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/vecscatterimpl.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 MatMatMult_MPIAIJ_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill, Mat *C) 19 { 20 PetscErrorCode ierr; 21 #if defined(PETSC_HAVE_HYPRE) 22 const char *algTypes[4] = {"scalable","nonscalable","seqmpi","hypre"}; 23 PetscInt nalg = 4; 24 #else 25 const char *algTypes[3] = {"scalable","nonscalable","seqmpi"}; 26 PetscInt nalg = 3; 27 #endif 28 PetscInt alg = 1; /* set nonscalable algorithm as default */ 29 MPI_Comm comm; 30 PetscBool flg; 31 32 PetscFunctionBegin; 33 if (scall == MAT_INITIAL_MATRIX) { 34 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 35 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); 36 37 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MatMatMult","Mat");CHKERRQ(ierr); 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 if (alg == 0 || alg == 1) { 79 Mat_MPIAIJ *c = (Mat_MPIAIJ*)(*C)->data; 80 Mat_APMPI *ap = c->ap; 81 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)(*C)),((PetscObject)(*C))->prefix,"MatFreeIntermediateDataStructures","Mat");CHKERRQ(ierr); 82 ap->freestruct = PETSC_FALSE; 83 ierr = PetscOptionsBool("-mat_freeintermediatedatastructures","Free intermediate data structures", "MatFreeIntermediateDataStructures",ap->freestruct,&ap->freestruct, NULL);CHKERRQ(ierr); 84 ierr = PetscOptionsEnd();CHKERRQ(ierr); 85 } 86 } 87 88 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 89 ierr = (*(*C)->ops->matmultnumeric)(A,B,*C);CHKERRQ(ierr); 90 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 91 PetscFunctionReturn(0); 92 } 93 94 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat A) 95 { 96 PetscErrorCode ierr; 97 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 98 Mat_APMPI *ptap = a->ap; 99 100 PetscFunctionBegin; 101 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 102 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 103 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 104 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 105 ierr = MatDestroy(&ptap->Pt);CHKERRQ(ierr); 106 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 107 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 108 ierr = PetscFree(ptap->apa);CHKERRQ(ierr); 109 ierr = ptap->destroy(A);CHKERRQ(ierr); 110 ierr = PetscFree(ptap);CHKERRQ(ierr); 111 PetscFunctionReturn(0); 112 } 113 114 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,Mat C) 115 { 116 PetscErrorCode ierr; 117 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 118 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 119 Mat_SeqAIJ *cd =(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 120 PetscScalar *cda=cd->a,*coa=co->a; 121 Mat_SeqAIJ *p_loc,*p_oth; 122 PetscScalar *apa,*ca; 123 PetscInt cm =C->rmap->n; 124 Mat_APMPI *ptap=c->ap; 125 PetscInt *api,*apj,*apJ,i,k; 126 PetscInt cstart=C->cmap->rstart; 127 PetscInt cdnz,conz,k0,k1; 128 MPI_Comm comm; 129 PetscMPIInt size; 130 131 PetscFunctionBegin; 132 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 133 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 134 135 if (!ptap) SETERRQ(comm,PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatFreeIntermediateDataStructures() or use '-mat_freeintermediatedatastructures'"); 136 137 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 138 /*-----------------------------------------------------*/ 139 /* update numerical values of P_oth and P_loc */ 140 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 141 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 142 143 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 144 /*----------------------------------------------------------*/ 145 /* get data from symbolic products */ 146 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 147 p_oth = NULL; 148 if (size >1) { 149 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 150 } 151 152 /* get apa for storing dense row A[i,:]*P */ 153 apa = ptap->apa; 154 155 api = ptap->api; 156 apj = ptap->apj; 157 for (i=0; i<cm; i++) { 158 /* compute apa = A[i,:]*P */ 159 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 160 161 /* set values in C */ 162 apJ = apj + api[i]; 163 cdnz = cd->i[i+1] - cd->i[i]; 164 conz = co->i[i+1] - co->i[i]; 165 166 /* 1st off-diagoanl part of C */ 167 ca = coa + co->i[i]; 168 k = 0; 169 for (k0=0; k0<conz; k0++) { 170 if (apJ[k] >= cstart) break; 171 ca[k0] = apa[apJ[k]]; 172 apa[apJ[k++]] = 0.0; 173 } 174 175 /* diagonal part of C */ 176 ca = cda + cd->i[i]; 177 for (k1=0; k1<cdnz; k1++) { 178 ca[k1] = apa[apJ[k]]; 179 apa[apJ[k++]] = 0.0; 180 } 181 182 /* 2nd off-diagoanl part of C */ 183 ca = coa + co->i[i]; 184 for (; k0<conz; k0++) { 185 ca[k0] = apa[apJ[k]]; 186 apa[apJ[k++]] = 0.0; 187 } 188 } 189 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 190 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 191 192 if (ptap->freestruct) { 193 ierr = MatFreeIntermediateDataStructures(C);CHKERRQ(ierr); 194 } 195 PetscFunctionReturn(0); 196 } 197 198 