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 Mat_Product *product = C->product; 21 Mat A=product->A,B=product->B; 22 MatProductAlgorithm alg=product->alg; 23 PetscReal fill=product->fill; 24 PetscBool flg; 25 26 PetscFunctionBegin; 27 /* scalable */ 28 PetscCall(PetscStrcmp(alg,"scalable",&flg)); 29 if (flg) { 30 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C)); 31 PetscFunctionReturn(0); 32 } 33 34 /* nonscalable */ 35 PetscCall(PetscStrcmp(alg,"nonscalable",&flg)); 36 if (flg) { 37 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C)); 38 PetscFunctionReturn(0); 39 } 40 41 /* seqmpi */ 42 PetscCall(PetscStrcmp(alg,"seqmpi",&flg)); 43 if (flg) { 44 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(A,B,fill,C)); 45 PetscFunctionReturn(0); 46 } 47 48 /* backend general code */ 49 PetscCall(PetscStrcmp(alg,"backend",&flg)); 50 if (flg) { 51 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 52 PetscFunctionReturn(0); 53 } 54 55 #if defined(PETSC_HAVE_HYPRE) 56 PetscCall(PetscStrcmp(alg,"hypre",&flg)); 57 if (flg) { 58 PetscCall(MatMatMultSymbolic_AIJ_AIJ_wHYPRE(A,B,fill,C)); 59 PetscFunctionReturn(0); 60 } 61 #endif 62 SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_SUP,"Mat Product Algorithm is not supported"); 63 } 64 65 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(void *data) 66 { 67 Mat_APMPI *ptap = (Mat_APMPI*)data; 68 69 PetscFunctionBegin; 70 PetscCall(PetscFree2(ptap->startsj_s,ptap->startsj_r)); 71 PetscCall(PetscFree(ptap->bufa)); 72 PetscCall(MatDestroy(&ptap->P_loc)); 73 PetscCall(MatDestroy(&ptap->P_oth)); 74 PetscCall(MatDestroy(&ptap->Pt)); 75 PetscCall(PetscFree(ptap->api)); 76 PetscCall(PetscFree(ptap->apj)); 77 PetscCall(PetscFree(ptap->apa)); 78 PetscCall(PetscFree(ptap)); 79 PetscFunctionReturn(0); 80 } 81 82 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,Mat C) 83 { 84 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 85 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 86 Mat_SeqAIJ *cd =(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 87 PetscScalar *cda=cd->a,*coa=co->a; 88 Mat_SeqAIJ *p_loc,*p_oth; 89 PetscScalar *apa,*ca; 90 PetscInt cm =C->rmap->n; 91 Mat_APMPI *ptap; 92 PetscInt *api,*apj,*apJ,i,k; 93 PetscInt cstart=C->cmap->rstart; 94 PetscInt cdnz,conz,k0,k1; 95 const PetscScalar *dummy; 96 MPI_Comm comm; 97 PetscMPIInt size; 98 99 PetscFunctionBegin; 100 MatCheckProduct(C,3); 101 ptap = (Mat_APMPI*)C->product->data; 102 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 103 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 104 PetscCallMPI(MPI_Comm_size(comm,&size)); 105 106 PetscCheckFalse(!ptap->P_oth && size>1,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatProductClear()"); 107 108 /* flag CPU mask for C */ 109 #if defined(PETSC_HAVE_DEVICE) 110 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 111 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 112 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 113 #endif 114 115 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 116 /*-----------------------------------------------------*/ 117 /* update numerical values of P_oth and P_loc */ 118 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 119 PetscCall(MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc)); 120 121 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 122 /*----------------------------------------------------------*/ 123 /* get data from symbolic products */ 124 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 125 p_oth = NULL; 126 if (size >1) { 127 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 128 } 129 130 /* get apa for storing dense row A[i,:]*P */ 131 apa = ptap->apa; 132 133 api = ptap->api; 134 apj = ptap->apj; 135 /* trigger copy to CPU */ 136 PetscCall(MatSeqAIJGetArrayRead(a->A,&dummy)); 137 PetscCall(MatSeqAIJRestoreArrayRead(a->A,&dummy)); 138 PetscCall(MatSeqAIJGetArrayRead(a->B,&dummy)); 139 PetscCall(MatSeqAIJRestoreArrayRead(a->B,&dummy)); 140 for (i=0; i<cm; i++) { 141 /* compute apa = A[i,:]*P */ 142 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 143 144 /* set values in C */ 145 apJ = apj + api[i]; 146 cdnz = cd->i[i+1] - cd->i[i]; 147 conz = co->i[i+1] - co->i[i]; 148 149 /* 1st off-diagonal part of C */ 150 ca = coa + co->i[i]; 151 k = 0; 152 for (k0=0; k0<conz; k0++) { 153 if (apJ[k] >= cstart) break; 154 ca[k0] = apa[apJ[k]]; 155 apa[apJ[k++]] = 0.0; 156 } 157 158 /* diagonal part of C */ 159 ca = cda + cd->i[i]; 160 for (k1=0; k1<cdnz; k1++) { 161 ca[k1] = apa[apJ[k]]; 162 apa[apJ[k++]] = 0.0; 163 } 164 165 /* 2nd off-diagonal part of C */ 166 ca = coa + co->i[i]; 167 for (; k0<conz; k0++) { 168 ca[k0] = apa[apJ[k]]; 169 apa[apJ[k++]] = 0.0; 170 } 171 } 172 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 173 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 174 PetscFunctionReturn(0); 175 } 176 177 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,PetscReal fill,Mat C) 178 { 179 PetscErrorCode ierr; 180 MPI_Comm comm; 181 PetscMPIInt size; 182 Mat_APMPI *ptap; 183 PetscFreeSpaceList free_space=NULL,current_space=NULL; 184 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 185 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 186 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 187 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 188 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 189 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 190 PetscBT lnkbt; 191 PetscReal afill; 192 MatType mtype; 193 194 PetscFunctionBegin; 195 MatCheckProduct(C,4); 196 PetscCheck(!C->product->data,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 197 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 198 PetscCallMPI(MPI_Comm_size(comm,&size)); 199 200 /* create struct Mat_APMPI and attached it to C later */ 201 PetscCall(PetscNew(&ptap)); 202 203 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 204 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 205 206 /* get P_loc by taking all local rows of P */ 207 PetscCall(MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc)); 208 209 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 210 pi_loc = p_loc->i; pj_loc = p_loc->j; 211 if (size > 1) { 212 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 213 pi_oth = p_oth->i; pj_oth = p_oth->j; 214 } else { 215 p_oth = NULL; 216 pi_oth = NULL; pj_oth = NULL; 217 } 218 219 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 220 /*-------------------------------------------------------------------*/ 221 PetscCall(PetscMalloc1(am+2,&api)); 222 ptap->api = api; 223 api[0] = 0; 224 225 /* create and initialize a linked list */ 226 PetscCall(PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt)); 227 228 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 229 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space)); 230 current_space = free_space; 231 232 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);PetscCall(ierr); 233 for (i=0; i<am; i++) { 234 /* diagonal portion of A */ 235 nzi = adi[i+1] - adi[i]; 236 for (j=0; j<nzi; j++) { 237 row = *adj++; 238 pnz = pi_loc[row+1] - pi_loc[row]; 239 Jptr = pj_loc + pi_loc[row]; 240 /* add non-zero cols of P into the sorted linked list lnk */ 241 PetscCall(PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt)); 242 } 243 /* off-diagonal portion of A */ 244 nzi = aoi[i+1] - aoi[i]; 245 for (j=0; j<nzi; j++) { 246 row = *aoj++; 247 pnz = pi_oth[row+1] - pi_oth[row]; 248 Jptr = pj_oth + pi_oth[row]; 249 PetscCall(PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt)); 250 } 251 /* add possible missing diagonal entry */ 252 if (C->force_diagonals) { 253 j = i + rstart; /* column index */ 254 PetscCall(PetscLLCondensedAddSorted(1,&j,lnk,lnkbt)); 255 } 256 257 apnz = lnk[0]; 258 api[i+1] = api[i] + apnz; 259 260 /* if free space is not available, double the total space in the list */ 261 if (current_space->local_remaining<apnz) { 262 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space)); 263 nspacedouble++; 264 } 265 266 /* Copy data into free space, then initialize lnk */ 267 PetscCall(PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt)); 268 PetscCall(MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz)); 269 270 current_space->array += apnz; 271 current_space->local_used += apnz; 272 current_space->local_remaining -= apnz; 273 } 274 275 /* Allocate space for apj, initialize apj, and */ 276 /* destroy list of free space and other temporary array(s) */ 277 PetscCall(PetscMalloc1(api[am]+1,&ptap->apj)); 278 apj = ptap->apj; 279 PetscCall(PetscFreeSpaceContiguous(&free_space,ptap->apj)); 280 PetscCall(PetscLLDestroy(lnk,lnkbt)); 281 282 /* malloc apa to store dense row A[i,:]*P */ 283 PetscCall(PetscCalloc1(pN,&ptap->apa)); 284 285 /* set and assemble symbolic parallel matrix C */ 286 /*---------------------------------------------*/ 287 PetscCall(MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE)); 288 PetscCall(MatSetBlockSizesFromMats(C,A,P)); 289 290 PetscCall(MatGetType(A,&mtype)); 291 PetscCall(MatSetType(C,mtype)); 292 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 293 ierr = MatPreallocateFinalize(dnz,onz);PetscCall(ierr); 294 295 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 296 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 297 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 298 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 299 300 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 301 C->ops->productnumeric = MatProductNumeric_AB; 302 303 /* attach the supporting struct to C for reuse */ 304 C->product->data = ptap; 305 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 306 307 /* set MatInfo */ 308 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 309 if (afill < 1.0) afill = 1.0; 310 C->info.mallocs = nspacedouble; 311 C->info.fill_ratio_given = fill; 312 C->info.fill_ratio_needed = afill; 313 314 #if defined(PETSC_USE_INFO) 315 if (api[am]) { 316 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 317 PetscCall(PetscInfo(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill)); 318 } else { 319 PetscCall(PetscInfo(C,"Empty matrix product\n")); 320 } 321 #endif 322 PetscFunctionReturn(0); 323 } 324 325 /* ------------------------------------------------------- */ 326 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat,Mat,PetscReal,Mat); 327 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat,Mat,Mat); 328 329 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AB(Mat C) 330 { 331 Mat_Product *product = C->product; 332 Mat A = product->A,B=product->B; 333 334 PetscFunctionBegin; 335 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) 336 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 337 338 C->ops->matmultsymbolic = MatMatMultSymbolic_MPIAIJ_MPIDense; 339 C->ops->productsymbolic = MatProductSymbolic_AB; 340 PetscFunctionReturn(0); 341 } 342 /* -------------------------------------------------------------------- */ 343 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(Mat C) 344 { 345 Mat_Product *product = C->product; 346 Mat A = product->A,B=product->B; 347 348 PetscFunctionBegin; 349 if (A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend) 350 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->rmap->rstart,A->rmap->rend,B->rmap->rstart,B->rmap->rend); 351 352 C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIAIJ_MPIDense; 353 C->ops->productsymbolic = MatProductSymbolic_AtB; 354 PetscFunctionReturn(0); 355 } 356 357 /* --------------------------------------------------------------------- */ 358 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense(Mat C) 359 { 360 Mat_Product *product = C->product; 361 362 PetscFunctionBegin; 363 switch (product->type) { 364 case MATPRODUCT_AB: 365 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AB(C)); 366 break; 367 case MATPRODUCT_AtB: 368 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(C)); 369 break; 370 default: 371 break; 372 } 373 PetscFunctionReturn(0); 374 } 375 /* ------------------------------------------------------- */ 376 377 typedef struct { 378 Mat workB,workB1; 379 MPI_Request *rwaits,*swaits; 380 PetscInt nsends,nrecvs; 381 MPI_Datatype *stype,*rtype; 382 PetscInt blda; 383 } MPIAIJ_MPIDense; 384 385 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 386 { 387 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*)ctx; 388 PetscInt i; 389 390 PetscFunctionBegin; 391 PetscCall(MatDestroy(&contents->workB)); 392 PetscCall(MatDestroy(&contents->workB1)); 393 for (i=0; i<contents->nsends; i++) { 394 PetscCallMPI(MPI_Type_free(&contents->stype[i])); 395 } 396 for (i=0; i<contents->nrecvs; i++) { 397 PetscCallMPI(MPI_Type_free(&contents->rtype[i])); 398 } 399 PetscCall(PetscFree4(contents->stype,contents->rtype,contents->rwaits,contents->swaits)); 400 PetscCall(PetscFree(contents)); 401 PetscFunctionReturn(0); 402 } 403 404 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat C) 405 { 406 PetscErrorCode ierr; 407 Mat_MPIAIJ *aij=(Mat_MPIAIJ*)A->data; 408 PetscInt nz=aij->B->cmap->n,nsends,nrecvs,i,nrows_to,j,blda,m,M,n,N; 409 MPIAIJ_MPIDense *contents; 410 VecScatter ctx=aij->Mvctx; 411 PetscInt Am=A->rmap->n,Bm=B->rmap->n,BN=B->cmap->N,Bbn,Bbn1,bs,nrows_from,numBb; 412 MPI_Comm comm; 413 MPI_Datatype type1,*stype,*rtype; 414 const PetscInt *sindices,*sstarts,*rstarts; 415 PetscMPIInt *disp; 416 PetscBool cisdense; 417 418 PetscFunctionBegin; 419 MatCheckProduct(C,4); 420 PetscCheck(!C->product->data,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data not empty"); 421 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 422 PetscCall(PetscObjectBaseTypeCompare((PetscObject)C,MATMPIDENSE,&cisdense)); 423 if (!cisdense) { 424 PetscCall(MatSetType(C,((PetscObject)B)->type_name)); 425 } 426 PetscCall(MatGetLocalSize(C,&m,&n)); 427 PetscCall(MatGetSize(C,&M,&N)); 428 if (m == PETSC_DECIDE || n == PETSC_DECIDE || M == PETSC_DECIDE || N == PETSC_DECIDE) { 429 PetscCall(MatSetSizes(C,Am,B->cmap->n,A->rmap->N,BN)); 430 } 431 PetscCall(MatSetBlockSizesFromMats(C,A,B)); 432 PetscCall(MatSetUp(C)); 433 PetscCall(MatDenseGetLDA(B,&blda)); 434 PetscCall(PetscNew(&contents)); 435 436 PetscCall(VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,NULL,NULL)); 437 PetscCall(VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,NULL,NULL)); 438 439 /* Create column block of B and C for memory scalability when BN is too large */ 440 /* Estimate Bbn, column size of Bb */ 441 if (nz) { 442 Bbn1 = 2*Am*BN/nz; 443 if (!Bbn1) Bbn1 = 1; 444 } else Bbn1 = BN; 445 446 bs = PetscAbs(B->cmap->bs); 447 Bbn1 = Bbn1/bs *bs; /* Bbn1 is a multiple of bs */ 448 if (Bbn1 > BN) Bbn1 = BN; 449 PetscCallMPI(MPI_Allreduce(&Bbn1,&Bbn,1,MPIU_INT,MPI_MAX,comm)); 450 451 /* Enable runtime option for Bbn */ 452 ierr = PetscOptionsBegin(comm,((PetscObject)C)->prefix,"MatMatMult","Mat");PetscCall(ierr); 453 PetscCall(PetscOptionsInt("-matmatmult_Bbn","Number of columns in Bb","MatMatMult",Bbn,&Bbn,NULL)); 454 ierr = PetscOptionsEnd();PetscCall(ierr); 455 Bbn = PetscMin(Bbn,BN); 456 457 if (Bbn > 0 && Bbn < BN) { 458 numBb = BN/Bbn; 459 Bbn1 = BN - numBb*Bbn; 460 } else numBb = 0; 461 462 if (numBb) { 463 PetscCall(PetscInfo(C,"use Bb, BN=%" PetscInt_FMT ", Bbn=%" PetscInt_FMT "; numBb=%" PetscInt_FMT "\n",BN,Bbn,numBb)); 464 if (Bbn1) { /* Create workB1 for the remaining columns */ 465 PetscCall(PetscInfo(C,"use Bb1, BN=%" PetscInt_FMT ", Bbn1=%" PetscInt_FMT "\n",BN,Bbn1)); 466 /* Create work matrix used to store off processor rows of B needed for local product */ 467 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nz,Bbn1,NULL,&contents->workB1)); 468 } else contents->workB1 = NULL; 469 } 470 471 /* Create work matrix used to store off processor rows of B needed for local product */ 472 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nz,Bbn,NULL,&contents->workB)); 473 474 /* Use MPI derived data type to reduce memory required by the send/recv buffers */ 475 PetscCall(PetscMalloc4(nsends,&stype,nrecvs,&rtype,nrecvs,&contents->rwaits,nsends,&contents->swaits)); 476 contents->stype = stype; 477 contents->nsends = nsends; 478 479 contents->rtype = rtype; 480 contents->nrecvs = nrecvs; 481 contents->blda = blda; 482 483 PetscCall(PetscMalloc1(Bm+1,&disp)); 484 for (i=0; i<nsends; i++) { 485 nrows_to = sstarts[i+1]-sstarts[i]; 486 for (j=0; j<nrows_to; j++) disp[j] = sindices[sstarts[i]+j]; /* rowB to be sent */ 487 PetscCallMPI(MPI_Type_create_indexed_block(nrows_to,1,disp,MPIU_SCALAR,&type1)); 488 PetscCallMPI(MPI_Type_create_resized(type1,0,blda*sizeof(PetscScalar),&stype[i])); 489 PetscCallMPI(MPI_Type_commit(&stype[i])); 490 PetscCallMPI(MPI_Type_free(&type1)); 491 } 492 493 for (i=0; i<nrecvs; i++) { 494 /* received values from a process form a (nrows_from x Bbn) row block in workB (column-wise) */ 495 nrows_from = rstarts[i+1]-rstarts[i]; 496 disp[0] = 0; 497 PetscCallMPI(MPI_Type_create_indexed_block(1,nrows_from,disp,MPIU_SCALAR,&type1)); 498 PetscCallMPI(MPI_Type_create_resized(type1,0,nz*sizeof(PetscScalar),&rtype[i])); 499 PetscCallMPI(MPI_Type_commit(&rtype[i])); 500 PetscCallMPI(MPI_Type_free(&type1)); 501 } 502 503 PetscCall(PetscFree(disp)); 504 PetscCall(VecScatterRestoreRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,NULL,NULL)); 505 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,NULL,NULL)); 506 PetscCall(MatSetOption(C,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE)); 507 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 508 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 509 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 510 511 C->product->data = contents; 512 C->product->destroy = MatMPIAIJ_MPIDenseDestroy; 513 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 514 PetscFunctionReturn(0); 515 } 516 517 PETSC_INTERN PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat,const PetscBool); 518 /* 519 Performs an efficient scatter on the rows of B needed by this process; this is 520 a modification of the VecScatterBegin_() routines. 521 522 Input: Bbidx = 0: B = Bb 523 = 1: B = Bb1, see MatMatMultSymbolic_MPIAIJ_MPIDense() 524 */ 525 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,PetscInt Bbidx,Mat C,Mat *outworkB) 526 { 527 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 528 const PetscScalar *b; 529 PetscScalar *rvalues; 530 VecScatter ctx = aij->Mvctx; 531 const PetscInt *sindices,*sstarts,*rstarts; 532 const PetscMPIInt *sprocs,*rprocs; 533 PetscInt i,nsends,nrecvs; 534 MPI_Request *swaits,*rwaits; 535 MPI_Comm comm; 536 PetscMPIInt tag=((PetscObject)ctx)->tag,ncols=B->cmap->N,nrows=aij->B->cmap->n,nsends_mpi,nrecvs_mpi; 537 MPIAIJ_MPIDense *contents; 538 Mat workB; 539 MPI_Datatype *stype,*rtype; 540 PetscInt blda; 541 542 PetscFunctionBegin; 543 MatCheckProduct(C,4); 544 PetscCheck(C->product->data,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data empty"); 545 contents = (MPIAIJ_MPIDense*)C->product->data; 546 PetscCall(VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL/*bs*/)); 547 PetscCall(VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,&rprocs,NULL/*bs*/)); 548 PetscCall(PetscMPIIntCast(nsends,&nsends_mpi)); 549 PetscCall(PetscMPIIntCast(nrecvs,&nrecvs_mpi)); 550 if (Bbidx == 0) workB = *outworkB = contents->workB; 551 else workB = *outworkB = contents->workB1; 552 PetscCheckFalse(nrows != workB->rmap->n,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Number of rows of workB %" PetscInt_FMT " not equal to columns of aij->B %d",workB->cmap->n,nrows); 553 swaits = contents->swaits; 554 rwaits = contents->rwaits; 555 556 PetscCall(MatDenseGetArrayRead(B,&b)); 557 PetscCall(MatDenseGetLDA(B,&blda)); 558 PetscCheckFalse(blda != contents->blda,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot reuse an input matrix with lda %" PetscInt_FMT " != %" PetscInt_FMT,blda,contents->blda); 559 PetscCall(MatDenseGetArray(workB,&rvalues)); 560 561 /* Post recv, use MPI derived data type to save memory */ 562 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 563 rtype = contents->rtype; 564 for (i=0; i<nrecvs; i++) { 565 PetscCallMPI(MPI_Irecv(rvalues+(rstarts[i]-rstarts[0]),ncols,rtype[i],rprocs[i],tag,comm,rwaits+i)); 566 } 567 568 stype = contents->stype; 569 for (i=0; i<nsends; i++) { 570 PetscCallMPI(MPI_Isend(b,ncols,stype[i],sprocs[i],tag,comm,swaits+i)); 571 } 572 573 if (nrecvs) PetscCallMPI(MPI_Waitall(nrecvs_mpi,rwaits,MPI_STATUSES_IGNORE)); 574 if (nsends) PetscCallMPI(MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE)); 575 576 PetscCall(VecScatterRestoreRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL)); 577 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,&rprocs,NULL)); 578 PetscCall(MatDenseRestoreArrayRead(B,&b)); 579 PetscCall(MatDenseRestoreArray(workB,&rvalues)); 580 PetscFunctionReturn(0); 581 } 582 583 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 584 { 585 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 586 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 587 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 588 Mat workB; 589 MPIAIJ_MPIDense *contents; 590 591 PetscFunctionBegin; 592 MatCheckProduct(C,3); 593 PetscCheck(C->product->data,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data empty"); 594 contents = (MPIAIJ_MPIDense*)C->product->data; 595 /* diagonal block of A times all local rows of B */ 596 /* TODO: this calls a symbolic multiplication every time, which could be avoided */ 597 PetscCall(MatMatMult(aij->A,bdense->A,MAT_REUSE_MATRIX,PETSC_DEFAULT,&cdense->A)); 598 if (contents->workB->cmap->n == B->cmap->N) { 599 /* get off processor parts of B needed to complete C=A*B */ 600 PetscCall(MatMPIDenseScatter(A,B,0,C,&workB)); 601 602 /* off-diagonal block of A times nonlocal rows of B */ 603 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A,PETSC_TRUE)); 604 } else { 605 Mat Bb,Cb; 606 PetscInt BN=B->cmap->N,n=contents->workB->cmap->n,i; 607 PetscBool ccpu; 608 609 PetscCheckFalse(n <= 0,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Column block size %" PetscInt_FMT " must be positive",n); 610 /* Prevent from unneeded copies back and forth from the GPU 611 when getting and restoring the submatrix 612 We need a