1 2 /* 3 Defines matrix-matrix product routines for pairs of MPIAIJ matrices 4 C = A * B 5 */ 6 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 7 #include <../src/mat/utils/freespace.h> 8 #include <../src/mat/impls/aij/mpi/mpiaij.h> 9 #include <petscbt.h> 10 #include <../src/mat/impls/dense/mpi/mpidense.h> 11 #include <petsc/private/vecimpl.h> 12 13 #if defined(PETSC_HAVE_HYPRE) 14 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat*); 15 #endif 16 17 PETSC_INTERN PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill, Mat *C) 18 { 19 PetscErrorCode ierr; 20 #if defined(PETSC_HAVE_HYPRE) 21 const char *algTypes[4] = {"scalable","nonscalable","seqmpi","hypre"}; 22 PetscInt nalg = 4; 23 #else 24 const char *algTypes[3] = {"scalable","nonscalable","seqmpi"}; 25 PetscInt nalg = 3; 26 #endif 27 PetscInt alg = 1; /* set nonscalable algorithm as default */ 28 MPI_Comm comm; 29 PetscBool flg; 30 31 PetscFunctionBegin; 32 if (scall == MAT_INITIAL_MATRIX) { 33 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 34 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 35 36 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 37 PetscOptionsObject->alreadyprinted = PETSC_FALSE; /* a hack to ensure the option shows in '-help' */ 38 ierr = PetscOptionsEList("-matmatmult_via","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[1],&alg,&flg);CHKERRQ(ierr); 39 ierr = PetscOptionsEnd();CHKERRQ(ierr); 40 41 if (!flg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 42 MatInfo Ainfo,Binfo; 43 PetscInt nz_local; 44 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 45 46 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 47 ierr = MatGetInfo(B,MAT_LOCAL,&Binfo);CHKERRQ(ierr); 48 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 49 50 if (B->cmap->N > fill*nz_local) alg_scalable_loc = PETSC_TRUE; 51 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 52 53 if (alg_scalable) { 54 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 55 ierr = PetscInfo2(B,"Use scalable algorithm, BN %D, fill*nz_allocated %g\n",B->cmap->N,fill*nz_local);CHKERRQ(ierr); 56 } 57 } 58 59 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 60 switch (alg) { 61 case 1: 62 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A,B,fill,C);CHKERRQ(ierr); 63 break; 64 case 2: 65 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(A,B,fill,C);CHKERRQ(ierr); 66 break; 67 #if defined(PETSC_HAVE_HYPRE) 68 case 3: 69 ierr = MatMatMultSymbolic_AIJ_AIJ_wHYPRE(A,B,fill,C);CHKERRQ(ierr); 70 break; 71 #endif 72 default: 73 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 74 break; 75 } 76 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 77 } 78 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 79 ierr = (*(*C)->ops->matmultnumeric)(A,B,*C);CHKERRQ(ierr); 80 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 81 PetscFunctionReturn(0); 82 } 83 84 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat A) 85 { 86 PetscErrorCode ierr; 87 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 88 Mat_PtAPMPI *ptap = a->ptap; 89 90 PetscFunctionBegin; 91 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 92 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 93 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 94 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 95 ierr = MatDestroy(&ptap->Pt);CHKERRQ(ierr); 96 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 97 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 98 ierr = PetscFree(ptap->apa);CHKERRQ(ierr); 99 ierr = ptap->destroy(A);CHKERRQ(ierr); 100 ierr = PetscFree(ptap);CHKERRQ(ierr); 101 PetscFunctionReturn(0); 102 } 103 104 PetscErrorCode MatDuplicate_MPIAIJ_MatMatMult(Mat A, MatDuplicateOption op, Mat *M) 105 { 106 PetscErrorCode ierr; 107 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 108 Mat_PtAPMPI *ptap = a->ptap; 109 110 PetscFunctionBegin; 111 ierr = (*ptap->duplicate)(A,op,M);CHKERRQ(ierr); 112 113 (*M)->ops->destroy = ptap->destroy; /* = MatDestroy_MPIAIJ, *M doesn't duplicate A's special structure! */ 114 (*M)->ops->duplicate = ptap->duplicate; /* = MatDuplicate_MPIAIJ */ 115 PetscFunctionReturn(0); 116 } 117 118 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,Mat C) 119 { 120 PetscErrorCode ierr; 121 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 122 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 123 Mat_SeqAIJ *cd =(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 124 PetscScalar *cda=cd->a,*coa=co->a; 125 Mat_SeqAIJ *p_loc,*p_oth; 126 PetscScalar *apa,*ca; 127 PetscInt cm =C->rmap->n; 128 Mat_PtAPMPI *ptap=c->ptap; 129 PetscInt *api,*apj,*apJ,i,k; 130 PetscInt cstart=C->cmap->rstart; 131 PetscInt cdnz,conz,k0,k1; 132 MPI_Comm comm; 133 PetscMPIInt size; 134 135 PetscFunctionBegin; 136 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 137 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 138 139 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 140 /*-----------------------------------------------------*/ 141 /* update numerical values of P_oth and P_loc */ 142 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 143 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 144 145 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 146 /*----------------------------------------------------------*/ 147 /* get data from symbolic products */ 148 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 149 p_oth = NULL; 150 if (size >1) { 151 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 152 } 153 154 /* get apa for storing dense row A[i,:]*P */ 155 apa = ptap->apa; 156 157 api = ptap->api; 158 apj = ptap->apj; 159 for (i=0; i<cm; i++) { 160 /* compute apa = A[i,:]*P */ 161 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 162 163 /* set values in C */ 164 apJ = apj + api[i]; 165 cdnz = cd->i[i+1] - cd->i[i]; 166 conz = co->i[i+1] - co->i[i]; 167 168 /* 1st off-diagoanl part of C */ 169 ca = coa + co->i[i]; 170 k = 0; 171 for (k0=0; k0<conz; k0++) { 172 if (apJ[k] >= cstart) break; 173 ca[k0] = apa[apJ[k]]; 174 apa[apJ[k++]] = 0.0; 175 } 176 177 /* diagonal part of C */ 178 ca = cda + cd->i[i]; 179 for (k1=0; k1<cdnz; k1++) { 180 ca[k1] = apa[apJ[k]]; 181 apa[apJ[k++]] = 0.0; 182 } 183 184 /* 2nd off-diagoanl part of C */ 185 ca = coa + co->i[i]; 186 for (; k0<conz; k0++) { 187 ca[k0] = apa[apJ[k]]; 188 apa[apJ[k++]] = 0.0; 189 } 190 } 191 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 192 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 193 PetscFunctionReturn(0); 194 } 195 196 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,PetscReal fill,Mat *C) 197 { 198 PetscErrorCode ierr; 199 MPI_Comm comm; 200 PetscMPIInt size; 201 Mat Cmpi; 202 Mat_PtAPMPI *ptap; 203 PetscFreeSpaceList free_space=NULL,current_space=NULL; 204 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c; 205 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 206 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz,*dnz_new,*onz_new; 207 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 208 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 209 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 210 PetscBT lnkbt; 211 PetscScalar *apa; 212 PetscReal afill; 213 214 PetscFunctionBegin; 215 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 216 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 217 218 /* create struct Mat_PtAPMPI and attached it to C later */ 219 ierr = PetscNew(&ptap);CHKERRQ(ierr); 220 221 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 222 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 223 224 /* get P_loc by taking all local rows of P */ 225 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 226 227 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 228 pi_loc = p_loc->i; pj_loc = p_loc->j; 229 if (size > 1) { 230 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 231 pi_oth = p_oth->i; pj_oth = p_oth->j; 232 } else { 233 p_oth = NULL; 234 pi_oth = NULL; pj_oth = NULL; 235 } 236 237 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 238 /*-------------------------------------------------------------------*/ 239 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 240 ptap->api = api; 241 api[0] = 0; 242 243 /* create and initialize a linked list */ 244 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 245 246 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 247 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 248 current_space = free_space; 249 250 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 251 for (i=0; i<am; i++) { 252 /* diagonal portion of A */ 253 nzi = adi[i+1] - adi[i]; 254 for (j=0; j<nzi; j++) { 255 row = *adj++; 256 pnz = pi_loc[row+1] - pi_loc[row]; 257 Jptr = pj_loc + pi_loc[row]; 258 /* add non-zero cols of P into the sorted linked list lnk */ 259 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 260 } 261 /* off-diagonal portion of A */ 262 nzi = aoi[i+1] - aoi[i]; 263 for (j=0; j<nzi; j++) { 264 row = *aoj++; 265 pnz = pi_oth[row+1] - pi_oth[row]; 266 Jptr = pj_oth + pi_oth[row]; 267 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 268 } 269 270 apnz = lnk[0]; 271 api[i+1] = api[i] + apnz; 272 273 /* if free space is not available, double the total space in the list */ 274 if (current_space->local_remaining<apnz) { 275 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 276 nspacedouble++; 277 } 278 279 /* Copy data into free space, then initialize lnk */ 