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A,Mat P,PetscReal fill,Mat *C) 199 { 200 PetscErrorCode ierr; 201 MPI_Comm comm; 202 PetscMPIInt size; 203 Mat Cmpi; 204 Mat_APMPI *ptap; 205 PetscFreeSpaceList free_space=NULL,current_space=NULL; 206 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*c; 207 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 208 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 209 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 210 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 211 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 212 PetscBT lnkbt; 213 PetscScalar *apa; 214 PetscReal afill; 215 MatType mtype; 216 217 PetscFunctionBegin; 218 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 219 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 220 221 /* create struct Mat_APMPI and attached it to C later */ 222 ierr = PetscNew(&ptap);CHKERRQ(ierr); 223 224 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 225 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 226 227 /* get P_loc by taking all local rows of P */ 228 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 229 230 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 231 pi_loc = p_loc->i; pj_loc = p_loc->j; 232 if (size > 1) { 233 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 234 pi_oth = p_oth->i; pj_oth = p_oth->j; 235 } else { 236 p_oth = NULL; 237 pi_oth = NULL; pj_oth = NULL; 238 } 239 240 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 241 /*-------------------------------------------------------------------*/ 242 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 243 ptap->api = api; 244 api[0] = 0; 245 246 /* create and initialize a linked list */ 247 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 248 249 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 250 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 251 current_space = free_space; 252 253 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 254 for (i=0; i<am; i++) { 255 /* diagonal portion of A */ 256 nzi = adi[i+1] - adi[i]; 257 for (j=0; j<nzi; j++) { 258 row = *adj++; 259 pnz = pi_loc[row+1] - pi_loc[row]; 260 Jptr = pj_loc + pi_loc[row]; 261 /* add non-zero cols of P into the sorted linked list lnk */ 262 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 263 } 264 /* off-diagonal portion of A */ 265 nzi = aoi[i+1] - aoi[i]; 266 for (j=0; j<nzi; j++) { 267 row = *aoj++; 268 pnz = pi_oth[row+1] - pi_oth[row]; 269 Jptr = pj_oth + pi_oth[row]; 270 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 271 } 272 273 apnz = lnk[0]; 274 api[i+1] = api[i] + apnz; 275 276 /* if free space is not available, double the total space in the list */ 277 if (current_space->local_remaining<apnz) { 278 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 279 nspacedouble++; 280 } 281 282 /* Copy data into free space, then initialize lnk */ 283 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 284 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 285 286 current_space->array += apnz; 287 current_space->local_used += apnz; 288 current_space->local_remaining -= apnz; 289 } 290 291 /* Allocate space for apj, initialize apj, and */ 292 /* destroy list of free space and other temporary array(s) */ 293 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 294 apj = ptap->apj; 295 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 296 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 297 298 /* malloc apa to store dense row A[i,:]*P */ 299 ierr = PetscCalloc1(pN,&apa);CHKERRQ(ierr); 300 301 ptap->apa = apa; 302 303 /* create and assemble symbolic parallel matrix Cmpi */ 304 /*----------------------------------------------------*/ 305 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 306 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 307 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 308 309 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 310 ierr = MatSetType(Cmpi,mtype);CHKERRQ(ierr); 311 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 312 313 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Cmpi, apj, api);CHKERRQ(ierr); 314 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 315 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 316 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 317 318 ptap->destroy = Cmpi->ops->destroy; 319 ptap->duplicate = Cmpi->ops->duplicate; 320 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 321 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 322 Cmpi->ops->freeintermediatedatastructures = MatFreeIntermediateDataStructures_MPIAIJ_AP; 323 324 /* attach the supporting struct to Cmpi for reuse */ 325 c = (Mat_MPIAIJ*)Cmpi->data; 326 c->ap = ptap; 327 328 *C = Cmpi; 329 330 /* set MatInfo */ 331 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 332 if (afill < 1.0) afill = 1.0; 333 Cmpi->info.mallocs = nspacedouble; 334 Cmpi->info.fill_ratio_given = fill; 335 Cmpi->info.fill_ratio_needed = afill; 336 337 #if defined(PETSC_USE_INFO) 338 if (api[am]) { 