*/ 647 /* update numerical values of P_oth and P_loc */ 648 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 649 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 650 651 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 652 /*----------------------------------------------------------*/ 653 /* get data from symbolic products */ 654 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 655 pi_loc = p_loc->i; pj_loc = p_loc->j; pa_loc = p_loc->a; 656 if (size >1) { 657 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 658 pi_oth = p_oth->i; pj_oth = p_oth->j; pa_oth = p_oth->a; 659 } else { 660 p_oth = NULL; pi_oth = NULL; pj_oth = NULL; pa_oth = NULL; 661 } 662 663 api = ptap->api; 664 apj = ptap->apj; 665 for (i=0; i<cm; i++) { 666 apJ = apj + api[i]; 667 668 /* diagonal portion of A */ 669 anz = adi[i+1] - adi[i]; 670 adj = ad->j + adi[i]; 671 ada = ad->a + adi[i]; 672 for (j=0; j<anz; j++) { 673 row = adj[j]; 674 pnz = pi_loc[row+1] - pi_loc[row]; 675 pj = pj_loc + pi_loc[row]; 676 pa = pa_loc + pi_loc[row]; 677 /* perform sparse axpy */ 678 valtmp = ada[j]; 679 nextp = 0; 680 for (k=0; nextp<pnz; k++) { 681 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 682 apa_sparse[k] += valtmp*pa[nextp++]; 683 } 684 } 685 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 686 } 687 688 /* off-diagonal portion of A */ 689 anz = aoi[i+1] - aoi[i]; 690 aoj = ao->j + aoi[i]; 691 aoa = ao->a + aoi[i]; 692 for (j=0; j<anz; j++) { 693 row = aoj[j]; 694 pnz = pi_oth[row+1] - pi_oth[row]; 695 pj = pj_oth + pi_oth[row]; 696 pa = pa_oth + pi_oth[row]; 697 /* perform sparse axpy */ 698 valtmp = aoa[j]; 699 nextp = 0; 700 for (k=0; nextp<pnz; k++) { 701 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 702 apa_sparse[k] += valtmp*pa[nextp++]; 703 } 704 } 705 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 706 } 707 708 /* set values in C */ 709 cdnz = cd->i[i+1] - cd->i[i]; 710 conz = co->i[i+1] - co->i[i]; 711 712 /* 1st off-diagonal part of C */ 713 ca = coa + co->i[i]; 714 k = 0; 715 for (k0=0; k0<conz; k0++) { 716 if (apJ[k] >= cstart) break; 717 ca[k0] = apa_sparse[k]; 718 apa_sparse[k] = 0.0; 719 k++; 720 } 721 722 /* diagonal part of C */ 723 ca = cda + cd->i[i]; 724 for (k1=0; k1<cdnz; k1++) { 725 ca[k1] = apa_sparse[k]; 726 apa_sparse[k] = 0.0; 727 k++; 728 } 729 730 /* 2nd off-diagonal part of C */ 731 ca = coa + co->i[i]; 732 for (; k0<conz; k0++) { 733 ca[k0] = apa_sparse[k]; 734 apa_sparse[k] = 0.0; 735 k++; 736 } 737 } 738 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 739 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 740 PetscFunctionReturn(0); 741 } 742 743 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 744 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat C) 745 { 746 PetscErrorCode ierr; 747 MPI_Comm comm; 748 PetscMPIInt size; 749 Mat_APMPI *ptap; 750 PetscFreeSpaceList free_space = NULL,current_space=NULL; 751 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 752 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 753 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 754 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 755 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=1; 756 PetscInt am=A->rmap->n,pn=P->cmap->n,pm=P->rmap->n,lsize=pn+20; 757 PetscReal afill; 758 MatType mtype; 759 760 PetscFunctionBegin; 761 MatCheckProduct(C,4); 762 if (C->product->data) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 763 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 764 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 765 766 /* create struct Mat_APMPI and attached it to C later */ 767 ierr = PetscNew(&ptap);CHKERRQ(ierr); 768 769 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 770 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 771 772 /* get P_loc by taking all local rows of P */ 773 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 774 775 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 776 pi_loc = p_loc->i; pj_loc = p_loc->j; 777 if (size > 1) { 778 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 779 pi_oth = p_oth->i; pj_oth = p_oth->j; 780 } else { 781 p_oth = NULL; 782 pi_oth = NULL; pj_oth = NULL; 783 } 784 785 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 786 /*-------------------------------------------------------------------*/ 787 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 788 ptap->api = api; 789 api[0] = 0; 790 791 ierr = PetscLLCondensedCreate_Scalable(lsize,&lnk);CHKERRQ(ierr); 792 793 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 794 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 795 current_space = free_space; 796 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 797 for (i=0; i<am; i++) { 798 /* diagonal portion of A */ 799 nzi = adi[i+1] - adi[i]; 800 for (j=0; j<nzi; j++) { 801 row = *adj++; 802 pnz = pi_loc[row+1] - pi_loc[row]; 803 Jptr = pj_loc + pi_loc[row]; 804 /* Expand list if it is not long enough */ 805 if (pnz+apnz_max > lsize) { 806 lsize = pnz+apnz_max; 807 ierr = PetscLLCondensedExpand_Scalable(lsize, &lnk);CHKERRQ(ierr); 808 } 809 /* add non-zero cols of P into the sorted linked list lnk */ 810 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 811 apnz = *lnk; /* The first element in the list is the number of items in the list */ 812 api[i+1] = api[i] + apnz; 813 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 814 } 815 /* off-diagonal portion of A */ 816 nzi = aoi[i+1] - aoi[i]; 817 for (j=0; j<nzi; j++) { 818 row = *aoj++; 819 pnz = pi_oth[row+1] - pi_oth[row]; 820 Jptr = pj_oth + pi_oth[row]; 821 /* Expand list if it is not long enough */ 822 if (pnz+apnz_max > lsize) { 823 lsize = pnz + apnz_max; 824 ierr = PetscLLCondensedExpand_Scalable(lsize, &lnk);CHKERRQ(ierr); 825 } 826 /* add non-zero cols of P into the sorted linked list lnk */ 827 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 828 apnz = *lnk; /* The first element in the list is the number of items in the list */ 829 api[i+1] = api[i] + apnz; 830 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 831 } 832 833 /* add missing diagonal entry */ 834 if (C->force_diagonals) { 835 j = i + rstart; /* column index */ 836 ierr = PetscLLCondensedAddSorted_Scalable(1,&j,lnk);CHKERRQ(ierr); 837 } 838 839 apnz = *lnk; 840 api[i+1] = api[i] + apnz; 841 if (apnz > apnz_max) apnz_max = apnz; 842 843 /* if free space is not available, double the total space in the list */ 844 if (current_space->local_remaining<apnz) { 845 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 846 nspacedouble++; 847 } 848 849 /* Copy data into free space, then initialize lnk */ 850 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 851 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 852 853 current_space->array += apnz; 854 current_space->local_used += apnz; 855 current_space->local_remaining -= apnz; 856 } 857 858 /* Allocate space for apj, initialize apj, and */ 859 /* destroy list of free space and other temporary array(s) */ 860 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 861 apj = ptap->apj; 862 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 863 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 864 865 /* create and assemble symbolic parallel matrix C */ 866 /*----------------------------------------------------*/ 867 ierr = MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 868 ierr = MatSetBlockSizesFromMats(C,A,P);CHKERRQ(ierr); 869 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 870 ierr = MatSetType(C,mtype);CHKERRQ(ierr); 871 ierr = MatMPIAIJSetPreallocation(C,0,dnz,0,onz);CHKERRQ(ierr); 872 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 873 874 /* malloc apa for assembly C */ 875 ierr = PetscCalloc1(apnz_max,&ptap->apa);CHKERRQ(ierr); 876 877 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api);CHKERRQ(ierr); 878 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 879 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 880 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 881 882 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 883 C->ops->productnumeric = MatProductNumeric_AB; 884 885 /* attach the supporting struct to C for reuse */ 886 C->product->data = ptap; 887 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 888 889 /* set MatInfo */ 890 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 891 if (afill < 1.0) afill = 1.0; 892 C->info.mallocs = nspacedouble; 893 C->info.fill_ratio_given = fill; 894 C->info.fill_ratio_needed = afill; 895 896 #if defined(PETSC_USE_INFO) 897 if (api[am]) { 898 ierr = PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 899 ierr = PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 900 } else { 901 ierr = PetscInfo(C,"Empty matrix product\n");CHKERRQ(ierr); 902 } 903 #endif 904 PetscFunctionReturn(0); 905 } 906 907 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 908 /* Three input arrays are merged to one output array. The size of the */ 909 /* output array is also output. Duplicate entries only show up once. */ 910 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, 911 PetscInt size2, PetscInt *in2, 912 PetscInt size3, PetscInt *in3, 913 PetscInt *size4, PetscInt *out) 914 { 915 int i = 0, j = 0, k = 0, l = 0; 916 917 /* Traverse all three arrays */ 918 while (i<size1 && j<size2 && k<size3) { 919 if (in1[i] < in2[j] && in1[i] < in3[k]) { 920 out[l++] = in1[i++]; 921 } 922 else if (in2[j] < in1[i] && in2[j] < in3[k]) { 923 out[l++] = in2[j++]; 924 } 925 else if (in3[k] < in1[i] && in3[k] < in2[j]) { 926 out[l++] = in3[k++]; 927 } 928 else if (in1[i] == in2[j] && in1[i] < in3[k]) { 929 out[l++] = in1[i]; 930 i++, j++; 931 } 932 else if (in1[i] == in3[k] && in1[i] < in2[j]) { 933 out[l++] = in1[i]; 934 i++, k++; 935 } 936 else if (in3[k] == in2[j] && in2[j] < in1[i]) { 937 out[l++] = in2[j]; 938 k++, j++; 939 } 940 else if (in1[i] == in2[j] && in1[i] == in3[k]) { 941 out[l++] = in1[i]; 942 i++, j++, k++; 943 } 944 } 945 946 /* Traverse two remaining arrays */ 947 while (i<size1 && j<size2) { 948 if (in1[i] < in2[j]) { 949 out[l++] = in1[i++]; 950 } 951 else if (in1[i] > in2[j]) { 952 out[l++] = in2[j++]; 953 } 954 else { 955 out[l++] = in1[i]; 956 i++, j++; 957 } 958 } 959 960 while (i<size1 && k<size3) { 961 if (in1[i] < in3[k]) { 962 out[l++] = in1[i++]; 963 } 964 else if (in1[i] > in3[k]) { 965 out[l++] = in3[k++]; 966 } 967 else { 968 out[l++] = in1[i]; 969 i++, k++; 970 } 971 } 972 973 while (k<size3 && j<size2) { 974 if (in3[k] < in2[j]) { 975 out[l++] = in3[k++]; 976 } 977 else if (in3[k] > in2[j]) { 978 out[l++] = in2[j++]; 979 } 980 else { 981 out[l++] = in3[k]; 982 k++, j++; 983 } 984 } 985 986 /* Traverse one remaining array */ 987 while (i<size1) out[l++] = in1[i++]; 988 while (j<size2) out[l++] = in2[j++]; 989 while (k<size3) out[l++] = in3[k++]; 990 991 *size4 = l; 992 } 993 994 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 995 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 996 /* matrix-matrix multiplications. */ 997 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat C) 998 { 999 PetscErrorCode ierr; 1000 MPI_Comm comm; 1001 PetscMPIInt size; 1002 Mat_APMPI *ptap; 1003 PetscFreeSpaceList free_space_diag=NULL, current_space=NULL; 1004 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 1005 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc; 1006 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 1007 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 1008 PetscInt adponz, adpdnz; 1009 PetscInt *pi_loc,*dnz,*onz; 1010 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,rstart=A->rmap->rstart; 1011 PetscInt *lnk,i, i1=0,pnz,row,*adpoi,*adpoj, *api, *adpoJ, *aopJ, *apJ,*Jptr, aopnz, nspacedouble=0,j,nzi, 1012 *apj,apnz, *adpdi, *adpdj, *adpdJ, *poff_i, *poff_j, *j_temp, *aopothi, *aopothj; 1013 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n, p_colstart, p_colend; 1014 PetscBT lnkbt; 1015 PetscReal afill; 1016 PetscMPIInt rank; 1017 Mat adpd, aopoth; 1018 MatType mtype; 1019 const char *prefix; 1020 1021 PetscFunctionBegin; 1022 MatCheckProduct(C,4); 1023 if (C->product->data) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 1024 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1025 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 1026 ierr = MPI_Comm_rank(comm, &rank);CHKERRMPI(ierr); 1027 ierr = MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend);CHKERRQ(ierr); 1028 1029 /* create struct Mat_APMPI and attached it to C later */ 1030 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1031 1032 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1033 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1034 1035 /* get P_loc by taking all local rows of P */ 1036 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1037 1038 1039 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1040 pi_loc = p_loc->i; 1041 1042 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 1043 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 1044 ierr = PetscMalloc1(am+2,&adpoi);CHKERRQ(ierr); 1045 1046 adpoi[0] = 0; 1047 ptap->api = api; 1048 api[0] = 0; 1049 1050 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1051 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 1052 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 1053 1054 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1055 ierr = MatGetOptionsPrefix(A,&prefix);CHKERRQ(ierr); 1056 ierr = MatProductCreate(a->A,p->A,NULL,&adpd);CHKERRQ(ierr); 1057 ierr = MatGetOptionsPrefix(A,&prefix);CHKERRQ(ierr); 1058 ierr = MatSetOptionsPrefix(adpd,prefix);CHKERRQ(ierr); 1059 ierr = MatAppendOptionsPrefix(adpd,"inner_diag_");CHKERRQ(ierr); 1060 1061 ierr = MatProductSetType(adpd,MATPRODUCT_AB);CHKERRQ(ierr); 1062 ierr = MatProductSetAlgorithm(adpd,"sorted");CHKERRQ(ierr); 1063 ierr = MatProductSetFill(adpd,fill);CHKERRQ(ierr); 1064 ierr = MatProductSetFromOptions(adpd);CHKERRQ(ierr); 1065 1066 adpd->force_diagonals = C->force_diagonals; 1067 ierr = MatProductSymbolic(adpd);CHKERRQ(ierr); 1068 1069 adpd_seq = (Mat_SeqAIJ*)((adpd)->data); 1070 adpdi = adpd_seq->i; adpdj = adpd_seq->j; 1071 p_off = (Mat_SeqAIJ*)((p->B)->data); 1072 poff_i = p_off->i; poff_j = p_off->j; 1073 1074 /* j_temp stores indices of a result row before they are added to the linked list */ 1075 ierr = PetscMalloc1(pN+2,&j_temp);CHKERRQ(ierr); 1076 1077 1078 /* Symbolic calc of the A_diag * p_loc_off */ 1079 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1080 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space_diag);CHKERRQ(ierr); 1081 current_space = free_space_diag; 1082 1083 for (i=0; i<am; i++) { 1084 /* A_diag * P_loc_off */ 1085 nzi = adi[i+1] - adi[i]; 1086 for (j=0; j<nzi; j++) { 1087 row = *adj++; 1088 pnz = poff_i[row+1] - poff_i[row]; 1089 Jptr = poff_j + poff_i[row]; 1090 for (i1 = 0; i1 < pnz; i1++) { 1091 j_temp[i1] = p->garray[Jptr[i1]]; 1092 } 1093 /* add non-zero cols of P into the sorted linked list lnk */ 1094 ierr = PetscLLCondensedAddSorted(pnz,j_temp,lnk,lnkbt);CHKERRQ(ierr); 1095 } 1096 1097 adponz = lnk[0]; 1098 adpoi[i+1] = adpoi[i] + adponz; 1099 1100 /* if free space is not available, double the total space in the list */ 1101 if (current_space->local_remaining<adponz) { 1102 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(adponz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1103 nspacedouble++; 1104 } 1105 1106 /* Copy data into free space, then initialize lnk */ 1107 ierr = PetscLLCondensedClean(pN,adponz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1108 1109 current_space->array += adponz; 1110 current_space->local_used += adponz; 1111 current_space->local_remaining -= adponz; 1112 } 1113 1114 /* Symbolic calc of A_off * P_oth */ 1115 ierr = MatSetOptionsPrefix(a->B,prefix);CHKERRQ(ierr); 1116 ierr = MatAppendOptionsPrefix(a->B,"inner_offdiag_");CHKERRQ(ierr); 1117 ierr = MatCreate(PETSC_COMM_SELF,&aopoth);CHKERRQ(ierr); 1118 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, aopoth);CHKERRQ(ierr); 1119 aopoth_seq = (Mat_SeqAIJ*)((aopoth)->data); 1120 aopothi = aopoth_seq->i; aopothj = aopoth_seq->j; 1121 1122 /* Allocate space for apj, adpj, aopj, ... */ 1123 /* destroy lists of free space and other temporary