280 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 281 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 282 283 current_space->array += apnz; 284 current_space->local_used += apnz; 285 current_space->local_remaining -= apnz; 286 } 287 288 /* Allocate space for apj, initialize apj, and */ 289 /* destroy list of free space and other temporary array(s) */ 290 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 291 apj = ptap->apj; 292 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 293 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 294 295 /* malloc apa to store dense row A[i,:]*P */ 296 ierr = PetscCalloc1(pN,&apa);CHKERRQ(ierr); 297 298 ptap->apa = apa; 299 300 /* create and assemble symbolic parallel matrix Cmpi */ 301 /*----------------------------------------------------*/ 302 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 303 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 304 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 305 306 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 307 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 308 309 /* dnz and onz are freed in MatPreallocateFinalize, so we need a copy of them */ 310 /* for MatSetValues_MPIAIJ_Symbolic(). */ 311 ierr = PetscMalloc2(am,&dnz_new,am,&onz_new);CHKERRQ(ierr); 312 ierr = PetscMemcpy(dnz_new,dnz,am*sizeof(PetscInt));CHKERRQ(ierr); 313 ierr = PetscMemcpy(onz_new,onz,am*sizeof(PetscInt));CHKERRQ(ierr); 314 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 315 316 ierr = MatSetValues_MPIAIJ_Symbolic(Cmpi, apj, api, dnz_new, onz_new);CHKERRQ(ierr); 317 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 318 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 319 320 ptap->destroy = Cmpi->ops->destroy; 321 ptap->duplicate = Cmpi->ops->duplicate; 322 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 323 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 324 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 325 326 /* attach the supporting struct to Cmpi for reuse */ 327 c = (Mat_MPIAIJ*)Cmpi->data; 328 c->ptap = ptap; 329 330 *C = Cmpi; 331 332 /* set MatInfo */ 333 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 334 if (afill < 1.0) afill = 1.0; 335 Cmpi->info.mallocs = nspacedouble; 336 Cmpi->info.fill_ratio_given = fill; 337 Cmpi->info.fill_ratio_needed = afill; 338 339 #if defined(PETSC_USE_INFO) 340 if (api[am]) { 341 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 342 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 343 } else { 344 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 345 } 346 #endif 347 ierr = PetscFree2(dnz_new,onz_new);CHKERRQ(ierr); 348 PetscFunctionReturn(0); 349 } 350 351 PETSC_INTERN PetscErrorCode MatMatMult_MPIAIJ_MPIDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 352 { 353 PetscErrorCode ierr; 354 355 PetscFunctionBegin; 356 if (scall == MAT_INITIAL_MATRIX) { 357 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 358 ierr = MatMatMultSymbolic_MPIAIJ_MPIDense(A,B,fill,C);CHKERRQ(ierr); 359 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 360 } 361 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 362 ierr = MatMatMultNumeric_MPIAIJ_MPIDense(A,B,*C);CHKERRQ(ierr); 363 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 364 PetscFunctionReturn(0); 365 } 366 367 typedef struct { 368 Mat workB; 369 PetscScalar *rvalues,*svalues; 370 MPI_Request *rwaits,*swaits; 371 } MPIAIJ_MPIDense; 372 373 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 374 { 375 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*) ctx; 376 PetscErrorCode ierr; 377 378 PetscFunctionBegin; 379 ierr = MatDestroy(&contents->workB);CHKERRQ(ierr); 380 ierr = PetscFree4(contents->rvalues,contents->svalues,contents->rwaits,contents->swaits);CHKERRQ(ierr); 381 ierr = PetscFree(contents);CHKERRQ(ierr); 382 PetscFunctionReturn(0); 383 } 384 385 /* 386 This is a "dummy function" that handles the case where matrix C was created as a dense matrix 387 directly by the user and passed to MatMatMult() with the MAT_REUSE_MATRIX option 388 389 It is the same as MatMatMultSymbolic_MPIAIJ_MPIDense() except does not create C 390 */ 391 PetscErrorCode MatMatMultNumeric_MPIDense(Mat A,Mat B,Mat C) 392 { 393 PetscErrorCode ierr; 394 PetscBool flg; 395 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 396 PetscInt nz = aij->B->cmap->n; 397 PetscContainer container; 398 MPIAIJ_MPIDense *contents; 399 VecScatter ctx = aij->Mvctx; 400 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 401 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 402 403 PetscFunctionBegin; 404 ierr = PetscObjectTypeCompare((PetscObject)B,MATMPIDENSE,&flg);CHKERRQ(ierr); 405 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Second matrix must be mpidense"); 406 407 /* Handle case where where user provided the final C matrix rather than calling MatMatMult() with MAT_INITIAL_MATRIX*/ 408 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 409 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"First matrix must be MPIAIJ"); 410 411 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 412 413 ierr = PetscNew(&contents);CHKERRQ(ierr); 414 /* Create work matrix used to store off processor rows of B needed for local product */ 415 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,NULL,&contents->workB);CHKERRQ(ierr); 416 /* Create work arrays needed */ 417 ierr = PetscMalloc4(B->cmap->N*from->starts[from->n],&contents->rvalues, 418 B->cmap->N*to->starts[to->n],&contents->svalues, 419 from->n,&contents->rwaits, 420 to->n,&contents->swaits);CHKERRQ(ierr); 421 422 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);CHKERRQ(ierr); 423 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 424 ierr = PetscContainerSetUserDestroy(container,MatMPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 425 ierr = PetscObjectCompose((PetscObject)C,"workB",(PetscObject)container);CHKERRQ(ierr); 426 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 427 428 ierr = (*C->ops->matmultnumeric)(A,B,C);CHKERRQ(ierr); 429 PetscFunctionReturn(0); 430 } 431 432 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C) 433 { 434 PetscErrorCode ierr; 435 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 436 PetscInt nz = aij->B->cmap->n; 437 PetscContainer container; 438 MPIAIJ_MPIDense *contents; 439 VecScatter ctx = aij->Mvctx; 440 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 441 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 442 PetscInt m = A->rmap->n,n=B->cmap->n; 443 444 PetscFunctionBegin; 445 ierr = MatCreate(PetscObjectComm((PetscObject)B),C);CHKERRQ(ierr); 446 ierr = MatSetSizes(*C,m,n,A->rmap->N,B->cmap->N);CHKERRQ(ierr); 447 ierr = MatSetBlockSizesFromMats(*C,A,B);CHKERRQ(ierr); 448 ierr = MatSetType(*C,MATMPIDENSE);CHKERRQ(ierr); 449 ierr = MatMPIDenseSetPreallocation(*C,NULL);CHKERRQ(ierr); 450 ierr = MatAssemblyBegin(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 451 ierr = MatAssemblyEnd(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 452 453 (*C)->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 454 455 ierr = PetscNew(&contents);CHKERRQ(ierr); 456 /* Create work matrix used to store off processor rows of B needed for local product */ 457 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,NULL,&contents->workB);CHKERRQ(ierr); 458 /* Create work arrays needed */ 459 ierr = PetscMalloc4(B->cmap->N*from->starts[from->n],&contents->rvalues, 460 B->cmap->N*to->starts[to->n],&contents->svalues, 461 from->n,&contents->rwaits, 462 to->n,&contents->swaits);CHKERRQ(ierr); 463 464 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);CHKERRQ(ierr); 465 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 466 ierr = PetscContainerSetUserDestroy(container,MatMPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 467 ierr = PetscObjectCompose((PetscObject)(*C),"workB",(PetscObject)container);CHKERRQ(ierr); 468 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 469 PetscFunctionReturn(0); 470 } 471 472 /* 473 Performs an efficient scatter on the rows of B needed by this process; this is 474 a modification of the VecScatterBegin_() routines. 