339 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 340 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 341 } else { 342 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 343 } 344 #endif 345 PetscFunctionReturn(0); 346 } 347 348 PETSC_INTERN PetscErrorCode MatMatMult_MPIAIJ_MPIDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 349 { 350 PetscErrorCode ierr; 351 352 PetscFunctionBegin; 353 if (scall == MAT_INITIAL_MATRIX) { 354 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 355 ierr = MatMatMultSymbolic_MPIAIJ_MPIDense(A,B,fill,C);CHKERRQ(ierr); 356 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 357 } 358 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 359 ierr = MatMatMultNumeric_MPIAIJ_MPIDense(A,B,*C);CHKERRQ(ierr); 360 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 361 PetscFunctionReturn(0); 362 } 363 364 typedef struct { 365 Mat workB; 366 PetscScalar *rvalues,*svalues; 367 MPI_Request *rwaits,*swaits; 368 } MPIAIJ_MPIDense; 369 370 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 371 { 372 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*) ctx; 373 PetscErrorCode ierr; 374 375 PetscFunctionBegin; 376 ierr = MatDestroy(&contents->workB);CHKERRQ(ierr); 377 ierr = PetscFree4(contents->rvalues,contents->svalues,contents->rwaits,contents->swaits);CHKERRQ(ierr); 378 ierr = PetscFree(contents);CHKERRQ(ierr); 379 PetscFunctionReturn(0); 380 } 381 382 /* 383 This is a "dummy function" that handles the case where matrix C was created as a dense matrix 384 directly by the user and passed to MatMatMult() with the MAT_REUSE_MATRIX option 385 386 It is the same as MatMatMultSymbolic_MPIAIJ_MPIDense() except does not create C 387 */ 388 PetscErrorCode MatMatMultNumeric_MPIDense(Mat A,Mat B,Mat C) 389 { 390 PetscErrorCode ierr; 391 PetscBool flg; 392 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 393 PetscInt nz = aij->B->cmap->n,to_n,to_entries,from_n,from_entries; 394 PetscContainer container; 395 MPIAIJ_MPIDense *contents; 396 VecScatter ctx = aij->Mvctx; 397 398 PetscFunctionBegin; 399 ierr = PetscObjectTypeCompare((PetscObject)B,MATMPIDENSE,&flg);CHKERRQ(ierr); 400 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Second matrix must be mpidense"); 401 402 /* Handle case where where user provided the final C matrix rather than calling MatMatMult() with MAT_INITIAL_MATRIX*/ 403 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 404 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"First matrix must be MPIAIJ"); 405 406 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 407 408 ierr = PetscNew(&contents);CHKERRQ(ierr); 409 /* Create work matrix used to store off processor rows of B needed for local product */ 410 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,NULL,&contents->workB);CHKERRQ(ierr); 411 /* Create work arrays needed */ 412 ierr = VecScatterGetRemoteCount_Private(ctx,PETSC_TRUE/*send*/,&to_n,&to_entries);CHKERRQ(ierr); 413 ierr = VecScatterGetRemoteCount_Private(ctx,PETSC_FALSE/*recv*/,&from_n,&from_entries);CHKERRQ(ierr); 414 ierr = PetscMalloc4(B->cmap->N*from_entries,&contents->rvalues,B->cmap->N*to_entries,&contents->svalues,from_n,&contents->rwaits,to_n,&contents->swaits);CHKERRQ(ierr); 415 416 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);CHKERRQ(ierr); 417 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 418 ierr = PetscContainerSetUserDestroy(container,MatMPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 419 ierr = PetscObjectCompose((PetscObject)C,"workB",(PetscObject)container);CHKERRQ(ierr); 420 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 421 422 ierr = (*C->ops->matmultnumeric)(A,B,C);CHKERRQ(ierr); 423 PetscFunctionReturn(0); 424 } 425 426 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C) 427 { 428 PetscErrorCode ierr; 429 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 430 PetscInt nz = aij->B->cmap->n,to_n,to_entries,from_n,from_entries; 431 PetscContainer container; 432 MPIAIJ_MPIDense *contents; 433 VecScatter ctx = aij->Mvctx; 434 PetscInt m = A->rmap->n,n=B->cmap->n; 435 436 PetscFunctionBegin; 437 ierr = MatCreate(PetscObjectComm((PetscObject)B),C);CHKERRQ(ierr); 438 ierr = MatSetSizes(*C,m,n,A->rmap->N,B->cmap->N);CHKERRQ(ierr); 439 ierr = MatSetBlockSizesFromMats(*C,A,B);CHKERRQ(ierr); 440 ierr = MatSetType(*C,MATMPIDENSE);CHKERRQ(ierr); 441 ierr = MatMPIDenseSetPreallocation(*C,NULL);CHKERRQ(ierr); 442 ierr = MatAssemblyBegin(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 443 ierr = MatAssemblyEnd(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 444 445 (*C)->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 446 447 ierr = PetscNew(&contents);CHKERRQ(ierr); 448 /* Create work matrix used to store off processor rows of B needed for local product */ 449 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,NULL,&contents->workB);CHKERRQ(ierr); 450 /* Create work arrays needed */ 451 ierr = VecScatterGetRemoteCount_Private(ctx,PETSC_TRUE/*send*/,&to_n,&to_entries);CHKERRQ(ierr); 452 ierr = VecScatterGetRemoteCount_Private(ctx,PETSC_FALSE/*recv*/,&from_n,&from_entries);CHKERRQ(ierr); 453 ierr = PetscMalloc4(B->cmap->N*from_entries,&contents->rvalues,B->cmap->N*to_entries,&contents->svalues,from_n,&contents->rwaits,to_n,&contents->swaits);CHKERRQ(ierr); 454 455 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);CHKERRQ(ierr); 456 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 457 ierr = PetscContainerSetUserDestroy(container,MatMPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 458 ierr = PetscObjectCompose((PetscObject)(*C),"workB",(PetscObject)container);CHKERRQ(ierr); 459 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 460 PetscFunctionReturn(0); 461 } 462 463 /* 464 Performs an efficient scatter on the rows of B needed by this process; this is 465 a modification of the VecScatterBegin_() routines. 466 */ 467 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,Mat C,Mat *outworkB) 468 { 469 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 470 PetscErrorCode ierr; 471 PetscScalar *b,*w,*svalues,*rvalues; 472 VecScatter ctx = aij->Mvctx; 473 PetscInt i,j,k; 474 const PetscInt *sindices,*sstarts,*rindices,*rstarts; 475 const PetscMPIInt *sprocs,*rprocs; 476 PetscInt nsends,nrecvs,nrecvs2; 477 MPI_Request *swaits,*rwaits; 478 MPI_Comm comm; 479 PetscMPIInt tag = ((PetscObject)ctx)->tag,ncols = B->cmap->N, nrows = aij->B->cmap->n,imdex,nrowsB = B->rmap->n,nsends_mpi,nrecvs_mpi; 480 MPI_Status status; 481 MPIAIJ_MPIDense *contents; 482 PetscContainer container; 483 Mat workB; 484 485 PetscFunctionBegin; 486 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 487 ierr = PetscObjectQuery((PetscObject)C,"workB",(PetscObject*)&container);CHKERRQ(ierr); 488 if (!container) SETERRQ(comm,PETSC_ERR_PLIB,"Container does not exist"); 489 ierr = PetscContainerGetPointer(container,(void**)&contents);CHKERRQ(ierr); 490 491 workB = *outworkB = contents->workB; 492 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); 493 ierr = VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL/*bs*/);CHKERRQ(ierr); 494 ierr = VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,&rindices,&rprocs,NULL/*bs*/);CHKERRQ(ierr); 495 ierr = PetscMPIIntCast(nsends,&nsends_mpi);CHKERRQ(ierr); 496 ierr = PetscMPIIntCast(nrecvs,&nrecvs_mpi);CHKERRQ(ierr); 497 svalues = contents->svalues; 498 rvalues = contents->rvalues; 499 swaits = contents->swaits; 500 rwaits = contents->rwaits; 501 502 ierr = MatDenseGetArray(B,&b);CHKERRQ(ierr); 503 ierr = MatDenseGetArray(workB,&w);CHKERRQ(ierr); 504 505 for (i=0; i<nrecvs; i++) { 506 ierr = MPI_Irecv(rvalues+ncols*(rstarts[i]-rstarts[0]),ncols*(rstarts[i+1]-rstarts[i]),MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 507 } 508 509 for (i=0; i<nsends; i++) { 510 /* pack a message at a time */ 511 for (j=0; j<sstarts[i+1]-sstarts[i]; j++) { 512 for (k=0; k<ncols; k++) { 513 svalues[ncols*(sstarts[i]-sstarts[0]+j) + k] = b[sindices[sstarts[i]+j] + nrowsB*k]; 514 } 515 } 516 ierr = MPI_Isend(svalues+ncols*(sstarts[i]-sstarts[0]),ncols*(sstarts[i+1]-sstarts[i]),MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 517 } 518 519 nrecvs2 = nrecvs; 520 while (nrecvs2) { 521 ierr = MPI_Waitany(nrecvs_mpi,rwaits,&imdex,&status);CHKERRQ(ierr); 522 nrecvs2--; 523 /* unpack a message at a time */ 524 for (j=0; j<rstarts[imdex+1]-rstarts[imdex]; j++) { 525 for (k=0; k<ncols; k++) { 526 w[rindices[rstarts[imdex]+j] + nrows*k] = rvalues[ncols*(rstarts[imdex]-rstarts[0]+j) + k]; 527 } 528 } 529 } 530 if (nsends) {ierr = MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);CHKERRQ(ierr);} 531 532 ierr = VecScatterRestoreRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&sindices,&sprocs,NULL/*bs*/);CHKERRQ(ierr); 533 ierr = VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,&rindices,&rprocs,NULL/*bs*/);CHKERRQ(ierr); 534 ierr = MatDenseRestoreArray(B,&b);CHKERRQ(ierr); 535 ierr = MatDenseRestoreArray(workB,&w);CHKERRQ(ierr); 536 ierr = MatAssemblyBegin(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 537 ierr = MatAssemblyEnd(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 538 PetscFunctionReturn(0); 539 540 } 541 extern PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat); 542 543 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 544 { 545 PetscErrorCode ierr; 546 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 547 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 548 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 549 Mat workB; 550 551 PetscFunctionBegin; 552 /* diagonal block of A times all local rows of B*/ 553 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(aij->A,bdense->A,cdense->A);CHKERRQ(ierr); 554 555 /* get off processor parts of B needed to complete the product */ 556 ierr = MatMPIDenseScatter(A,B,C,&workB);CHKERRQ(ierr); 557 558 /* off-diagonal block of A times nonlocal rows of B */ 559 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A);CHKERRQ(ierr); 560 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 561 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 