array(s) */ 1124 1125 ierr = PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am]+2, &ptap->apj);CHKERRQ(ierr); 1126 ierr = PetscMalloc1(adpoi[am]+2, &adpoj);CHKERRQ(ierr); 1127 1128 /* Copy from linked list to j-array */ 1129 ierr = PetscFreeSpaceContiguous(&free_space_diag,adpoj);CHKERRQ(ierr); 1130 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1131 1132 adpoJ = adpoj; 1133 adpdJ = adpdj; 1134 aopJ = aopothj; 1135 apj = ptap->apj; 1136 apJ = apj; /* still empty */ 1137 1138 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1139 /* A_diag * P_loc_diag to get A*P */ 1140 for (i = 0; i < am; i++) { 1141 aopnz = aopothi[i+1] - aopothi[i]; 1142 adponz = adpoi[i+1] - adpoi[i]; 1143 adpdnz = adpdi[i+1] - adpdi[i]; 1144 1145 /* Correct indices from A_diag*P_diag */ 1146 for (i1 = 0; i1 < adpdnz; i1++) { 1147 adpdJ[i1] += p_colstart; 1148 } 1149 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1150 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1151 ierr = MatPreallocateSet(i+rstart, apnz, apJ, dnz, onz);CHKERRQ(ierr); 1152 1153 aopJ += aopnz; 1154 adpoJ += adponz; 1155 adpdJ += adpdnz; 1156 apJ += apnz; 1157 api[i+1] = api[i] + apnz; 1158 } 1159 1160 /* malloc apa to store dense row A[i,:]*P */ 1161 ierr = PetscCalloc1(pN+2,&ptap->apa);CHKERRQ(ierr); 1162 1163 /* create and assemble symbolic parallel matrix C */ 1164 ierr = MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1165 ierr = MatSetBlockSizesFromMats(C,A,P);CHKERRQ(ierr); 1166 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 1167 ierr = MatSetType(C,mtype);CHKERRQ(ierr); 1168 ierr = MatMPIAIJSetPreallocation(C,0,dnz,0,onz);CHKERRQ(ierr); 1169 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1170 1171 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api);CHKERRQ(ierr); 1172 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1173 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1174 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1175 1176 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1177 C->ops->productnumeric = MatProductNumeric_AB; 1178 1179 /* attach the supporting struct to C for reuse */ 1180 C->product->data = ptap; 1181 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 1182 1183 /* set MatInfo */ 1184 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 1185 if (afill < 1.0) afill = 1.0; 1186 C->info.mallocs = nspacedouble; 1187 C->info.fill_ratio_given = fill; 1188 C->info.fill_ratio_needed = afill; 1189 1190 #if defined(PETSC_USE_INFO) 1191 if (api[am]) { 1192 ierr = PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 1193 ierr = PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 1194 } else { 1195 ierr = PetscInfo(C,"Empty matrix product\n");CHKERRQ(ierr); 1196 } 1197 #endif 1198 1199 ierr = MatDestroy(&aopoth);CHKERRQ(ierr); 1200 ierr = MatDestroy(&adpd);CHKERRQ(ierr); 1201 ierr = PetscFree(j_temp);CHKERRQ(ierr); 1202 ierr = PetscFree(adpoj);CHKERRQ(ierr); 1203 ierr = PetscFree(adpoi);CHKERRQ(ierr); 1204 PetscFunctionReturn(0); 1205 } 1206 1207 /*-------------------------------------------------------------------------*/ 1208 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1209 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P,Mat A,Mat C) 1210 { 1211 PetscErrorCode ierr; 1212 Mat_APMPI *ptap; 1213 Mat Pt; 1214 1215 PetscFunctionBegin; 1216 MatCheckProduct(C,3); 1217 ptap = (Mat_APMPI*)C->product->data; 1218 if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1219 if (!ptap->Pt) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1220 1221 Pt = ptap->Pt; 1222 ierr = MatTranspose(P,MAT_REUSE_MATRIX,&Pt);CHKERRQ(ierr); 1223 ierr = MatMatMultNumeric_MPIAIJ_MPIAIJ(Pt,A,C);CHKERRQ(ierr); 1224 PetscFunctionReturn(0); 1225 } 1226 1227 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1228 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,PetscReal fill,Mat C) 1229 { 1230 PetscErrorCode ierr; 1231 Mat_APMPI *ptap; 1232 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 1233 MPI_Comm comm; 1234 PetscMPIInt size,rank; 1235 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1236 PetscInt pn=P->cmap->n,aN=A->cmap->N,an=A->cmap->n; 1237 PetscInt *lnk,i,k,nsend,rstart; 1238 PetscBT lnkbt; 1239 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,nrecv; 1240 PETSC_UNUSED PetscMPIInt icompleted=0; 1241 PetscInt **buf_rj,**buf_ri,**buf_ri_k,row,ncols,*cols; 1242 PetscInt len,proc,*dnz,*onz,*owners,nzi; 1243 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1244 MPI_Request *swaits,*rwaits; 1245 MPI_Status *sstatus,rstatus; 1246 PetscLayout rowmap; 1247 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1248 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 1249 PetscInt *Jptr,*prmap=p->garray,con,j,Crmax; 1250 Mat_SeqAIJ *a_loc,*c_loc,*c_oth; 1251 PetscTable ta; 1252 MatType mtype; 1253 const char *prefix; 1254 1255 PetscFunctionBegin; 1256 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1257 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 1258 ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 1259 1260 /* create symbolic parallel matrix C */ 1261 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 1262 ierr = MatSetType(C,mtype);CHKERRQ(ierr); 1263 1264 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1265 1266 /* create struct Mat_APMPI and attached it to C later */ 1267 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1268 ptap->reuse = MAT_INITIAL_MATRIX; 1269 1270 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1271 /* --------------------------------- */ 1272 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1273 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1274 1275 /* (1) compute symbolic A_loc */ 1276 /* ---------------------------*/ 1277 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1278 1279 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1280 /* ------------------------------------ */ 1281 ierr = MatGetOptionsPrefix(A,&prefix);CHKERRQ(ierr); 1282 ierr = MatSetOptionsPrefix(ptap->Ro,prefix);CHKERRQ(ierr); 1283 ierr = MatAppendOptionsPrefix(ptap->Ro,"inner_offdiag_");CHKERRQ(ierr); 1284 ierr = MatCreate(PETSC_COMM_SELF,&ptap->C_oth);CHKERRQ(ierr); 1285 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->A_loc,fill,ptap->C_oth);CHKERRQ(ierr); 1286 1287 /* (3) send coj of C_oth to other processors */ 1288 /* ------------------------------------------ */ 1289 /* determine row ownership */ 1290 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 1291 rowmap->n = pn; 1292 rowmap->bs = 1; 1293 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 1294 owners = rowmap->range; 1295 1296 /* determine the number of messages to send, their lengths */ 1297 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 1298 ierr = PetscArrayzero(len_s,size);CHKERRQ(ierr); 1299 ierr = PetscArrayzero(len_si,size);CHKERRQ(ierr); 1300 1301 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 1302 coi = c_oth->i; coj = c_oth->j; 1303 con = ptap->C_oth->rmap->n; 1304 proc = 0; 1305 for (i=0; i<con; i++) { 1306 while (prmap[i] >= owners[proc+1]) proc++; 1307 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1308 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1309 } 1310 1311 len = 0; /* max length of buf_si[], see (4) */ 1312 owners_co[0] = 0; 1313 nsend = 0; 1314 for (proc=0; proc<size; proc++) { 1315 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1316 if (len_s[proc]) { 1317 nsend++; 1318 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1319 len += len_si[proc]; 1320 } 1321 } 1322 1323 /* determine the number and length of messages to receive for coi and coj */ 1324 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 1325 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 1326 1327 /* post the Irecv and Isend of coj */ 1328 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1329 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1330 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 1331 for (proc=0, k=0; proc<size; proc++) { 1332 if (!len_s[proc]) continue; 1333 i = owners_co[proc]; 1334 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRMPI(ierr); 