475 */ 476 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,Mat C,Mat *outworkB) 477 { 478 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 479 PetscErrorCode ierr; 480 PetscScalar *b,*w,*svalues,*rvalues; 481 VecScatter ctx = aij->Mvctx; 482 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 483 VecScatter_MPI_General *to = (VecScatter_MPI_General*) ctx->todata; 484 PetscInt i,j,k; 485 PetscInt *sindices,*sstarts,*rindices,*rstarts; 486 PetscMPIInt *sprocs,*rprocs,nrecvs; 487 MPI_Request *swaits,*rwaits; 488 MPI_Comm comm; 489 PetscMPIInt tag = ((PetscObject)ctx)->tag,ncols = B->cmap->N, nrows = aij->B->cmap->n,imdex,nrowsB = B->rmap->n; 490 MPI_Status status; 491 MPIAIJ_MPIDense *contents; 492 PetscContainer container; 493 Mat workB; 494 495 PetscFunctionBegin; 496 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 497 ierr = PetscObjectQuery((PetscObject)C,"workB",(PetscObject*)&container);CHKERRQ(ierr); 498 if (!container) SETERRQ(comm,PETSC_ERR_PLIB,"Container does not exist"); 499 ierr = PetscContainerGetPointer(container,(void**)&contents);CHKERRQ(ierr); 500 501 workB = *outworkB = contents->workB; 502 if (nrows != workB->rmap->n) SETERRQ2(comm,PETSC_ERR_PLIB,"Number of rows of workB %D not equal to columns of aij->B %D",nrows,workB->cmap->n); 503 sindices = to->indices; 504 sstarts = to->starts; 505 sprocs = to->procs; 506 swaits = contents->swaits; 507 svalues = contents->svalues; 508 509 rindices = from->indices; 510 rstarts = from->starts; 511 rprocs = from->procs; 512 rwaits = contents->rwaits; 513 rvalues = contents->rvalues; 514 515 ierr = MatDenseGetArray(B,&b);CHKERRQ(ierr); 516 ierr = MatDenseGetArray(workB,&w);CHKERRQ(ierr); 517 518 for (i=0; i<from->n; i++) { 519 ierr = MPI_Irecv(rvalues+ncols*rstarts[i],ncols*(rstarts[i+1]-rstarts[i]),MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 520 } 521 522 for (i=0; i<to->n; i++) { 523 /* pack a message at a time */ 524 for (j=0; j<sstarts[i+1]-sstarts[i]; j++) { 525 for (k=0; k<ncols; k++) { 526 svalues[ncols*(sstarts[i] + j) + k] = b[sindices[sstarts[i]+j] + nrowsB*k]; 527 } 528 } 529 ierr = MPI_Isend(svalues+ncols*sstarts[i],ncols*(sstarts[i+1]-sstarts[i]),MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 530 } 531 532 nrecvs = from->n; 533 while (nrecvs) { 534 ierr = MPI_Waitany(from->n,rwaits,&imdex,&status);CHKERRQ(ierr); 535 nrecvs--; 536 /* unpack a message at a time */ 537 for (j=0; j<rstarts[imdex+1]-rstarts[imdex]; j++) { 538 for (k=0; k<ncols; k++) { 539 w[rindices[rstarts[imdex]+j] + nrows*k] = rvalues[ncols*(rstarts[imdex] + j) + k]; 540 } 541 } 542 } 543 if (to->n) {ierr = MPI_Waitall(to->n,swaits,to->sstatus);CHKERRQ(ierr);} 544 545 ierr = MatDenseRestoreArray(B,&b);CHKERRQ(ierr); 546 ierr = MatDenseRestoreArray(workB,&w);CHKERRQ(ierr); 547 ierr = MatAssemblyBegin(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 548 ierr = MatAssemblyEnd(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 549 PetscFunctionReturn(0); 550 } 551 extern PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat); 552 553 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 554 { 555 PetscErrorCode ierr; 556 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 557 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 558 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 559 Mat workB; 560 561 PetscFunctionBegin; 562 /* diagonal block of A times all local rows of B*/ 563 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(aij->A,bdense->A,cdense->A);CHKERRQ(ierr); 564 565 /* get off processor parts of B needed to complete the product */ 566 ierr = MatMPIDenseScatter(A,B,C,&workB);CHKERRQ(ierr); 567 568 /* off-diagonal block of A times nonlocal rows of B */ 569 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A);CHKERRQ(ierr); 570 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 571 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 572 PetscFunctionReturn(0); 573 } 574 575 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 576 { 577 PetscErrorCode ierr; 578 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 579 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 580 Mat_SeqAIJ *cd = (Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 581 PetscInt *adi = ad->i,*adj,*aoi=ao->i,*aoj; 582 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 583 Mat_SeqAIJ *p_loc,*p_oth; 584 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 585 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 586 PetscInt cm = C->rmap->n,anz,pnz; 587 Mat_PtAPMPI *ptap = c->ptap; 588 PetscScalar *apa_sparse = ptap->apa; 589 PetscInt *api,*apj,*apJ,i,j,k,row; 590 PetscInt cstart = C->cmap->rstart; 591 PetscInt cdnz,conz,k0,k1,nextp; 592 MPI_Comm comm; 593 PetscMPIInt size; 594 595 PetscFunctionBegin; 596 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 597 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 598 599 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 600 /*-----------------------------------------------------*/ 601 /* update numerical values of P_oth and P_loc */ 602 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 603 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 604 605 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 606 /*----------------------------------------------------------*/ 607 /* get data from symbolic products */ 608 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 609 pi_loc = p_loc->i; pj_loc = p_loc->j; pa_loc = p_loc->a; 610 if (size >1) { 611 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 612 pi_oth = p_oth->i; pj_oth = p_oth->j; pa_oth = p_oth->a; 613 } else { 614 p_oth = NULL; pi_oth = NULL; pj_oth = NULL; pa_oth = NULL; 615 } 616 617 api = ptap->api; 618 apj = ptap->apj; 619 for (i=0; i<cm; i++) { 620 apJ = apj + api[i]; 621 622 /* diagonal portion of A */ 623 anz = adi[i+1] - adi[i]; 624 adj = ad->j + adi[i]; 625 ada = ad->a + adi[i]; 626 for (j=0; j<anz; j++) { 627 row = adj[j]; 628 pnz = pi_loc[row+1] - pi_loc[row]; 629 pj = pj_loc + pi_loc[row]; 630 pa = pa_loc + pi_loc[row]; 631 /* perform sparse axpy */ 632 valtmp = ada[j]; 633 nextp = 0; 634 for (k=0; nextp<pnz; k++) { 635 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 636 apa_sparse[k] += valtmp*pa[nextp++]; 637 } 638 } 639 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 640 } 641 642 /* off-diagonal portion of A */ 643 anz = aoi[i+1] - aoi[i]; 644 aoj = ao->j + aoi[i]; 645 aoa = ao->a + aoi[i]; 646 for (j=0; j<anz; j++) { 647 row = aoj[j]; 648 pnz = pi_oth[row+1] - pi_oth[row]; 649 pj = pj_oth + pi_oth[row]; 650 pa = pa_oth + pi_oth[row]; 651 /* perform sparse axpy */ 652 valtmp = aoa[j]; 653 nextp = 0; 654 for (k=0; nextp<pnz; k++) { 655 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 656 apa_sparse[k] += valtmp*pa[nextp++]; 657 } 658 } 659 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 660 } 661 662 /* set values in C */ 663 cdnz = cd->i[i+1] - cd->i[i]; 664 conz = co->i[i+1] - co->i[i]; 665 666 /* 1st off-diagoanl part of C */ 667 ca = coa + co->i[i]; 668 k = 0; 669 for (k0=0; k0<conz; k0++) { 670 if (apJ[k] >= cstart) break; 671 ca[k0] = apa_sparse[k]; 672 apa_sparse[k] = 0.0; 673 k++; 674 } 675 676 /* diagonal part of C */ 677 ca = cda + cd->i[i]; 678 for (k1=0; k1<cdnz; k1++) { 679 ca[k1] = apa_sparse[k]; 680 apa_sparse[k] = 0.0; 681 k++; 682 } 683 684 /* 2nd off-diagoanl part of C */ 685 ca = coa + co->i[i]; 686 for (; k0<conz; k0++) { 687 ca[k0] = apa_sparse[k]; 688 apa_sparse[k] = 0.0; 689 k++; 690 } 691 } 692 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 693 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 694 PetscFunctionReturn(0); 695 } 696 697 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 698 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 699 { 700 PetscErrorCode ierr; 701 MPI_Comm comm; 702 PetscMPIInt size; 703 Mat Cmpi; 704 Mat_PtAPMPI *ptap; 705 PetscFreeSpaceList free_space = NULL,current_space=NULL; 706 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c; 707 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 708 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz,*dnz_new,*onz_new; 709 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 710 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max; 711 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 712 PetscReal afill; 713 PetscScalar *apa; 714 PetscTable ta; 715 716 PetscFunctionBegin; 717 718 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 719 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 720 721 /* create struct Mat_PtAPMPI and attached it to C later */ 722 ierr = PetscNew(&ptap);CHKERRQ(ierr); 723 724 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 725 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 726 727 /* get P_loc by taking all local rows of P */ 728 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 729 730 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 731 pi_loc = p_loc->i; pj_loc = p_loc->j; 732 if (size > 1) { 733 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 734 pi_oth = p_oth->i; pj_oth = p_oth->j; 735 } else { 736 p_oth = NULL; 737 pi_oth = NULL; pj_oth = NULL; 738 } 739 740 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 741 /*-------------------------------------------------------------------*/ 742 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 743 ptap->api = api; 744 api[0] = 0; 745 746 /* create and initialize a linked list */ 747 ierr = PetscTableCreate(pn,pN,&ta);CHKERRQ(ierr); 748 749 /* Calculate apnz_max */ 750 apnz_max = 0; 751 for (i=0; i<am; i++) { 752 ierr = PetscTableRemoveAll(ta);CHKERRQ(ierr); 753 /* diagonal portion of A */ 754 nzi = adi[i+1] - adi[i]; 755 Jptr = adj+adi[i]; /* cols of A_diag */ 756 MatMergeRows_SeqAIJ(p_loc,nzi,Jptr,ta); 757 ierr = PetscTableGetCount(ta,&apnz);CHKERRQ(ierr); 758 if (apnz_max < apnz) apnz_max = apnz; 759 760 /* off-diagonal portion of A */ 761 nzi = aoi[i+1] - aoi[i]; 762 Jptr = aoj+aoi[i]; /* cols of A_off */ 763 MatMergeRows_SeqAIJ(p_oth,nzi,Jptr,ta); 764 ierr = PetscTableGetCount(ta,&apnz);CHKERRQ(ierr); 765 if (apnz_max < apnz) apnz_max = apnz; 766 } 767 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 768 769 ierr = PetscLLCondensedCreate_Scalable(apnz_max,&lnk);CHKERRQ(ierr); 770 771 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 772 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 773 current_space = free_space; 774 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 775 for (i=0; i<am; i++) { 776 /* diagonal portion of A */ 777 nzi = adi[i+1] - adi[i]; 778 for (j=0; j<nzi; j++) { 779 row = *adj++; 780 pnz = pi_loc[row+1] - pi_loc[row]; 781 Jptr = pj_loc + pi_loc[row]; 782 /* add non-zero cols of P into the sorted linked list lnk */ 