562 PetscFunctionReturn(0); 563 } 564 565 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 566 { 567 PetscErrorCode ierr; 568 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 569 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 570 Mat_SeqAIJ *cd = (Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 571 PetscInt *adi = ad->i,*adj,*aoi=ao->i,*aoj; 572 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 573 Mat_SeqAIJ *p_loc,*p_oth; 574 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 575 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 576 PetscInt cm = C->rmap->n,anz,pnz; 577 Mat_APMPI *ptap = c->ap; 578 PetscScalar *apa_sparse; 579 PetscInt *api,*apj,*apJ,i,j,k,row; 580 PetscInt cstart = C->cmap->rstart; 581 PetscInt cdnz,conz,k0,k1,nextp; 582 MPI_Comm comm; 583 PetscMPIInt size; 584 585 PetscFunctionBegin; 586 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 587 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 588 589 if (!ptap) { 590 SETERRQ(comm,PETSC_ERR_ARG_WRONGSTATE,"AP cannot be reused. Do not call MatFreeIntermediateDataStructures() or use '-mat_freeintermediatedatastructures'"); 591 } 592 apa_sparse = ptap->apa; 593 594 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 595 /*-----------------------------------------------------*/ 596 /* update numerical values of P_oth and P_loc */ 597 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 598 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 599 600 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 601 /*----------------------------------------------------------*/ 602 /* get data from symbolic products */ 603 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 604 pi_loc = p_loc->i; pj_loc = p_loc->j; pa_loc = p_loc->a; 605 if (size >1) { 606 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 607 pi_oth = p_oth->i; pj_oth = p_oth->j; pa_oth = p_oth->a; 608 } else { 609 p_oth = NULL; pi_oth = NULL; pj_oth = NULL; pa_oth = NULL; 610 } 611 612 api = ptap->api; 613 apj = ptap->apj; 614 for (i=0; i<cm; i++) { 615 apJ = apj + api[i]; 616 617 /* diagonal portion of A */ 618 anz = adi[i+1] - adi[i]; 619 adj = ad->j + adi[i]; 620 ada = ad->a + adi[i]; 621 for (j=0; j<anz; j++) { 622 row = adj[j]; 623 pnz = pi_loc[row+1] - pi_loc[row]; 624 pj = pj_loc + pi_loc[row]; 625 pa = pa_loc + pi_loc[row]; 626 /* perform sparse axpy */ 627 valtmp = ada[j]; 628 nextp = 0; 629 for (k=0; nextp<pnz; k++) { 630 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 631 apa_sparse[k] += valtmp*pa[nextp++]; 632 } 633 } 634 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 635 } 636 637 /* off-diagonal portion of A */ 638 anz = aoi[i+1] - aoi[i]; 639 aoj = ao->j + aoi[i]; 640 aoa = ao->a + aoi[i]; 641 for (j=0; j<anz; j++) { 642 row = aoj[j]; 643 pnz = pi_oth[row+1] - pi_oth[row]; 644 pj = pj_oth + pi_oth[row]; 645 pa = pa_oth + pi_oth[row]; 646 /* perform sparse axpy */ 647 valtmp = aoa[j]; 648 nextp = 0; 649 for (k=0; nextp<pnz; k++) { 650 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 651 apa_sparse[k] += valtmp*pa[nextp++]; 652 } 653 } 654 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 655 } 656 657 /* set values in C */ 658 cdnz = cd->i[i+1] - cd->i[i]; 659 conz = co->i[i+1] - co->i[i]; 660 661 /* 1st off-diagoanl part of C */ 662 ca = coa + co->i[i]; 663 k = 0; 664 for (k0=0; k0<conz; k0++) { 665 if (apJ[k] >= cstart) break; 666 ca[k0] = apa_sparse[k]; 667 apa_sparse[k] = 0.0; 668 k++; 669 } 670 671 /* diagonal part of C */ 672 ca = cda + cd->i[i]; 673 for (k1=0; k1<cdnz; k1++) { 674 ca[k1] = apa_sparse[k]; 675 apa_sparse[k] = 0.0; 676 k++; 677 } 678 679 /* 2nd off-diagoanl part of C */ 680 ca = coa + co->i[i]; 681 for (; k0<conz; k0++) { 682 ca[k0] = apa_sparse[k]; 683 apa_sparse[k] = 0.0; 684 k++; 685 } 686 } 687 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 688 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 689 690 if (ptap->freestruct) { 691 ierr = MatFreeIntermediateDataStructures(C);CHKERRQ(ierr); 692 } 693 PetscFunctionReturn(0); 694 } 695 696 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 697 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 698 { 699 PetscErrorCode ierr; 700 MPI_Comm comm; 701 PetscMPIInt size; 702 Mat Cmpi; 703 Mat_APMPI *ptap; 704 PetscFreeSpaceList free_space = NULL,current_space=NULL; 705 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data,*c; 706 Mat_SeqAIJ *ad = (Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 707 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 708 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 709 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=0; 710 PetscInt am=A->rmap->n,pn=P->cmap->n,pm=P->rmap->n,lsize=pn+20; 711 PetscReal afill; 712 PetscScalar *apa; 713 MatType mtype; 714 715 PetscFunctionBegin; 716 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 717 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 718 719 /* create struct Mat_APMPI and attached it to C later */ 720 ierr = PetscNew(&ptap);CHKERRQ(ierr); 721 722 