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_ABt_MPIAIJ_MPIAIJ(Mat C) 19 { 20 Mat_Product *product = C->product; 21 Mat B=product->B; 22 23 PetscFunctionBegin; 24 PetscCall(MatTranspose(B,MAT_INITIAL_MATRIX,&product->B)); 25 PetscCall(MatDestroy(&B)); 26 PetscCall(MatProductSymbolic_AB_MPIAIJ_MPIAIJ(C)); 27 PetscFunctionReturn(0); 28 } 29 30 PETSC_INTERN PetscErrorCode MatProductSymbolic_AB_MPIAIJ_MPIAIJ(Mat C) 31 { 32 Mat_Product *product = C->product; 33 Mat A=product->A,B=product->B; 34 MatProductAlgorithm alg=product->alg; 35 PetscReal fill=product->fill; 36 PetscBool flg; 37 38 PetscFunctionBegin; 39 /* scalable */ 40 PetscCall(PetscStrcmp(alg,"scalable",&flg)); 41 if (flg) { 42 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C)); 43 PetscFunctionReturn(0); 44 } 45 46 /* nonscalable */ 47 PetscCall(PetscStrcmp(alg,"nonscalable",&flg)); 48 if (flg) { 49 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C)); 50 PetscFunctionReturn(0); 51 } 52 53 /* seqmpi */ 54 PetscCall(PetscStrcmp(alg,"seqmpi",&flg)); 55 if (flg) { 56 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(A,B,fill,C)); 57 PetscFunctionReturn(0); 58 } 59 60 /* backend general code */ 61 PetscCall(PetscStrcmp(alg,"backend",&flg)); 62 if (flg) { 63 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 64 PetscFunctionReturn(0); 65 } 66 67 #if defined(PETSC_HAVE_HYPRE) 68 PetscCall(PetscStrcmp(alg,"hypre",&flg)); 69 if (flg) { 70 PetscCall(MatMatMultSymbolic_AIJ_AIJ_wHYPRE(A,B,fill,C)); 71 PetscFunctionReturn(0); 72 } 73 #endif 74 SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_SUP,"Mat Product Algorithm is not supported"); 75 } 76 77 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(void *data) 78 { 79 Mat_APMPI *ptap = (Mat_APMPI*)data; 80 81 PetscFunctionBegin; 82 PetscCall(PetscFree2(ptap->startsj_s,ptap->startsj_r)); 83 PetscCall(PetscFree(ptap->bufa)); 84 PetscCall(MatDestroy(&ptap->P_loc)); 85 PetscCall(MatDestroy(&ptap->P_oth)); 86 PetscCall(MatDestroy(&ptap->Pt)); 87 PetscCall(PetscFree(ptap->api)); 88 PetscCall(PetscFree(ptap->apj)); 89 PetscCall(PetscFree(ptap->apa)); 90 PetscCall(PetscFree(ptap)); 91 PetscFunctionReturn(0); 92 } 93 94 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,Mat C) 95 { 96 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 97 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 98 Mat_SeqAIJ *cd =(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 99 PetscScalar *cda=cd->a,*coa=co->a; 100 Mat_SeqAIJ *p_loc,*p_oth; 101 PetscScalar *apa,*ca; 102 PetscInt cm =C->rmap->n; 103 Mat_APMPI *ptap; 104 PetscInt *api,*apj,*apJ,i,k; 105 PetscInt cstart=C->cmap->rstart; 106 PetscInt cdnz,conz,k0,k1; 107 const PetscScalar *dummy; 108 MPI_Comm comm; 109 PetscMPIInt size; 110 111 PetscFunctionBegin; 112 MatCheckProduct(C,3); 113 ptap = (Mat_APMPI*)C->product->data; 114 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 115 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 116 PetscCallMPI(MPI_Comm_size(comm,&size)); 117 PetscCheck(ptap->P_oth || size <= 1,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatProductClear()"); 118 119 /* flag CPU mask for C */ 120 #if defined(PETSC_HAVE_DEVICE) 121 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 122 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 123 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 124 #endif 125 126 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 127 /*-----------------------------------------------------*/ 128 /* update numerical values of P_oth and P_loc */ 129 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 130 PetscCall(MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc)); 131 132 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 133 /*----------------------------------------------------------*/ 134 /* get data from symbolic products */ 135 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 136 p_oth = NULL; 137 if (size >1) { 138 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 139 } 140 141 /* get apa for storing dense row A[i,:]*P */ 142 apa = ptap->apa; 143 144 api = ptap->api; 145 apj = ptap->apj; 146 /* trigger copy to CPU */ 147 PetscCall(MatSeqAIJGetArrayRead(a->A,&dummy)); 148 PetscCall(MatSeqAIJRestoreArrayRead(a->A,&dummy)); 149 PetscCall(MatSeqAIJGetArrayRead(a->B,&dummy)); 150 PetscCall(MatSeqAIJRestoreArrayRead(a->B,&dummy)); 151 for (i=0; i<cm; i++) { 152 /* compute apa = A[i,:]*P */ 153 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 154 155 /* set values in C */ 156 apJ = apj + api[i]; 157 cdnz = cd->i[i+1] - cd->i[i]; 158 conz = co->i[i+1] - co->i[i]; 159 160 /* 1st off-diagonal part of C */ 161 ca = coa + co->i[i]; 162 k = 0; 163 for (k0=0; k0<conz; k0++) { 164 if (apJ[k] >= cstart) break; 165 ca[k0] = apa[apJ[k]]; 166 apa[apJ[k++]] = 0.0; 167 } 168 169 /* diagonal part of C */ 170 ca = cda + cd->i[i]; 171 for (k1=0; k1<cdnz; k1++) { 172 ca[k1] = apa[apJ[k]]; 173 apa[apJ[k++]] = 0.0; 174 } 175 176 /* 2nd off-diagonal part of C */ 177 ca = coa + co->i[i]; 178 for (; k0<conz; k0++) { 179 ca[k0] = apa[apJ[k]]; 180 apa[apJ[k++]] = 0.0; 181 } 182 } 183 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 184 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 185 PetscFunctionReturn(0); 186 } 187 188 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,PetscReal fill,Mat C) 189 { 190 MPI_Comm comm; 191 PetscMPIInt size; 192 Mat_APMPI *ptap; 193 PetscFreeSpaceList free_space=NULL,current_space=NULL; 194 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 195 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 196 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 197 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 198 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 199 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 200 PetscBT lnkbt; 201 PetscReal afill; 202 MatType mtype; 203 204 PetscFunctionBegin; 205 MatCheckProduct(C,4); 206 PetscCheck(!C->product->data,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 207 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 208 PetscCallMPI(MPI_Comm_size(comm,&size)); 209 210 /* create struct Mat_APMPI and attached it to C later */ 211 PetscCall(PetscNew(&ptap)); 212 213 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 214 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 215 216 /* get P_loc by taking all local rows of P */ 217 PetscCall(MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc)); 218 219 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 220 pi_loc = p_loc->i; pj_loc = p_loc->j; 221 if (size > 1) { 222 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 223 pi_oth = p_oth->i; pj_oth = p_oth->j; 224 } else { 225 p_oth = NULL; 226 pi_oth = NULL; pj_oth = NULL; 227 } 228 229 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 230 /*-------------------------------------------------------------------*/ 231 PetscCall(PetscMalloc1(am+2,&api)); 232 ptap->api = api; 233 api[0] = 0; 234 235 /* create and initialize a linked list */ 236 PetscCall(PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt)); 237 238 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 239 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space)); 240 current_space = free_space; 241 242 MatPreallocateBegin(comm,am,pn,dnz,onz); 243 for (i=0; i<am; i++) { 244 /* diagonal portion of A */ 245 nzi = adi[i+1] - adi[i]; 246 for (j=0; j<nzi; j++) { 247 row = *adj++; 248 pnz = pi_loc[row+1] - pi_loc[row]; 249 Jptr = pj_loc + pi_loc[row]; 250 /* add non-zero cols of P into the sorted linked list lnk */ 251 PetscCall(PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt)); 252 } 253 /* off-diagonal portion of A */ 254 nzi = aoi[i+1] - aoi[i]; 255 for (j=0; j<nzi; j++) { 256 row = *aoj++; 257 pnz = pi_oth[row+1] - pi_oth[row]; 258 Jptr = pj_oth + pi_oth[row]; 259 PetscCall(PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt)); 260 } 261 /* add possible missing diagonal entry */ 262 if (C->force_diagonals) { 263 j = i + rstart; /* column index */ 264 PetscCall(PetscLLCondensedAddSorted(1,&j,lnk,lnkbt)); 265 } 266 267 apnz = lnk[0]; 268 api[i+1] = api[i] + apnz; 269 270 /* if free space is not available, double the total space in the list */ 271 if (current_space->local_remaining<apnz) { 272 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space)); 273 nspacedouble++; 274 } 275 276 /* Copy data into free space, then initialize lnk */ 277 PetscCall(PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt)); 278 PetscCall(MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz)); 279 280 current_space->array += apnz; 281 current_space->local_used += apnz; 282 current_space->local_remaining -= apnz; 283 } 284 285 /* Allocate space for apj, initialize apj, and */ 286 /* destroy list of free space and other temporary array(s) */ 287 PetscCall(PetscMalloc1(api[am]+1,&ptap->apj)); 288 apj = ptap->apj; 289 PetscCall(PetscFreeSpaceContiguous(&free_space,ptap->apj)); 290 PetscCall(PetscLLDestroy(lnk,lnkbt)); 291 292 /* malloc apa to store dense row A[i,:]*P */ 293 PetscCall(PetscCalloc1(pN,&ptap->apa)); 294 295 /* set and assemble symbolic parallel matrix C */ 296 /*---------------------------------------------*/ 297 PetscCall(MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE)); 298 PetscCall(MatSetBlockSizesFromMats(C,A,P)); 299 300 PetscCall(MatGetType(A,&mtype)); 301 PetscCall(MatSetType(C,mtype)); 302 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 303 MatPreallocateEnd(dnz,onz); 304 305 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 306 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 307 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 308 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 309 310 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 311 C->ops->productnumeric = MatProductNumeric_AB; 312 313 /* attach the supporting struct to C for reuse */ 314 C->product->data = ptap; 315 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 316 317 /* set MatInfo */ 318 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 319 if (afill < 1.0) afill = 1.0; 320 C->info.mallocs = nspacedouble; 321 C->info.fill_ratio_given = fill; 322 C->info.fill_ratio_needed = afill; 323 324 #if defined(PETSC_USE_INFO) 325 if (api[am]) { 326 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 327 PetscCall(PetscInfo(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill)); 328 } else { 329 PetscCall(PetscInfo(C,"Empty matrix product\n")); 330 } 331 #endif 332 PetscFunctionReturn(0); 333 } 334 335 /* ------------------------------------------------------- */ 336 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat,Mat,PetscReal,Mat); 337 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat,Mat,Mat); 338 339 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AB(Mat C) 340 { 341 Mat_Product *product = C->product; 342 Mat A = product->A,B=product->B; 343 344 PetscFunctionBegin; 345 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) 