proper GPU code for AIJ * dense in parallel */ 613 PetscCall(MatBoundToCPU(C,&ccpu)); 614 PetscCall(MatBindToCPU(C,PETSC_TRUE)); 615 for (i=0; i<BN; i+=n) { 616 PetscCall(MatDenseGetSubMatrix(B,i,PetscMin(i+n,BN),&Bb)); 617 PetscCall(MatDenseGetSubMatrix(C,i,PetscMin(i+n,BN),&Cb)); 618 619 /* get off processor parts of B needed to complete C=A*B */ 620 PetscCall(MatMPIDenseScatter(A,Bb,(i+n)>BN,C,&workB)); 621 622 /* off-diagonal block of A times nonlocal rows of B */ 623 cdense = (Mat_MPIDense*)Cb->data; 624 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A,PETSC_TRUE)); 625 PetscCall(MatDenseRestoreSubMatrix(B,&Bb)); 626 PetscCall(MatDenseRestoreSubMatrix(C,&Cb)); 627 } 628 PetscCall(MatBindToCPU(C,ccpu)); 629 } 630 PetscFunctionReturn(0); 631 } 632 633 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 634 { 635 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 636 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 637 Mat_SeqAIJ *cd = (Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 638 PetscInt *adi = ad->i,*adj,*aoi=ao->i,*aoj; 639 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 640 Mat_SeqAIJ *p_loc,*p_oth; 641 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 642 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 643 PetscInt cm = C->rmap->n,anz,pnz; 644 Mat_APMPI *ptap; 645 PetscScalar *apa_sparse; 646 const PetscScalar *dummy; 647 PetscInt *api,*apj,*apJ,i,j,k,row; 648 PetscInt cstart = C->cmap->rstart; 649 PetscInt cdnz,conz,k0,k1,nextp; 650 MPI_Comm comm; 651 PetscMPIInt size; 652 653 PetscFunctionBegin; 654 MatCheckProduct(C,3); 655 ptap = (Mat_APMPI*)C->product->data; 656 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 657 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 658 PetscCallMPI(MPI_Comm_size(comm,&size)); 659 PetscCheckFalse(!ptap->P_oth && size>1,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatProductClear()"); 660 661 /* flag CPU mask for C */ 662 #if defined(PETSC_HAVE_DEVICE) 663 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 664 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 665 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 666 #endif 667 apa_sparse = ptap->apa; 668 669 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 670 /*-----------------------------------------------------*/ 671 /* update numerical values of P_oth and P_loc */ 672 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 673 PetscCall(MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc)); 674 675 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 676 /*----------------------------------------------------------*/ 677 /* get data from symbolic products */ 678 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 679 pi_loc = p_loc->i; pj_loc = p_loc->j; pa_loc = p_loc->a; 680 if (size >1) { 681 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 682 pi_oth = p_oth->i; pj_oth = p_oth->j; pa_oth = p_oth->a; 683 } else { 684 p_oth = NULL; pi_oth = NULL; pj_oth = NULL; pa_oth = NULL; 685 } 686 687 /* trigger copy to CPU */ 688 PetscCall(MatSeqAIJGetArrayRead(a->A,&dummy)); 689 PetscCall(MatSeqAIJRestoreArrayRead(a->A,&dummy)); 690 PetscCall(MatSeqAIJGetArrayRead(a->B,&dummy)); 691 PetscCall(MatSeqAIJRestoreArrayRead(a->B,&dummy)); 692 api = ptap->api; 693 apj = ptap->apj; 694 for (i=0; i<cm; i++) { 695 apJ = apj + api[i]; 696 697 /* diagonal portion of A */ 698 anz = adi[i+1] - adi[i]; 699 adj = ad->j + adi[i]; 700 ada = ad->a + adi[i]; 701 for (j=0; j<anz; j++) { 702 row = adj[j]; 703 pnz = pi_loc[row+1] - pi_loc[row]; 704 pj = pj_loc + pi_loc[row]; 705 pa = pa_loc + pi_loc[row]; 706 /* perform sparse axpy */ 707 valtmp = ada[j]; 708 nextp = 0; 709 for (k=0; nextp<pnz; k++) { 710 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 711 apa_sparse[k] += valtmp*pa[nextp++]; 712 } 713 } 714 PetscCall(PetscLogFlops(2.0*pnz)); 715 } 716 717 /* off-diagonal portion of A */ 718 anz = aoi[i+1] - aoi[i]; 719 aoj = ao->j + aoi[i]; 720 aoa = ao->a + aoi[i]; 721 for (j=0; j<anz; j++) { 722 row = aoj[j]; 723 pnz = pi_oth[row+1] - pi_oth[row]; 724 pj = pj_oth + pi_oth[row]; 725 pa = pa_oth + pi_oth[row]; 726 /* perform sparse axpy */ 727 valtmp = aoa[j]; 728 nextp = 0; 729 for (k=0; nextp<pnz; k++) { 730 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 731 apa_sparse[k] += valtmp*pa[nextp++]; 732 } 733 } 734 PetscCall(PetscLogFlops(2.0*pnz)); 735 } 736 737 /* set values in C */ 738 cdnz = cd->i[i+1] - cd->i[i]; 739 conz = co->i[i+1] - co->i[i]; 740 741 /* 1st off-diagonal part of C */ 742 ca = coa + co->i[i]; 743 k = 0; 744 for (k0=0; k0<conz; k0++) { 745 if (apJ[k] >= cstart) break; 746 ca[k0] = apa_sparse[k]; 747 apa_sparse[k] = 0.0; 748 k++; 749 } 750 751 /* diagonal part of C */ 752 ca = cda + cd->i[i]; 753 for (k1=0; k1<cdnz; k1++) { 754 ca[k1] = apa_sparse[k]; 755 apa_sparse[k] = 0.0; 756 k++; 757 } 758 759 /* 2nd off-diagonal part of C */ 760 ca = coa + co->i[i]; 761 for (; k0<conz; k0++) { 762 ca[k0] = apa_sparse[k]; 763 apa_sparse[k] = 0.0; 764 k++; 765 } 766 } 767 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 768 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 769 PetscFunctionReturn(0); 770 } 771 772 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 773 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat C) 774 { 775 PetscErrorCode ierr; 776 MPI_Comm comm; 777 PetscMPIInt size; 778 Mat_APMPI *ptap; 779 PetscFreeSpaceList free_space = NULL,current_space=NULL; 780 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 781 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 782 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 783 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 784 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=1; 785 PetscInt am=A->rmap->n,pn=P->cmap->n,pm=P->rmap->n,lsize=pn+20; 786 PetscReal afill; 787 MatType mtype; 788 789 PetscFunctionBegin; 790 MatCheckProduct(C,4); 791 PetscCheck(!C->product->data,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 792 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 793 PetscCallMPI(MPI_Comm_size(comm,&size)); 794 795 /* create struct Mat_APMPI and attached it to C later */ 796 PetscCall(PetscNew(&ptap)); 797 798 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 799 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 800 801 /* get P_loc by taking all local rows of P */ 802 PetscCall(MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc)); 803 804 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 805 pi_loc = p_loc->i; pj_loc = p_loc->j; 806 if (size > 1) { 807 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 808 pi_oth = p_oth->i; pj_oth = p_oth->j; 809 } else { 810 p_oth = NULL; 811 pi_oth = NULL; pj_oth = NULL; 812 } 813 814 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 815 /*-------------------------------------------------------------------*/ 816 PetscCall(PetscMalloc1(am+2,&api)); 817 ptap->api = api; 818 api[0] = 0; 819 820 PetscCall(PetscLLCondensedCreate_Scalable(lsize,&lnk)); 821 822 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 823 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space)); 824 current_space = free_space; 825 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);PetscCall(ierr); 826 for (i=0; i<am; i++) { 827 /* diagonal portion of A */ 828 nzi = adi[i+1] - adi[i]; 829 for (j=0; j<nzi; j++) { 830 row = *adj++; 831 pnz = pi_loc[row+1] - pi_loc[row]; 832 Jptr = pj_loc + pi_loc[row]; 833 /* Expand list if it is not long enough */ 834 if (pnz+apnz_max > lsize) { 835 lsize = pnz+apnz_max; 836 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 837 } 838 /* add non-zero cols of P into the sorted linked list lnk */ 839 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk)); 840 apnz = *lnk; /* The first element in the list is the number of items in the list */ 841 api[i+1] = api[i] + apnz; 842 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 843 } 844 /* off-diagonal portion of A */ 845 nzi = aoi[i+1] - aoi[i]; 846 for (j=0; j<nzi; j++) { 847 row = *aoj++; 848 pnz = pi_oth[row+1] - pi_oth[row]; 849 Jptr = pj_oth + pi_oth[row]; 850 /* Expand list if it is not long enough */ 851 if (pnz+apnz_max > lsize) { 852 lsize = pnz + apnz_max; 853 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 854 } 855 /* add non-zero cols of P into the sorted linked list lnk */ 856 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk)); 857 apnz = *lnk; /* The first element in the list is the number of items in the list */ 858 api[i+1] = api[i] + apnz; 859 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 860 } 861 862 /* add missing diagonal entry */ 863 if (C->force_diagonals) { 864 j = i + rstart; /* column index */ 865 PetscCall(PetscLLCondensedAddSorted_Scalable(1,&j,lnk)); 866 } 867 868 apnz = *lnk; 869 api[i+1] = api[i] + apnz; 870 if (apnz > apnz_max) apnz_max = apnz; 871 872 /* if free space is not available, double the total space in the list */ 873 if (current_space->local_remaining<apnz) { 874 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space)); 875 nspacedouble++; 876 } 877 878 /* Copy data into free space, then initialize lnk */ 879 PetscCall(PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk)); 880 PetscCall(MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz)); 881 882 current_space->array += apnz; 883 current_space->local_used += apnz; 884 current_space->local_remaining -= apnz; 885 } 886 887 /* Allocate space for apj, initialize apj, and */ 888 /* destroy list of free space and other temporary array(s) */ 889 PetscCall(PetscMalloc1(api[am]+1,&ptap->apj)); 890 apj = ptap->apj; 891 PetscCall(PetscFreeSpaceContiguous(&free_space,ptap->apj)); 892 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 893 894 /* create and assemble symbolic parallel matrix C */ 895 /*----------------------------------------------------*/ 896 PetscCall(MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE)); 897 PetscCall(MatSetBlockSizesFromMats(C,A,P)); 898 PetscCall(MatGetType(A,&mtype)); 899 PetscCall(MatSetType(C,mtype)); 900 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 901 ierr = MatPreallocateFinalize(dnz,onz);PetscCall(ierr); 902 903 /* malloc apa for assembly C */ 904 PetscCall(PetscCalloc1(apnz_max,&ptap->apa)); 905 906 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 907 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 908 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 909 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 910 911 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 912 C->ops->productnumeric = MatProductNumeric_AB; 913 914 /* attach the supporting struct to C for reuse */ 915 C->product->data = ptap; 916 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 917 918 /* set MatInfo */ 919 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 920 if (afill < 1.0) afill = 1.0; 921 C->info.mallocs = nspacedouble; 922 C->info.fill_ratio_given = fill; 923 C->info.fill_ratio_needed = afill; 924 925 #if defined(PETSC_USE_INFO) 926 if (api[am]) { 927 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 928 