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 723 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 724 725 /* get P_loc by taking all local rows of P */ 726 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 727 728 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 729 pi_loc = p_loc->i; pj_loc = p_loc->j; 730 if (size > 1) { 731 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 732 pi_oth = p_oth->i; pj_oth = p_oth->j; 733 } else { 734 p_oth = NULL; 735 pi_oth = NULL; pj_oth = NULL; 736 } 737 738 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 739 /*-------------------------------------------------------------------*/ 740 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 741 ptap->api = api; 742 api[0] = 0; 743 744 ierr = PetscLLCondensedCreate_Scalable(lsize,&lnk);CHKERRQ(ierr); 745 746 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 747 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 748 current_space = free_space; 749 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 750 for (i=0; i<am; i++) { 751 /* diagonal portion of A */ 752 nzi = adi[i+1] - adi[i]; 753 for (j=0; j<nzi; j++) { 754 row = *adj++; 755 pnz = pi_loc[row+1] - pi_loc[row]; 756 Jptr = pj_loc + pi_loc[row]; 757 /* Expand list if it is not long enough */ 758 if (pnz+apnz_max > lsize) { 759 lsize = pnz+apnz_max; 760 ierr = PetscLLCondensedExpand_Scalable(lsize, &lnk);CHKERRQ(ierr); 761 } 762 /* add non-zero cols of P into the sorted linked list lnk */ 763 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 764 apnz = *lnk; /* The first element in the list is the number of items in the list */ 765 api[i+1] = api[i] + apnz; 766 if (apnz > apnz_max) apnz_max = apnz; 767 } 768 /* off-diagonal portion of A */ 769 nzi = aoi[i+1] - aoi[i]; 770 for (j=0; j<nzi; j++) { 771 row = *aoj++; 772 pnz = pi_oth[row+1] - pi_oth[row]; 773 Jptr = pj_oth + pi_oth[row]; 774 /* Expand list if it is not long enough */ 775 if (pnz+apnz_max > lsize) { 776 lsize = pnz + apnz_max; 777 ierr = PetscLLCondensedExpand_Scalable(lsize, &lnk);CHKERRQ(ierr); 778 } 779 /* add non-zero cols of P into the sorted linked list lnk */ 780 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 781 apnz = *lnk; /* The first element in the list is the number of items in the list */ 782 api[i+1] = api[i] + apnz; 783 if (apnz > apnz_max) apnz_max = apnz; 784 } 785 apnz = *lnk; 786 api[i+1] = api[i] + apnz; 787 if (apnz > apnz_max) apnz_max = apnz; 788 789 /* if free space is not available, double the total space in the list */ 790 if (current_space->local_remaining<apnz) { 791 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 792 nspacedouble++; 793 } 794 795 /* Copy data into free space, then initialize lnk */ 796 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 797 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 798 799 current_space->array += apnz; 800 current_space->local_used += apnz; 801 current_space->local_remaining -= apnz; 802 } 803 804 /* Allocate space for apj, initialize apj, and */ 805 /* destroy list of free space and other temporary array(s) */ 806 ierr = PetscMalloc1(api[am]+1,&ptap->apj);CHKERRQ(ierr); 807 apj = ptap->apj; 808 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 809 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 810 811 /* create and assemble symbolic parallel matrix Cmpi */ 812 /*----------------------------------------------------*/ 813 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 814 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 815 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 816 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 817 ierr = MatSetType(Cmpi,mtype);CHKERRQ(ierr); 818 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 819 820 /* malloc apa for assembly Cmpi */ 821 ierr = PetscCalloc1(apnz_max,&apa);CHKERRQ(ierr); 822 ptap->apa = apa; 823 824 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Cmpi, apj, api);CHKERRQ(ierr); 825 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 826 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 827 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 828 829 ptap->destroy = Cmpi->ops->destroy; 830 ptap->duplicate = Cmpi->ops->duplicate; 831 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 832 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 833 Cmpi->ops->freeintermediatedatastructures = MatFreeIntermediateDataStructures_MPIAIJ_AP; 834 835 /* attach the supporting struct to Cmpi for reuse */ 836 c = (Mat_MPIAIJ*)Cmpi->data; 837 c->ap = ptap; 838 *C = Cmpi; 839 840 /* set MatInfo */ 841 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 842 if (afill < 1.0) afill = 1.0; 843 Cmpi->info.mallocs = nspacedouble; 844 Cmpi->info.fill_ratio_given = fill; 845 Cmpi->info.fill_ratio_needed = afill; 846 847 #if defined(PETSC_USE_INFO) 848 if (api[am]) { 849 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 850 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 851 } else { 852 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 853 } 854 #endif 