783 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 784 } 785 /* off-diagonal portion of A */ 786 nzi = aoi[i+1] - aoi[i]; 787 for (j=0; j<nzi; j++) { 788 row = *aoj++; 789 pnz = pi_oth[row+1] - pi_oth[row]; 790 Jptr = pj_oth + pi_oth[row]; 791 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 792 } 793 794 apnz = *lnk; 795 api[i+1] = api[i] + apnz; 796 797 /* if free space is not available, double the total space in the list */ 798 if (current_space->local_remaining<apnz) { 799 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 800 nspacedouble++; 801 } 802 803 /* Copy data into free space, then initialize lnk */ 804 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 805 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 806 807 current_space->array += apnz; 808 current_space->local_used += apnz; 809 current_space->local_remaining -= apnz; 810 } 811 812 /* Allocate space for apj, initialize apj, and */ 813 /* destroy list of free space and other temporary array(s) */ 814 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 815 apj = ptap->apj; 816 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 817 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 818 819 /* create and assemble symbolic parallel matrix Cmpi */ 820 /*----------------------------------------------------*/ 821 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 822 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 823 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 824 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 825 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 826 827 /* malloc apa for assembly Cmpi */ 828 ierr = PetscCalloc1(apnz_max,&apa);CHKERRQ(ierr); 829 830 ptap->apa = apa; 831 832 /* dnz and onz are freed in MatPreallocateFinalize, so we need a copy of them */ 833 /* for MatSetValues_MPIAIJ_Symbolic(). */ 834 ierr = PetscMalloc2(am,&dnz_new,am,&onz_new);CHKERRQ(ierr); 835 ierr = PetscMemcpy(dnz_new,dnz,am*sizeof(PetscInt));CHKERRQ(ierr); 836 ierr = PetscMemcpy(onz_new,onz,am*sizeof(PetscInt));CHKERRQ(ierr); 837 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 838 839 ierr = MatSetValues_MPIAIJ_Symbolic(Cmpi, apj, api, dnz_new, onz_new);CHKERRQ(ierr); 840 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 841 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 842 843 ptap->destroy = Cmpi->ops->destroy; 844 ptap->duplicate = Cmpi->ops->duplicate; 845 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 846 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 847 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 848 849 /* attach the supporting struct to Cmpi for reuse */ 850 c = (Mat_MPIAIJ*)Cmpi->data; 851 c->ptap = ptap; 852 853 *C = Cmpi; 854 855 /* set MatInfo */ 856 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 857 if (afill < 1.0) afill = 1.0; 858 Cmpi->info.mallocs = nspacedouble; 859 Cmpi->info.fill_ratio_given = fill; 860 Cmpi->info.fill_ratio_needed = afill; 861 862 #if defined(PETSC_USE_INFO) 863 if (api[am]) { 864 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 865 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 866 } else { 867 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 868 } 869 #endif 870 ierr = PetscFree2(dnz_new,onz_new);CHKERRQ(ierr); 871 PetscFunctionReturn(0); 872 } 873 874 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 875 /* Three input arrays are merged to one output array. The size of the */ 876 /* output array is also output. Duplicate entries only show up once. */ 877 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, 878 PetscInt size2, PetscInt *in2, 879 PetscInt size3, PetscInt *in3, 880 PetscInt *size4, PetscInt *out) 881 { 882 int i = 0, j = 0, k = 0, l = 0; 883 884 /* Traverse all three arrays */ 885 while (i<size1 && j<size2 && k<size3) { 886 if (in1[i] < in2[j] && in1[i] < in3[k]) { 887 out[l++] = in1[i++]; 888 } 889 else if(in2[j] < in1[i] && in2[j] < in3[k]) { 890 out[l++] = in2[j++]; 891 } 892 else if(in3[k] < in1[i] && in3[k] < in2[j]) { 893 out[l++] = in3[k++]; 894 } 895 else if(in1[i] == in2[j] && in1[i] < in3[k]) { 896 out[l++] = in1[i]; 897 i++, j++; 898 } 899 else if(in1[i] == in3[k] && in1[i] < in2[j]) { 900 out[l++] = in1[i]; 901 i++, k++; 902 } 903 else if(in3[k] == in2[j] && in2[j] < in1[i]) { 904 out[l++] = in2[j]; 905 k++, j++; 906 } 907 else if(in1[i] == in2[j] && in1[i] == in3[k]) { 908 out[l++] = in1[i]; 909 i++, j++, k++; 910 } 911 } 912 913 /* Traverse two remaining arrays */ 914 while (i<size1 && j<size2) { 915 if (in1[i] < in2[j]) { 916 out[l++] = in1[i++]; 917 } 918 else if(in1[i] > in2[j]) { 919 out[l++] = in2[j++]; 920 } 921 else { 922 out[l++] = in1[i]; 923 i++, j++; 924 } 925 } 926 927 while (i<size1 && k<size3) { 928 if (in1[i] < in3[k]) { 929 out[l++] = in1[i++]; 930 } 931 else if(in1[i] > in3[k]) { 932 out[l++] = in3[k++]; 933 } 934 else { 935 out[l++] = in1[i]; 936 i++, k++; 937 } 938 } 939 940 while (k<size3 && j<size2) { 941 if (in3[k] < in2[j]) { 942 out[l++] = in3[k++]; 943 } 944 else if(in3[k] > in2[j]) { 945 out[l++] = in2[j++]; 946 } 947 else { 948 out[l++] = in3[k]; 949 k++, j++; 950 } 951 } 952 953 /* Traverse one remaining array */ 954 while (i<size1) out[l++] = in1[i++]; 955 while (j<size2) out[l++] = in2[j++]; 956 while (k<size3) out[l++] = in3[k++]; 957 958 *size4 = l; 959 } 960 961 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 962 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 963 /* matrix-matrix multiplications. */ 964 #undef __FUNCT__ 965 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI" 966 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat *C) 967 { 968 PetscErrorCode ierr; 969 MPI_Comm comm; 970 PetscMPIInt size; 971 Mat Cmpi; 972 Mat_PtAPMPI *ptap; 973 PetscFreeSpaceList free_space_diag=NULL, current_space=NULL; 974 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 975 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc; 976 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 977 Mat_MPIAIJ *c; 978 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 979 PetscInt adponz, adpdnz; 980 PetscInt *pi_loc,*dnz,*onz,*dnz_new,*onz_new; 981 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,rstart=A->rmap->rstart; 982 PetscInt *lnk,i, i1=0,pnz,row,*adpoi,*adpoj, *api, *adpoJ, *aopJ, *apJ,*Jptr, aopnz, nspacedouble=0,j,nzi, 983 *apj,apnz, *adpdi, *adpdj, *adpdJ, *poff_i, *poff_j, *j_temp, *aopothi, *aopothj; 984 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n, p_colstart, p_colend; 985 PetscBT lnkbt; 986 PetscScalar *apa; 987 PetscReal afill; 988 PetscMPIInt rank; 989 Mat adpd, aopoth; 990 991 PetscFunctionBegin; 992 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 993 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 994 ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr); 995 ierr = MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend); CHKERRQ(ierr); 996 997 /* create struct Mat_PtAPMPI and attached it to C later */ 998 ierr = PetscNew(&ptap);CHKERRQ(ierr); 999 1000 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1001 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1002 1003 /* get P_loc by taking all local rows of P */ 1004 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1005 1006 1007 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1008 pi_loc = p_loc->i; 1009 1010 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 1011 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 1012 ierr = PetscMalloc1(am+2,&adpoi);CHKERRQ(ierr); 1013 1014 adpoi[0] = 0; 1015 ptap->api = api; 1016 api[0] = 0; 1017 1018 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1019 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 1020 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 1021 1022 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1023 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->A, p->A, fill, &adpd);CHKERRQ(ierr); 1024 adpd_seq = (Mat_SeqAIJ*)((adpd)->data); 1025 adpdi = adpd_seq->i; adpdj = adpd_seq->j; 1026 p_off = (Mat_SeqAIJ*)((p->B)->data); 1027 poff_i = p_off->i; poff_j = p_off->j; 1028 1029 /* j_temp stores indices of a result row before they are added to the linked list */ 1030 ierr = PetscMalloc1(pN+2,&j_temp);CHKERRQ(ierr); 1031 1032 1033 /* Symbolic calc of the A_diag * p_loc_off */ 1034 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1035 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space_diag);CHKERRQ(ierr); 1036 current_space = free_space_diag; 1037 1038 for (i=0; i<am; i++) { 1039 /* A_diag * P_loc_off */ 1040 nzi = adi[i+1] - adi[i]; 1041 for (j=0; j<nzi; j++) { 1042 row = *adj++; 1043 pnz = poff_i[row+1] - poff_i[row]; 1044 Jptr = poff_j + poff_i[row]; 1045 for(i1 = 0; i1 < pnz; i1++) { 1046 j_temp[i1] = p->garray[Jptr[i1]]; 1047 } 1048 /* add non-zero cols of P into the sorted linked list lnk */ 1049 ierr = PetscLLCondensedAddSorted(pnz,j_temp,lnk,lnkbt);CHKERRQ(ierr); 1050 } 1051 1052 adponz = lnk[0]; 1053 adpoi[i+1] = adpoi[i] + adponz; 1054 1055 /* if free space is not available, double the total space in the list */ 1056 if (current_space->local_remaining<adponz) { 1057 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(adponz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1058 nspacedouble++; 1059 } 1060 1061 /* Copy data into free space, then initialize lnk */ 1062 ierr = PetscLLCondensedClean(pN,adponz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1063 1064 current_space->array += adponz; 1065 current_space->local_used += adponz; 1066 current_space->local_remaining -= adponz; 1067 } 1068 1069 /* Symbolic calc of A_off * P_oth */ 1070 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, &aopoth);CHKERRQ(ierr); 1071 aopoth_seq = (Mat_SeqAIJ*)((aopoth)->data); 1072 aopothi = aopoth_seq->i; aopothj = aopoth_seq->j; 1073 1074 /* Allocate space for apj, adpj, aopj, ... */ 1075 /* destroy lists of free space and other temporary array(s) */ 1076 1077 ierr = PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am]+2, &ptap->apj);CHKERRQ(ierr); 1078 ierr = PetscMalloc1(adpoi[am]+2, &adpoj);CHKERRQ(ierr); 1079 1080 /* Copy from linked list to j-array */ 1081 ierr = PetscFreeSpaceContiguous(&free_space_diag,adpoj);CHKERRQ(ierr); 1082 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1083 1084 adpoJ = adpoj; 1085 adpdJ = adpdj; 1086 aopJ = aopothj; 1087 apj = ptap->apj; 1088 apJ = apj; /* still empty */ 1089 1090 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1091 /* A_diag * P_loc_diag to get A*P */ 1092 for (i = 0; i < am; i++) { 1093 aopnz = aopothi[i+1] - aopothi[i]; 1094 adponz = adpoi[i+1] - adpoi[i]; 1095 adpdnz = adpdi[i+1] - adpdi[i]; 1096 1097 /* Correct indices from A_diag*P_diag */ 1098 for(i1 = 0; i1 < adpdnz; i1++) { 1099 adpdJ[i1] += p_colstart; 1100 } 1101 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1102 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1103 ierr = MatPreallocateSet(i+rstart, apnz, apJ, dnz, onz); CHKERRQ(ierr); 1104 1105 aopJ += aopnz; 1106 adpoJ += adponz; 1107 adpdJ += adpdnz; 1108 apJ += apnz; 1109 api[i+1] = api[i] + apnz; 1110 } 1111 1112 /* malloc apa to store dense row A[i,:]*P */ 1113 ierr = PetscCalloc1(pN+2,&apa);CHKERRQ(ierr); 1114 1115 ptap->apa = apa; 1116 /* create and assemble symbolic parallel matrix Cmpi */ 1117 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1118 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1119 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 1120 1121 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1122 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1123 1124 /* dnz and onz are freed in MatPreallocateFinalize, so we need a copy of them */ 1125 /* for MatSetValues_MPIAIJ_Symbolic(). */ 1126 ierr = PetscMalloc2(am,&dnz_new,am,&onz_new);CHKERRQ(ierr); 1127 ierr = PetscMemcpy(dnz_new,dnz,am*sizeof(PetscInt));CHKERRQ(ierr); 1128 ierr = PetscMemcpy(onz_new,onz,am*sizeof(PetscInt));CHKERRQ(ierr); 1129 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1130 1131 1132 ierr = MatSetValues_MPIAIJ_Symbolic(Cmpi, apj, api, dnz_new, onz_new);CHKERRQ(ierr); 1133 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1134 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1135 1136 1137 ptap->destroy = Cmpi->ops->destroy; 1138 ptap->duplicate = Cmpi->ops->duplicate; 1139 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1140 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 1141 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 1142 1143 /* attach the supporting struct to Cmpi for reuse */ 1144 c = (Mat_MPIAIJ*)Cmpi->data; 1145 c->ptap = ptap; 1146 *C = Cmpi; 1147 1148 /* set MatInfo */ 1149 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 1150 if (afill < 1.0) afill = 1.0; 1151 Cmpi->info.mallocs = nspacedouble; 1152 Cmpi->info.fill_ratio_given = fill; 1153 Cmpi->info.fill_ratio_needed = afill; 1154 1155 #if defined(PETSC_USE_INFO) 1156 if (api[am]) { 1157 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 1158 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 1159 } else { 1160 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1161 } 1162 #endif 1163 1164 ierr = MatDestroy(&aopoth);CHKERRQ(ierr); 1165 ierr = MatDestroy(&adpd);CHKERRQ(ierr); 1166 ierr = PetscFree(j_temp);CHKERRQ(ierr); 1167 ierr = PetscFree(adpoj);CHKERRQ(ierr); 1168 ierr = PetscFree(adpoi);CHKERRQ(ierr); 1169 ierr = PetscFree2(dnz_new,onz_new);CHKERRQ(ierr); 1170 PetscFunctionReturn(0); 1171 } 1172 1173 1174 /*-------------------------------------------------------------------------*/ 1175 PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat P,Mat A,MatReuse scall,PetscReal fill,Mat *C) 1176 { 1177 PetscErrorCode ierr; 1178 const char *algTypes[3] = {"scalable","nonscalable","matmatmult"}; 1179 PetscInt aN=A->cmap->N,alg=1; /* set default algorithm */ 1180 PetscBool flg; 1181 1182 PetscFunctionBegin; 1183 if (scall == MAT_INITIAL_MATRIX) { 1184 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 1185 PetscOptionsObject->alreadyprinted = PETSC_FALSE; /* a hack to ensure the option shows in '-help' */ 1186 ierr = PetscOptionsEList("-mattransposematmult_via","Algorithmic approach","MatTransposeMatMult",algTypes,3,algTypes[1],&alg,&flg);CHKERRQ(ierr); 1187 ierr = PetscOptionsEnd();CHKERRQ(ierr); 1188 1189 ierr = PetscLogEventBegin(MAT_TransposeMatMultSymbolic,P,A,0,0);CHKERRQ(ierr); 1190 switch (alg) { 1191 case 1: 1192 if (!flg && aN > 100000) { /* may switch to scalable algorithm as default */ 1193 MatInfo Ainfo,Pinfo; 1194 PetscInt nz_local; 1195 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 1196 MPI_Comm comm; 1197 1198 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 1199 ierr = MatGetInfo(P,MAT_LOCAL,&Pinfo);CHKERRQ(ierr); 1200 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); /* estimated local nonzero entries */ 1201 1202 if (aN > fill*nz_local) alg_scalable_loc = PETSC_TRUE; 1203 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1204 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 1205 1206 if (alg_scalable) { 1207 alg = 0; /* scalable algorithm would slower than nonscalable algorithm */ 1208 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 1209 break; 1210 } 1211 } 1212 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(P,A,fill,C);CHKERRQ(ierr); 1213 break; 1214 case 2: 1215 { 1216 Mat Pt; 1217 Mat_PtAPMPI *ptap; 1218 Mat_MPIAIJ *c; 1219 ierr = MatTranspose(P,MAT_INITIAL_MATRIX,&Pt);CHKERRQ(ierr); 1220 ierr = MatMatMult(Pt,A,MAT_INITIAL_MATRIX,fill,C);CHKERRQ(ierr); 1221 c = (Mat_MPIAIJ*)(*C)->data; 1222 ptap = c->ptap; 1223 ptap->Pt = Pt; 1224 (*C)->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 1225 PetscFunctionReturn(0); 1226 } 1227 break; 1228 default: /* scalable algorithm */ 1229 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 1230 break; 1231 } 1232 ierr = PetscLogEventEnd(MAT_TransposeMatMultSymbolic,P,A,0,0);CHKERRQ(ierr); 1233 } 1234 ierr = PetscLogEventBegin(MAT_TransposeMatMultNumeric,P,A,0,0);CHKERRQ(ierr); 1235 ierr = (*(*C)->ops->mattransposemultnumeric)(P,A,*C);CHKERRQ(ierr); 1236 ierr = PetscLogEventEnd(MAT_TransposeMatMultNumeric,P,A,0,0);CHKERRQ(ierr); 1237 PetscFunctionReturn(0); 1238 } 1239 1240 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1241 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P,Mat A,Mat C) 1242 { 1243 PetscErrorCode ierr; 1244 Mat_MPIAIJ *c=(Mat_MPIAIJ*)C->data; 1245 Mat_PtAPMPI *ptap= c->ptap; 1246 Mat Pt=ptap->Pt; 1247 1248 PetscFunctionBegin; 1249 ierr = MatTranspose(P,MAT_REUSE_MATRIX,&Pt);CHKERRQ(ierr); 1250 ierr = MatMatMultNumeric(Pt,A,C);CHKERRQ(ierr); 1251 PetscFunctionReturn(0); 1252 } 1253 1254 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat,MatDuplicateOption,Mat*); 1255 1256 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1257 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,PetscReal fill,Mat *C) 1258 { 1259 PetscErrorCode ierr; 1260 Mat_PtAPMPI *ptap; 1261 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c; 1262 MPI_Comm comm; 1263 PetscMPIInt size,rank; 1264 Mat Cmpi; 1265 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1266 PetscInt pn=P->cmap->n,aN=A->cmap->N,an=A->cmap->n; 1267 PetscInt *lnk,i,k,nsend; 1268 PetscBT lnkbt; 1269 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 1270 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1271 PetscInt len,proc,*dnz,*onz,*owners,nzi; 1272 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1273 MPI_Request *swaits,*rwaits; 1274 MPI_Status *sstatus,rstatus; 1275 PetscLayout rowmap; 1276 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1277 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 1278 PetscInt *Jptr,*prmap=p->garray,con,j,Crmax; 1279 Mat_SeqAIJ *a_loc,*c_loc,*c_oth; 1280 PetscTable ta; 1281 1282 PetscFunctionBegin; 1283 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1284 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1285 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1286 1287 /* create symbolic parallel matrix Cmpi */ 1288 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1289 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1290 1291 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1292 1293 /* create struct Mat_PtAPMPI and attached it to C later */ 1294 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1295 ptap->reuse = MAT_INITIAL_MATRIX; 1296 1297 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1298 /* --------------------------------- */ 1299 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1300 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1301 1302 /* (1) compute symbolic A_loc */ 1303 /* ---------------------------*/ 1304 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1305 1306 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1307 /* ------------------------------------ */ 1308 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->A_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 