PetscCall(PetscInfo(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill)); 929 } else { 930 PetscCall(PetscInfo(C,"Empty matrix product\n")); 931 } 932 #endif 933 PetscFunctionReturn(0); 934 } 935 936 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 937 /* Three input arrays are merged to one output array. The size of the */ 938 /* output array is also output. Duplicate entries only show up once. */ 939 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, 940 PetscInt size2, PetscInt *in2, 941 PetscInt size3, PetscInt *in3, 942 PetscInt *size4, PetscInt *out) 943 { 944 int i = 0, j = 0, k = 0, l = 0; 945 946 /* Traverse all three arrays */ 947 while (i<size1 && j<size2 && k<size3) { 948 if (in1[i] < in2[j] && in1[i] < in3[k]) { 949 out[l++] = in1[i++]; 950 } 951 else if (in2[j] < in1[i] && in2[j] < in3[k]) { 952 out[l++] = in2[j++]; 953 } 954 else if (in3[k] < in1[i] && in3[k] < in2[j]) { 955 out[l++] = in3[k++]; 956 } 957 else if (in1[i] == in2[j] && in1[i] < in3[k]) { 958 out[l++] = in1[i]; 959 i++, j++; 960 } 961 else if (in1[i] == in3[k] && in1[i] < in2[j]) { 962 out[l++] = in1[i]; 963 i++, k++; 964 } 965 else if (in3[k] == in2[j] && in2[j] < in1[i]) { 966 out[l++] = in2[j]; 967 k++, j++; 968 } 969 else if (in1[i] == in2[j] && in1[i] == in3[k]) { 970 out[l++] = in1[i]; 971 i++, j++, k++; 972 } 973 } 974 975 /* Traverse two remaining arrays */ 976 while (i<size1 && j<size2) { 977 if (in1[i] < in2[j]) { 978 out[l++] = in1[i++]; 979 } 980 else if (in1[i] > in2[j]) { 981 out[l++] = in2[j++]; 982 } 983 else { 984 out[l++] = in1[i]; 985 i++, j++; 986 } 987 } 988 989 while (i<size1 && k<size3) { 990 if (in1[i] < in3[k]) { 991 out[l++] = in1[i++]; 992 } 993 else if (in1[i] > in3[k]) { 994 out[l++] = in3[k++]; 995 } 996 else { 997 out[l++] = in1[i]; 998 i++, k++; 999 } 1000 } 1001 1002 while (k<size3 && j<size2) { 1003 if (in3[k] < in2[j]) { 1004 out[l++] = in3[k++]; 1005 } 1006 else if (in3[k] > in2[j]) { 1007 out[l++] = in2[j++]; 1008 } 1009 else { 1010 out[l++] = in3[k]; 1011 k++, j++; 1012 } 1013 } 1014 1015 /* Traverse one remaining array */ 1016 while (i<size1) out[l++] = in1[i++]; 1017 while (j<size2) out[l++] = in2[j++]; 1018 while (k<size3) out[l++] = in3[k++]; 1019 1020 *size4 = l; 1021 } 1022 1023 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 1024 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 1025 /* matrix-matrix multiplications. */ 1026 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat C) 1027 { 1028 PetscErrorCode ierr; 1029 MPI_Comm comm; 1030 PetscMPIInt size; 1031 Mat_APMPI *ptap; 1032 PetscFreeSpaceList free_space_diag=NULL, current_space=NULL; 1033 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 1034 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc; 1035 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 1036 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 1037 PetscInt adponz, adpdnz; 1038 PetscInt *pi_loc,*dnz,*onz; 1039 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,rstart=A->rmap->rstart; 1040 PetscInt *lnk,i, i1=0,pnz,row,*adpoi,*adpoj, *api, *adpoJ, *aopJ, *apJ,*Jptr, aopnz, nspacedouble=0,j,nzi, 1041 *apj,apnz, *adpdi, *adpdj, *adpdJ, *poff_i, *poff_j, *j_temp, *aopothi, *aopothj; 1042 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n, p_colstart, p_colend; 1043 PetscBT lnkbt; 1044 PetscReal afill; 1045 PetscMPIInt rank; 1046 Mat adpd, aopoth; 1047 MatType mtype; 1048 const char *prefix; 1049 1050 PetscFunctionBegin; 1051 MatCheckProduct(C,4); 1052 PetscCheck(!C->product->data,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 1053 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 1054 PetscCallMPI(MPI_Comm_size(comm,&size)); 1055 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1056 PetscCall(MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend)); 1057 1058 /* create struct Mat_APMPI and attached it to C later */ 1059 PetscCall(PetscNew(&ptap)); 1060 1061 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1062 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 1063 1064 /* get P_loc by taking all local rows of P */ 1065 PetscCall(MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc)); 1066 1067 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1068 pi_loc = p_loc->i; 1069 1070 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 1071 PetscCall(PetscMalloc1(am+2,&api)); 1072 PetscCall(PetscMalloc1(am+2,&adpoi)); 1073 1074 adpoi[0] = 0; 1075 ptap->api = api; 1076 api[0] = 0; 1077 1078 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1079 PetscCall(PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt)); 1080 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);PetscCall(ierr); 1081 1082 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1083 PetscCall(MatGetOptionsPrefix(A,&prefix)); 1084 PetscCall(MatProductCreate(a->A,p->A,NULL,&adpd)); 1085 PetscCall(MatGetOptionsPrefix(A,&prefix)); 1086 PetscCall(MatSetOptionsPrefix(adpd,prefix)); 1087 PetscCall(MatAppendOptionsPrefix(adpd,"inner_diag_")); 1088 1089 PetscCall(MatProductSetType(adpd,MATPRODUCT_AB)); 1090 PetscCall(MatProductSetAlgorithm(adpd,"sorted")); 1091 PetscCall(MatProductSetFill(adpd,fill)); 1092 PetscCall(MatProductSetFromOptions(adpd)); 1093 1094 adpd->force_diagonals = C->force_diagonals; 1095 PetscCall(MatProductSymbolic(adpd)); 1096 1097 adpd_seq = (Mat_SeqAIJ*)((adpd)->data); 1098 adpdi = adpd_seq->i; adpdj = adpd_seq->j; 1099 p_off = (Mat_SeqAIJ*)((p->B)->data); 1100 poff_i = p_off->i; poff_j = p_off->j; 1101 1102 /* j_temp stores indices of a result row before they are added to the linked list */ 1103 PetscCall(PetscMalloc1(pN+2,&j_temp)); 1104 1105 /* Symbolic calc of the A_diag * p_loc_off */ 1106 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1107 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space_diag)); 1108 current_space = free_space_diag; 1109 1110 for (i=0; i<am; i++) { 1111 /* A_diag * P_loc_off */ 1112 nzi = adi[i+1] - adi[i]; 1113 for (j=0; j<nzi; j++) { 1114 row = *adj++; 1115 pnz = poff_i[row+1] - poff_i[row]; 1116 Jptr = poff_j + poff_i[row]; 1117 for (i1 = 0; i1 < pnz; i1++) { 1118 j_temp[i1] = p->garray[Jptr[i1]]; 1119 } 1120 /* add non-zero cols of P into the sorted linked list lnk */ 1121 PetscCall(PetscLLCondensedAddSorted(pnz,j_temp,lnk,lnkbt)); 1122 } 1123 1124 adponz = lnk[0]; 1125 adpoi[i+1] = adpoi[i] + adponz; 1126 1127 /* if free space is not available, double the total space in the list */ 1128 if (current_space->local_remaining<adponz) { 1129 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(adponz,current_space->total_array_size),¤t_space)); 1130 nspacedouble++; 1131 } 1132 1133 /* Copy data into free space, then initialize lnk */ 1134 PetscCall(PetscLLCondensedClean(pN,adponz,current_space->array,lnk,lnkbt)); 1135 1136 current_space->array += adponz; 1137 current_space->local_used += adponz; 1138 current_space->local_remaining -= adponz; 1139 } 1140 1141 /* Symbolic calc of A_off * P_oth */ 1142 PetscCall(MatSetOptionsPrefix(a->B,prefix)); 1143 PetscCall(MatAppendOptionsPrefix(a->B,"inner_offdiag_")); 1144 PetscCall(MatCreate(PETSC_COMM_SELF,&aopoth)); 1145 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, aopoth)); 1146 aopoth_seq = (Mat_SeqAIJ*)((aopoth)->data); 1147 aopothi = aopoth_seq->i; aopothj = aopoth_seq->j; 1148 1149 /* Allocate space for apj, adpj, aopj, ... */ 1150 /* destroy lists of free space and other temporary array(s) */ 1151 1152 PetscCall(PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am]+2, &ptap->apj)); 1153 PetscCall(PetscMalloc1(adpoi[am]+2, &adpoj)); 1154 1155 /* Copy from linked list to j-array */ 1156 PetscCall(PetscFreeSpaceContiguous(&free_space_diag,adpoj)); 1157 PetscCall(PetscLLDestroy(lnk,lnkbt)); 1158 1159 adpoJ = adpoj; 1160 adpdJ = adpdj; 1161 aopJ = aopothj; 1162 apj = ptap->apj; 1163 apJ = apj; /* still empty */ 1164 1165 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1166 /* A_diag * P_loc_diag to get A*P */ 1167 for (i = 0; i < am; i++) { 1168 aopnz = aopothi[i+1] - aopothi[i]; 1169 adponz = adpoi[i+1] - adpoi[i]; 1170 adpdnz = adpdi[i+1] - adpdi[i]; 1171 1172 /* Correct indices from A_diag*P_diag */ 1173 for (i1 = 0; i1 < adpdnz; i1++) { 1174 adpdJ[i1] += p_colstart; 1175 } 1176 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1177 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1178 PetscCall(MatPreallocateSet(i+rstart, apnz, apJ, dnz, onz)); 1179 1180 aopJ += aopnz; 1181 adpoJ += adponz; 1182 adpdJ += adpdnz; 1183 apJ += apnz; 1184 api[i+1] = api[i] + apnz; 1185 } 1186 1187 /* malloc apa to store dense row A[i,:]*P */ 1188 PetscCall(PetscCalloc1(pN+2,&ptap->apa)); 1189 1190 /* create and assemble symbolic parallel matrix C */ 1191 PetscCall(MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE)); 1192 PetscCall(MatSetBlockSizesFromMats(C,A,P)); 1193 PetscCall(MatGetType(A,&mtype)); 1194 PetscCall(MatSetType(C,mtype)); 1195 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 1196 ierr = MatPreallocateFinalize(dnz,onz);PetscCall(ierr); 1197 1198 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 1199 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1200 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1201 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 1202 1203 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1204 C->ops->productnumeric = MatProductNumeric_AB; 1205 1206 /* attach the supporting struct to C for reuse */ 1207 C->product->data = ptap; 1208 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 1209 1210 /* set MatInfo */ 1211 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 1212 if (afill < 1.0) afill = 1.0; 1213 C->info.mallocs = nspacedouble; 1214 C->info.fill_ratio_given = fill; 1215 C->info.fill_ratio_needed = afill; 1216 1217 #if defined(PETSC_USE_INFO) 1218 if (api[am]) { 1219 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 1220 PetscCall(PetscInfo(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill)); 1221 } else { 1222 PetscCall(PetscInfo(C,"Empty matrix product\n")); 1223 } 1224 #endif 1225 1226 PetscCall(MatDestroy(&aopoth)); 1227 PetscCall(MatDestroy(&adpd)); 1228 PetscCall(PetscFree(j_temp)); 1229 PetscCall(PetscFree(adpoj)); 1230 PetscCall(PetscFree(adpoi)); 1231 PetscFunctionReturn(0); 1232 } 1233 1234 /*-------------------------------------------------------------------------*/ 1235 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1236 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P,Mat A,Mat C) 1237 { 1238 Mat_APMPI *ptap; 1239 Mat Pt; 1240 1241 PetscFunctionBegin; 1242 MatCheckProduct(C,3); 1243 ptap = (Mat_APMPI*)C->product->data; 1244 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1245 PetscCheck(ptap->Pt,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1246 1247 Pt = ptap->Pt; 1248 PetscCall(MatTranspose(P,MAT_REUSE_MATRIX,&Pt)); 1249 PetscCall(MatMatMultNumeric_MPIAIJ_MPIAIJ(Pt,A,C)); 1250 PetscFunctionReturn(0); 1251 } 1252 1253 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1254 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,PetscReal fill,Mat C) 1255 { 1256 PetscErrorCode ierr; 1257 Mat_APMPI *ptap; 1258 