855 PetscFunctionReturn(0); 856 } 857 858 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 859 /* Three input arrays are merged to one output array. The size of the */ 860 /* output array is also output. Duplicate entries only show up once. */ 861 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, 862 PetscInt size2, PetscInt *in2, 863 PetscInt size3, PetscInt *in3, 864 PetscInt *size4, PetscInt *out) 865 { 866 int i = 0, j = 0, k = 0, l = 0; 867 868 /* Traverse all three arrays */ 869 while (i<size1 && j<size2 && k<size3) { 870 if (in1[i] < in2[j] && in1[i] < in3[k]) { 871 out[l++] = in1[i++]; 872 } 873 else if(in2[j] < in1[i] && in2[j] < in3[k]) { 874 out[l++] = in2[j++]; 875 } 876 else if(in3[k] < in1[i] && in3[k] < in2[j]) { 877 out[l++] = in3[k++]; 878 } 879 else if(in1[i] == in2[j] && in1[i] < in3[k]) { 880 out[l++] = in1[i]; 881 i++, j++; 882 } 883 else if(in1[i] == in3[k] && in1[i] < in2[j]) { 884 out[l++] = in1[i]; 885 i++, k++; 886 } 887 else if(in3[k] == in2[j] && in2[j] < in1[i]) { 888 out[l++] = in2[j]; 889 k++, j++; 890 } 891 else if(in1[i] == in2[j] && in1[i] == in3[k]) { 892 out[l++] = in1[i]; 893 i++, j++, k++; 894 } 895 } 896 897 /* Traverse two remaining arrays */ 898 while (i<size1 && j<size2) { 899 if (in1[i] < in2[j]) { 900 out[l++] = in1[i++]; 901 } 902 else if(in1[i] > in2[j]) { 903 out[l++] = in2[j++]; 904 } 905 else { 906 out[l++] = in1[i]; 907 i++, j++; 908 } 909 } 910 911 while (i<size1 && k<size3) { 912 if (in1[i] < in3[k]) { 913 out[l++] = in1[i++]; 914 } 915 else if(in1[i] > in3[k]) { 916 out[l++] = in3[k++]; 917 } 918 else { 919 out[l++] = in1[i]; 920 i++, k++; 921 } 922 } 923 924 while (k<size3 && j<size2) { 925 if (in3[k] < in2[j]) { 926 out[l++] = in3[k++]; 927 } 928 else if(in3[k] > in2[j]) { 929 out[l++] = in2[j++]; 930 } 931 else { 932 out[l++] = in3[k]; 933 k++, j++; 934 } 935 } 936 937 /* Traverse one remaining array */ 938 while (i<size1) out[l++] = in1[i++]; 939 while (j<size2) out[l++] = in2[j++]; 940 while (k<size3) out[l++] = in3[k++]; 941 942 *size4 = l; 943 } 944 945 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 946 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 947 /* matrix-matrix multiplications. */ 948 #undef __FUNCT__ 949 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI" 950 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat *C) 951 { 952 PetscErrorCode ierr; 953 MPI_Comm comm; 954 PetscMPIInt size; 955 Mat Cmpi; 956 Mat_APMPI *ptap; 957 PetscFreeSpaceList free_space_diag=NULL, current_space=NULL; 958 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 959 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc; 960 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data; 961 Mat_MPIAIJ *c; 962 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 963 PetscInt adponz, adpdnz; 964 PetscInt *pi_loc,*dnz,*onz; 965 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,rstart=A->rmap->rstart; 966 PetscInt *lnk,i, i1=0,pnz,row,*adpoi,*adpoj, *api, *adpoJ, *aopJ, *apJ,*Jptr, aopnz, nspacedouble=0,j,nzi, 967 *apj,apnz, *adpdi, *adpdj, *adpdJ, *poff_i, *poff_j, *j_temp, *aopothi, *aopothj; 968 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n, p_colstart, p_colend; 969 PetscBT lnkbt; 970 PetscScalar *apa; 971 PetscReal afill; 972 PetscMPIInt rank; 973 Mat adpd, aopoth; 974 MatType mtype; 975 976 PetscFunctionBegin; 977 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 978 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 979 ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr); 980 ierr = MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend); CHKERRQ(ierr); 981 982 /* create struct Mat_APMPI and attached it to C later */ 983 ierr = PetscNew(&ptap);CHKERRQ(ierr); 984 985 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 986 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 987 988 /* get P_loc by taking all local rows of P */ 989 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 990 991 992 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 993 pi_loc = p_loc->i; 994 995 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 996 ierr = PetscMalloc1(am+2,&api);CHKERRQ(ierr); 997 ierr = PetscMalloc1(am+2,&adpoi);CHKERRQ(ierr); 998 999 adpoi[0] = 0; 1000 ptap->api = api; 1001 api[0] = 0; 1002 1003 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1004 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 1005 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 1006 1007 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1008 ierr = MatSetOptionsPrefix(a->A,"inner_diag_");CHKERRQ(ierr); 1009 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->A, p->A, fill, &adpd);CHKERRQ(ierr); 1010 adpd_seq = (Mat_SeqAIJ*)((adpd)->data); 1011 adpdi = adpd_seq->i; adpdj = adpd_seq->j; 1012 p_off = (Mat_SeqAIJ*)((p->B)->data); 1013 poff_i = p_off->i; poff_j = p_off->j; 1014 1015 /* j_temp stores indices of a result row before they are added to the linked list */ 1016 ierr = PetscMalloc1(pN+2,&j_temp);CHKERRQ(ierr); 1017 1018 1019 /* Symbolic calc of the