346 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); 347 348 C->ops->matmultsymbolic = MatMatMultSymbolic_MPIAIJ_MPIDense; 349 C->ops->productsymbolic = MatProductSymbolic_AB; 350 PetscFunctionReturn(0); 351 } 352 /* -------------------------------------------------------------------- */ 353 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(Mat C) 354 { 355 Mat_Product *product = C->product; 356 Mat A = product->A,B=product->B; 357 358 PetscFunctionBegin; 359 if (A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend) 360 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); 361 362 C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIAIJ_MPIDense; 363 C->ops->productsymbolic = MatProductSymbolic_AtB; 364 PetscFunctionReturn(0); 365 } 366 367 /* --------------------------------------------------------------------- */ 368 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense(Mat C) 369 { 370 Mat_Product *product = C->product; 371 372 PetscFunctionBegin; 373 switch (product->type) { 374 case MATPRODUCT_AB: 375 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AB(C)); 376 break; 377 case MATPRODUCT_AtB: 378 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(C)); 379 break; 380 default: 381 break; 382 } 383 PetscFunctionReturn(0); 384 } 385 /* ------------------------------------------------------- */ 386 387 typedef struct { 388 Mat workB,workB1; 389 MPI_Request *rwaits,*swaits; 390 PetscInt nsends,nrecvs; 391 MPI_Datatype *stype,*rtype; 392 PetscInt blda; 393 } MPIAIJ_MPIDense; 394 395 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 396 { 397 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*)ctx; 398 PetscInt i; 399 400 PetscFunctionBegin; 401 PetscCall(MatDestroy(&contents->workB)); 402 PetscCall(MatDestroy(&contents->workB1)); 403 for (i=0; i<contents->nsends; i++) { 404 PetscCallMPI(MPI_Type_free(&contents->stype[i])); 405 } 406 for (i=0; i<contents->nrecvs; i++) { 407 PetscCallMPI(MPI_Type_free(&contents->rtype[i])); 408 } 409 PetscCall(PetscFree4(contents->stype,contents->rtype,contents->rwaits,contents->swaits)); 410 PetscCall(PetscFree(contents)); 411 PetscFunctionReturn(0); 412 } 413 414 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat C) 415 { 416 Mat_MPIAIJ *aij=(Mat_MPIAIJ*)A->data; 417 PetscInt nz=aij->B->cmap->n,nsends,nrecvs,i,nrows_to,j,blda,m,M,n,N; 418 MPIAIJ_MPIDense *contents; 419 VecScatter ctx=aij->Mvctx; 420 PetscInt Am=A->rmap->n,Bm=B->rmap->n,BN=B->cmap->N,Bbn,Bbn1,bs,nrows_from,numBb; 421 MPI_Comm comm; 422 MPI_Datatype type1,*stype,*rtype; 423 const PetscInt *sindices,*sstarts,*rstarts; 424 PetscMPIInt *disp; 425 PetscBool cisdense; 426 427 PetscFunctionBegin; 428 MatCheckProduct(C,4); 429 PetscCheck(!C->product->data,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data not empty"); 430 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 431 PetscCall(PetscObjectBaseTypeCompare((PetscObject)C,MATMPIDENSE,&cisdense)); 432 if (!cisdense) { 433 PetscCall(MatSetType(C,((PetscObject)B)->type_name)); 434 } 435 PetscCall(MatGetLocalSize(C,&m,&n)); 436 PetscCall(MatGetSize(C,&M,&N)); 437 if (m == PETSC_DECIDE || n == PETSC_DECIDE || M == PETSC_DECIDE || N == PETSC_DECIDE) { 438 PetscCall(MatSetSizes(C,Am,B->cmap->n,A->rmap->N,BN)); 439 } 440 PetscCall(MatSetBlockSizesFromMats(C,A,B)); 441 PetscCall(MatSetUp(C)); 442 PetscCall(MatDenseGetLDA(B,&blda)); 443 PetscCall(PetscNew(&contents)); 444 445 PetscCall(VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,NULL,NULL)); 446 PetscCall(VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,NULL,NULL)); 447 448 /* Create column block of B and C for memory scalability when BN is too large */ 449 /* Estimate Bbn, column size of Bb */ 450 if (nz) { 451 Bbn1 = 2*Am*BN/nz; 452 if (!Bbn1) Bbn1 = 1; 453 } else Bbn1 = BN; 454 455 bs = PetscAbs(B->cmap->bs); 456 Bbn1 = Bbn1/bs *bs; /* Bbn1 is a multiple of bs */ 457 if (Bbn1 > BN) Bbn1 = BN; 458 PetscCallMPI(MPI_Allreduce(&Bbn1,&Bbn,1,MPIU_INT,MPI_MAX,comm)); 459 460 /* Enable runtime option for Bbn */ 461 PetscOptionsBegin(comm,((PetscObject)C)->prefix,"MatMatMult","Mat"); 462 PetscCall(PetscOptionsInt("-matmatmult_Bbn","Number of columns in Bb","MatMatMult",Bbn,&Bbn,NULL)); 463 PetscOptionsEnd(); 464 Bbn = PetscMin(Bbn,BN); 465 466 if (Bbn > 0 && Bbn < BN) { 467 numBb = BN/Bbn; 468 Bbn1 = BN - numBb*Bbn; 469 } else numBb = 0; 470 471 if (numBb) { 472 PetscCall(PetscInfo(C,"use Bb, BN=%" PetscInt_FMT ", Bbn=%" PetscInt_FMT "; numBb=%" PetscInt_FMT "\n",BN,Bbn,numBb)); 473 if (Bbn1) { /* Create workB1 for the remaining columns */ 474 PetscCall(PetscInfo(C,"use Bb1, BN=%" PetscInt_FMT ", Bbn1=%" PetscInt_FMT "\n",BN,Bbn1)); 475 /* Create work matrix used to store off processor rows of B needed for local product */ 476 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nz,Bbn1,NULL,&contents->workB1)); 477 } else contents->workB1 = NULL; 478 } 479 480 /* Create work matrix used to store off processor rows of B needed for local product */ 481 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nz,Bbn,NULL,&contents->workB)); 482 483 /* Use MPI derived data type to reduce memory required by the send/recv buffers */ 484 PetscCall(PetscMalloc4(nsends,&stype,nrecvs,&rtype,nrecvs,&contents->rwaits,nsends,&contents->swaits)); 485 contents->stype = stype; 486 contents->nsends = nsends; 487 488 contents->rtype = rtype; 489 contents->nrecvs = nrecvs; 490 contents->blda = blda; 491 492 PetscCall(PetscMalloc1(Bm+1,&disp)); 493 for (i=0; i<nsends; i++) { 494 nrows_to = sstarts[i+1]-sstarts[i]; 495 for (j=0; j<nrows_to; j++) disp[j] = sindices[sstarts[i]+j]; /* rowB to be sent */ 496 PetscCallMPI(MPI_Type_create_indexed_block(nrows_to,1,disp,MPIU_SCALAR,&type1)); 497 PetscCallMPI(MPI_Type_create_resized(type1,0,blda*sizeof(PetscScalar),&stype[i])); 498 PetscCallMPI(MPI_Type_commit(&stype[i])); 499 PetscCallMPI(MPI_Type_free(&type1)); 500 } 501 502 for (i=0; i<nrecvs; i++) { 503 /* received values from a process form a (nrows_from x Bbn) row block in workB (column-wise) */ 504 nrows_from = rstarts[i+1]-rstarts[i]; 505 disp[0] = 0; 506 PetscCallMPI(MPI_Type_create_indexed_block(1,nrows_from,disp,MPIU_SCALAR,&type1)); 507 PetscCallMPI(MPI_Type_create_resized(type1,0,nz*sizeof(PetscScalar),&rtype[i])); 508 PetscCallMPI(MPI_Type_commit(&rtype[i])); 509 PetscCallMPI(MPI_Type_free(&type1)); 510 } 511 512 PetscCall(PetscFree(disp)); 513 PetscCall(VecScatterRestoreRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,NULL,NULL)); 514 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,NULL,NULL)); 515 PetscCall(MatSetOption(C,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE)); 516 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 517 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 518 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 519 520 C->product->data = contents; 521 C->product->destroy = MatMPIAIJ_MPIDenseDestroy; 522 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 523 PetscFunctionReturn(0); 524 } 525 526 PETSC_INTERN PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat,const PetscBool); 527 /* 528 Performs an efficient scatter on the rows of B needed by this process; this is 529 a modification of the VecScatterBegin_() routines. 530 531 Input: Bbidx = 0: B = Bb 532 = 1: B = Bb1, see MatMatMultSymbolic_MPIAIJ_MPIDense() 533 */ 534 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,PetscInt Bbidx,Mat C,Mat *outworkB) 535 { 536 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 537 const PetscScalar *b; 538 PetscScalar *rvalues; 539 VecScatter ctx = aij->Mvctx; 540 const PetscInt *sindices,*sstarts,*rstarts; 541 const PetscMPIInt *sprocs,*rprocs; 542 PetscInt i,nsends,nrecvs; 543 MPI_Request *swaits,*rwaits; 544 MPI_Comm comm; 545 PetscMPIInt tag=((PetscObject)ctx)->tag,ncols=B->cmap->N,nrows=aij->B->cmap->n,nsends_mpi,nrecvs_mpi; 546 MPIAIJ_MPIDense *contents; 547 Mat workB; 548 MPI_Datatype *stype,*rtype; 549 PetscInt blda; 550 551 PetscFunctionBegin; 552 MatCheckProduct(C,4); 553 PetscCheck(C->product->data,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data empty"); 554 contents = (MPIAIJ_MPIDense*)C->product->data; 555 PetscCall(VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL/*bs*/)); 556 PetscCall(VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,&rprocs,NULL/*bs*/)); 557 PetscCall(PetscMPIIntCast(nsends,&nsends_mpi)); 558 PetscCall(PetscMPIIntCast(nrecvs,&nrecvs_mpi)); 559 if (Bbidx == 0) workB = *outworkB = contents->workB; 560 else workB = *outworkB = contents->workB1; 561 PetscCheck(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); 562 swaits = contents->swaits; 563 rwaits = contents->rwaits; 564 565 PetscCall(MatDenseGetArrayRead(B,&b)); 566 PetscCall(MatDenseGetLDA(B,&blda)); 567 PetscCheck(blda == contents->blda,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot reuse an input matrix with lda %" PetscInt_FMT " != %" PetscInt_FMT,blda,contents->blda); 568 PetscCall(MatDenseGetArray(workB,&rvalues)); 569 570 /* Post recv, use MPI derived data type to save memory */ 571 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 572 rtype = contents->rtype; 573 for (i=0; i<nrecvs; i++) { 574 PetscCallMPI(MPI_Irecv(rvalues+(rstarts[i]-rstarts[0]),ncols,rtype[i],rprocs[i],tag,comm,rwaits+i)); 575 } 576 577 stype = contents->stype; 578 for (i=0; i<nsends; i++) { 579 PetscCallMPI(MPI_Isend(b,ncols,stype[i],sprocs[i],tag,comm,swaits+i)); 580 } 581 582 if (nrecvs) PetscCallMPI(MPI_Waitall(nrecvs_mpi,rwaits,MPI_STATUSES_IGNORE)); 583 if (nsends) PetscCallMPI(MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE)); 584 585 PetscCall(VecScatterRestoreRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL)); 586 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL,&rprocs,NULL)); 587 PetscCall(MatDenseRestoreArrayRead(B,&b)); 588 PetscCall(MatDenseRestoreArray(workB,&rvalues)); 589 PetscFunctionReturn(0); 590 } 591 592 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 593 { 594 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 595 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 596 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 597 Mat workB; 598 MPIAIJ_MPIDense *contents; 599 600 PetscFunctionBegin; 601 MatCheckProduct(C,3); 602 PetscCheck(C->product->data,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Product data empty"); 603 contents = (MPIAIJ_MPIDense*)C->product->data; 604 /* diagonal block of A times all local rows of B */ 605 /* TODO: this calls a symbolic multiplication every time, which could be avoided */ 606 PetscCall(MatMatMult(aij->A,bdense->A,MAT_REUSE_MATRIX,PETSC_DEFAULT,&cdense->A)); 607 if (contents->workB->cmap->n == B->cmap->N) { 608 /* get off processor parts of B needed to complete C=A*B */ 609 PetscCall(MatMPIDenseScatter(A,B,0,C,&workB)); 610 611 /* off-diagonal block of A times nonlocal rows of B */ 612 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A,PETSC_TRUE)); 613 } else { 614 Mat Bb,Cb; 615 PetscInt BN=B->cmap->N,n=contents->workB->cmap->n,i; 616 PetscBool ccpu; 617 618 PetscCheck(n > 0,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Column block size %" PetscInt_FMT " must be positive",n); 619 /* Prevent from unneeded copies back and forth from the GPU 620 when getting and restoring the submatrix 621 We need a proper GPU code for AIJ * dense in parallel */ 622 PetscCall(MatBoundToCPU(C,&ccpu)); 623 PetscCall(MatBindToCPU(C,PETSC_TRUE)); 624 for (i=0; i<BN; i+=n) { 625 PetscCall(MatDenseGetSubMatrix(B,PETSC_DECIDE,PETSC_DECIDE,i,PetscMin(i+n,BN),&Bb)); 626 PetscCall(MatDenseGetSubMatrix(C,PETSC_DECIDE,PETSC_DECIDE,i,PetscMin(i+n,BN),&Cb)); 627 628 /* get off processor parts of B needed to complete C=A*B */ 629 PetscCall(MatMPIDenseScatter(A,Bb,(i+n)>BN,C,&workB)); 630 631 /* off-diagonal block of A times nonlocal rows of B */ 632 cdense = (Mat_MPIDense*)Cb->data; 633 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A,PETSC_TRUE)); 634 PetscCall(MatDenseRestoreSubMatrix(B,&Bb)); 635 PetscCall(MatDenseRestoreSubMatrix(C,&Cb)); 636 } 637 PetscCall(MatBindToCPU(C,ccpu)); 638 } 639 PetscFunctionReturn(0); 640 } 641 642 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 643 { 644 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 645 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 646 Mat_SeqAIJ *cd = (Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 647 PetscInt *adi = ad->i,*adj,*aoi=ao->i,*aoj; 648 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 649 Mat_SeqAIJ *p_loc,*p_oth; 650 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 651 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 652 PetscInt cm = C->rmap->n,anz,pnz; 653 Mat_APMPI *ptap; 654 PetscScalar *apa_sparse; 655 const PetscScalar *dummy; 656 PetscInt *api,*apj,*apJ,i,j,k,row; 657 PetscInt cstart = C->cmap->rstart; 658 PetscInt cdnz,conz,k0,k1,nextp; 659 MPI_Comm comm; 660 PetscMPIInt size; 661 662 PetscFunctionBegin; 663 MatCheckProduct(C,3); 664 ptap = (Mat_APMPI*)C->product->data; 665 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 666 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 667 PetscCallMPI(MPI_Comm_size(comm,&size)); 668 PetscCheck(ptap->P_oth || size <= 1,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatProductClear()"); 669 670 /* flag CPU mask for C */ 671 #if defined(PETSC_HAVE_DEVICE) 672 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 673 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 674 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 675 #endif 676 apa_sparse = ptap->apa; 677 678 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 679 /*-----------------------------------------------------*/ 680 /* update numerical values of P_oth and P_loc */ 681 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 682 PetscCall(MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc)); 683 684 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 685 /*----------------------------------------------------------*/ 686 /* get data from symbolic products */ 687 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 688 pi_loc = p_loc->i; pj_loc = p_loc->j; pa_loc = p_loc->a; 689 if (size >1) { 690 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 691 pi_oth = p_oth->i; pj_oth = p_oth->j; pa_oth = p_oth->a; 692 } else { 693 p_oth = NULL; pi_oth = NULL; pj_oth = NULL; pa_oth = NULL; 694 } 695 696 /* trigger copy to CPU */ 697 PetscCall(MatSeqAIJGetArrayRead(a->A,&dummy)); 698 PetscCall(MatSeqAIJRestoreArrayRead(a->A,&dummy)); 699 PetscCall(MatSeqAIJGetArrayRead(a->B,&dummy)); 700 PetscCall(MatSeqAIJRestoreArrayRead(a->B,&dummy)); 701 api = ptap->api; 702 apj = ptap->apj; 703 for (i=0; i<cm; i++) { 704 apJ = apj + api[i]; 705 706 /* diagonal portion of A */ 707 anz = adi[i+1] - adi[i]; 708 adj = ad->j + adi[i]; 709 ada = ad->a + adi[i]; 710 for (j=0; j<anz; j++) { 711 row = adj[j]; 712 pnz = pi_loc[row+1] - pi_loc[row]; 713 pj = pj_loc + pi_loc[row]; 714 pa = pa_loc + pi_loc[row]; 715 /* perform sparse axpy */ 716 valtmp = ada[j]; 717 nextp = 0; 718 for (k=0; nextp<pnz; k++) { 719 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 720 apa_sparse[k] += valtmp*pa[nextp++]; 721 } 722 } 723 PetscCall(PetscLogFlops(2.0*pnz)); 724 } 725 726 /* off-diagonal portion of A */ 727 anz = aoi[i+1] - aoi[i]; 728 aoj = ao->j + aoi[i]; 729 aoa = ao->a + aoi[i]; 730 for (j=0; j<anz; j++) { 731 row = aoj[j]; 732 pnz = pi_oth[row+1] - pi_oth[row]; 733 pj = pj_oth + pi_oth[row]; 734 pa = pa_oth + pi_oth[row]; 735 /* perform sparse axpy */ 736 valtmp = aoa[j]; 737 nextp = 0; 738 for (k=0; nextp<pnz; k++) { 739 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 740 apa_sparse[k] += valtmp*pa[nextp++]; 741 } 742 } 743 PetscCall(PetscLogFlops(2.0*pnz)); 744 } 745 746 /* set values in C */ 747 cdnz = cd->i[i+1] - cd->i[i]; 748 conz = co->i[i+1] - co->i[i]; 749 750 /* 1st off-diagonal part of C */ 751 ca = coa + co->i[i]; 752 k = 0; 753 for (k0=0; k0<conz; k0++) { 754 if (apJ[k] >= cstart) break; 755 ca[k0] = apa_sparse[k]; 756 apa_sparse[k] = 0.0; 757 k++; 758 } 759 760 /* diagonal part of C */ 761 ca = cda + cd->i[i]; 762 for (k1=0; k1<cdnz; k1++) { 763 ca[k1] = apa_sparse[k]; 764 apa_sparse[k] = 0.0; 765 k++; 766 } 767 768 /* 2nd off-diagonal part of C */ 769 ca = coa + co->i[i]; 770 for (; k0<conz; k0++) { 771 ca[k0] = apa_sparse[k]; 772 apa_sparse[k] = 0.0; 773 k++; 774 } 775 } 776 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 777 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 778 PetscFunctionReturn(0); 779 } 780 781 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 782 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat C) 783 { 784 MPI_Comm comm; 785 PetscMPIInt size; 786 Mat_APMPI *ptap; 787 PetscFreeSpaceList free_space = NULL,current_space=NULL; 788 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 789 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 790 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 791 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 792 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=1; 793 PetscInt am=A->rmap->n,pn=P->cmap->n,pm=P->rmap->n,lsize=pn+20; 794 PetscReal afill; 795 MatType mtype; 796 797 PetscFunctionBegin; 798 MatCheckProduct(C,4); 799 PetscCheck(!C->product->data,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 800 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 801 PetscCallMPI(MPI_Comm_size(comm,&size)); 802 803 /* create struct Mat_APMPI and attached it to C later */ 804 PetscCall(PetscNew(&ptap)); 805 806 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 807 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 808 809 /* get P_loc by taking all local rows of P */ 810 PetscCall(MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc)); 811 812 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 813 pi_loc = p_loc->i; pj_loc = p_loc->j; 814 if (size > 1) { 815 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 816 pi_oth = p_oth->i; pj_oth = p_oth->j; 817 } else { 818 p_oth = NULL; 819 pi_oth = NULL; pj_oth = NULL; 820 } 821 822 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 823 /*-------------------------------------------------------------------*/ 824 PetscCall(PetscMalloc1(am+2,&api)); 825 ptap->api = api; 826 api[0] = 0; 827 828 PetscCall(PetscLLCondensedCreate_Scalable(lsize,&lnk)); 829 830 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 831 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space)); 832 current_space = free_space; 833 MatPreallocateBegin(comm,am,pn,dnz,onz); 834 for (i=0; i<am; i++) { 835 /* diagonal portion of A */ 836 nzi = adi[i+1] - adi[i]; 837 for (j=0; j<nzi; j++) { 838 row = *adj++; 839 pnz = pi_loc[row+1] - pi_loc[row]; 840 Jptr = pj_loc + pi_loc[row]; 841 /* Expand list if it is not long enough */ 842 if (pnz+apnz_max > lsize) { 843 lsize = pnz+apnz_max; 844 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 845 } 846 /* add non-zero cols of P into the sorted linked list lnk */ 847 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk)); 848 apnz = *lnk; /* The first element in the list is the number of items in the list */ 849 api[i+1] = api[i] + apnz; 850 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 851 } 852 /* off-diagonal portion of A */ 853 nzi = aoi[i+1] - aoi[i]; 854 for (j=0; j<nzi; j++) { 855 row = *aoj++; 856 pnz = pi_oth[row+1] - pi_oth[row]; 857 Jptr = pj_oth + pi_oth[row]; 858 /* Expand list if it is not long enough */ 859 if (pnz+apnz_max > lsize) { 860 lsize = pnz + apnz_max; 861 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 862 } 863 /* add non-zero cols of P into the sorted linked list lnk */ 864 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk)); 865 apnz = *lnk; /* The first element in the list is the number of items in the list */ 866 api[i+1] = api[i] + apnz; 867 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 868 } 869 870 /* add missing diagonal entry */ 871 if (C->force_diagonals) { 872 j = i + rstart; /* column index */ 873 PetscCall(PetscLLCondensedAddSorted_Scalable(1,&j,lnk)); 874 } 875 876 apnz = *lnk; 877 api[i+1] = api[i] + apnz; 878 if (apnz > apnz_max) apnz_max = apnz; 879 880 /* if free space is not available, double the total space in the list */ 881 if (current_space->local_remaining<apnz) { 882 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space)); 883 nspacedouble++; 884 } 885 886 /* Copy data into free space, then initialize lnk */ 887 PetscCall(PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk)); 888 PetscCall(MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz)); 889 890 current_space->array += apnz; 891 current_space->local_used += apnz; 892 current_space->local_remaining -= apnz; 893 } 894 895 /* Allocate space for apj, initialize apj, and */ 896 /* destroy list of free space and other temporary array(s) */ 897 PetscCall(PetscMalloc1(api[am]+1,&ptap->apj)); 898 apj = ptap->apj; 899 PetscCall(PetscFreeSpaceContiguous(&free_space,ptap->apj)); 900 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 901 902 /* create and assemble symbolic parallel matrix C */ 903 /*----------------------------------------------------*/ 904 PetscCall(MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE)); 905 PetscCall(MatSetBlockSizesFromMats(C,A,P)); 906 PetscCall(MatGetType(A,&mtype)); 907 PetscCall(MatSetType(C,mtype)); 908 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 909 MatPreallocateEnd(dnz,onz); 910 911 /* malloc apa for assembly C */ 912 PetscCall(PetscCalloc1(apnz_max,&ptap->apa)); 913 914 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 915 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 916 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 917 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 918 919 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 920 C->ops->productnumeric = MatProductNumeric_AB; 921 922 /* attach the supporting struct