1309 1310 /* (3) send coj of C_oth to other processors */ 1311 /* ------------------------------------------ */ 1312 /* determine row ownership */ 1313 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 1314 rowmap->n = pn; 1315 rowmap->bs = 1; 1316 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 1317 owners = rowmap->range; 1318 1319 /* determine the number of messages to send, their lengths */ 1320 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 1321 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1322 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1323 1324 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 1325 coi = c_oth->i; coj = c_oth->j; 1326 con = ptap->C_oth->rmap->n; 1327 proc = 0; 1328 for (i=0; i<con; i++) { 1329 while (prmap[i] >= owners[proc+1]) proc++; 1330 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1331 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1332 } 1333 1334 len = 0; /* max length of buf_si[], see (4) */ 1335 owners_co[0] = 0; 1336 nsend = 0; 1337 for (proc=0; proc<size; proc++) { 1338 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1339 if (len_s[proc]) { 1340 nsend++; 1341 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1342 len += len_si[proc]; 1343 } 1344 } 1345 1346 /* determine the number and length of messages to receive for coi and coj */ 1347 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 1348 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 1349 1350 /* post the Irecv and Isend of coj */ 1351 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1352 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1353 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 1354 for (proc=0, k=0; proc<size; proc++) { 1355 if (!len_s[proc]) continue; 1356 i = owners_co[proc]; 1357 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1358 k++; 1359 } 1360 1361 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1362 /* ---------------------------------------- */ 1363 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->A_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 1364 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 1365 1366 /* receives coj are complete */ 1367 for (i=0; i<nrecv; i++) { 1368 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1369 } 1370 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1371 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 1372 1373 /* add received column indices into ta to update Crmax */ 1374 a_loc = (Mat_SeqAIJ*)(ptap->A_loc)->data; 1375 1376 /* create and initialize a linked list */ 1377 ierr = PetscTableCreate(an,aN,&ta);CHKERRQ(ierr); /* for compute Crmax */ 1378 MatRowMergeMax_SeqAIJ(a_loc,ptap->A_loc->rmap->N,ta); 1379 1380 for (k=0; k<nrecv; k++) {/* k-th received message */ 1381 Jptr = buf_rj[k]; 1382 for (j=0; j<len_r[k]; j++) { 1383 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1384 } 1385 } 1386 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 1387 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 1388 1389 /* (4) send and recv coi */ 1390 /*-----------------------*/ 1391 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1392 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1393 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1394 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1395 for (proc=0,k=0; proc<size; proc++) { 1396 if (!len_s[proc]) continue; 1397 /* form outgoing message for i-structure: 1398 buf_si[0]: nrows to be sent 1399 [1:nrows]: row index (global) 1400 [nrows+1:2*nrows+1]: i-structure index 1401 */ 1402 /*-------------------------------------------*/ 1403 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 1404 buf_si_i = buf_si + nrows+1; 1405 buf_si[0] = nrows; 1406 buf_si_i[0] = 0; 1407 nrows = 0; 1408 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1409 nzi = coi[i+1] - coi[i]; 1410 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1411 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1412 nrows++; 1413 } 1414 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1415 k++; 1416 buf_si += len_si[proc]; 1417 } 1418 for (i=0; i<nrecv; i++) { 1419 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1420 } 1421 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1422 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 1423 1424 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 1425 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1426 ierr = PetscFree(swaits);CHKERRQ(ierr); 1427 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1428 1429 /* (5) compute the local portion of Cmpi */ 1430 /* ------------------------------------------ */ 1431 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of Cmpi */ 1432 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 1433 current_space = free_space; 1434 1435 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 1436 for (k=0; k<nrecv; k++) { 1437 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1438 nrows = *buf_ri_k[k]; 1439 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1440 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1441 } 1442 1443 ierr = MatPreallocateInitialize(comm,pn,an,dnz,onz);CHKERRQ(ierr); 1444 ierr = PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt);CHKERRQ(ierr); 1445 for (i=0; i<pn; i++) { 1446 /* add C_loc into Cmpi */ 1447 nzi = c_loc->i[i+1] - c_loc->i[i]; 1448 Jptr = c_loc->j + c_loc->i[i]; 1449 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1450 1451 /* add received col data into lnk */ 1452 for (k=0; k<nrecv; k++) { /* k-th received message */ 1453 if (i == *nextrow[k]) { /* i-th row */ 1454 nzi = *(nextci[k]+1) - *nextci[k]; 1455 Jptr = buf_rj[k] + *nextci[k]; 1456 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1457 nextrow[k]++; nextci[k]++; 1458 } 1459 } 1460 nzi = lnk[0]; 1461 1462 /* copy data into free space, then initialize lnk */ 1463 ierr = PetscLLCondensedClean(aN,nzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1464 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 1465 } 1466 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1467 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1468 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 1469 1470 /* local sizes and preallocation */ 1471 ierr = MatSetSizes(Cmpi,pn,an,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1472 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 1473 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1474 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1475 1476 /* members in merge */ 1477 ierr = PetscFree(id_r);CHKERRQ(ierr); 1478 ierr = PetscFree(len_r);CHKERRQ(ierr); 1479 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 1480 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 1481 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 1482 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 1483 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 1484 1485 /* attach the supporting struct to Cmpi for reuse */ 1486 c = (Mat_MPIAIJ*)Cmpi->data; 1487 c->ptap = ptap; 1488 ptap->duplicate = Cmpi->ops->duplicate; 1489 ptap->destroy = Cmpi->ops->destroy; 1490 1491 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 1492 Cmpi->assembled = PETSC_FALSE; 1493 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1494 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 1495 *C = Cmpi; 1496 PetscFunctionReturn(0); 1497 } 1498 1499 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,Mat C) 1500 { 1501 PetscErrorCode ierr; 1502 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1503 Mat_SeqAIJ *c_seq; 1504 Mat_PtAPMPI *ptap = c->ptap; 1505 Mat A_loc,C_loc,C_oth; 1506 PetscInt i,rstart,rend,cm,ncols,row; 1507 const PetscInt *cols; 1508 const PetscScalar *vals; 1509 1510 PetscFunctionBegin; 1511 ierr = MatZeroEntries(C);CHKERRQ(ierr); 1512 1513 if (ptap->reuse == MAT_REUSE_MATRIX) { 1514 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1515 /* 1) get R = Pd^T, Ro = Po^T */ 1516 /*----------------------------*/ 1517 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1518 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1519 1520 /* 2) compute numeric A_loc */ 1521 /*--------------------------*/ 1522 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1523 } 1524 1525 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1526 A_loc = ptap->A_loc; 1527 ierr = ((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,A_loc,ptap->C_loc);CHKERRQ(ierr); 1528 ierr = ((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,A_loc,ptap->C_oth);CHKERRQ(ierr); 1529 C_loc = ptap->C_loc; 1530 C_oth = ptap->C_oth; 1531 1532 /* add C_loc and Co to to C */ 1533 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 1534 1535 /* C_loc -> C */ 1536 cm = C_loc->rmap->N; 1537 c_seq = (Mat_SeqAIJ*)C_loc->data; 1538 cols = c_seq->j; 1539 vals = c_seq->a; 1540 for (i=0; i<cm; i++) { 1541 ncols = c_seq->i[i+1] - c_seq->i[i]; 1542 row = rstart + i; 1543 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1544 cols += ncols; vals += ncols; 1545 } 1546 1547 /* Co -> C, off-processor part */ 1548 cm = C_oth->rmap->N; 1549 c_seq = (Mat_SeqAIJ*)C_oth->data; 1550 cols = c_seq->j; 1551 vals = c_seq->a; 1552 for (i=0; i<cm; i++) { 1553 ncols = c_seq->i[i+1] - c_seq->i[i]; 1554 row = p->garray[i]; 1555 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1556 cols += ncols; vals += ncols; 1557 } 1558 