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 1259 MPI_Comm comm; 1260 PetscMPIInt size,rank; 1261 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1262 PetscInt pn=P->cmap->n,aN=A->cmap->N,an=A->cmap->n; 1263 PetscInt *lnk,i,k,nsend,rstart; 1264 PetscBT lnkbt; 1265 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,nrecv; 1266 PETSC_UNUSED PetscMPIInt icompleted=0; 1267 PetscInt **buf_rj,**buf_ri,**buf_ri_k,row,ncols,*cols; 1268 PetscInt len,proc,*dnz,*onz,*owners,nzi; 1269 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1270 MPI_Request *swaits,*rwaits; 1271 MPI_Status *sstatus,rstatus; 1272 PetscLayout rowmap; 1273 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1274 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 1275 PetscInt *Jptr,*prmap=p->garray,con,j,Crmax; 1276 Mat_SeqAIJ *a_loc,*c_loc,*c_oth; 1277 PetscTable ta; 1278 MatType mtype; 1279 const char *prefix; 1280 1281 PetscFunctionBegin; 1282 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 1283 PetscCallMPI(MPI_Comm_size(comm,&size)); 1284 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 1285 1286 /* create symbolic parallel matrix C */ 1287 PetscCall(MatGetType(A,&mtype)); 1288 PetscCall(MatSetType(C,mtype)); 1289 1290 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1291 1292 /* create struct Mat_APMPI and attached it to C later */ 1293 PetscCall(PetscNew(&ptap)); 1294 ptap->reuse = MAT_INITIAL_MATRIX; 1295 1296 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1297 /* --------------------------------- */ 1298 PetscCall(MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd)); 1299 PetscCall(MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro)); 1300 1301 /* (1) compute symbolic A_loc */ 1302 /* ---------------------------*/ 1303 PetscCall(MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&ptap->A_loc)); 1304 1305 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1306 /* ------------------------------------ */ 1307 PetscCall(MatGetOptionsPrefix(A,&prefix)); 1308 PetscCall(MatSetOptionsPrefix(ptap->Ro,prefix)); 1309 PetscCall(MatAppendOptionsPrefix(ptap->Ro,"inner_offdiag_")); 1310 PetscCall(MatCreate(PETSC_COMM_SELF,&ptap->C_oth)); 1311 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->A_loc,fill,ptap->C_oth)); 1312 1313 /* (3) send coj of C_oth to other processors */ 1314 /* ------------------------------------------ */ 1315 /* determine row ownership */ 1316 PetscCall(PetscLayoutCreate(comm,&rowmap)); 1317 rowmap->n = pn; 1318 rowmap->bs = 1; 1319 PetscCall(PetscLayoutSetUp(rowmap)); 1320 owners = rowmap->range; 1321 1322 /* determine the number of messages to send, their lengths */ 1323 PetscCall(PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co)); 1324 PetscCall(PetscArrayzero(len_s,size)); 1325 PetscCall(PetscArrayzero(len_si,size)); 1326 1327 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 1328 coi = c_oth->i; coj = c_oth->j; 1329 con = ptap->C_oth->rmap->n; 1330 proc = 0; 1331 for (i=0; i<con; i++) { 1332 while (prmap[i] >= owners[proc+1]) proc++; 1333 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1334 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1335 } 1336 1337 len = 0; /* max length of buf_si[], see (4) */ 1338 owners_co[0] = 0; 1339 nsend = 0; 1340 for (proc=0; proc<size; proc++) { 1341 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1342 if (len_s[proc]) { 1343 nsend++; 1344 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1345 len += len_si[proc]; 1346 } 1347 } 1348 1349 /* determine the number and length of messages to receive for coi and coj */ 1350 PetscCall(PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv)); 1351 PetscCall(PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri)); 1352 1353 /* post the Irecv and Isend of coj */ 1354 PetscCall(PetscCommGetNewTag(comm,&tagj)); 1355 PetscCall(PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits)); 1356 PetscCall(PetscMalloc1(nsend+1,&swaits)); 1357 for (proc=0, k=0; proc<size; proc++) { 1358 if (!len_s[proc]) continue; 1359 i = owners_co[proc]; 1360 PetscCallMPI(MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k)); 1361 k++; 1362 } 1363 1364 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1365 /* ---------------------------------------- */ 1366 PetscCall(MatSetOptionsPrefix(ptap->Rd,prefix)); 1367 PetscCall(MatAppendOptionsPrefix(ptap->Rd,"inner_diag_")); 1368 PetscCall(MatCreate(PETSC_COMM_SELF,&ptap->C_loc)); 1369 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->A_loc,fill,ptap->C_loc)); 1370 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 1371 1372 /* receives coj are complete */ 1373 for (i=0; i<nrecv; i++) { 1374 PetscCallMPI(MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus)); 1375 } 1376 PetscCall(PetscFree(rwaits)); 1377 if (nsend) PetscCallMPI(MPI_Waitall(nsend,swaits,sstatus)); 1378 1379 /* add received column indices into ta to update Crmax */ 1380 a_loc = (Mat_SeqAIJ*)(ptap->A_loc)->data; 1381 1382 /* create and initialize a linked list */ 1383 PetscCall(PetscTableCreate(an,aN,&ta)); /* for compute Crmax */ 1384 MatRowMergeMax_SeqAIJ(a_loc,ptap->A_loc->rmap->N,ta); 1385 1386 for (k=0; k<nrecv; k++) {/* k-th received message */ 1387 Jptr = buf_rj[k]; 1388 for (j=0; j<len_r[k]; j++) { 1389 PetscCall(PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES)); 1390 } 1391 } 1392 PetscCall(PetscTableGetCount(ta,&Crmax)); 1393 PetscCall(PetscTableDestroy(&ta)); 1394 1395 /* (4) send and recv coi */ 1396 /*-----------------------*/ 1397 PetscCall(PetscCommGetNewTag(comm,&tagi)); 1398 PetscCall(PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits)); 1399 PetscCall(PetscMalloc1(len+1,&buf_s)); 1400 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1401 for (proc=0,k=0; proc<size; proc++) { 1402 if (!len_s[proc]) continue; 1403 /* form outgoing message for i-structure: 1404 buf_si[0]: nrows to be sent 1405 [1:nrows]: row index (global) 1406 [nrows+1:2*nrows+1]: i-structure index 1407 */ 1408 /*-------------------------------------------*/ 1409 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 1410 buf_si_i = buf_si + nrows+1; 1411 buf_si[0] = nrows; 1412 buf_si_i[0] = 0; 1413 nrows = 0; 1414 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1415 nzi = coi[i+1] - coi[i]; 1416 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1417 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1418 nrows++; 1419 } 1420 PetscCallMPI(MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k)); 1421 k++; 1422 buf_si += len_si[proc]; 1423 } 1424 for (i=0; i<nrecv; i++) { 1425 PetscCallMPI(MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus)); 1426 } 1427 PetscCall(PetscFree(rwaits)); 1428 if (nsend) PetscCallMPI(MPI_Waitall(nsend,swaits,sstatus)); 1429 1430 PetscCall(PetscFree4(len_s,len_si,sstatus,owners_co)); 1431 PetscCall(PetscFree(len_ri)); 1432 PetscCall(PetscFree(swaits)); 1433 PetscCall(PetscFree(buf_s)); 1434 1435 /* (5) compute the local portion of C */ 1436 /* ------------------------------------------ */ 1437 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of C */ 1438 PetscCall(PetscFreeSpaceGet(Crmax,&free_space)); 1439 current_space = free_space; 1440 1441 PetscCall(PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci)); 1442 for (k=0; k<nrecv; k++) { 1443 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1444 nrows = *buf_ri_k[k]; 1445 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1446 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1447 } 1448 1449 ierr = MatPreallocateInitialize(comm,pn,an,dnz,onz);PetscCall(ierr); 1450 PetscCall(PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt)); 1451 for (i=0; i<pn; i++) { /* for each local row of C */ 1452 /* add C_loc into C */ 1453 nzi = c_loc->i[i+1] - c_loc->i[i]; 1454 Jptr = c_loc->j + c_loc->i[i]; 1455 PetscCall(PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt)); 1456 1457 /* add received col data into lnk */ 1458 for (k=0; k<nrecv; k++) { /* k-th received message */ 1459 if (i == *nextrow[k]) { /* i-th row */ 1460 nzi = *(nextci[k]+1) - *nextci[k]; 1461 Jptr = buf_rj[k] + *nextci[k]; 1462 PetscCall(PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt)); 1463 nextrow[k]++; nextci[k]++; 1464 } 1465 } 1466 1467 /* add missing diagonal entry */ 1468 if (C->force_diagonals) { 1469 k = i + owners[rank]; /* column index */ 1470 PetscCall(PetscLLCondensedAddSorted(1,&k,lnk,lnkbt)); 1471 } 1472 1473 nzi = lnk[0]; 1474 1475 /* copy data into free space, then initialize lnk */ 1476 PetscCall(PetscLLCondensedClean(aN,nzi,current_space->array,lnk,lnkbt)); 1477 PetscCall(MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz)); 1478 } 1479 PetscCall(PetscFree3(buf_ri_k,nextrow,nextci)); 1480 PetscCall(PetscLLDestroy(lnk,lnkbt)); 1481 PetscCall(PetscFreeSpaceDestroy(free_space)); 1482 1483 /* local sizes and preallocation */ 1484 PetscCall(MatSetSizes(C,pn,an,PETSC_DETERMINE,PETSC_DETERMINE)); 1485 if (P->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->rmap,P->cmap->bs)); 1486 if (A->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->cmap,A->cmap->bs)); 1487 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 1488 ierr = MatPreallocateFinalize(dnz,onz);PetscCall(ierr); 1489 1490 /* add C_loc and C_oth to C */ 1491 PetscCall(MatGetOwnershipRange(C,&rstart,NULL)); 1492 for (i=0; i<pn; i++) { 1493 ncols = c_loc->i[i+1] - c_loc->i[i]; 1494 cols = c_loc->j + c_loc->i[i]; 1495 row = rstart + i; 1496 PetscCall(MatSetValues(C,1,(const PetscInt*)&row,ncols,(const PetscInt*)cols,NULL,INSERT_VALUES)); 1497 1498 if (C->force_diagonals) { 1499 PetscCall(MatSetValues(C,1,(const PetscInt*)&row,1,(const PetscInt*)&row,NULL,INSERT_VALUES)); 1500 } 1501 } 1502 for (i=0; i<con; i++) { 1503 ncols = c_oth->i[i+1] - c_oth->i[i]; 1504 cols = c_oth->j + c_oth->i[i]; 1505 row = prmap[i]; 1506 PetscCall(MatSetValues(C,1,(const PetscInt*)&row,ncols,(const PetscInt*)cols,NULL,INSERT_VALUES)); 1507 } 1508 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1509 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1510 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 1511 1512 /* members in merge */ 1513 PetscCall(PetscFree(id_r)); 1514 PetscCall(PetscFree(len_r)); 1515 PetscCall(PetscFree(buf_ri[0])); 1516 PetscCall(PetscFree(buf_ri)); 1517 PetscCall(PetscFree(buf_rj[0])); 1518 PetscCall(PetscFree(buf_rj)); 1519 PetscCall(PetscLayoutDestroy(&rowmap)); 1520 1521 /* attach the supporting struct to C for reuse */ 1522 C->product->data = ptap; 1523 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 1524 PetscFunctionReturn(0); 1525 } 1526 1527 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,Mat C) 1528 { 1529 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 1530 Mat_SeqAIJ *c_seq; 1531 Mat_APMPI *ptap; 1532 Mat A_loc,C_loc,C_oth; 1533 PetscInt i,rstart,rend,cm,ncols,row; 1534 const PetscInt *cols; 1535 const PetscScalar *vals; 1536 1537 PetscFunctionBegin; 1538 MatCheckProduct(C,3); 1539 ptap = (Mat_APMPI*)C->product->data; 1540 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1541 PetscCheck(ptap->A_loc,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1542 PetscCall(MatZeroEntries(C)); 