to C for reuse */ 923 C->product->data = ptap; 924 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 925 926 /* set MatInfo */ 927 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 928 if (afill < 1.0) afill = 1.0; 929 C->info.mallocs = nspacedouble; 930 C->info.fill_ratio_given = fill; 931 C->info.fill_ratio_needed = afill; 932 933 #if defined(PETSC_USE_INFO) 934 if (api[am]) { 935 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 936 PetscCall(PetscInfo(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill)); 937 } else { 938 PetscCall(PetscInfo(C,"Empty matrix product\n")); 939 } 940 #endif 941 PetscFunctionReturn(0); 942 } 943 944 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 945 /* Three input arrays are merged to one output array. The size of the */ 946 /* output array is also output. Duplicate entries only show up once. */ 947 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, 948 PetscInt size2, PetscInt *in2, 949 PetscInt size3, PetscInt *in3, 950 PetscInt *size4, PetscInt *out) 951 { 952 int i = 0, j = 0, k = 0, l = 0; 953 954 /* Traverse all three arrays */ 955 while (i<size1 && j<size2 && k<size3) { 956 if (in1[i] < in2[j] && in1[i] < in3[k]) { 957 out[l++] = in1[i++]; 958 } 959 else if (in2[j] < in1[i] && in2[j] < in3[k]) { 960 out[l++] = in2[j++]; 961 } 962 else if (in3[k] < in1[i] && in3[k] < in2[j]) { 963 out[l++] = in3[k++]; 964 } 965 else if (in1[i] == in2[j] && in1[i] < in3[k]) { 966 out[l++] = in1[i]; 967 i++, j++; 968 } 969 else if (in1[i] == in3[k] && in1[i] < in2[j]) { 970 out[l++] = in1[i]; 971 i++, k++; 972 } 973 else if (in3[k] == in2[j] && in2[j] < in1[i]) { 974 out[l++] = in2[j]; 975 k++, j++; 976 } 977 else if (in1[i] == in2[j] && in1[i] == in3[k]) { 978 out[l++] = in1[i]; 979 i++, j++, k++; 980 } 981 } 982 983 /* Traverse two remaining arrays */ 984 while (i<size1 && j<size2) { 985 if (in1[i] < in2[j]) { 986 out[l++] = in1[i++]; 987 } 988 else if (in1[i] > in2[j]) { 989 out[l++] = in2[j++]; 990 } 991 else { 992 out[l++] = in1[i]; 993 i++, j++; 994 } 995 } 996 997 while (i<size1 && k<size3) { 998 if (in1[i] < in3[k]) { 999 out[l++] = in1[i++]; 1000 } 1001 else if (in1[i] > in3[k]) { 1002 out[l++] = in3[k++]; 1003 } 1004 else { 1005 out[l++] = in1[i]; 1006 i++, k++; 1007 } 1008 } 1009 1010 while (k<size3 && j<size2) { 1011 if (in3[k] < in2[j]) { 1012 out[l++] = in3[k++]; 1013 } 1014 else if (in3[k] > in2[j]) { 1015 out[l++] = in2[j++]; 1016 } 1017 else { 1018 out[l++] = in3[k]; 1019 k++, j++; 1020 } 1021 } 1022 1023 /* Traverse one remaining array */ 1024 while (i<size1) out[l++] = in1[i++]; 1025 while (j<size2) out[l++] = in2[j++]; 1026 while (k<size3) out[l++] = in3[k++]; 1027 1028 *size4 = l; 1029 } 1030 1031 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 1032 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 1033 /* matrix-matrix multiplications. */ 1034 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat C) 1035 { 1036 MPI_Comm comm; 1037 PetscMPIInt size; 1038 Mat_APMPI *ptap; 1039 PetscFreeSpaceList free_space_diag=NULL, current_space=NULL; 1040 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 1041 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc; 1042 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 1043 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 1044 PetscInt adponz, adpdnz; 1045 PetscInt *pi_loc,*dnz,*onz; 1046 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,rstart=A->rmap->rstart; 1047 PetscInt *lnk,i, i1=0,pnz,row,*adpoi,*adpoj, *api, *adpoJ, *aopJ, *apJ,*Jptr, aopnz, nspacedouble=0,j,nzi, 1048 *apj,apnz, *adpdi, *adpdj, *adpdJ, *poff_i, *poff_j, *j_temp, *aopothi, *aopothj; 1049 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n, p_colstart, p_colend; 1050 PetscBT lnkbt; 1051 PetscReal afill; 1052 PetscMPIInt rank; 1053 Mat adpd, aopoth; 1054 MatType mtype; 1055 const char *prefix; 1056 1057 PetscFunctionBegin; 1058 MatCheckProduct(C,4); 1059 PetscCheck(!C->product->data,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty"); 1060 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 1061 PetscCallMPI(MPI_Comm_size(comm,&size)); 1062 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1063 PetscCall(MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend)); 1064 1065 /* create struct Mat_APMPI and attached it to C later */ 1066 PetscCall(PetscNew(&ptap)); 1067 1068 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1069 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth)); 1070 1071 /* get P_loc by taking all local rows of P */ 1072 PetscCall(MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc)); 1073 1074 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1075 pi_loc = p_loc->i; 1076 1077 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 1078 PetscCall(PetscMalloc1(am+2,&api)); 1079 PetscCall(PetscMalloc1(am+2,&adpoi)); 1080 1081 adpoi[0] = 0; 1082 ptap->api = api; 1083 api[0] = 0; 1084 1085 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1086 PetscCall(PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt)); 1087 MatPreallocateBegin(comm,am,pn,dnz,onz); 1088 1089 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1090 PetscCall(MatGetOptionsPrefix(A,&prefix)); 1091 PetscCall(MatProductCreate(a->A,p->A,NULL,&adpd)); 1092 PetscCall(MatGetOptionsPrefix(A,&prefix)); 1093 PetscCall(MatSetOptionsPrefix(adpd,prefix)); 1094 PetscCall(MatAppendOptionsPrefix(adpd,"inner_diag_")); 1095 1096 PetscCall(MatProductSetType(adpd,MATPRODUCT_AB)); 1097 PetscCall(MatProductSetAlgorithm(adpd,"sorted")); 1098 PetscCall(MatProductSetFill(adpd,fill)); 1099 PetscCall(MatProductSetFromOptions(adpd)); 1100 1101 adpd->force_diagonals = C->force_diagonals; 1102 PetscCall(MatProductSymbolic(adpd)); 1103 1104 adpd_seq = (Mat_SeqAIJ*)((adpd)->data); 1105 adpdi = adpd_seq->i; adpdj = adpd_seq->j; 1106 p_off = (Mat_SeqAIJ*)((p->B)->data); 1107 poff_i = p_off->i; poff_j = p_off->j; 1108 1109 /* j_temp stores indices of a result row before they are added to the linked list */ 1110 PetscCall(PetscMalloc1(pN+2,&j_temp)); 1111 1112 /* Symbolic calc of the A_diag * p_loc_off */ 1113 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1114 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space_diag)); 1115 current_space = free_space_diag; 1116 1117 for (i=0; i<am; i++) { 1118 /* A_diag * P_loc_off */ 1119 nzi = adi[i+1] - adi[i]; 1120 for (j=0; j<nzi; j++) { 1121 row = *adj++; 1122 pnz = poff_i[row+1] - poff_i[row]; 1123 Jptr = poff_j + poff_i[row]; 1124 for (i1 = 0; i1 < pnz; i1++) { 1125 j_temp[i1] = p->garray[Jptr[i1]]; 1126 } 1127 /* add non-zero cols of P into the sorted linked list lnk */ 1128 PetscCall(PetscLLCondensedAddSorted(pnz,j_temp,lnk,lnkbt)); 1129 } 1130 1131 adponz = lnk[0]; 1132 adpoi[i+1] = adpoi[i] + adponz; 1133 1134 /* if free space is not available, double the total space in the list */ 1135 if (current_space->local_remaining<adponz) { 1136 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(adponz,current_space->total_array_size),¤t_space)); 1137 nspacedouble++; 1138 } 1139 1140 /* Copy data into free space, then initialize lnk */ 1141 PetscCall(PetscLLCondensedClean(pN,adponz,current_space->array,lnk,lnkbt)); 1142 1143 current_space->array += adponz; 1144 current_space->local_used += adponz; 1145 current_space->local_remaining -= adponz; 1146 } 1147 1148 /* Symbolic calc of A_off * P_oth */ 1149 PetscCall(MatSetOptionsPrefix(a->B,prefix)); 1150 PetscCall(MatAppendOptionsPrefix(a->B,"inner_offdiag_")); 1151 PetscCall(MatCreate(PETSC_COMM_SELF,&aopoth)); 1152 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, aopoth)); 1153 aopoth_seq = (Mat_SeqAIJ*)((aopoth)->data); 1154 aopothi = aopoth_seq->i; aopothj = aopoth_seq->j; 1155 1156 /* Allocate space for apj, adpj, aopj, ... */ 1157 /* destroy lists of free space and other temporary array(s) */ 1158 1159 PetscCall(PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am]+2, &ptap->apj)); 1160 PetscCall(PetscMalloc1(adpoi[am]+2, &adpoj)); 1161 1162 /* Copy from linked list to j-array */ 1163 PetscCall(PetscFreeSpaceContiguous(&free_space_diag,adpoj)); 1164 PetscCall(PetscLLDestroy(lnk,lnkbt)); 1165 1166 adpoJ = adpoj; 1167 adpdJ = adpdj; 1168 aopJ = aopothj; 1169 apj = ptap->apj; 1170 apJ = apj; /* still empty */ 1171 1172 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1173 /* A_diag * P_loc_diag to get A*P */ 1174 for (i = 0; i < am; i++) { 1175 aopnz = aopothi[i+1] - aopothi[i]; 1176 adponz = adpoi[i+1] - adpoi[i]; 1177 adpdnz = adpdi[i+1] - adpdi[i]; 1178 1179 /* Correct indices from A_diag*P_diag */ 1180 for (i1 = 0; i1 < adpdnz; i1++) { 1181 adpdJ[i1] += p_colstart; 1182 } 1183 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1184 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1185 PetscCall(MatPreallocateSet(i+rstart, apnz, apJ, dnz, onz)); 1186 1187 aopJ += aopnz; 1188 adpoJ += adponz; 1189 adpdJ += adpdnz; 1190 apJ += apnz; 1191 api[i+1] = api[i] + apnz; 1192 } 1193 1194 /* malloc apa to store dense row A[i,:]*P */ 1195 PetscCall(PetscCalloc1(pN+2,&ptap->apa)); 1196 1197 /* create and assemble symbolic parallel matrix C */ 1198 PetscCall(MatSetSizes(C,am,pn,PETSC_DETERMINE,PETSC_DETERMINE)); 1199 PetscCall(MatSetBlockSizesFromMats(C,A,P)); 1200 PetscCall(MatGetType(A,&mtype)); 1201 PetscCall(MatSetType(C,mtype)); 1202 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 1203 MatPreallocateEnd(dnz,onz); 1204 1205 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 1206 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1207 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1208 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 1209 1210 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1211 C->ops->productnumeric = MatProductNumeric_AB; 1212 1213 /* attach the supporting struct to C for reuse */ 1214 C->product->data = ptap; 1215 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 1216 1217 /* set MatInfo */ 1218 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 1219 if (afill < 1.0) afill = 1.0; 1220 C->info.mallocs = nspacedouble; 1221 C->info.fill_ratio_given = fill; 1222 C->info.fill_ratio_needed = afill; 1223 1224 #if defined(PETSC_USE_INFO) 1225 if (api[am]) { 1226 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 1227 PetscCall(PetscInfo(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill)); 1228 } else { 1229 PetscCall(PetscInfo(C,"Empty matrix product\n")); 1230 } 1231 #endif 1232 1233 PetscCall(MatDestroy(&aopoth)); 1234 PetscCall(MatDestroy(&adpd)); 1235 PetscCall(PetscFree(j_temp)); 1236 PetscCall(PetscFree(adpoj)); 1237 PetscCall(PetscFree(adpoi)); 1238 PetscFunctionReturn(0); 1239 } 1240 1241 /*-------------------------------------------------------------------------*/ 1242 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1243 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P,Mat A,Mat C) 1244 { 1245 Mat_APMPI *ptap; 1246 Mat Pt; 1247 1248 PetscFunctionBegin; 1249 MatCheckProduct(C,3); 1250 ptap = (Mat_APMPI*)C->product->data; 1251 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1252 PetscCheck(ptap->Pt,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1253 1254 Pt = ptap->Pt; 1255 PetscCall(MatTranspose(P,MAT_REUSE_MATRIX,&Pt)); 1256 PetscCall(MatMatMultNumeric_MPIAIJ_MPIAIJ(Pt,A,C)); 1257 PetscFunctionReturn(0); 1258 } 1259 1260 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1261 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,PetscReal fill,Mat C) 1262 { 1263 Mat_APMPI *ptap; 1264 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 1265 MPI_Comm comm; 1266 PetscMPIInt size,rank; 1267 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1268 PetscInt pn=P->cmap->n,aN=A->cmap->N,an=A->cmap->n; 1269 PetscInt *lnk,i,k,nsend,rstart; 1270 PetscBT lnkbt; 1271 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,nrecv; 1272 PETSC_UNUSED PetscMPIInt icompleted=0; 1273 PetscInt **buf_rj,**buf_ri,**buf_ri_k,row,ncols,*cols; 1274 PetscInt len,proc,*dnz,*onz,*owners,nzi; 1275 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1276 MPI_Request *swaits,*rwaits; 1277 MPI_Status *sstatus,rstatus; 1278 PetscLayout rowmap; 1279 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1280 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 1281 PetscInt *Jptr,*prmap=p->garray,con,j,Crmax; 1282 Mat_SeqAIJ *a_loc,*c_loc,*c_oth; 1283 PetscTable ta; 1284 MatType mtype; 1285 const char *prefix; 1286 1287 PetscFunctionBegin; 1288 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 1289 PetscCallMPI(MPI_Comm_size(comm,&size)); 1290 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 1291 1292 /* create symbolic parallel matrix C */ 1293 PetscCall(MatGetType(A,&mtype)); 1294 PetscCall(MatSetType(C,mtype)); 1295 1296 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1297 1298 /* create struct Mat_APMPI and attached it to C later */ 1299 PetscCall(PetscNew(&ptap)); 1300 ptap->reuse = MAT_INITIAL_MATRIX; 1301 1302 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1303 /* --------------------------------- */ 1304 PetscCall(MatTranspose(p->A,MAT_INITIAL_MATRIX,&ptap->Rd)); 1305 PetscCall(MatTranspose(p->B,MAT_INITIAL_MATRIX,&ptap->Ro)); 1306 1307 /* (1) compute symbolic A_loc */ 1308 /* ---------------------------*/ 1309 PetscCall(MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&ptap->A_loc)); 1310 1311 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1312 /* ------------------------------------ */ 1313 PetscCall(MatGetOptionsPrefix(A,&prefix)); 1314 PetscCall(MatSetOptionsPrefix(ptap->Ro,prefix)); 1315 PetscCall(MatAppendOptionsPrefix(ptap->Ro,"inner_offdiag_")); 1316 PetscCall(MatCreate(PETSC_COMM_SELF,&ptap->C_oth)); 1317 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->A_loc,fill,ptap->C_oth)); 1318 1319 /* (3) send coj of C_oth to other processors */ 1320 /* ------------------------------------------ */ 1321 /* determine row ownership */ 1322 PetscCall(PetscLayoutCreate(comm,&rowmap)); 1323 rowmap->n = pn; 1324 rowmap->bs = 1; 1325 PetscCall(PetscLayoutSetUp(rowmap)); 1326 owners = rowmap->range; 1327 1328 /* determine the number of messages to send, their lengths */ 1329 PetscCall(PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co)); 1330 PetscCall(PetscArrayzero(len_s,size)); 1331 PetscCall(PetscArrayzero(len_si,size)); 1332 1333 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 1334 coi = c_oth->i; coj = c_oth->j; 1335 con = ptap->C_oth->rmap->n; 1336 proc = 0; 1337 for (i=0; i<con; i++) { 1338 while (prmap[i] >= owners[proc+1]) proc++; 1339 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1340 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1341 } 1342 1343 len = 0; /* max length of buf_si[], see (4) */ 1344 owners_co[0] = 0; 1345 nsend = 0; 1346 for (proc=0; proc<size; proc++) { 1347 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1348 if (len_s[proc]) { 1349 nsend++; 1350 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1351 len += len_si[proc]; 1352 } 1353 } 1354 1355 /* determine the number and length of messages to receive for coi and coj */ 1356 PetscCall(PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv)); 1357 PetscCall(PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri)); 1358 1359 /* post the Irecv and Isend of coj */ 1360 PetscCall(PetscCommGetNewTag(comm,&tagj)); 1361 PetscCall(PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits)); 1362 PetscCall(PetscMalloc1(nsend+1,&swaits)); 1363 for (proc=0, k=0; proc<size; proc++) { 1364 if (!len_s[proc]) continue; 1365 i = owners_co[proc]; 1366 PetscCallMPI(MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k)); 1367 k++; 1368 } 1369 1370 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1371 /* ---------------------------------------- */ 1372 PetscCall(MatSetOptionsPrefix(ptap->Rd,prefix)); 1373 PetscCall(MatAppendOptionsPrefix(ptap->Rd,"inner_diag_")); 1374 PetscCall(MatCreate(PETSC_COMM_SELF,&ptap->C_loc)); 1375 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->A_loc,fill,ptap->C_loc)); 1376 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 1377 1378 /* receives coj are complete */ 1379 for (i=0; i<nrecv; i++) { 1380 PetscCallMPI(MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus)); 1381 } 1382 PetscCall(PetscFree(rwaits)); 1383 if (nsend) PetscCallMPI(MPI_Waitall(nsend,swaits,sstatus)); 1384 1385 /* add received column indices into ta to update Crmax */ 1386 a_loc = (Mat_SeqAIJ*)(ptap->A_loc)->data; 1387 1388 /* create and initialize a linked list */ 1389 PetscCall(PetscTableCreate(an,aN,&ta)); /* for compute Crmax */ 1390 MatRowMergeMax_SeqAIJ(a_loc,ptap->A_loc->rmap->N,ta); 1391 1392 for (k=0; k<nrecv; k++) {/* k-th received message */ 1393 Jptr = buf_rj[k]; 1394 for (j=0; j<len_r[k]; j++) { 1395 PetscCall(PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES)); 1396 } 1397 } 1398 PetscCall(PetscTableGetCount(ta,&Crmax)); 1399 PetscCall(PetscTableDestroy(&ta)); 1400 1401 /* (4) send and recv coi */ 1402 /*-----------------------*/ 1403 PetscCall(PetscCommGetNewTag(comm,&tagi)); 1404 PetscCall(PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits)); 1405 PetscCall(PetscMalloc1(len+1,&buf_s)); 1406 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1407 for (proc=0,k=0; proc<size; proc++) { 1408 if (!len_s[proc]) continue; 1409 /* form outgoing message for i-structure: 1410 buf_si[0]: nrows to be sent 1411 [1:nrows]: row index (global) 1412 [nrows+1:2*nrows+1]: i-structure index 1413 */ 1414 /*-------------------------------------------*/ 1415 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 1416 buf_si_i = buf_si + nrows+1; 1417 buf_si[0] = nrows; 1418 buf_si_i[0] = 0; 1419 nrows = 0; 1420 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1421 nzi = coi[i+1] - coi[i]; 1422 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1423 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1424 nrows++; 1425 } 1426 PetscCallMPI(MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k)); 1427 k++; 1428 buf_si += len_si[proc]; 1429 } 1430 for (i=0; i<nrecv; i++) { 1431 PetscCallMPI(MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus)); 1432 } 1433 PetscCall(PetscFree(rwaits)); 1434 if (nsend) PetscCallMPI(MPI_Waitall(nsend,swaits,sstatus)); 1435 1436 PetscCall(PetscFree4(len_s,len_si,sstatus,owners_co)); 1437 PetscCall(PetscFree(len_ri)); 1438 PetscCall(PetscFree(swaits)); 1439 PetscCall(PetscFree(buf_s)); 1440 1441 /* (5) compute the local portion of C */ 1442 /* ------------------------------------------ */ 1443 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of C */ 1444 PetscCall(PetscFreeSpaceGet(Crmax,&free_space)); 1445 current_space = free_space; 1446 1447 PetscCall(PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci)); 1448 for (k=0; k<nrecv; k++) { 1449 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1450 nrows = *buf_ri_k[k]; 1451 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1452 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1453 } 1454 1455 MatPreallocateBegin(comm,pn,an,dnz,onz); 1456 PetscCall(PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt)); 1457 for (i=0; i<pn; i++) { /* for each local row of C */ 1458 /* add C_loc into C */ 1459 nzi = c_loc->i[i+1] - c_loc->i[i]; 1460 Jptr = c_loc->j + c_loc->i[i]; 1461 PetscCall(PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt)); 1462 1463 /* add received col data into lnk */ 1464 for (k=0; k<nrecv; k++) { /* k-th received message */ 1465 if (i == *nextrow[k]) { /* i-th row */ 1466 nzi = *(nextci[k]+1) - *nextci[k]; 1467 Jptr = buf_rj[k] + *nextci[k]; 1468 PetscCall(PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt)); 1469 nextrow[k]++; nextci[k]++; 1470 } 1471 } 1472 1473 /* add missing diagonal entry */ 1474 if (C->force_diagonals) { 1475 k = i + owners[rank]; /* column index */ 1476 PetscCall(PetscLLCondensedAddSorted(1,&k,lnk,lnkbt)); 1477 } 1478 1479 nzi = lnk[0]; 1480 1481 /* copy data into free space, then initialize lnk */ 1482 PetscCall(PetscLLCondensedClean(aN,nzi,current_space->array,lnk,lnkbt)); 1483 PetscCall(MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz)); 1484 } 1485 PetscCall(PetscFree3(buf_ri_k,nextrow,nextci)); 1486 PetscCall(PetscLLDestroy(lnk,lnkbt)); 1487 PetscCall(PetscFreeSpaceDestroy(free_space)); 1488 1489 /* local sizes and preallocation */ 1490 PetscCall(MatSetSizes(C,pn,an,PETSC_DETERMINE,PETSC_DETERMINE)); 1491 if (P->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->rmap,P->cmap->bs)); 1492 if (A->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->cmap,A->cmap->bs)); 1493 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 1494 MatPreallocateEnd(dnz,onz); 1495 1496 /* add C_loc and C_oth to C */ 1497 PetscCall(MatGetOwnershipRange(C,&rstart,NULL)); 1498 for (i=0; i<pn; i++) { 1499 ncols = c_loc->i[i+1] - c_loc->i[i]; 1500 cols = c_loc->j + c_loc->i[i]; 1501 row = rstart + i; 1502 PetscCall(MatSetValues(C,1,(const PetscInt*)&row,ncols,(const PetscInt*)cols,NULL,INSERT_VALUES)); 1503 1504 if (C->force_diagonals) { 1505 PetscCall(MatSetValues(C,1,(const PetscInt*)&row,1,(const PetscInt*)&row,NULL,INSERT_VALUES)); 1506 } 1507 } 1508 for (i=0; i<con; i++) { 1509 ncols = c_oth->i[i+1] - c_oth->i[i]; 1510 cols = c_oth->j + c_oth->i[i]; 1511 row = prmap[i]; 1512 PetscCall(MatSetValues(C,1,(const PetscInt*)&row,ncols,(const PetscInt*)cols,NULL,INSERT_VALUES)); 1513 } 1514 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1515 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1516 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 1517 1518 /* members in merge */ 1519 PetscCall(PetscFree(id_r)); 1520 PetscCall(PetscFree(len_r)); 1521 PetscCall(PetscFree(buf_ri[0])); 1522 PetscCall(PetscFree(buf_ri)); 1523 PetscCall(PetscFree(buf_rj[0])); 1524 PetscCall(PetscFree(buf_rj)); 1525 PetscCall(PetscLayoutDestroy(&rowmap)); 1526 1527 /* attach the supporting struct to C for reuse */ 1528 C->product->data = ptap; 1529 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 