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1559 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1560 1561 ptap->reuse = MAT_REUSE_MATRIX; 1562 PetscFunctionReturn(0); 1563 } 1564 1565 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 1566 { 1567 PetscErrorCode ierr; 1568 Mat_Merge_SeqsToMPI *merge; 1569 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1570 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1571 Mat_PtAPMPI *ptap; 1572 PetscInt *adj; 1573 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1574 MatScalar *ada,*ca,valtmp; 1575 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1576 MPI_Comm comm; 1577 PetscMPIInt size,rank,taga,*len_s; 1578 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1579 PetscInt **buf_ri,**buf_rj; 1580 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1581 MPI_Request *s_waits,*r_waits; 1582 MPI_Status *status; 1583 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1584 PetscInt *ai,*aj,*coi,*coj,*poJ,*pdJ; 1585 Mat A_loc; 1586 Mat_SeqAIJ *a_loc; 1587 1588 PetscFunctionBegin; 1589 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1590 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1591 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1592 1593 ptap = c->ptap; 1594 merge = ptap->merge; 1595 1596 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1597 /*------------------------------------------*/ 1598 /* get data from symbolic products */ 1599 coi = merge->coi; coj = merge->coj; 1600 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1601 bi = merge->bi; bj = merge->bj; 1602 owners = merge->rowmap->range; 1603 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); 1604 1605 /* get A_loc by taking all local rows of A */ 1606 A_loc = ptap->A_loc; 1607 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 1608 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1609 ai = a_loc->i; 1610 aj = a_loc->j; 1611 1612 for (i=0; i<am; i++) { 1613 anz = ai[i+1] - ai[i]; 1614 adj = aj + ai[i]; 1615 ada = a_loc->a + ai[i]; 1616 1617 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1618 /*-------------------------------------------------------------*/ 1619 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1620 pnz = po->i[i+1] - po->i[i]; 1621 poJ = po->j + po->i[i]; 1622 pA = po->a + po->i[i]; 1623 for (j=0; j<pnz; j++) { 1624 row = poJ[j]; 1625 cj = coj + coi[row]; 1626 ca = coa + coi[row]; 1627 /* perform sparse axpy */ 1628 nexta = 0; 1629 valtmp = pA[j]; 1630 for (k=0; nexta<anz; k++) { 1631 if (cj[k] == adj[nexta]) { 1632 ca[k] += valtmp*ada[nexta]; 1633 nexta++; 1634 } 1635 } 1636 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1637 } 1638 1639 /* put the value into Cd (diagonal part) */ 1640 pnz = pd->i[i+1] - pd->i[i]; 1641 pdJ = pd->j + pd->i[i]; 1642 pA = pd->a + pd->i[i]; 1643 for (j=0; j<pnz; j++) { 1644 row = pdJ[j]; 1645 cj = bj + bi[row]; 1646 ca = ba + bi[row]; 1647 /* perform sparse axpy */ 1648 nexta = 0; 1649 valtmp = pA[j]; 1650 for (k=0; nexta<anz; k++) { 1651 if (cj[k] == adj[nexta]) { 1652 ca[k] += valtmp*ada[nexta]; 1653 nexta++; 1654 } 1655 } 1656 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1657 } 1658 } 1659 1660 /* 3) send and recv matrix values coa */ 1661 /*------------------------------------*/ 1662 buf_ri = merge->buf_ri; 1663 buf_rj = merge->buf_rj; 1664 len_s = merge->len_s; 1665 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1666 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1667 1668 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1669 for (proc=0,k=0; proc<size; proc++) { 1670 if (!len_s[proc]) continue; 1671 i = merge->owners_co[proc]; 1672 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1673 k++; 1674 } 1675 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1676 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1677 1678 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1679 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1680 ierr = PetscFree(coa);CHKERRQ(ierr); 1681 1682 /* 4) insert local Cseq and received values into Cmpi */ 1683 /*----------------------------------------------------*/ 1684 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1685 for (k=0; k<merge->nrecv; k++) { 1686 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1687 nrows = *(buf_ri_k[k]); 1688 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1689 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1690 } 1691 1692 for (i=0; i<cm; i++) { 1693 row = owners[rank] + i; /* global row index of C_seq */ 1694 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1695 ba_i = ba + bi[i]; 1696 bnz = bi[i+1] - bi[i]; 1697 /* add received vals into ba */ 1698 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1699 /* i-th row */ 1700 if (i == *nextrow[k]) { 1701 cnz = *(nextci[k]+1) - *nextci[k]; 1702 cj = buf_rj[k] + *(nextci[k]); 1703 ca = abuf_r[k] + *(nextci[k]); 1704 nextcj = 0; 1705 for (j=0; nextcj<cnz; j++) { 1706 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1707 ba_i[j] += ca[nextcj++]; 1708 } 1709 } 1710 nextrow[k]++; nextci[k]++; 1711 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1712 } 1713 } 1714 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1715 } 1716 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1717 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1718 1719 ierr = PetscFree(ba);CHKERRQ(ierr); 1720 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1721 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1722 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1723 PetscFunctionReturn(0); 1724 } 1725 1726 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat *C) 1727 { 1728 PetscErrorCode ierr; 1729 Mat Cmpi,A_loc,POt,PDt; 1730 Mat_PtAPMPI *ptap; 1731 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1732 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*a=(Mat_MPIAIJ*)A->data,*c; 1733 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1734 PetscInt nnz; 1735 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1736 PetscInt am =A->rmap->n,pn=P->cmap->n; 1737 MPI_Comm comm; 1738 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1739 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1740 PetscInt len,proc,*dnz,*onz,*owners; 1741 PetscInt nzi,*bi,*bj; 1742 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1743 MPI_Request *swaits,*rwaits; 1744 MPI_Status *sstatus,rstatus; 1745 Mat_Merge_SeqsToMPI *merge; 1746 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1747 PetscReal afill =1.0,afill_tmp; 1748 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Armax; 1749 PetscScalar *vals; 1750 Mat_SeqAIJ *a_loc,*pdt,*pot; 1751 PetscTable ta; 1752 1753 PetscFunctionBegin; 1754 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1755 /* check if matrix local sizes are compatible */ 1756 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A (%D, %D) != P (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 1757 1758 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1759 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1760 1761 /* create struct Mat_PtAPMPI and attached it to C later */ 1762 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1763 1764 /* get A_loc by taking all local rows of A */ 1765 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1766 1767 ptap->A_loc = A_loc; 1768 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1769 ai = a_loc->i; 1770 aj = a_loc->j; 1771 1772 /* determine symbolic Co=(p->B)^T*A - send to others */ 1773 /*----------------------------------------------------*/ 1774 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1775 pdt = (Mat_SeqAIJ*)PDt->data; 1776 pdti = pdt->i; pdtj = pdt->j; 1777 1778 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1779 pot = (Mat_SeqAIJ*)POt->data; 1780 poti = pot->i; potj = pot->j; 1781 1782 /* then, compute symbolic Co = (p->B)^T*A */ 1783 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1784 >= (num of nonzero rows of C_seq) - pn */ 1785 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 1786 coi[0] = 0; 1787 1788 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1789 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(poti[pon],ai[am])); 1790 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1791 current_space = free_space; 1792 1793 /* create and initialize a linked list */ 1794 ierr = PetscTableCreate(A->cmap->n + a->B->cmap->N,aN,&ta);CHKERRQ(ierr); 1795 MatRowMergeMax_SeqAIJ(a_loc,am,ta); 1796 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1797 1798 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1799 1800 for (i=0; i<pon; i++) { 1801 pnz = poti[i+1] - poti[i]; 1802 ptJ = potj + poti[i]; 1803 for (j=0; j<pnz; j++) { 1804 row = ptJ[j]; /* row of A_loc == col of Pot */ 1805 anz = ai[row+1] - ai[row]; 1806 Jptr = aj + ai[row]; 1807 /* add non-zero cols of AP into the sorted linked list lnk */ 1808 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1809 } 1810 nnz = lnk[0]; 1811 1812 /* If free space is not available, double the total space in the list */ 1813 if (current_space->local_remaining<nnz) { 1814 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1815 nspacedouble++; 1816 } 1817 1818 /* Copy data into free space, and zero out denserows */ 1819 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1820 1821 current_space->array += nnz; 1822 current_space->local_used += nnz; 1823 current_space->local_remaining -= nnz; 1824 1825 coi[i+1] = coi[i] + nnz; 1826 } 1827 