1543 1544 if (ptap->reuse == MAT_REUSE_MATRIX) { 1545 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1546 /* 1) get R = Pd^T, Ro = Po^T */ 1547 /*----------------------------*/ 1548 PetscCall(MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd)); 1549 PetscCall(MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro)); 1550 1551 /* 2) compute numeric A_loc */ 1552 /*--------------------------*/ 1553 PetscCall(MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&ptap->A_loc)); 1554 } 1555 1556 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1557 A_loc = ptap->A_loc; 1558 PetscCall(((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,A_loc,ptap->C_loc)); 1559 PetscCall(((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,A_loc,ptap->C_oth)); 1560 C_loc = ptap->C_loc; 1561 C_oth = ptap->C_oth; 1562 1563 /* add C_loc and C_oth to C */ 1564 PetscCall(MatGetOwnershipRange(C,&rstart,&rend)); 1565 1566 /* C_loc -> C */ 1567 cm = C_loc->rmap->N; 1568 c_seq = (Mat_SeqAIJ*)C_loc->data; 1569 cols = c_seq->j; 1570 vals = c_seq->a; 1571 for (i=0; i<cm; i++) { 1572 ncols = c_seq->i[i+1] - c_seq->i[i]; 1573 row = rstart + i; 1574 PetscCall(MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES)); 1575 cols += ncols; vals += ncols; 1576 } 1577 1578 /* Co -> C, off-processor part */ 1579 cm = C_oth->rmap->N; 1580 c_seq = (Mat_SeqAIJ*)C_oth->data; 1581 cols = c_seq->j; 1582 vals = c_seq->a; 1583 for (i=0; i<cm; i++) { 1584 ncols = c_seq->i[i+1] - c_seq->i[i]; 1585 row = p->garray[i]; 1586 PetscCall(MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES)); 1587 cols += ncols; vals += ncols; 1588 } 1589 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1590 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1591 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 1592 1593 ptap->reuse = MAT_REUSE_MATRIX; 1594 PetscFunctionReturn(0); 1595 } 1596 1597 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 1598 { 1599 Mat_Merge_SeqsToMPI *merge; 1600 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 1601 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1602 Mat_APMPI *ptap; 1603 PetscInt *adj; 1604 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1605 MatScalar *ada,*ca,valtmp; 1606 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1607 MPI_Comm comm; 1608 PetscMPIInt size,rank,taga,*len_s; 1609 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1610 PetscInt **buf_ri,**buf_rj; 1611 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1612 MPI_Request *s_waits,*r_waits; 1613 MPI_Status *status; 1614 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1615 const PetscScalar *dummy; 1616 PetscInt *ai,*aj,*coi,*coj,*poJ,*pdJ; 1617 Mat A_loc; 1618 Mat_SeqAIJ *a_loc; 1619 1620 PetscFunctionBegin; 1621 MatCheckProduct(C,3); 1622 ptap = (Mat_APMPI*)C->product->data; 1623 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1624 PetscCheck(ptap->A_loc,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1625 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 1626 PetscCallMPI(MPI_Comm_size(comm,&size)); 1627 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 1628 1629 merge = ptap->merge; 1630 1631 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1632 /*------------------------------------------*/ 1633 /* get data from symbolic products */ 1634 coi = merge->coi; coj = merge->coj; 1635 PetscCall(PetscCalloc1(coi[pon]+1,&coa)); 1636 bi = merge->bi; bj = merge->bj; 1637 owners = merge->rowmap->range; 1638 PetscCall(PetscCalloc1(bi[cm]+1,&ba)); 1639 1640 /* get A_loc by taking all local rows of A */ 1641 A_loc = ptap->A_loc; 1642 PetscCall(MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc)); 1643 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1644 ai = a_loc->i; 1645 aj = a_loc->j; 1646 1647 /* trigger copy to CPU */ 1648 PetscCall(MatSeqAIJGetArrayRead(p->A,&dummy)); 1649 PetscCall(MatSeqAIJRestoreArrayRead(p->A,&dummy)); 1650 PetscCall(MatSeqAIJGetArrayRead(p->B,&dummy)); 1651 PetscCall(MatSeqAIJRestoreArrayRead(p->B,&dummy)); 1652 for (i=0; i<am; i++) { 1653 anz = ai[i+1] - ai[i]; 1654 adj = aj + ai[i]; 1655 ada = a_loc->a + ai[i]; 1656 1657 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1658 /*-------------------------------------------------------------*/ 1659 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1660 pnz = po->i[i+1] - po->i[i]; 1661 poJ = po->j + po->i[i]; 1662 pA = po->a + po->i[i]; 1663 for (j=0; j<pnz; j++) { 1664 row = poJ[j]; 1665 cj = coj + coi[row]; 1666 ca = coa + coi[row]; 1667 /* perform sparse axpy */ 1668 nexta = 0; 1669 valtmp = pA[j]; 1670 for (k=0; nexta<anz; k++) { 1671 if (cj[k] == adj[nexta]) { 1672 ca[k] += valtmp*ada[nexta]; 1673 nexta++; 1674 } 1675 } 1676 PetscCall(PetscLogFlops(2.0*anz)); 1677 } 1678 1679 /* put the value into Cd (diagonal part) */ 1680 pnz = pd->i[i+1] - pd->i[i]; 1681 pdJ = pd->j + pd->i[i]; 1682 pA = pd->a + pd->i[i]; 1683 for (j=0; j<pnz; j++) { 1684 row = pdJ[j]; 1685 cj = bj + bi[row]; 1686 ca = ba + bi[row]; 1687 /* perform sparse axpy */ 1688 nexta = 0; 1689 valtmp = pA[j]; 1690 for (k=0; nexta<anz; k++) { 1691 if (cj[k] == adj[nexta]) { 1692 ca[k] += valtmp*ada[nexta]; 1693 nexta++; 1694 } 1695 } 1696 PetscCall(PetscLogFlops(2.0*anz)); 1697 } 1698 } 1699 1700 /* 3) send and recv matrix values coa */ 1701 /*------------------------------------*/ 1702 buf_ri = merge->buf_ri; 1703 buf_rj = merge->buf_rj; 1704 len_s = merge->len_s; 1705 PetscCall(PetscCommGetNewTag(comm,&taga)); 1706 PetscCall(PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits)); 1707 1708 PetscCall(PetscMalloc2(merge->nsend+1,&s_waits,size,&status)); 1709 for (proc=0,k=0; proc<size; proc++) { 1710 if (!len_s[proc]) continue; 1711 i = merge->owners_co[proc]; 1712 PetscCallMPI(MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k)); 1713 k++; 1714 } 1715 if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv,r_waits,status)); 1716 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend,s_waits,status)); 1717 1718 PetscCall(PetscFree2(s_waits,status)); 1719 PetscCall(PetscFree(r_waits)); 1720 PetscCall(PetscFree(coa)); 1721 1722 /* 4) insert local Cseq and received values into Cmpi */ 1723 /*----------------------------------------------------*/ 1724 PetscCall(PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci)); 1725 for (k=0; k<merge->nrecv; k++) { 1726 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1727 nrows = *(buf_ri_k[k]); 1728 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1729 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1730 } 1731 1732 for (i=0; i<cm; i++) { 1733 row = owners[rank] + i; /* global row index of C_seq */ 1734 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1735 ba_i = ba + bi[i]; 1736 bnz = bi[i+1] - bi[i]; 1737 /* add received vals into ba */ 1738 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1739 /* i-th row */ 1740 if (i == *nextrow[k]) { 1741 cnz = *(nextci[k]+1) - *nextci[k]; 1742 cj = buf_rj[k] + *(nextci[k]); 1743 ca = abuf_r[k] + *(nextci[k]); 1744 nextcj = 0; 1745 for (j=0; nextcj<cnz; j++) { 1746 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1747 ba_i[j] += ca[nextcj++]; 1748 } 1749 } 1750 nextrow[k]++; nextci[k]++; 1751 PetscCall(PetscLogFlops(2.0*cnz)); 1752 } 1753 } 1754 PetscCall(MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES)); 1755 } 1756 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1757 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1758 1759 PetscCall(PetscFree(ba)); 1760 PetscCall(PetscFree(abuf_r[0])); 1761 PetscCall(PetscFree(abuf_r)); 1762 PetscCall(PetscFree3(buf_ri_k,nextrow,nextci)); 1763 PetscFunctionReturn(0); 1764 } 1765 1766 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat C) 1767 { 1768 PetscErrorCode ierr; 1769 Mat A_loc; 1770 Mat_APMPI *ptap; 1771 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1772 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*a=(Mat_MPIAIJ*)A->data; 1773 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1774 PetscInt nnz; 1775 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1776 PetscInt am =A->rmap->n,pn=P->cmap->n; 1777 MPI_Comm comm; 1778 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1779 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1780 PetscInt len,proc,*dnz,*onz,*owners; 1781 PetscInt nzi,*bi,*bj; 1782 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1783 MPI_Request *swaits,*rwaits; 1784 MPI_Status *sstatus,rstatus; 1785 Mat_Merge_SeqsToMPI *merge; 1786 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1787 PetscReal afill =1.0,afill_tmp; 1788 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Armax; 1789 Mat_SeqAIJ *a_loc; 1790 PetscTable ta; 1791 MatType mtype; 1792 1793 PetscFunctionBegin; 1794 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 1795 /* check if matrix local sizes are compatible */ 1796 PetscCheckFalse(A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend,comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A (%" PetscInt_FMT ", %" PetscInt_FMT ") != P (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 1797 1798 PetscCallMPI(MPI_Comm_size(comm,&size)); 1799 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 1800 1801 /* create struct Mat_APMPI and attached it to C later */ 1802 PetscCall(PetscNew(&ptap)); 1803 1804 /* get A_loc by taking all local rows of A */ 1805 PetscCall(MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc)); 1806 1807 ptap->A_loc = A_loc; 1808 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1809 ai = a_loc->i; 1810 aj = a_loc->j; 1811 1812 /* determine symbolic Co=(p->B)^T*A - send to others */ 1813 /*----------------------------------------------------*/ 1814 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj)); 1815 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj)); 1816 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1817 >= (num of nonzero rows of C_seq) - pn */ 1818 PetscCall(PetscMalloc1(pon+1,&coi)); 1819 coi[0] = 0; 1820 1821 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1822 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(poti[pon],ai[am])); 1823 PetscCall(PetscFreeSpaceGet(nnz,&free_space)); 1824 current_space = free_space; 1825 1826 /* create and initialize a linked list */ 1827 PetscCall(PetscTableCreate(A->cmap->n + a->B->cmap->N,aN,&ta)); 1828 MatRowMergeMax_SeqAIJ(a_loc,am,ta); 1829 PetscCall(PetscTableGetCount(ta,&Armax)); 1830 1831 PetscCall(PetscLLCondensedCreate_Scalable(Armax,&lnk)); 1832 1833 for (i=0; i<pon; i++) { 1834 pnz = poti[i+1] - poti[i]; 1835 ptJ = potj + poti[i]; 1836 for (j=0; j<pnz; j++) { 1837 row = ptJ[j]; /* row of A_loc == col of Pot */ 1838 anz = ai[row+1] - ai[row]; 1839 Jptr = aj + ai[row]; 1840 /* add non-zero cols of AP into the sorted linked list lnk */ 1841 PetscCall(PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk)); 1842 } 1843 nnz = lnk[0]; 1844 1845 /* If free space is not available, double the total space in the list */ 1846 if (current_space->local_remaining<nnz) { 1847 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space)); 1848 nspacedouble++; 1849 } 1850 1851 /* Copy