1530 PetscFunctionReturn(0); 1531 } 1532 1533 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,Mat C) 1534 { 1535 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 1536 Mat_SeqAIJ *c_seq; 1537 Mat_APMPI *ptap; 1538 Mat A_loc,C_loc,C_oth; 1539 PetscInt i,rstart,rend,cm,ncols,row; 1540 const PetscInt *cols; 1541 const PetscScalar *vals; 1542 1543 PetscFunctionBegin; 1544 MatCheckProduct(C,3); 1545 ptap = (Mat_APMPI*)C->product->data; 1546 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1547 PetscCheck(ptap->A_loc,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1548 PetscCall(MatZeroEntries(C)); 1549 1550 if (ptap->reuse == MAT_REUSE_MATRIX) { 1551 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1552 /* 1) get R = Pd^T, Ro = Po^T */ 1553 /*----------------------------*/ 1554 PetscCall(MatTranspose(p->A,MAT_REUSE_MATRIX,&ptap->Rd)); 1555 PetscCall(MatTranspose(p->B,MAT_REUSE_MATRIX,&ptap->Ro)); 1556 1557 /* 2) compute numeric A_loc */ 1558 /*--------------------------*/ 1559 PetscCall(MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&ptap->A_loc)); 1560 } 1561 1562 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1563 A_loc = ptap->A_loc; 1564 PetscCall(((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,A_loc,ptap->C_loc)); 1565 PetscCall(((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,A_loc,ptap->C_oth)); 1566 C_loc = ptap->C_loc; 1567 C_oth = ptap->C_oth; 1568 1569 /* add C_loc and C_oth to C */ 1570 PetscCall(MatGetOwnershipRange(C,&rstart,&rend)); 1571 1572 /* C_loc -> C */ 1573 cm = C_loc->rmap->N; 1574 c_seq = (Mat_SeqAIJ*)C_loc->data; 1575 cols = c_seq->j; 1576 vals = c_seq->a; 1577 for (i=0; i<cm; i++) { 1578 ncols = c_seq->i[i+1] - c_seq->i[i]; 1579 row = rstart + i; 1580 PetscCall(MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES)); 1581 cols += ncols; vals += ncols; 1582 } 1583 1584 /* Co -> C, off-processor part */ 1585 cm = C_oth->rmap->N; 1586 c_seq = (Mat_SeqAIJ*)C_oth->data; 1587 cols = c_seq->j; 1588 vals = c_seq->a; 1589 for (i=0; i<cm; i++) { 1590 ncols = c_seq->i[i+1] - c_seq->i[i]; 1591 row = p->garray[i]; 1592 PetscCall(MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES)); 1593 cols += ncols; vals += ncols; 1594 } 1595 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1596 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1597 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 1598 1599 ptap->reuse = MAT_REUSE_MATRIX; 1600 PetscFunctionReturn(0); 1601 } 1602 1603 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 1604 { 1605 Mat_Merge_SeqsToMPI *merge; 1606 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 1607 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1608 Mat_APMPI *ptap; 1609 PetscInt *adj; 1610 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1611 MatScalar *ada,*ca,valtmp; 1612 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1613 MPI_Comm comm; 1614 PetscMPIInt size,rank,taga,*len_s; 1615 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1616 PetscInt **buf_ri,**buf_rj; 1617 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1618 MPI_Request *s_waits,*r_waits; 1619 MPI_Status *status; 1620 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1621 const PetscScalar *dummy; 1622 PetscInt *ai,*aj,*coi,*coj,*poJ,*pdJ; 1623 Mat A_loc; 1624 Mat_SeqAIJ *a_loc; 1625 1626 PetscFunctionBegin; 1627 MatCheckProduct(C,3); 1628 ptap = (Mat_APMPI*)C->product->data; 1629 PetscCheck(ptap,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be computed. Missing data"); 1630 PetscCheck(ptap->A_loc,PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatProductClear()"); 1631 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 1632 PetscCallMPI(MPI_Comm_size(comm,&size)); 1633 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 1634 1635 merge = ptap->merge; 1636 1637 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1638 /*------------------------------------------*/ 1639 /* get data from symbolic products */ 1640 coi = merge->coi; coj = merge->coj; 1641 PetscCall(PetscCalloc1(coi[pon]+1,&coa)); 1642 bi = merge->bi; bj = merge->bj; 1643 owners = merge->rowmap->range; 1644 PetscCall(PetscCalloc1(bi[cm]+1,&ba)); 1645 1646 /* get A_loc by taking all local rows of A */ 1647 A_loc = ptap->A_loc; 1648 PetscCall(MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc)); 1649 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1650 ai = a_loc->i; 1651 aj = a_loc->j; 1652 1653 /* trigger copy to CPU */ 1654 PetscCall(MatSeqAIJGetArrayRead(p->A,&dummy)); 1655 PetscCall(MatSeqAIJRestoreArrayRead(p->A,&dummy)); 1656 PetscCall(MatSeqAIJGetArrayRead(p->B,&dummy)); 1657 PetscCall(MatSeqAIJRestoreArrayRead(p->B,&dummy)); 1658 for (i=0; i<am; i++) { 1659 anz = ai[i+1] - ai[i]; 1660 adj = aj + ai[i]; 1661 ada = a_loc->a + ai[i]; 1662 1663 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1664 /*-------------------------------------------------------------*/ 1665 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1666 pnz = po->i[i+1] - po->i[i]; 1667 poJ = po->j + po->i[i]; 1668 pA = po->a + po->i[i]; 1669 for (j=0; j<pnz; j++) { 1670 row = poJ[j]; 1671 cj = coj + coi[row]; 1672 ca = coa + coi[row]; 1673 /* perform sparse axpy */ 1674 nexta = 0; 1675 valtmp = pA[j]; 1676 for (k=0; nexta<anz; k++) { 1677 if (cj[k] == adj[nexta]) { 1678 ca[k] += valtmp*ada[nexta]; 1679 nexta++; 1680 } 1681 } 1682 PetscCall(PetscLogFlops(2.0*anz)); 1683 } 1684 1685 /* put the value into Cd (diagonal part) */ 1686 pnz = pd->i[i+1] - pd->i[i]; 1687 pdJ = pd->j + pd->i[i]; 1688 pA = pd->a + pd->i[i]; 1689 for (j=0; j<pnz; j++) { 1690 row = pdJ[j]; 1691 cj = bj + bi[row]; 1692 ca = ba + bi[row]; 1693 /* perform sparse axpy */ 1694 nexta = 0; 1695 valtmp = pA[j]; 1696 for (k=0; nexta<anz; k++) { 1697 if (cj[k] == adj[nexta]) { 1698 ca[k] += valtmp*ada[nexta]; 1699 nexta++; 1700 } 1701 } 1702 PetscCall(PetscLogFlops(2.0*anz)); 1703 } 1704 } 1705 1706 /* 3) send and recv matrix values coa */ 1707 /*------------------------------------*/ 1708 buf_ri = merge->buf_ri; 1709 buf_rj = merge->buf_rj; 1710 len_s = merge->len_s; 1711 PetscCall(PetscCommGetNewTag(comm,&taga)); 1712 PetscCall(PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits)); 1713 1714 PetscCall(PetscMalloc2(merge->nsend+1,&s_waits,size,&status)); 1715 for (proc=0,k=0; proc<size; proc++) { 1716 if (!len_s[proc]) continue; 1717 i = merge->owners_co[proc]; 1718 PetscCallMPI(MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k)); 1719 k++; 1720 } 1721 if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv,r_waits,status)); 1722 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend,s_waits,status)); 1723 1724 PetscCall(PetscFree2(s_waits,status)); 1725 PetscCall(PetscFree(r_waits)); 1726 PetscCall(PetscFree(coa)); 1727 1728 /* 4) insert local Cseq and received values into Cmpi */ 1729 /*----------------------------------------------------*/ 1730 PetscCall(PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci)); 1731 for (k=0; k<merge->nrecv; k++) { 1732 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1733 nrows = *(buf_ri_k[k]); 1734 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1735 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1736 } 1737 1738 for (i=0; i<cm; i++) { 1739 row = owners[rank] + i; /* global row index of C_seq */ 1740 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1741 ba_i = ba + bi[i]; 1742 bnz = bi[i+1] - bi[i]; 1743 /* add received vals into ba */ 1744 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1745 /* i-th row */ 1746 if (i == *nextrow[k]) { 1747 cnz = *(nextci[k]+1) - *nextci[k]; 1748 cj = buf_rj[k] + *(nextci[k]); 1749 ca = abuf_r[k] + *(nextci[k]); 1750 nextcj = 0; 1751 for (j=0; nextcj<cnz; j++) { 1752 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1753 ba_i[j] += ca[nextcj++]; 1754 } 1755 } 1756 nextrow[k]++; nextci[k]++; 1757 PetscCall(PetscLogFlops(2.0*cnz)); 1758 } 1759 } 1760 PetscCall(MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES)); 1761 } 1762 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 1763 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 1764 1765 PetscCall(PetscFree(ba)); 1766 PetscCall(PetscFree(abuf_r[0])); 1767 PetscCall(PetscFree(abuf_r)); 1768 PetscCall(PetscFree3(buf_ri_k,nextrow,nextci)); 1769 PetscFunctionReturn(0); 1770 } 1771 1772 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat C) 1773 { 1774 Mat A_loc; 1775 Mat_APMPI *ptap; 1776 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1777 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*a=(Mat_MPIAIJ*)A->data; 1778 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1779 PetscInt nnz; 1780 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1781 PetscInt am =A->rmap->n,pn=P->cmap->n; 1782 MPI_Comm comm; 1783 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1784 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1785 PetscInt len,proc,*dnz,*onz,*owners; 1786 PetscInt nzi,*bi,*bj; 1787 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1788 MPI_Request *swaits,*rwaits; 1789 MPI_Status *sstatus,rstatus; 1790 Mat_Merge_SeqsToMPI *merge; 1791 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1792 PetscReal afill =1.0,afill_tmp; 1793 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Armax; 1794 Mat_SeqAIJ *a_loc; 1795 PetscTable ta; 1796 MatType mtype; 1797 1798 PetscFunctionBegin; 1799 PetscCall(PetscObjectGetComm((PetscObject)A,&comm)); 1800 /* check if matrix local sizes are compatible */ 1801 PetscCheck(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); 1802 1803 PetscCallMPI(MPI_Comm_size(comm,&size)); 1804 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 1805 1806 /* create struct Mat_APMPI and attached it to C later */ 1807 PetscCall(PetscNew(&ptap)); 1808 1809 /* get A_loc by taking all local rows of A */ 1810 PetscCall(MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc)); 1811 1812 ptap->A_loc = A_loc; 1813 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1814 ai = a_loc->i; 1815 aj = a_loc->j; 1816 1817 /* determine symbolic Co=(p->B)^T*A - send to others */ 1818 /*----------------------------------------------------*/ 1819 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj)); 1820 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj)); 1821 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1822 >= (num of nonzero rows of C_seq) - pn */ 1823 PetscCall(PetscMalloc1(pon+1,&coi)); 1824 coi[0] = 0; 1825 1826 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1827 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(poti[pon],ai[am])); 1828 PetscCall(PetscFreeSpaceGet(nnz,&free_space)); 1829 current_space = free_space; 1830 1831 /* create and initialize a linked list */ 1832 PetscCall(PetscTableCreate(A->cmap->n + a->B->cmap->N,aN,&ta)); 1833 MatRowMergeMax_SeqAIJ(a_loc,am,ta); 1834 PetscCall(PetscTableGetCount(ta,&Armax)); 1835 1836 PetscCall(PetscLLCondensedCreate_Scalable(Armax,&lnk)); 1837 1838 for (i=0; i<pon; i++) { 1839 pnz = poti[i+1] - poti[i]; 1840 ptJ = potj + poti[i]; 1841 for (j=0; j<pnz; j++) { 1842 row = ptJ[j]; /* row of A_loc == col of Pot */ 1843 anz = ai[row+1] - ai[row]; 1844 Jptr = aj + ai[row]; 1845 /* add non-zero cols of AP into the sorted linked list lnk */ 1846 PetscCall(PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk)); 1847 } 1848 nnz = lnk[0]; 1849 1850 /* If free space is not available, double the total space in the list */ 1851 if (current_space->local_remaining<nnz) { 1852 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space)); 1853 nspacedouble++; 1854 } 1855 1856 /* Copy data into free space, and zero out denserows */ 1857 PetscCall(PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk)); 1858 1859 current_space->array += nnz; 1860 current_space->local_used += nnz; 1861 current_space->local_remaining -= nnz; 1862 1863 coi[i+1] = coi[i] + nnz; 1864 } 1865 1866 PetscCall(PetscMalloc1(coi[pon]+1,&coj)); 1867 PetscCall(PetscFreeSpaceContiguous(&free_space,coj)); 1868 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); /* must destroy to get a new one for C */ 1869 1870 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1871 if (afill_tmp > afill) afill = afill_tmp; 1872 1873 /* send j-array (coj) of Co to other processors */ 1874 /*----------------------------------------------*/ 1875 /* determine row ownership */ 1876 PetscCall(PetscNew(&merge)); 1877 PetscCall(PetscLayoutCreate(comm,&merge->rowmap)); 1878 1879 merge->rowmap->n = pn; 1880 merge->rowmap->bs = 1; 1881 1882 PetscCall(PetscLayoutSetUp(merge->rowmap)); 1883 owners = merge->rowmap->range; 1884 1885 /* determine the number of messages to send, their lengths */ 1886 PetscCall(PetscCalloc1(size,&len_si)); 1887 PetscCall(PetscCalloc1(size,&merge->len_s)); 1888 1889 len_s = merge->len_s; 1890 merge->nsend = 0; 1891 1892 PetscCall(PetscMalloc1(size+2,&owners_co)); 1893 1894 proc = 0; 1895 for (i=0; i<pon; i++) { 1896 while (prmap[i] >= owners[proc+1]) proc++; 1897 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1898 len_s[proc] += coi[i+1] - coi[i]; 1899 } 1900 1901 len = 0; /* max length of buf_si[] */ 1902 owners_co[0] = 0; 1903 for (proc=0; proc<size; proc++) { 1904 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1905 if (len_si[proc]) { 1906 merge->nsend++; 1907 len_si[proc] = 2*(len_si[proc] + 1); 1908 len += len_si[proc]; 1909 } 1910 } 1911 1912 /* determine the number and length of messages to receive for coi and coj */ 1913 PetscCall(PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv)); 1914 PetscCall(PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri)); 1915 1916 /* post the Irecv and Isend of coj */ 1917 PetscCall(PetscCommGetNewTag(comm,&tagj)); 1918 PetscCall(PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits)); 1919 PetscCall(PetscMalloc1(merge->nsend+1,&swaits)); 1920 for (proc=0, k=0; proc<size; proc++) { 1921 if (!len_s[proc]) continue; 1922 i = owners_co[proc]; 1923 PetscCallMPI(MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k)); 1924 k++; 1925 } 1926 1927 /* receives and sends of coj are complete */ 1928 PetscCall(PetscMalloc1(size,&sstatus)); 1929 for (i=0; i<merge->nrecv; i++) { 1930 PETSC_UNUSED PetscMPIInt icompleted; 1931 PetscCallMPI(MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus)); 1932 } 1933 PetscCall(PetscFree(rwaits)); 1934 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend,swaits,sstatus)); 1935 1936 /* add received column indices into table to update Armax */ 1937 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/tutorials/ex56.c! */ 1938 for (k=0; k<merge->nrecv; k++) {/* k-th received message */ 1939 Jptr = buf_rj[k]; 1940 for (j=0; j<merge->len_r[k]; j++) { 1941 PetscCall(PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES)); 1942 } 1943 } 1944 PetscCall(PetscTableGetCount(ta,&Armax)); 1945 1946 /* send and recv coi */ 1947 /*-------------------*/ 1948 PetscCall(PetscCommGetNewTag(comm,&tagi)); 1949 PetscCall(PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits)); 1950 PetscCall(PetscMalloc1(len+1,&buf_s)); 1951 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1952 for (proc=0,k=0; proc<size; proc++) { 1953 if (!len_s[proc]) continue; 1954 /* form outgoing message for i-structure: 1955 buf_si[0]: nrows to be sent 1956 [1:nrows]: row index (global) 1957 [nrows+1:2*nrows+1]: i-structure index 1958 */ 1959 /*-------------------------------------------*/ 1960 nrows = len_si[proc]/2 - 1; 1961 buf_si_i = buf_si + nrows+1; 1962 buf_si[0] = nrows; 1963 buf_si_i[0] = 0; 1964 nrows = 0; 1965 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1966 nzi = coi[i+1] - coi[i]; 1967 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1968 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1969 nrows++; 1970 } 1971 PetscCallMPI(MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k)); 1972 k++; 1973 buf_si += len_si[proc]; 1974 } 1975 i = merge->nrecv; 1976 while (i--) { 1977 PETSC_UNUSED PetscMPIInt icompleted; 1978 PetscCallMPI(MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus)); 1979 } 1980 PetscCall(PetscFree(rwaits)); 1981 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend,swaits,sstatus)); 1982 PetscCall(PetscFree(len_si)); 1983 PetscCall(PetscFree(len_ri)); 1984 PetscCall(PetscFree(swaits)); 1985 PetscCall(PetscFree(sstatus)); 1986 PetscCall(PetscFree(buf_s)); 1987 1988 /* compute the local portion of C (mpi mat) */ 1989 /*------------------------------------------*/ 1990 /* allocate bi array and free space for accumulating nonzero column info */ 1991 PetscCall(PetscMalloc1(pn+1,&bi)); 1992 bi[0] = 0; 1993 1994 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1995 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(pdti[pn],PetscIntSumTruncate(poti[pon],ai[am]))); 1996 PetscCall(PetscFreeSpaceGet(nnz,&free_space)); 1997 current_space = free_space; 1998 1999 PetscCall(PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci)); 2000 for (k=0; k<merge->nrecv; k++) { 2001 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 2002 nrows = *buf_ri_k[k]; 2003 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 2004 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th received i-structure */ 2005 } 2006 2007 PetscCall(PetscLLCondensedCreate_Scalable(Armax,&lnk)); 2008 MatPreallocateBegin(comm,pn,A->cmap->n,dnz,onz); 2009 rmax = 0; 2010 for (i=0; i<pn; i++) { 2011 /* add pdt[i,:]*AP into lnk */ 2012 pnz = pdti[i+1] - pdti[i]; 2013 ptJ = pdtj + pdti[i]; 2014 for (j=0; j<pnz; j++) { 2015 row = ptJ[j]; /* row of AP == col of Pt */ 2016 anz = ai[row+1] - ai[row]; 2017 Jptr = aj + ai[row]; 2018 /* add non-zero cols of AP into the sorted linked list lnk */ 2019 PetscCall(PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk)); 2020 } 2021 2022 /* add received col data into lnk */ 2023 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 2024 if (i == *nextrow[k]) { /* i-th row */ 2025 nzi = *(nextci[k]+1) - *nextci[k]; 2026 Jptr = buf_rj[k] + *nextci[k]; 2027 PetscCall(PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk)); 2028 nextrow[k]++; nextci[k]++; 2029 } 2030 } 2031 2032 /* add missing diagonal entry */ 2033 if (C->force_diagonals) { 2034 k = i + owners[rank]; /* column index */ 2035 PetscCall(PetscLLCondensedAddSorted_Scalable(1,&k,lnk)); 2036 } 2037 2038 nnz = lnk[0]; 2039 2040 /* if free space is not available, make more free space */ 2041 if (current_space->local_remaining<nnz) { 2042 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space)); 2043 nspacedouble++; 2044 } 2045 /* copy data into free space, then initialize lnk */ 2046 PetscCall(PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk)); 2047 PetscCall(MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz)); 2048 2049 current_space->array += nnz; 2050 current_space->local_used += nnz; 2051 current_space->local_remaining -= nnz; 2052 2053 bi[i+1] = bi[i] + nnz; 2054 if (nnz > rmax) rmax = nnz; 2055 } 2056 PetscCall(PetscFree3(buf_ri_k,nextrow,nextci)); 2057 2058 PetscCall(PetscMalloc1(bi[pn]+1,&bj)); 2059 PetscCall(PetscFreeSpaceContiguous(&free_space,bj)); 2060 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 2061 if (afill_tmp > afill) afill = afill_tmp; 2062 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 2063 PetscCall(PetscTableDestroy(&ta)); 2064 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj)); 2065 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj)); 2066 2067 /* create symbolic parallel matrix C - why cannot be assembled in Numeric part */ 2068 /*-------------------------------------------------------------------------------*/ 2069 PetscCall(MatSetSizes(C,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE)); 2070 PetscCall(MatSetBlockSizes(C,PetscAbs(P->cmap->bs),PetscAbs(A->cmap->bs))); 2071 PetscCall(MatGetType(A,&mtype)); 2072 PetscCall(MatSetType(C,mtype)); 2073 PetscCall(MatMPIAIJSetPreallocation(C,0,dnz,0,onz)); 2074 MatPreallocateEnd(dnz,onz); 2075 PetscCall(MatSetBlockSize(C,1)); 2076 PetscCall(MatSetOption(C,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE)); 2077 for (i=0; i<pn; i++) { 2078 row = i + rstart; 2079 nnz = bi[i+1] - bi[i]; 2080 Jptr = bj + bi[i]; 2081 PetscCall(MatSetValues(C,1,&row,nnz,Jptr,NULL,INSERT_VALUES)); 2082 } 2083 PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY)); 2084 PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY)); 2085 PetscCall(MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE)); 2086 merge->bi = bi; 2087 merge->bj = bj; 2088 merge->coi = coi; 2089 merge->coj = coj; 2090 merge->buf_ri = buf_ri; 2091 merge->buf_rj = buf_rj; 2092 merge->owners_co = owners_co; 2093 2094 /* attach the supporting struct to C for reuse */ 2095 C->product->data = ptap; 2096 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2097 ptap->merge = merge; 2098 2099 C->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2100 2101 #if defined(PETSC_USE_INFO) 2102 if (bi[pn] != 0) { 2103 PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill)); 2104 PetscCall(PetscInfo(C,"Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n",(double)afill)); 2105 } else { 2106 PetscCall(PetscInfo(C,"Empty matrix product\n")); 2107 } 2108 #endif 2109 PetscFunctionReturn(0); 2110 } 2111 2112 /* ---------------------------------------------------------------- */ 2113 static PetscErrorCode MatProductSymbolic_AtB_MPIAIJ_MPIAIJ(Mat C) 2114 { 2115 Mat_Product *product = C->product; 2116 Mat A=product->A,B=product->B; 2117 PetscReal fill=product->fill; 2118 PetscBool flg; 2119 2120 PetscFunctionBegin; 2121 /* scalable */ 2122 PetscCall(PetscStrcmp(product->alg,"scalable",&flg)); 2123 if (flg) { 2124 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C)); 2125 goto next; 2126 } 2127 2128 /* nonscalable */ 2129 PetscCall(PetscStrcmp(product->alg,"nonscalable",&flg)); 2130 if (flg) { 2131 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C)); 2132 goto next; 2133 } 2134 2135 /* matmatmult */ 2136 PetscCall(PetscStrcmp(product->alg,"at*b",&flg)); 