1828 ierr = PetscMalloc1(coi[pon]+1,&coj);CHKERRQ(ierr); 1829 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1830 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); /* must destroy to get a new one for C */ 1831 1832 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1833 if (afill_tmp > afill) afill = afill_tmp; 1834 1835 /* send j-array (coj) of Co to other processors */ 1836 /*----------------------------------------------*/ 1837 /* determine row ownership */ 1838 ierr = PetscNew(&merge);CHKERRQ(ierr); 1839 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1840 1841 merge->rowmap->n = pn; 1842 merge->rowmap->bs = 1; 1843 1844 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1845 owners = merge->rowmap->range; 1846 1847 /* determine the number of messages to send, their lengths */ 1848 ierr = PetscCalloc1(size,&len_si);CHKERRQ(ierr); 1849 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 1850 1851 len_s = merge->len_s; 1852 merge->nsend = 0; 1853 1854 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 1855 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1856 1857 proc = 0; 1858 for (i=0; i<pon; i++) { 1859 while (prmap[i] >= owners[proc+1]) proc++; 1860 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1861 len_s[proc] += coi[i+1] - coi[i]; 1862 } 1863 1864 len = 0; /* max length of buf_si[] */ 1865 owners_co[0] = 0; 1866 for (proc=0; proc<size; proc++) { 1867 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1868 if (len_si[proc]) { 1869 merge->nsend++; 1870 len_si[proc] = 2*(len_si[proc] + 1); 1871 len += len_si[proc]; 1872 } 1873 } 1874 1875 /* determine the number and length of messages to receive for coi and coj */ 1876 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1877 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1878 1879 /* post the Irecv and Isend of coj */ 1880 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1881 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1882 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 1883 for (proc=0, k=0; proc<size; proc++) { 1884 if (!len_s[proc]) continue; 1885 i = owners_co[proc]; 1886 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1887 k++; 1888 } 1889 1890 /* receives and sends of coj are complete */ 1891 ierr = PetscMalloc1(size,&sstatus);CHKERRQ(ierr); 1892 for (i=0; i<merge->nrecv; i++) { 1893 PetscMPIInt icompleted; 1894 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1895 } 1896 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1897 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1898 1899 /* add received column indices into table to update Armax */ 1900 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/examples/tutorials/ex56.c! */ 1901 for (k=0; k<merge->nrecv; k++) {/* k-th received message */ 1902 Jptr = buf_rj[k]; 1903 for (j=0; j<merge->len_r[k]; j++) { 1904 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1905 } 1906 } 1907 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1908 /* printf("Armax %d, an %d + Bn %d = %d, aN %d\n",Armax,A->cmap->n,a->B->cmap->N,A->cmap->n+a->B->cmap->N,aN); */ 1909 1910 /* send and recv coi */ 1911 /*-------------------*/ 1912 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1913 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1914 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1915 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1916 for (proc=0,k=0; proc<size; proc++) { 1917 if (!len_s[proc]) continue; 1918 /* form outgoing message for i-structure: 1919 buf_si[0]: nrows to be sent 1920 [1:nrows]: row index (global) 1921 [nrows+1:2*nrows+1]: i-structure index 1922 */ 1923 /*-------------------------------------------*/ 1924 nrows = len_si[proc]/2 - 1; 1925 buf_si_i = buf_si + nrows+1; 1926 buf_si[0] = nrows; 1927 buf_si_i[0] = 0; 1928 nrows = 0; 1929 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1930 nzi = coi[i+1] - coi[i]; 1931 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1932 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1933 nrows++; 1934 } 1935 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1936 k++; 1937 buf_si += len_si[proc]; 1938 } 1939 i = merge->nrecv; 1940 while (i--) { 1941 PetscMPIInt icompleted; 1942 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1943 } 1944 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1945 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1946 ierr = PetscFree(len_si);CHKERRQ(ierr); 1947 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1948 ierr = PetscFree(swaits);CHKERRQ(ierr); 1949 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1950 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1951 1952 /* compute the local portion of C (mpi mat) */ 1953 /*------------------------------------------*/ 1954 /* allocate bi array and free space for accumulating nonzero column info */ 1955 ierr = PetscMalloc1(pn+1,&bi);CHKERRQ(ierr); 1956 bi[0] = 0; 1957 1958 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1959 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(pdti[pn],PetscIntSumTruncate(poti[pon],ai[am]))); 1960 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1961 current_space = free_space; 1962 1963 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1964 for (k=0; k<merge->nrecv; k++) { 1965 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1966 nrows = *buf_ri_k[k]; 1967 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1968 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recieved i-structure */ 1969 } 1970 1971 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1972 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1973 rmax = 0; 1974 for (i=0; i<pn; i++) { 1975 /* add pdt[i,:]*AP into lnk */ 1976 pnz = pdti[i+1] - pdti[i]; 1977 ptJ = pdtj + pdti[i]; 1978 for (j=0; j<pnz; j++) { 1979 row = ptJ[j]; /* row of AP == col of Pt */ 1980 anz = ai[row+1] - ai[row]; 1981 Jptr = aj + ai[row]; 1982 /* add non-zero cols of AP into the sorted linked list lnk */ 1983 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1984 } 1985 1986 /* add received col data into lnk */ 1987 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1988 if (i == *nextrow[k]) { /* i-th row */ 1989 nzi = *(nextci[k]+1) - *nextci[k]; 1990 Jptr = buf_rj[k] + *nextci[k]; 1991 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 1992 nextrow[k]++; nextci[k]++; 1993 } 1994 } 1995 nnz = lnk[0]; 1996 1997 /* if free space is not available, make more free space */ 1998 if (current_space->local_remaining<nnz) { 1999 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 2000 nspacedouble++; 2001 } 2002 /* copy data into free space, then initialize lnk */ 2003 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 2004 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 2005 2006 current_space->array += nnz; 2007 current_space->local_used += nnz; 2008 current_space->local_remaining -= nnz; 2009 2010 bi[i+1] = bi[i] + nnz; 2011 if (nnz > rmax) rmax = nnz; 2012 } 2013 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 2014 2015 ierr = PetscMalloc1(bi[pn]+1,&bj);CHKERRQ(ierr); 2016 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 2017 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 2018 if (afill_tmp > afill) afill = afill_tmp; 2019 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 2020 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 2021 2022 ierr = MatDestroy(&POt);CHKERRQ(ierr); 2023 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 2024 2025 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 2026 /*----------------------------------------------------------------------------------*/ 2027 ierr = PetscCalloc1(rmax+1,&vals);CHKERRQ(ierr); 2028 2029 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 2030 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2031 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(A->cmap->bs));CHKERRQ(ierr); 2032 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 2033 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 2034 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 2035 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 2036 for (i=0; i<pn; i++) { 2037 row = i + rstart; 2038 nnz = bi[i+1] - bi[i]; 2039 Jptr = bj + bi[i]; 2040 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 2041 } 2042 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2043 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2044 ierr = PetscFree(vals);CHKERRQ(ierr); 2045 2046 merge->bi = bi; 2047 merge->bj = bj; 2048 merge->coi = coi; 2049 merge->coj = coj; 2050 merge->buf_ri = buf_ri; 2051 merge->buf_rj = buf_rj; 2052 merge->owners_co = owners_co; 2053 2054 /* attach the supporting struct to Cmpi for reuse */ 2055 c = (Mat_MPIAIJ*)Cmpi->data; 2056 2057 c->ptap = ptap; 2058 ptap->api = NULL; 2059 ptap->apj = NULL; 2060 ptap->merge = merge; 2061 ptap->apa = NULL; 2062 ptap->destroy = Cmpi->ops->destroy; 2063 ptap->duplicate = Cmpi->ops->duplicate; 2064 2065 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2066 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 2067 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 2068 2069 *C = Cmpi; 2070 #if defined(PETSC_USE_INFO) 2071 if (bi[pn] != 0) { 2072 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 2073 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 2074 } else { 2075 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 2076 } 2077 #endif 2078 PetscFunctionReturn(0); 2079 } 2080