1335 k++; 1336 } 1337 1338 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1339 /* ---------------------------------------- */ 1340 ierr = MatSetOptionsPrefix(ptap->Rd,prefix);CHKERRQ(ierr); 1341 ierr = MatAppendOptionsPrefix(ptap->Rd,"inner_diag_");CHKERRQ(ierr); 1342 ierr = MatCreate(PETSC_COMM_SELF,&ptap->C_loc);CHKERRQ(ierr); 1343 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->A_loc,fill,ptap->C_loc);CHKERRQ(ierr); 1344 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 1345 1346 /* receives coj are complete */ 1347 for (i=0; i<nrecv; i++) { 1348 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRMPI(ierr); 1349 } 1350 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1351 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRMPI(ierr);} 1352 1353 /* add received column indices into ta to update Crmax */ 1354 a_loc = (Mat_SeqAIJ*)(ptap->A_loc)->data; 1355 1356 /* create and initialize a linked list */ 1357 ierr = PetscTableCreate(an,aN,&ta);CHKERRQ(ierr); /* for compute Crmax */ 1358 MatRowMergeMax_SeqAIJ(a_loc,ptap->A_loc->rmap->N,ta); 1359 1360 for (k=0; k<nrecv; k++) {/* k-th received message */ 1361 Jptr = buf_rj[k]; 1362 for (j=0; j<len_r[k]; j++) { 1363 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1364 } 1365 } 1366 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 1367 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 1368 1369 /* (4) send and recv coi */ 1370 /*-----------------------*/ 1371 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1372 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1373 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1374 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1375 for (proc=0,k=0; proc<size; proc++) { 1376 if (!len_s[proc]) continue; 1377 /* form outgoing message for i-structure: 1378 buf_si[0]: nrows to be sent 1379 [1:nrows]: row index (global) 1380 [nrows+1:2*nrows+1]: i-structure index 1381 */ 1382 /*-------------------------------------------*/ 1383 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 1384 buf_si_i = buf_si + nrows+1; 1385 buf_si[0] = nrows; 1386 buf_si_i[0] = 0; 1387 nrows = 0; 1388 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1389 nzi = coi[i+1] - coi[i]; 1390 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1391 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1392 nrows++; 1393 } 1394 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRMPI(ierr); 1395 k++; 1396 buf_si += len_si[proc]; 1397 } 1398 for (i=0; i<nrecv; i++) { 1399 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRMPI(ierr); 1400 } 1401 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1402 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRMPI(ierr);} 1403 1404 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 1405 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1406 ierr = PetscFree(swaits);CHKERRQ(ierr); 1407 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1408 1409 /* (5) compute the local portion of C */ 1410 /* ------------------------------------------ */ 1411 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of C */ 1412 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 1413 current_space = free_space; 1414 1415 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 1416 for (k=0; k<nrecv; k++) { 1417 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1418 nrows = *buf_ri_k[k]; 1419 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1420 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1421 } 1422 1423 ierr = MatPreallocateInitialize(comm,pn,an,dnz,onz);CHKERRQ(ierr); 1424 ierr = PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt);CHKERRQ(ierr); 1425 for (i=0; i<pn; i++) { /* for each local row of C */ 1426 /* add C_loc into C */ 1427 nzi = c_loc->i[i+1] - c_loc->i[i]; 1428 Jptr = c_loc->j + c_loc->i[i]; 1429 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1430 1431 /* add received col data into lnk */ 1432 for (k=0; k<nrecv; k++) { /* k-th received message */ 1433 if (i == *nextrow[k]) { /* i-th row */ 1434 nzi = *(nextci[k]+1) - *nextci[k]; 1435 Jptr = buf_rj[k] + *nextci[k]; 1436 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1437 nextrow[k]++; nextci[k]++; 1438 } 1439 } 1440 1441 /* add missing diagonal entry */ 1442 if (C->force_diagonals) { 1443 k = i + owners[rank]; /* column index */ 1444 ierr = PetscLLCondensedAddSorted(1,&k,lnk,lnkbt);CHKERRQ(ierr); 1445 } 1446 1447 nzi = lnk[0]; 1448 1449 /* copy data into free space, then initialize lnk */ 1450 ierr = PetscLLCondensedClean(aN,nzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1451 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 1452 } 1453 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1454 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1455 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 1456 1457 /* local sizes and preallocation */ 1458 ierr = MatSetSizes(C,pn,an,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1459 if (P->cmap->bs > 0) {ierr = PetscLayoutSetBlockSize(C->rmap,P->cmap->bs);CHKERRQ(ierr);} 1460 if (A->cmap->bs > 0) {ierr = PetscLayoutSetBlockSize(C->cmap,A->cmap->bs);CHKERRQ(ierr);} 1461 ierr = MatMPIAIJSetPreallocation(C,0,dnz,0,onz);CHKERRQ(ierr); 1462 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1463 1464 /* add C_loc and C_oth to C */ 1465 ierr = MatGetOwnershipRange(C,&rstart,NULL);CHKERRQ(ierr); 1466 for (i=0; i<pn; i++) { 1467 ncols = c_loc->i[i+1] - c_loc->i[i]; 1468 cols = c_loc->j + c_loc->i[i]; 1469 row = rstart + i; 1470 ierr = MatSetValues(C,1,(const PetscInt*)&row,ncols,(const PetscInt*)cols,NULL,INSERT_VALUES);CHKERRQ(ierr); 1471 1472 if (C->force_diagonals) { 1473 ierr = MatSetValues(C,1,(const PetscInt*)&row,1,(const PetscInt*)&row,NULL,INSERT_VALUES);CHKERRQ(ierr); 1474 } 1475 } 1476 for (i=0; i<con; i++) { 1477 ncols = c_oth->i[i+1] - c_oth->i[i]; 1478 cols = c_oth->j + c_oth->i[i]; 1479 row = prmap[i]; 1480 ierr = MatSetValues(C,1,(const PetscInt*)&row,ncols,(const PetscInt*)cols,NULL,INSERT_VALUES);CHKERRQ(ierr); 1481 } 1482 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1483 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1484 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1485 1486 /* members in merge */ 1487 ierr = PetscFree(id_r);CHKERRQ(ierr); 1488 ierr = PetscFree(len_r);CHKERRQ(ierr); 1489 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 1490 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 1491 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 1492 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 1493 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 1494 1495 /* attach the supporting struct to C for reuse */ 1496 C->product->data = ptap; 1497 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 1498 PetscFunctionReturn(0); 1499 } 1500 1501 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,Mat C) 1502 { 1503 PetscErrorCode ierr; 1504 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 1505 Mat_SeqAIJ *c_seq; 1506 Mat_APMPI *ptap; 1507 Mat A_loc,C_loc,C_oth; 1508 PetscInt i,rstart,rend,cm,ncols,row; 1509 const PetscInt *cols; 1510 const PetscScalar *vals; 1511 1512 PetscFunctionBegin; 1513 MatCheckProduct(C,3); 1514 ptap = (Mat_APMPI*)C->product->data; 1515 if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1516 if (!ptap->A_loc) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1517 ierr = MatZeroEntries(C);CHKERRQ(ierr); 1518 1519 if (ptap->reuse == MAT_REUSE_MATRIX) { 1520 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1521 /* 1) get R = Pd^T, Ro = Po^T */ 1522 /*----------------------------*/ 1523 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1524 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1525 1526 /* 2) compute numeric A_loc */ 1527 /*--------------------------*/ 1528 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1529 } 1530 1531 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1532 A_loc = ptap->A_loc; 1533 ierr = ((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,A_loc,ptap->C_loc);CHKERRQ(ierr); 1534 ierr = ((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,A_loc,ptap->C_oth);CHKERRQ(ierr); 1535 C_loc = ptap->C_loc; 1536 C_oth = ptap->C_oth; 1537 1538 /* add C_loc and C_oth to C */ 1539 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 1540 1541 /* C_loc -> C */ 1542 cm = C_loc->rmap->N; 1543 c_seq = (Mat_SeqAIJ*)C_loc->data; 1544 cols = c_seq->j; 1545 vals = c_seq->a; 1546 for (i=0; i<cm; i++) { 1547 ncols = c_seq->i[i+1] - c_seq->i[i]; 1548 row = rstart + i; 1549 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1550 cols += ncols; vals += ncols; 1551 } 1552 1553 /* Co -> C, off-processor part */ 1554 cm = C_oth->rmap->N; 1555 c_seq = (Mat_SeqAIJ*)C_oth->data; 1556 cols = c_seq->j; 1557 vals = c_seq->a; 1558 for (i=0; i<cm; i++) { 1559 ncols = c_seq->i[i+1] - c_seq->i[i]; 1560 row = p->garray[i]; 1561 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1562 cols += ncols; vals += ncols; 1563 } 1564 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1565 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1566 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1567 1568 ptap->reuse = MAT_REUSE_MATRIX; 1569 PetscFunctionReturn(0); 1570 } 1571 1572 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 1573 { 1574 PetscErrorCode ierr; 1575 Mat_Merge_SeqsToMPI *merge; 1576 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 1577 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1578 Mat_APMPI *ptap; 1579 PetscInt *adj; 1580 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1581 MatScalar *ada,*ca,valtmp; 1582 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1583 MPI_Comm comm; 1584 PetscMPIInt size,rank,taga,*len_s; 1585 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1586 PetscInt **buf_ri,**buf_rj; 1587 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1588 MPI_Request *s_waits,*r_waits; 1589 MPI_Status *status; 1590 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1591 PetscInt *ai,*aj,*coi,*coj,*poJ,*pdJ; 1592 Mat A_loc; 1593 Mat_SeqAIJ *a_loc; 1594 1595 PetscFunctionBegin; 1596 MatCheckProduct(C,3); 1597 ptap = (Mat_APMPI*)C->product->data; 1598 if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1599 if (!ptap->A_loc) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1600 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1601 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 1602 ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 1603 1604 merge = ptap->merge; 1605 1606 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1607 /*------------------------------------------*/ 1608 /* get data from symbolic products */ 1609 coi = merge->coi; coj = merge->coj; 1610 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1611 bi = merge->bi; bj = merge->bj; 1612 owners = merge->rowmap->range; 1613 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); 1614 1615 /* get A_loc by taking all local rows of A */ 1616 A_loc = ptap->A_loc; 1617 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 1618 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1619 ai = a_loc->i; 1620 aj = a_loc->j; 1621 1622 for (i=0; i<am; i++) { 1623 anz = ai[i+1] - ai[i]; 1624 adj = aj + ai[i]; 1625 ada = a_loc->a + ai[i]; 1626 1627 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1628 /*-------------------------------------------------------------*/ 1629 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1630 pnz = po->i[i+1] - po->i[i]; 1631 poJ = po->j + po->i[i]; 1632 pA = po->a + po->i[i]; 1633 for (j=0; j<pnz; j++) { 1634 row = poJ[j]; 1635 cj = coj + coi[row]; 1636 ca = coa + coi[row]; 1637 /* perform sparse axpy */ 1638 nexta = 0; 1639 valtmp = pA[j]; 1640 for (k=0; nexta<anz; k++) { 1641 if (cj[k] == adj[nexta]) { 1642 ca[k] += valtmp*ada[nexta]; 1643 nexta++; 1644 } 1645 } 1646 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1647 } 1648 1649 /* put the value into Cd (diagonal part) */ 1650 pnz = pd->i[i+1] - pd->i[i]; 1651 pdJ = pd->j + pd->i[i]; 1652 pA = pd->a + pd->i[i]; 1653 for (j=0; j<pnz; j++) { 1654 row = pdJ[j]; 1655 cj = bj + bi[row]; 1656 ca = ba + bi[row]; 1657 /* perform sparse axpy */ 1658 nexta = 0; 1659 valtmp = pA[j]; 1660 for (k=0; nexta<anz; k++) { 1661 if (cj[k] == adj[nexta]) { 1662 ca[k] += valtmp*ada[nexta]; 1663 nexta++; 1664 } 1665 } 1666 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1667 } 1668 } 1669 1670 /* 3) send and recv matrix values coa */ 1671 /*------------------------------------*/ 1672 buf_ri = merge->buf_ri; 1673 buf_rj = merge->buf_rj; 1674 len_s = merge->len_s; 1675 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1676 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1677 1678 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1679 for (proc=0,k=0; proc<size; proc++) { 1680 if (!len_s[proc]) continue; 1681 i = merge->owners_co[proc]; 1682 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRMPI(ierr); 1683 k++; 1684 } 1685 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRMPI(ierr);} 1686 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRMPI(ierr);} 1687 1688 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1689 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1690 ierr = PetscFree(coa);CHKERRQ(ierr); 1691 1692 /* 4) insert local Cseq and received values into Cmpi */ 1693 /*----------------------------------------------------*/ 1694 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1695 for (k=0; k<merge->nrecv; k++) { 1696 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1697 nrows = *(buf_ri_k[k]); 1698 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1699 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1700 } 1701 1702 for (i=0; i<cm; i++) { 1703 row = owners[rank] + i; /* global row index of C_seq */ 1704 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1705 ba_i = ba + bi[i]; 1706 bnz = bi[i+1] - bi[i]; 1707 /* add received vals into ba */ 1708 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1709 /* i-th row */ 1710 if (i == *nextrow[k]) { 1711 cnz = *(nextci[k]+1) - *nextci[k]; 1712 cj = buf_rj[k] + *(nextci[k]); 1713 ca = abuf_r[k] + *(nextci[k]); 1714 nextcj = 0; 1715 for (j=0; nextcj<cnz; j++) { 1716 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1717 ba_i[j] += ca[nextcj++]; 1718 } 1719 } 1720 nextrow[k]++; nextci[k]++; 1721 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1722 } 1723 } 1724 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1725 } 1726 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1727 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1728 1729 ierr = PetscFree(ba);CHKERRQ(ierr); 1730 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1731 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1732 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1733 PetscFunctionReturn(0); 1734 } 1735 1736 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat C) 1737 { 1738 PetscErrorCode ierr; 1739 Mat A_loc,POt,PDt; 1740 Mat_APMPI *ptap; 1741 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1742 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*a=(Mat_MPIAIJ*)A->data; 1743 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1744 PetscInt nnz; 1745 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1746 PetscInt am =A->rmap->n,pn=P->cmap->n; 1747 MPI_Comm comm; 1748 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1749 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1750 PetscInt len,proc,*dnz,*onz,*owners; 1751 PetscInt nzi,*bi,*bj; 1752 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1753 MPI_Request *swaits,*rwaits; 1754 MPI_Status *sstatus,rstatus; 1755 Mat_Merge_SeqsToMPI *merge; 1756 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1757 PetscReal afill =1.0,afill_tmp; 1758 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Armax; 1759 Mat_SeqAIJ *a_loc,*pdt,*pot; 1760 PetscTable ta; 1761 MatType mtype; 1762 1763 PetscFunctionBegin; 1764 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1765 /* check if matrix local sizes are compatible */ 1766 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); 1767 1768 ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 1769 ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 1770 1771 /* create struct Mat_APMPI and attached it to C later */ 1772 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1773 1774 /* get A_loc by taking all local rows of A */ 1775 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1776 1777 ptap->A_loc = A_loc; 1778 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1779 ai = a_loc->i; 1780 aj = a_loc->j; 1781 1782 /* determine symbolic Co=(p->B)^T*A - send to others */ 1783 /*----------------------------------------------------*/ 1784 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1785 pdt = (Mat_SeqAIJ*)PDt->data; 1786 pdti = pdt->i; pdtj = pdt->j; 1787 1788 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1789 pot = (Mat_SeqAIJ*)POt->data; 1790 poti = pot->i; potj = pot->j; 1791 1792 /* then, compute symbolic Co = (p->B)^T*A */ 1793 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1794 >= (num of nonzero rows of C_seq) - pn */ 1795 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 1796 coi[0] = 0; 1797 1798 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1799 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(poti[pon],ai[am])); 1800 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1801 current_space = free_space; 1802 1803 /* create and initialize a linked list */ 1804 ierr = PetscTableCreate(A->cmap->n + a->B->cmap->N,aN,&ta);CHKERRQ(ierr); 1805 MatRowMergeMax_SeqAIJ(a_loc,am,ta); 1806 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1807 1808 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1809 1810 for (i=0; i<pon; i++) { 1811 pnz = poti[i+1] - poti[i]; 1812 ptJ = potj + poti[i]; 1813 for (j=0; j<pnz; j++) { 1814 row = ptJ[j]; /* row of A_loc == col of Pot */ 1815 anz = ai[row+1] - ai[row]; 1816 Jptr = aj + ai[row]; 1817 /* add non-zero cols of AP into the sorted linked list lnk */ 1818 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1819 } 1820 nnz = lnk[0]; 1821 1822 /* If free space is not available, double the total space in the list */ 1823 if (current_space->local_remaining<nnz) { 1824 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1825 nspacedouble++; 1826 } 1827 1828 /* Copy data into free space, and zero out denserows */ 1829 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1830 1831 current_space->array += nnz; 1832 current_space->local_used += nnz; 1833 current_space->local_remaining -= nnz; 1834 1835 coi[i+1] = coi[i] + nnz; 1836 } 1837 1838 ierr = PetscMalloc1(coi[pon]+1,&coj);CHKERRQ(ierr); 1839 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1840 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); /* must destroy to get a new one for C */ 1841 1842 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1843 if (afill_tmp > afill) afill = afill_tmp; 1844 1845 /* send j-array (coj) of Co to other processors */ 1846 /*----------------------------------------------*/ 1847 /* determine row ownership */ 1848 ierr = PetscNew(&merge);CHKERRQ(ierr); 1849 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1850 1851 merge->rowmap->n = pn; 1852 merge->rowmap->bs = 1; 1853 1854 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1855 owners = merge->rowmap->range; 1856 1857 /* determine the number of messages to send, their lengths */ 1858 ierr = PetscCalloc1(size,&len_si);CHKERRQ(ierr); 1859 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 1860 1861 len_s = merge->len_s; 1862 merge->nsend = 0; 1863 1864 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 1865 1866 proc = 0; 1867 for (i=0; i<pon; i++) { 1868 while (prmap[i] >= owners[proc+1]) proc++; 1869 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1870 len_s[proc] += coi[i+1] - coi[i]; 1871 } 1872 1873 len = 0; /* max length of buf_si[] */ 1874 owners_co[0] = 0; 1875 for (proc=0; proc<size; proc++) { 1876 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1877 if (len_si[proc]) { 1878 merge->nsend++; 1879 len_si[proc] = 2*(len_si[proc] + 1); 1880 len += len_si[proc]; 1881 } 1882 } 1883 1884 /* determine the number and length of messages to receive for coi and coj */ 1885 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1886 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1887 1888 /* post the Irecv and Isend of coj */ 1889 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1890 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1891 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 1892 for (proc=0, k=0; proc<size; proc++) { 1893 if (!len_s[proc]) continue; 1894 i = owners_co[proc]; 1895 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRMPI(ierr); 1896 k++; 1897 } 1898 1899 /* receives and sends of coj are complete */ 1900 ierr = PetscMalloc1(size,&sstatus);CHKERRQ(ierr); 1901 for (i=0; i<merge->nrecv; i++) { 1902 PETSC_UNUSED PetscMPIInt icompleted; 1903 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRMPI(ierr); 1904 } 1905 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1906 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRMPI(ierr);} 1907 1908 /* add received column indices into table to update Armax */ 1909 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/tutorials/ex56.c! */ 1910 for (k=0; k<merge->nrecv; k++) {/* k-th received message */ 1911 Jptr = buf_rj[k]; 1912 for (j=0; j<merge->len_r[k]; j++) { 1913 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1914 } 1915 } 1916 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1917 /* 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); */ 1918 1919 /* send and recv coi */ 1920 /*-------------------*/ 1921 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1922 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1923 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1924 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1925 for (proc=0,k=0; proc<size; proc++) { 1926 if (!len_s[proc]) continue; 1927 /* form outgoing message for i-structure: 1928 buf_si[0]: nrows to be sent 1929 [1:nrows]: row index (global) 1930 [nrows+1:2*nrows+1]: i-structure index 1931 */ 1932 /*-------------------------------------------*/ 1933 nrows = len_si[proc]/2 - 1; 1934 buf_si_i = buf_si + nrows+1; 1935 buf_si[0] = nrows; 1936 buf_si_i[0] = 0; 1937 nrows = 0; 1938 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1939 nzi = coi[i+1] - coi[i]; 1940 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1941 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1942 nrows++; 1943 } 1944 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRMPI(ierr); 1945 k++; 1946 buf_si += len_si[proc]; 1947 } 1948 i = merge->nrecv; 1949 while (i--) { 1950 PETSC_UNUSED PetscMPIInt icompleted; 1951 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRMPI(ierr); 1952 } 1953 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1954 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRMPI(ierr);} 1955 ierr = PetscFree(len_si);CHKERRQ(ierr); 1956 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1957 ierr = PetscFree(swaits);CHKERRQ(ierr); 1958 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1959 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1960 1961 /* compute the local portion of C (mpi mat) */ 1962 /*------------------------------------------*/ 1963 /* allocate bi array and free space for accumulating nonzero column info */ 1964 ierr = PetscMalloc1(pn+1,&bi);CHKERRQ(ierr); 1965 bi[0] = 0; 1966 1967 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1968 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(pdti[pn],PetscIntSumTruncate(poti[pon],ai[am]))); 1969 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1970 current_space = free_space; 1971 1972 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1973 for (k=0; k<merge->nrecv; k++) { 1974 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1975 nrows = *buf_ri_k[k]; 1976 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1977 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th received i-structure */ 1978 } 1979 1980 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1981 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1982 rmax = 0; 1983 for (i=0; i<pn; i++) { 1984 /* add pdt[i,:]*AP into lnk */ 1985 pnz = pdti[i+1] - pdti[i]; 1986 ptJ = pdtj + pdti[i]; 1987 for (j=0; j<pnz; j++) { 1988 row = ptJ[j]; /* row of AP == col of Pt */ 1989 anz = ai[row+1] - ai[row]; 1990 Jptr = aj + ai[row]; 1991 /* add non-zero cols of AP into the sorted linked list lnk */ 1992 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1993 } 1994 1995 /* add received col data into lnk */ 1996 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1997 if (i == *nextrow[k]) { /* i-th row */ 1998 nzi = *(nextci[k]+1) - *nextci[k]; 1999 Jptr = buf_rj[k] + *nextci[k]; 2000 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 2001 nextrow[k]++; nextci[k]++; 2002 } 2003 } 2004 2005 /* add missing diagonal entry */ 2006 if (C->force_diagonals) { 2007 k = i + owners[rank]; /* column index */ 2008 ierr = PetscLLCondensedAddSorted_Scalable(1,&k,lnk);CHKERRQ(ierr); 2009 } 2010 2011 nnz = lnk[0]; 2012 2013 /* if free space is not available, make more free space */ 2014 if (current_space->local_remaining<nnz) { 2015 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 2016 nspacedouble++; 2017 } 2018 /* copy data into free space, then initialize lnk */ 2019 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 2020 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 2021 2022 current_space->array += nnz; 2023 current_space->local_used += nnz; 2024 current_space->local_remaining -= nnz; 2025 2026 bi[i+1] = bi[i] + nnz; 2027 if (nnz > rmax) rmax = nnz; 2028 } 2029 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 2030 2031 ierr = PetscMalloc1(bi[pn]+1,&bj);CHKERRQ(ierr); 2032 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 2033 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 2034 if (afill_tmp > afill) afill = afill_tmp; 2035 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 2036 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 2037 2038 ierr = MatDestroy(&POt);CHKERRQ(ierr); 2039 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 2040 2041 /* create symbolic parallel matrix C - why cannot be assembled in Numeric part */ 2042 /*-------------------------------------------------------------------------------*/ 2043 ierr = MatSetSizes(C,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2044 ierr = MatSetBlockSizes(C,PetscAbs(P->cmap->bs),PetscAbs(A->cmap->bs));CHKERRQ(ierr); 2045 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 2046 ierr = MatSetType(C,mtype);CHKERRQ(ierr); 2047 ierr = MatMPIAIJSetPreallocation(C,0,dnz,0,onz);CHKERRQ(ierr); 2048 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 2049 ierr = MatSetBlockSize(C,1);CHKERRQ(ierr); 2050 ierr = MatSetOption(C,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 2051 for (i=0; i<pn; i++) { 2052 row = i + rstart; 2053 nnz = bi[i+1] - bi[i]; 2054 Jptr = bj + bi[i]; 2055 ierr = MatSetValues(C,1,&row,nnz,Jptr,NULL,INSERT_VALUES);CHKERRQ(ierr); 2056 } 2057 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2058 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2059 ierr = MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2060 merge->bi = bi; 2061 merge->bj = bj; 2062 merge->coi = coi; 2063 merge->coj = coj; 2064 merge->buf_ri = buf_ri; 2065 merge->buf_rj = buf_rj; 2066 merge->owners_co = owners_co; 2067 2068 /* attach the supporting struct to C for reuse */ 2069 C->product->data = ptap; 2070 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2071 ptap->merge = merge; 2072 2073 C->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2074 2075 #if defined(PETSC_USE_INFO) 2076 if (bi[pn] != 0) { 2077 ierr = PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 2078 ierr = PetscInfo1(C,"Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 2079 } else { 2080 ierr = PetscInfo(C,"Empty matrix product\n");CHKERRQ(ierr); 2081 } 2082 #endif 2083 PetscFunctionReturn(0); 2084 } 2085 2086 /* ---------------------------------------------------------------- */ 2087 static PetscErrorCode MatProductSymbolic_AtB_MPIAIJ_MPIAIJ(Mat C) 2088 { 2089 PetscErrorCode ierr; 2090 Mat_Product *product = C->product; 2091 Mat A=product->A,B=product->B; 2092 PetscReal fill=product->fill; 2093 PetscBool flg; 2094 2095 PetscFunctionBegin; 2096 /* scalable */ 2097 ierr = PetscStrcmp(product->alg,"scalable",&flg);CHKERRQ(ierr); 2098 if (flg) { 2099 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 2100 goto next; 2101 } 2102 2103 /* nonscalable */ 2104 ierr = PetscStrcmp(product->alg,"nonscalable",&flg);CHKERRQ(ierr); 2105 if (flg) { 2106 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C);CHKERRQ(ierr); 2107 goto next; 2108 } 2109 2110 /* matmatmult */ 2111 ierr = PetscStrcmp(product->alg,"at*b",&flg);CHKERRQ(ierr); 2112 if (flg) { 2113 Mat At; 2114 Mat_APMPI *ptap; 2115 2116 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 2117 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ(At,B,fill,C);CHKERRQ(ierr); 2118 ptap = (Mat_APMPI*)C->product->data; 2119 if (ptap) { 2120 ptap->Pt = At; 2121 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2122 } 2123 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 2124 goto next; 2125 } 2126 2127 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatProduct type is not supported"); 2128 2129 next: 2130 C->ops->productnumeric = MatProductNumeric_AtB; 2131 PetscFunctionReturn(0); 2132 } 2133 2134 /* ---------------------------------------------------------------- */ 2135 /* Set options for MatMatMultxxx_MPIAIJ_MPIAIJ */ 2136 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AB(Mat C) 2137 { 2138 PetscErrorCode ierr; 2139 Mat_Product *product = C->product; 2140 Mat A=product->A,B=product->B; 2141 #if defined(PETSC_HAVE_HYPRE) 2142 const char *algTypes[4] = {"scalable","nonscalable","seqmpi","hypre"}; 2143 PetscInt nalg = 4; 2144 #else 2145 const char *algTypes[3] = {"scalable","nonscalable","seqmpi"}; 2146 PetscInt nalg = 3; 2147 #endif 2148 PetscInt alg = 1; /* set nonscalable algorithm as default */ 2149 PetscBool flg; 2150 MPI_Comm comm; 2151 2152 PetscFunctionBegin; 2153 /* Check matrix local sizes */ 2154 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 2155 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) SETERRQ4(PETSC_COMM_SELF,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); 2156 2157 /* Set "nonscalable" as default algorithm */ 2158 ierr = PetscStrcmp(C->product->alg,"default",&flg);CHKERRQ(ierr); 2159 if (flg) { 2160 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2161 2162 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2163 if (B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2164 MatInfo Ainfo,Binfo; 2165 PetscInt nz_local; 2166 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2167 2168 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 2169 ierr = MatGetInfo(B,MAT_LOCAL,&Binfo);CHKERRQ(ierr); 2170 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2171 2172 if (B->cmap->N > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2173 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 2174 2175 if (alg_scalable) { 2176 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2177 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2178 ierr = PetscInfo2(B,"Use scalable algorithm, BN %D, fill*nz_allocated %g\n",B->cmap->N,product->fill*nz_local);CHKERRQ(ierr); 2179 } 2180 } 2181 } 2182 2183 /* Get runtime option */ 2184 if (product->api_user) { 2185 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMult","Mat");CHKERRQ(ierr); 2186 ierr = PetscOptionsEList("-matmatmult_via","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2187 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2188 } else { 2189 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AB","Mat");CHKERRQ(ierr); 2190 ierr = PetscOptionsEList("-matproduct_ab_via","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2191 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2192 } 2193 if (flg) { 2194 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2195 } 2196 2197 C->ops->productsymbolic = MatProductSymbolic_AB_MPIAIJ_MPIAIJ; 2198 PetscFunctionReturn(0); 2199 } 2200 2201 /* Set options for MatTransposeMatMultXXX_MPIAIJ_MPIAIJ */ 2202 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AtB(Mat C) 2203 { 2204 PetscErrorCode ierr; 2205 Mat_Product *product = C->product; 2206 Mat A=product->A,B=product->B; 2207 const char *algTypes[3] = {"scalable","nonscalable","at*b"}; 2208 PetscInt nalg = 3; 2209 PetscInt alg = 1; /* set default algorithm */ 2210 PetscBool flg; 2211 MPI_Comm comm; 2212 2213 PetscFunctionBegin; 2214 /* Check matrix local sizes */ 2215 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 2216 if (A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A (%D, %D) != B (%D,%D)",A->rmap->rstart,A->rmap->rend,B->rmap->rstart,B->rmap->rend); 2217 2218 /* Set default algorithm */ 2219 ierr = PetscStrcmp(C->product->alg,"default",&flg);CHKERRQ(ierr); 2220 if (flg) { 2221 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2222 } 2223 2224 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2225 if (alg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2226 MatInfo Ainfo,Binfo; 2227 PetscInt nz_local; 2228 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2229 2230 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 2231 ierr = MatGetInfo(B,MAT_LOCAL,&Binfo);CHKERRQ(ierr); 2232 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2233 2234 if (B->cmap->N > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2235 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 2236 2237 if (alg_scalable) { 2238 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2239 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2240 ierr = PetscInfo2(B,"Use scalable algorithm, BN %D, fill*nz_allocated %g\n",B->cmap->N,product->fill*nz_local);CHKERRQ(ierr); 2241 } 2242 } 2243 2244 /* Get runtime option */ 2245 if (product->api_user) { 2246 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatTransposeMatMult","Mat");CHKERRQ(ierr); 2247 ierr = PetscOptionsEList("-mattransposematmult_via","Algorithmic approach","MatTransposeMatMult",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2248 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2249 } else { 2250 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AtB","Mat");CHKERRQ(ierr); 2251 ierr = PetscOptionsEList("-matproduct_atb_via","Algorithmic approach","MatTransposeMatMult",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2252 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2253 } 2254 if (flg) { 2255 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2256 } 2257 2258 C->ops->productsymbolic = MatProductSymbolic_AtB_MPIAIJ_MPIAIJ; 2259 PetscFunctionReturn(0); 2260 } 2261 2262 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_PtAP(Mat C) 2263 { 2264 PetscErrorCode ierr; 2265 Mat_Product *product = C->product; 2266 Mat A=product->A,P=product->B; 2267 MPI_Comm comm; 2268 PetscBool flg; 2269 PetscInt alg=1; /* set default algorithm */ 2270 #if !defined(PETSC_HAVE_HYPRE) 2271 const char *algTypes[4] = {"scalable","nonscalable","allatonce","allatonce_merged"}; 2272 PetscInt nalg=4; 2273 #else 2274 const char *algTypes[5] = {"scalable","nonscalable","allatonce","allatonce_merged","hypre"}; 2275 PetscInt nalg=5; 2276 #endif 2277 PetscInt pN=P->cmap->N; 2278 2279 PetscFunctionBegin; 2280 /* Check matrix local sizes */ 2281 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 2282 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%D, %D) != Prow (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 2283 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%D, %D) != Prow (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 2284 2285 /* Set "nonscalable" as default algorithm */ 2286 ierr = PetscStrcmp(C->product->alg,"default",&flg);CHKERRQ(ierr); 2287 if (flg) { 2288 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2289 2290 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2291 if (pN > 100000) { 2292 MatInfo Ainfo,Pinfo; 2293 PetscInt nz_local; 2294 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2295 2296 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 2297 ierr = MatGetInfo(P,MAT_LOCAL,&Pinfo);CHKERRQ(ierr); 2298 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); 2299 2300 if (pN > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2301 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 2302 2303 if (alg_scalable) { 2304 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2305 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2306 } 2307 } 2308 } 2309 2310 /* Get runtime option */ 2311 if (product->api_user) { 2312 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatPtAP","Mat");CHKERRQ(ierr); 2313 ierr = PetscOptionsEList("-matptap_via","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2314 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2315 } else { 2316 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_PtAP","Mat");CHKERRQ(ierr); 2317 ierr = PetscOptionsEList("-matproduct_ptap_via","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2318 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2319 } 2320 if (flg) { 2321 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2322 } 2323 2324 C->ops->productsymbolic = MatProductSymbolic_PtAP_MPIAIJ_MPIAIJ; 2325 PetscFunctionReturn(0); 2326 } 2327 2328 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_RARt(Mat C) 2329 { 2330 Mat_Product *product = C->product; 2331 Mat A = product->A,R=product->B; 2332 2333 PetscFunctionBegin; 2334 /* Check matrix local sizes */ 2335 if (A->cmap->n != R->cmap->n || A->rmap->n != R->cmap->n) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A local (%D, %D), R local (%D,%D)",A->rmap->n,A->rmap->n,R->rmap->n,R->cmap->n); 2336 2337 C->ops->productsymbolic = MatProductSymbolic_RARt_MPIAIJ_MPIAIJ; 2338 PetscFunctionReturn(0); 2339 } 2340 2341 /* 2342 Set options for ABC = A*B*C = A*(B*C); ABC's algorithm must be chosen from AB's algorithm 2343 */ 2344 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABC(Mat C) 2345 { 2346 PetscErrorCode ierr; 2347 Mat_Product *product = C->product; 2348 PetscBool flg = PETSC_FALSE; 2349 PetscInt alg = 1; /* default algorithm */ 2350 const char *algTypes[3] = {"scalable","nonscalable","seqmpi"}; 2351 PetscInt nalg = 3; 2352 2353 PetscFunctionBegin; 2354 /* Set default algorithm */ 2355 ierr = PetscStrcmp(C->product->alg,"default",&flg);CHKERRQ(ierr); 2356 if (flg) { 2357 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2358 } 2359 2360 /* Get runtime option */ 2361 if (product->api_user) { 2362 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMatMult","Mat");CHKERRQ(ierr); 2363 ierr = PetscOptionsEList("-matmatmatmult_via","Algorithmic approach","MatMatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2364 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2365 } else { 2366 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_ABC","Mat");CHKERRQ(ierr); 2367 ierr = PetscOptionsEList("-matproduct_abc_via","Algorithmic approach","MatProduct_ABC",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 2368 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2369 } 2370 if (flg) { 2371 ierr = MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);CHKERRQ(ierr); 2372 } 2373 2374 C->ops->matmatmultsymbolic = MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ; 2375 C->ops->productsymbolic = MatProductSymbolic_ABC; 2376 PetscFunctionReturn(0); 2377 } 2378 2379 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ(Mat C) 2380 { 2381 PetscErrorCode ierr; 2382 Mat_Product *product = C->product; 2383 2384 PetscFunctionBegin; 2385 switch (product->type) { 2386 case MATPRODUCT_AB: 2387 ierr = MatProductSetFromOptions_MPIAIJ_AB(C);CHKERRQ(ierr); 2388 break; 2389 case MATPRODUCT_AtB: 2390 ierr = MatProductSetFromOptions_MPIAIJ_AtB(C);CHKERRQ(ierr); 2391 break; 2392 case MATPRODUCT_PtAP: 2393 ierr = MatProductSetFromOptions_MPIAIJ_PtAP(C);CHKERRQ(ierr); 2394 break; 2395 case MATPRODUCT_RARt: 2396 ierr = MatProductSetFromOptions_MPIAIJ_RARt(C);CHKERRQ(ierr); 2397 break; 2398 case MATPRODUCT_ABC: 2399 ierr = MatProductSetFromOptions_MPIAIJ_ABC(C);CHKERRQ(ierr); 2400 break; 2401 default: 2402 break; 2403 } 2404 PetscFunctionReturn(0); 2405 } 2406