2137 if (flg) { 2138 Mat At; 2139 Mat_APMPI *ptap; 2140 2141 PetscCall(MatTranspose(A,MAT_INITIAL_MATRIX,&At)); 2142 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(At,B,fill,C)); 2143 ptap = (Mat_APMPI*)C->product->data; 2144 if (ptap) { 2145 ptap->Pt = At; 2146 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2147 } 2148 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 2149 goto next; 2150 } 2151 2152 /* backend general code */ 2153 PetscCall(PetscStrcmp(product->alg,"backend",&flg)); 2154 if (flg) { 2155 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 2156 PetscFunctionReturn(0); 2157 } 2158 2159 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatProduct type is not supported"); 2160 2161 next: 2162 C->ops->productnumeric = MatProductNumeric_AtB; 2163 PetscFunctionReturn(0); 2164 } 2165 2166 /* ---------------------------------------------------------------- */ 2167 /* Set options for MatMatMultxxx_MPIAIJ_MPIAIJ */ 2168 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AB(Mat C) 2169 { 2170 Mat_Product *product = C->product; 2171 Mat A=product->A,B=product->B; 2172 #if defined(PETSC_HAVE_HYPRE) 2173 const char *algTypes[5] = {"scalable","nonscalable","seqmpi","backend","hypre"}; 2174 PetscInt nalg = 5; 2175 #else 2176 const char *algTypes[4] = {"scalable","nonscalable","seqmpi","backend",}; 2177 PetscInt nalg = 4; 2178 #endif 2179 PetscInt alg = 1; /* set nonscalable algorithm as default */ 2180 PetscBool flg; 2181 MPI_Comm comm; 2182 2183 PetscFunctionBegin; 2184 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 2185 2186 /* Set "nonscalable" as default algorithm */ 2187 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2188 if (flg) { 2189 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2190 2191 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2192 if (B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2193 MatInfo Ainfo,Binfo; 2194 PetscInt nz_local; 2195 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2196 2197 PetscCall(MatGetInfo(A,MAT_LOCAL,&Ainfo)); 2198 PetscCall(MatGetInfo(B,MAT_LOCAL,&Binfo)); 2199 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2200 2201 if (B->cmap->N > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2202 PetscCall(MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm)); 2203 2204 if (alg_scalable) { 2205 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2206 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2207 PetscCall(PetscInfo(B,"Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n",B->cmap->N,(double)(product->fill*nz_local))); 2208 } 2209 } 2210 } 2211 2212 /* Get runtime option */ 2213 if (product->api_user) { 2214 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMult","Mat"); 2215 PetscCall(PetscOptionsEList("-matmatmult_via","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2216 PetscOptionsEnd(); 2217 } else { 2218 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AB","Mat"); 2219 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2220 PetscOptionsEnd(); 2221 } 2222 if (flg) { 2223 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2224 } 2225 2226 C->ops->productsymbolic = MatProductSymbolic_AB_MPIAIJ_MPIAIJ; 2227 PetscFunctionReturn(0); 2228 } 2229 2230 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABt(Mat C) 2231 { 2232 PetscFunctionBegin; 2233 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2234 C->ops->productsymbolic = MatProductSymbolic_ABt_MPIAIJ_MPIAIJ; 2235 PetscFunctionReturn(0); 2236 } 2237 2238 /* Set options for MatTransposeMatMultXXX_MPIAIJ_MPIAIJ */ 2239 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AtB(Mat C) 2240 { 2241 Mat_Product *product = C->product; 2242 Mat A=product->A,B=product->B; 2243 const char *algTypes[4] = {"scalable","nonscalable","at*b","backend"}; 2244 PetscInt nalg = 4; 2245 PetscInt alg = 1; /* set default algorithm */ 2246 PetscBool flg; 2247 MPI_Comm comm; 2248 2249 PetscFunctionBegin; 2250 /* Check matrix local sizes */ 2251 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 2252 PetscCheck(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); 2253 2254 /* Set default algorithm */ 2255 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2256 if (flg) { 2257 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2258 } 2259 2260 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2261 if (alg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2262 MatInfo Ainfo,Binfo; 2263 PetscInt nz_local; 2264 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2265 2266 PetscCall(MatGetInfo(A,MAT_LOCAL,&Ainfo)); 2267 PetscCall(MatGetInfo(B,MAT_LOCAL,&Binfo)); 2268 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2269 2270 if (B->cmap->N > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2271 PetscCall(MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm)); 2272 2273 if (alg_scalable) { 2274 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2275 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2276 PetscCall(PetscInfo(B,"Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n",B->cmap->N,(double)(product->fill*nz_local))); 2277 } 2278 } 2279 2280 /* Get runtime option */ 2281 if (product->api_user) { 2282 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatTransposeMatMult","Mat"); 2283 PetscCall(PetscOptionsEList("-mattransposematmult_via","Algorithmic approach","MatTransposeMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2284 PetscOptionsEnd(); 2285 } else { 2286 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AtB","Mat"); 2287 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatTransposeMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2288 PetscOptionsEnd(); 2289 } 2290 if (flg) { 2291 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2292 } 2293 2294 C->ops->productsymbolic = MatProductSymbolic_AtB_MPIAIJ_MPIAIJ; 2295 PetscFunctionReturn(0); 2296 } 2297 2298 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_PtAP(Mat C) 2299 { 2300 Mat_Product *product = C->product; 2301 Mat A=product->A,P=product->B; 2302 MPI_Comm comm; 2303 PetscBool flg; 2304 PetscInt alg=1; /* set default algorithm */ 2305 #if !defined(PETSC_HAVE_HYPRE) 2306 const char *algTypes[5] = {"scalable","nonscalable","allatonce","allatonce_merged","backend"}; 2307 PetscInt nalg=5; 2308 #else 2309 const char *algTypes[6] = {"scalable","nonscalable","allatonce","allatonce_merged","backend","hypre"}; 2310 PetscInt nalg=6; 2311 #endif 2312 PetscInt pN=P->cmap->N; 2313 2314 PetscFunctionBegin; 2315 /* Check matrix local sizes */ 2316 PetscCall(PetscObjectGetComm((PetscObject)C,&comm)); 2317 PetscCheck(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); 2318 PetscCheck(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); 2319 2320 /* Set "nonscalable" as default algorithm */ 2321 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2322 if (flg) { 2323 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2324 2325 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2326 if (pN > 100000) { 2327 MatInfo Ainfo,Pinfo; 2328 PetscInt nz_local; 2329 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 2330 2331 PetscCall(MatGetInfo(A,MAT_LOCAL,&Ainfo)); 2332 PetscCall(MatGetInfo(P,MAT_LOCAL,&Pinfo)); 2333 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); 2334 2335 if (pN > product->fill*nz_local) alg_scalable_loc = PETSC_TRUE; 2336 PetscCall(MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm)); 2337 2338 if (alg_scalable) { 2339 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2340 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2341 } 2342 } 2343 } 2344 2345 /* Get runtime option */ 2346 if (product->api_user) { 2347 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatPtAP","Mat"); 2348 PetscCall(PetscOptionsEList("-matptap_via","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[alg],&alg,&flg)); 2349 PetscOptionsEnd(); 2350 } else { 2351 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_PtAP","Mat"); 2352 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[alg],&alg,&flg)); 2353 PetscOptionsEnd(); 2354 } 2355 if (flg) { 2356 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2357 } 2358 2359 C->ops->productsymbolic = MatProductSymbolic_PtAP_MPIAIJ_MPIAIJ; 2360 PetscFunctionReturn(0); 2361 } 2362 2363 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_RARt(Mat C) 2364 { 2365 Mat_Product *product = C->product; 2366 Mat A = product->A,R=product->B; 2367 2368 PetscFunctionBegin; 2369 /* Check matrix local sizes */ 2370 PetscCheck(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); 2371 2372 C->ops->productsymbolic = MatProductSymbolic_RARt_MPIAIJ_MPIAIJ; 2373 PetscFunctionReturn(0); 2374 } 2375 2376 /* 2377 Set options for ABC = A*B*C = A*(B*C); ABC's algorithm must be chosen from AB's algorithm 2378 */ 2379 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABC(Mat C) 2380 { 2381 Mat_Product *product = C->product; 2382 PetscBool flg = PETSC_FALSE; 2383 PetscInt alg = 1; /* default algorithm */ 2384 const char *algTypes[3] = {"scalable","nonscalable","seqmpi"}; 2385 PetscInt nalg = 3; 2386 2387 PetscFunctionBegin; 2388 /* Set default algorithm */ 2389 PetscCall(PetscStrcmp(C->product->alg,"default",&flg)); 2390 if (flg) { 2391 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2392 } 2393 2394 /* Get runtime option */ 2395 if (product->api_user) { 2396 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMatMult","Mat"); 2397 PetscCall(PetscOptionsEList("-matmatmatmult_via","Algorithmic approach","MatMatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg)); 2398 PetscOptionsEnd(); 2399 } else { 2400 PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_ABC","Mat"); 2401 PetscCall(PetscOptionsEList("-mat_product_algorithm","Algorithmic approach","MatProduct_ABC",algTypes,nalg,algTypes[alg],&alg,&flg)); 2402 PetscOptionsEnd(); 2403 } 2404 if (flg) { 2405 PetscCall(MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg])); 2406 } 2407 2408 C->ops->matmatmultsymbolic = MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ; 2409 C->ops->productsymbolic = MatProductSymbolic_ABC; 2410 PetscFunctionReturn(0); 2411 } 2412 2413 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ(Mat C) 2414 { 2415 Mat_Product *product = C->product; 2416 2417 PetscFunctionBegin; 2418 switch (product->type) { 2419 case MATPRODUCT_AB: 2420 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2421 break; 2422 case MATPRODUCT_ABt: 2423 PetscCall(MatProductSetFromOptions_MPIAIJ_ABt(C)); 2424 break; 2425 case MATPRODUCT_AtB: 2426 PetscCall(MatProductSetFromOptions_MPIAIJ_AtB(C)); 2427 break; 2428 case MATPRODUCT_PtAP: 2429 PetscCall(MatProductSetFromOptions_MPIAIJ_PtAP(C)); 2430 break; 2431 case MATPRODUCT_RARt: 2432 PetscCall(MatProductSetFromOptions_MPIAIJ_RARt(C)); 2433 break; 2434 case MATPRODUCT_ABC: 2435 PetscCall(MatProductSetFromOptions_MPIAIJ_ABC(C)); 2436 break; 2437 default: 2438 break; 2439 } 2440 PetscFunctionReturn(0); 2441 } 2442