data into free space, and zero out denserows */ 1852 PetscCall(PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk)); 1853 1854 current_space->array += nnz; 1855 current_space->local_used += nnz; 1856 current_space->local_remaining -= nnz; 1857 1858 coi[i+1] = coi[i] + nnz; 1859 } 1860 1861 PetscCall(PetscMalloc1(coi[pon]+1,&coj)); 1862 PetscCall(PetscFreeSpaceContiguous(&free_space,coj)); 1863 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); /* must destroy to get a new one for C */ 1864 1865 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1866 if (afill_tmp > afill) afill = afill_tmp; 1867 1868 /* send j-array (coj) of Co to other processors */ 1869 /*----------------------------------------------*/ 1870 /* determine row ownership */ 1871 PetscCall(PetscNew(&merge)); 1872 PetscCall(PetscLayoutCreate(comm,&merge->rowmap)); 1873 1874 merge->rowmap->n = pn; 1875 merge->rowmap->bs = 1; 1876 1877 PetscCall(PetscLayoutSetUp(merge->rowmap)); 1878 owners = merge->rowmap->range; 1879 1880 /* determine the number of messages to send, their lengths */ 1881 PetscCall(PetscCalloc1(size,&len_si)); 1882 PetscCall(PetscCalloc1(size,&merge->len_s)); 1883 1884 len_s = merge->len_s; 1885 merge->nsend = 0; 1886 1887 PetscCall(PetscMalloc1(size+2,&owners_co)); 1888 1889 proc = 0; 1890 for (i=0; i<pon; i++) { 1891 while (prmap[i] >= owners[proc+1]) proc++; 1892 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1893 len_s[proc] += coi[i+1] - coi[i]; 1894 } 1895 1896 len = 0; /* max length of buf_si[] */ 1897 owners_co[0] = 0; 1898 for (proc=0; proc<size; proc++) { 1899 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1900 if (len_si[proc]) { 1901 merge->nsend++; 1902 len_si[proc] = 2*(len_si[proc] + 1); 1903 len += len_si[proc]; 1904 } 1905 } 1906 1907 /* determine the number and length of messages to receive for coi and coj */ 1908 PetscCall(PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv)); 1909 PetscCall(PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri)); 1910 1911 /* post the Irecv and Isend of coj */ 1912 PetscCall(PetscCommGetNewTag(comm,&tagj)); 1913 PetscCall(PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits)); 1914 PetscCall(PetscMalloc1(merge->nsend+1,&swaits)); 1915 for (proc=0, k=0; proc<size; proc++) { 1916 if (!len_s[proc]) continue; 1917 i = owners_co[proc]; 1918 PetscCallMPI(MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k)); 1919 k++; 1920 } 1921 1922 /* receives and sends of coj are complete */ 1923 PetscCall(PetscMalloc1(size,&sstatus)); 1924 for (i=0; i<merge->nrecv; i++) { 1925 PETSC_UNUSED PetscMPIInt icompleted; 1926 PetscCallMPI(MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus)); 1927 } 1928 PetscCall(PetscFree(rwaits)); 1929 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend,swaits,sstatus)); 1930 1931 /* add received column indices into table to update Armax */ 1932 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/tutorials/ex56.c! */ 1933 for (k=0; k<merge->nrecv; k++) {/* k-th received message */ 1934 Jptr = buf_rj[k]; 1935 for (j=0; j<merge->len_r[k]; j++) { 1936 PetscCall(PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES)); 1937 } 1938 } 1939 PetscCall(PetscTableGetCount(ta,&Armax)); 1940 1941 /* send and recv coi */ 1942 /*-------------------*/ 1943 PetscCall(PetscCommGetNewTag(comm,&tagi)); 1944 PetscCall(PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits)); 1945 PetscCall(PetscMalloc1(len+1,&buf_s)); 1946 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1947 for (proc=0,k=0; proc<size; proc++) { 1948 if (!len_s[proc]) continue; 1949 /* form outgoing message for i-structure: 1950 buf_si[0]: nrows to be sent 1951 [1:nrows]: row index (global) 1952 [nrows+1:2*nrows+1]: i-structure index 1953 */ 1954 /*-------------------------------------------*/ 1955 nrows = len_si[proc]/2 - 1; 1956 buf_si_i = buf_si + nrows+1; 1957 buf_si[0] = nrows; 1958 buf_si_i[0] = 0; 1959 nrows = 0; 1960 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1961 nzi = coi[i+1] - coi[i]; 1962 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1963 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1964 nrows++; 1965 } 1966 PetscCallMPI(MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k)); 1967 k++; 1968 buf_si += len_si[proc]; 1969 } 1970 i = merge->nrecv; 1971 while (i--) { 1972 PETSC_UNUSED PetscMPIInt icompleted; 1973 PetscCallMPI(MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus)); 1974 } 1975 PetscCall(PetscFree(rwaits)); 1976 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend,swaits,sstatus)); 1977 PetscCall(PetscFree(len_si)); 1978 PetscCall(PetscFree(len_ri)); 1979 PetscCall(PetscFree(swaits)); 1980 PetscCall(PetscFree(sstatus)); 1981 PetscCall(PetscFree(buf_s)); 1982 1983 /* compute the local portion of C (mpi mat) */ 1984 /*------------------------------------------*/ 1985 /* allocate bi array and free space for accumulating nonzero column info */ 1986 PetscCall(PetscMalloc1(pn+1,&bi)); 1987 bi[0] = 0; 1988 1989 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1990 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(pdti[pn],PetscIntSumTruncate(poti[pon],ai[am]))); 1991 PetscCall(PetscFreeSpaceGet(nnz,&free_space)); 1992 current_space = free_space; 1993 1994 PetscCall(PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci)); 1995 for (k=0; k<merge->nrecv; k++) { 1996 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1997 nrows = *buf_ri_k[k]; 1998 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1999 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th received i-structure */ 2000 } 2001 2002 PetscCall(PetscLLCondensedCreate_Scalable(Armax,&lnk)); 2003 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);PetscCall(ierr); 2004 rmax = 0; 2005 for (i=0; i<pn; i++) { 2006 /* add pdt[i,:]*AP into lnk */ 2007 pnz = pdti[i+1] - pdti[i]; 2008 ptJ = pdtj + pdti[i]; 2009 for (j=0; j<pnz; j++) { 2010 row = ptJ[j]; /* row of AP == col of Pt */ 2011 anz = ai[row+1] - ai[row]; 2012 Jptr = aj + ai[row]; 2013 /* add non-zero cols of AP into the sorted linked list lnk */ 2014 PetscCall(PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk)); 2015 } 2016 2017 /* add received col data into lnk */ 2018 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 2019 if (i == *nextrow[k]) { /* i-th row */ 2020 nzi = *(nextci[k]+1) - *nextci[k]; 2021 Jptr = buf_rj[k] + *nextci[k]; 2022 PetscCall(PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk)); 2023 nextrow[k]++; nextci[k]++; 2024 } 2025 } 2026 2027 /* add missing diagonal entry */ 2028 if (C->force_diagonals) { 2029 k = i + owners[rank]; /* column index */ 2030 PetscCall(PetscLLCondensedAddSorted_Scalable(1,&k,lnk)); 2031 } 2032 2033 nnz = lnk[0]; 2034 2035 /* if free space is not available, make more free space */ 2036 if (current_space->local_remaining<nnz) { 2037 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space)); 2038 nspacedouble++; 2039 } 2040 /* copy data into free space, then initialize lnk */ 2041 PetscCall(PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk)); 2042 PetscCall(MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz)); 2043 2044 current_space->array += nnz; 2045 current_space->local_used += nnz; 2046 current_space->local_remaining -= nnz; 2047 2048 bi[i+1] = bi[i] + nnz; 2049 if (nnz > rmax) rmax = nnz; 2050 } 2051 PetscCall(PetscFree3(buf_ri_k,nextrow,nextci)); 2052 2053 PetscCall(PetscMalloc1(bi[pn]+1,&bj)); 2054 PetscCall(PetscFreeSpaceContiguous(&free_space,bj)); 2055 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 2056 if (afill_tmp > afill) afill = afill_tmp; 2057 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 2058 PetscCall(PetscTableDestroy(&ta)); 2059 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj)); 2060 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj)); 2061 2062 /* create symbolic parallel matrix C - why cannot be assembled in Numeric part */ 2063 /*-------------------------------------------------------------------------------*/ 2064 PetscCall(MatSetSizes(C,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE)); 2065 PetscCall(MatSetBlockSizes(C,PetscAbs(P->cmap->bs),PetscAbs(A->cmap->bs))); 2066 PetscCall(MatGetType(A,&mtype)); 2067 PetscCall(MatSetType(C,mtype)); 2068 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 2069 ierr = MatPreallocateFinalize(dnz,onz);PetscCall(ierr); 2070 PetscCall(MatSetBlockSize(C,1)); 2071 PetscCall(MatSetOption(C,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE)); 2072 for (i=0; i<pn; i++) { 2073 row = i + rstart; 2074 nnz = bi[i+1] - bi[i]; 2075 Jptr = bj + bi[i]; 2076 PetscCall(MatSetValues(C,1,&row,nnz,Jptr,NULL,INSERT_VALUES)); 2077 } 2078 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 2079 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 2080 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 2081 merge->bi = bi; 2082 merge->bj = bj; 2083 merge->coi = coi; 2084 merge->coj = coj; 2085 merge->buf_ri = buf_ri; 2086 merge->buf_rj = buf_rj; 2087 merge->owners_co = owners_co; 2088 2089 /* attach the supporting struct to C for reuse */ 2090 C->product->data = ptap; 2091 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2092 ptap->merge = merge; 2093 2094 C->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2095 2096 #if defined(PETSC_USE_INFO) 2097 if (bi[pn] != 0) { 2098 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 2099 PetscCall(PetscInfo(C,"Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n",(double)afill)); 2100 } else { 2101 PetscCall(PetscInfo(C,"Empty matrix product\n")); 2102 } 2103 #endif 2104 PetscFunctionReturn(0); 2105 } 2106 2107 /* ---------------------------------------------------------------- */ 2108 static PetscErrorCode MatProductSymbolic_AtB_MPIAIJ_MPIAIJ(Mat C) 2109 { 2110 Mat_Product *product = C->product; 2111 Mat A=product->A,B=product->B; 2112 PetscReal fill=product->fill; 2113 PetscBool flg; 2114 2115 PetscFunctionBegin; 2116 /* scalable */ 2117 PetscCall(PetscStrcmp(product->alg,"scalable",&flg)); 2118 if (flg) { 2119 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C)); 2120 goto next; 2121 } 2122 2123 /* nonscalable */ 2124 PetscCall(PetscStrcmp(product->alg,"nonscalable",&flg)); 2125 if (flg) { 2126 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C)); 2127 goto next; 2128 } 2129 2130 /* matmatmult */ 2131 PetscCall(PetscStrcmp(product->alg,"at*b",&flg)); 2132 if (flg) { 2133 Mat At; 2134 Mat_APMPI *ptap; 2135 2136 PetscCall(MatTranspose(A,MAT_INITIAL_MATRIX,&At)); 2137 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(At,B,fill,C)); 2138 ptap = (Mat_APMPI*)C->product->data; 2139 if (ptap) { 2140 ptap->Pt = At; 2141 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2142 } 2143 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 2144 goto next; 2145 } 2146 2147 /* backend general code */ 2148 PetscCall(PetscStrcmp(product->alg,"backend",&flg)); 2149 if (flg) { 2150 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 2151 PetscFunctionReturn(0); 2152 } 2153 2154 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatProduct type is not supported"); 2155 2156 next: 2157 C->ops->productnumeric = MatProductNumeric_AtB; 2158 PetscFunctionReturn(0); 2159 } 2160 2161 /* ---------------------------------------------------------------- */ 2162 /* Set options for MatMatMultxxx_MPIAIJ_MPIAIJ */ 2163 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AB(Mat C) 2164 { 2165 PetscErrorCode ierr; 2166 Mat_Product *product = C->product; 2167 Mat A=product->A,B=product->B; 2168 #if defined(PETSC_HAVE_HYPRE) 2169 const char *algTypes[5] = {"scalable","nonscalable","seqmpi","backend","hypre"}; 2170 PetscInt nalg = 5; 2171 #else 2172 const char *algTypes[4] = {"scalable","nonscalable","seqmpi","backend",}; 2173 PetscInt nalg = 4; 2174 #endif 2175 PetscInt alg = 1; /* set nonscalable algorithm as default */ 2176 PetscBool flg; 2177 MPI_Comm comm; 2178 2179 PetscFunctionBegin; 2180 /* Check matrix local sizes */ 2181 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 2182 PetscCheckFalse(A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend,PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 2183 2184 /* Set "nonscalable" as default algorithm */ 2185 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2186 if (flg) { 2187 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2188 2189 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2190 if (B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2191 MatInfo Ainfo,Binfo; 2192 PetscInt nz_local; 2193 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2194 2195 PetscCall(MatGetInfo(A,MAT_LOCAL,&Ainfo)); 2196 PetscCall(MatGetInfo(B,MAT_LOCAL,&Binfo)); 2197 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2198 2199 if (B->cmap->N > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2200 PetscCall(MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm)); 2201 2202 if (alg_scalable) { 2203 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2204 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2205 PetscCall(PetscInfo(B,"Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n",B->cmap->N,(double)(product->fill*nz_local))); 2206 } 2207 } 2208 } 2209 2210 /* Get runtime option */ 2211 if (product->api_user) { 2212 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMult","Mat");PetscCall(ierr); 2213 PetscCall(PetscOptionsEList("-matmatmult_via","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2214 ierr = PetscOptionsEnd();PetscCall(ierr); 2215 } else { 2216 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AB","Mat");PetscCall(ierr); 2217 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2218 ierr = PetscOptionsEnd();PetscCall(ierr); 2219 } 2220 if (flg) { 2221 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2222 } 2223 2224 C->ops->productsymbolic = MatProductSymbolic_AB_MPIAIJ_MPIAIJ; 2225 PetscFunctionReturn(0); 2226 } 2227 2228 /* Set options for MatTransposeMatMultXXX_MPIAIJ_MPIAIJ */ 2229 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AtB(Mat C) 2230 { 2231 PetscErrorCode ierr; 2232 Mat_Product *product = C->product; 2233 Mat A=product->A,B=product->B; 2234 const char *algTypes[4] = {"scalable","nonscalable","at*b","backend"}; 2235 PetscInt nalg = 4; 2236 PetscInt alg = 1; /* set default algorithm */ 2237 PetscBool flg; 2238 MPI_Comm comm; 2239 2240 PetscFunctionBegin; 2241 /* Check matrix local sizes */ 2242 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 2243 PetscCheckFalse(A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend,PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A (%" PetscInt_FMT ", %" PetscInt_FMT ") != B (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->rmap->rstart,A->rmap->rend,B->rmap->rstart,B->rmap->rend); 2244 2245 /* Set default algorithm */ 2246 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2247 if (flg) { 2248 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2249 } 2250 2251 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2252 if (alg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2253 MatInfo Ainfo,Binfo; 2254 PetscInt nz_local; 2255 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2256 2257 PetscCall(MatGetInfo(A,MAT_LOCAL,&Ainfo)); 2258 PetscCall(MatGetInfo(B,MAT_LOCAL,&Binfo)); 2259 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2260 2261 if (B->cmap->N > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2262 PetscCall(MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm)); 2263 2264 if (alg_scalable) { 2265 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2266 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2267 PetscCall(PetscInfo(B,"Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n",B->cmap->N,(double)(product->fill*nz_local))); 2268 } 2269 } 2270 2271 /* Get runtime option */ 2272 if (product->api_user) { 2273 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatTransposeMatMult","Mat");PetscCall(ierr); 2274 PetscCall(PetscOptionsEList("-mattransposematmult_via","Algorithmic approach","MatTransposeMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2275 ierr = PetscOptionsEnd();PetscCall(ierr); 2276 } else { 2277 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AtB","Mat");PetscCall(ierr); 2278 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatTransposeMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2279 ierr = PetscOptionsEnd();PetscCall(ierr); 2280 } 2281 if (flg) { 2282 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2283 } 2284 2285 C->ops->productsymbolic = MatProductSymbolic_AtB_MPIAIJ_MPIAIJ; 2286 PetscFunctionReturn(0); 2287 } 2288 2289 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_PtAP(Mat C) 2290 { 2291 PetscErrorCode ierr; 2292 Mat_Product *product = C->product; 2293 Mat A=product->A,P=product->B; 2294 MPI_Comm comm; 2295 PetscBool flg; 2296 PetscInt alg=1; /* set default algorithm */ 2297 #if !defined(PETSC_HAVE_HYPRE) 2298 const char *algTypes[5] = {"scalable","nonscalable","allatonce","allatonce_merged","backend"}; 2299 PetscInt nalg=5; 2300 #else 2301 const char *algTypes[6] = {"scalable","nonscalable","allatonce","allatonce_merged","backend","hypre"}; 2302 PetscInt nalg=6; 2303 #endif 2304 PetscInt pN=P->cmap->N; 2305 2306 PetscFunctionBegin; 2307 /* Check matrix local sizes */ 2308 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 2309 PetscCheckFalse(A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend,PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%" PetscInt_FMT ", %" PetscInt_FMT ") != Prow (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 2310 PetscCheckFalse(A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend,PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%" PetscInt_FMT ", %" PetscInt_FMT ") != Prow (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 2311 2312 /* Set "nonscalable" as default algorithm */ 2313 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2314 if (flg) { 2315 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2316 2317 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2318 if (pN > 100000) { 2319 MatInfo Ainfo,Pinfo; 2320 PetscInt nz_local; 2321 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2322 2323 PetscCall(MatGetInfo(A,MAT_LOCAL,&Ainfo)); 2324 PetscCall(MatGetInfo(P,MAT_LOCAL,&Pinfo)); 2325 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); 2326 2327 if (pN > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2328 PetscCall(MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm)); 2329 2330 if (alg_scalable) { 2331 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2332 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2333 } 2334 } 2335 } 2336 2337 /* Get runtime option */ 2338 if (product->api_user) { 2339 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatPtAP","Mat");PetscCall(ierr); 2340 PetscCall(PetscOptionsEList("-matptap_via","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[alg],&alg,&flg)); 2341 ierr = PetscOptionsEnd();PetscCall(ierr); 2342 } else { 2343 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_PtAP","Mat");PetscCall(ierr); 2344 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[alg],&alg,&flg)); 2345 ierr = PetscOptionsEnd();PetscCall(ierr); 2346 } 2347 if (flg) { 2348 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2349 } 2350 2351 C->ops->productsymbolic = MatProductSymbolic_PtAP_MPIAIJ_MPIAIJ; 2352 PetscFunctionReturn(0); 2353 } 2354 2355 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_RARt(Mat C) 2356 { 2357 Mat_Product *product = C->product; 2358 Mat A = product->A,R=product->B; 2359 2360 PetscFunctionBegin; 2361 /* Check matrix local sizes */ 2362 PetscCheckFalse(A->cmap->n != R->cmap->n || A->rmap->n != R->cmap->n,PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A local (%" PetscInt_FMT ", %" PetscInt_FMT "), R local (%" PetscInt_FMT ",%" PetscInt_FMT ")",A->rmap->n,A->rmap->n,R->rmap->n,R->cmap->n); 2363 2364 C->ops->productsymbolic = MatProductSymbolic_RARt_MPIAIJ_MPIAIJ; 2365 PetscFunctionReturn(0); 2366 } 2367 2368 /* 2369 Set options for ABC = A*B*C = A*(B*C); ABC's algorithm must be chosen from AB's algorithm 2370 */ 2371 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABC(Mat C) 2372 { 2373 PetscErrorCode ierr; 2374 Mat_Product *product = C->product; 2375 PetscBool flg = PETSC_FALSE; 2376 PetscInt alg = 1; /* default algorithm */ 2377 const char *algTypes[3] = {"scalable","nonscalable","seqmpi"}; 2378 PetscInt nalg = 3; 2379 2380 PetscFunctionBegin; 2381 /* Set default algorithm */ 2382 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2383 if (flg) { 2384 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2385 } 2386 2387 /* Get runtime option */ 2388 if (product->api_user) { 2389 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMatMult","Mat");PetscCall(ierr); 2390 PetscCall(PetscOptionsEList("-matmatmatmult_via","Algorithmic approach","MatMatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2391 ierr = PetscOptionsEnd();PetscCall(ierr); 2392 } else { 2393 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_ABC","Mat");PetscCall(ierr); 2394 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatProduct_ABC",algTypes,nalg,algTypes[alg],&alg,&flg)); 2395 ierr = PetscOptionsEnd();PetscCall(ierr); 2396 } 2397 if (flg) { 2398 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2399 } 2400 2401 C->ops->matmatmultsymbolic = MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ; 2402 C->ops->productsymbolic = MatProductSymbolic_ABC; 2403 PetscFunctionReturn(0); 2404 } 2405 2406 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ(Mat C) 2407 { 2408 Mat_Product *product = C->product; 2409 2410 PetscFunctionBegin; 2411 switch (product->type) { 2412 case MATPRODUCT_AB: 2413 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2414 break; 2415 case MATPRODUCT_AtB: 2416 PetscCall(MatProductSetFromOptions_MPIAIJ_AtB(C)); 2417 break; 2418 case MATPRODUCT_PtAP: 2419 PetscCall(MatProductSetFromOptions_MPIAIJ_PtAP(C)); 2420 break; 2421 case MATPRODUCT_RARt: 2422 PetscCall(MatProductSetFromOptions_MPIAIJ_RARt(C)); 2423 break; 2424 case MATPRODUCT_ABC: 2425 PetscCall(MatProductSetFromOptions_MPIAIJ_ABC(C)); 2426 break; 2427 default: 2428 break; 2429 } 2430 PetscFunctionReturn(0); 2431 } 2432