A_diag * p_loc_off */ 1020 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1021 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space_diag);CHKERRQ(ierr); 1022 current_space = free_space_diag; 1023 1024 for (i=0; i<am; i++) { 1025 /* A_diag * P_loc_off */ 1026 nzi = adi[i+1] - adi[i]; 1027 for (j=0; j<nzi; j++) { 1028 row = *adj++; 1029 pnz = poff_i[row+1] - poff_i[row]; 1030 Jptr = poff_j + poff_i[row]; 1031 for(i1 = 0; i1 < pnz; i1++) { 1032 j_temp[i1] = p->garray[Jptr[i1]]; 1033 } 1034 /* add non-zero cols of P into the sorted linked list lnk */ 1035 ierr = PetscLLCondensedAddSorted(pnz,j_temp,lnk,lnkbt);CHKERRQ(ierr); 1036 } 1037 1038 adponz = lnk[0]; 1039 adpoi[i+1] = adpoi[i] + adponz; 1040 1041 /* if free space is not available, double the total space in the list */ 1042 if (current_space->local_remaining<adponz) { 1043 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(adponz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1044 nspacedouble++; 1045 } 1046 1047 /* Copy data into free space, then initialize lnk */ 1048 ierr = PetscLLCondensedClean(pN,adponz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1049 1050 current_space->array += adponz; 1051 current_space->local_used += adponz; 1052 current_space->local_remaining -= adponz; 1053 } 1054 1055 /* Symbolic calc of A_off * P_oth */ 1056 ierr = MatSetOptionsPrefix(a->B,"inner_offdiag_");CHKERRQ(ierr); 1057 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, &aopoth);CHKERRQ(ierr); 1058 aopoth_seq = (Mat_SeqAIJ*)((aopoth)->data); 1059 aopothi = aopoth_seq->i; aopothj = aopoth_seq->j; 1060 1061 /* Allocate space for apj, adpj, aopj, ... */ 1062 /* destroy lists of free space and other temporary array(s) */ 1063 1064 ierr = PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am]+2, &ptap->apj);CHKERRQ(ierr); 1065 ierr = PetscMalloc1(adpoi[am]+2, &adpoj);CHKERRQ(ierr); 1066 1067 /* Copy from linked list to j-array */ 1068 ierr = PetscFreeSpaceContiguous(&free_space_diag,adpoj);CHKERRQ(ierr); 1069 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1070 1071 adpoJ = adpoj; 1072 adpdJ = adpdj; 1073 aopJ = aopothj; 1074 apj = ptap->apj; 1075 apJ = apj; /* still empty */ 1076 1077 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1078 /* A_diag * P_loc_diag to get A*P */ 1079 for (i = 0; i < am; i++) { 1080 aopnz = aopothi[i+1] - aopothi[i]; 1081 adponz = adpoi[i+1] - adpoi[i]; 1082 adpdnz = adpdi[i+1] - adpdi[i]; 1083 1084 /* Correct indices from A_diag*P_diag */ 1085 for(i1 = 0; i1 < adpdnz; i1++) { 1086 adpdJ[i1] += p_colstart; 1087 } 1088 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1089 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1090 ierr = MatPreallocateSet(i+rstart, apnz, apJ, dnz, onz); CHKERRQ(ierr); 1091 1092 aopJ += aopnz; 1093 adpoJ += adponz; 1094 adpdJ += adpdnz; 1095 apJ += apnz; 1096 api[i+1] = api[i] + apnz; 1097 } 1098 1099 /* malloc apa to store dense row A[i,:]*P */ 1100 ierr = PetscCalloc1(pN+2,&apa);CHKERRQ(ierr); 1101 1102 ptap->apa = apa; 1103 /* create and assemble symbolic parallel matrix Cmpi */ 1104 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1105 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1106 ierr = MatSetBlockSizesFromMats(Cmpi,A,P);CHKERRQ(ierr); 1107 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 1108 ierr = MatSetType(Cmpi,mtype);CHKERRQ(ierr); 1109 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1110 1111 1112 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Cmpi, apj, api);CHKERRQ(ierr); 1113 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1114 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1115 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1116 1117 1118 ptap->destroy = Cmpi->ops->destroy; 1119 ptap->duplicate = Cmpi->ops->duplicate; 1120 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1121 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 1122 1123 /* attach the supporting struct to Cmpi for reuse */ 1124 c = (Mat_MPIAIJ*)Cmpi->data; 1125 c->ap = ptap; 1126 *C = Cmpi; 1127 1128 /* set MatInfo */ 1129 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 1130 if (afill < 1.0) afill = 1.0; 1131 Cmpi->info.mallocs = nspacedouble; 1132 Cmpi->info.fill_ratio_given = fill; 1133 Cmpi->info.fill_ratio_needed = afill; 1134 1135 #if defined(PETSC_USE_INFO) 1136 if (api[am]) { 1137 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 1138 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);CHKERRQ(ierr); 1139 } else { 1140 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1141 } 1142 #endif 1143 1144 ierr = MatDestroy(&aopoth);CHKERRQ(ierr); 1145 ierr = MatDestroy(&adpd);CHKERRQ(ierr); 1146 ierr = PetscFree(j_temp);CHKERRQ(ierr); 1147 ierr = PetscFree(adpoj);CHKERRQ(ierr); 1148 ierr = PetscFree(adpoi);CHKERRQ(ierr); 1149 PetscFunctionReturn(0); 1150 } 1151 1152 1153 /*-------------------------------------------------------------------------*/ 1154 PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat P,Mat A,MatReuse scall,PetscReal fill,Mat *C) 1155 { 1156 PetscErrorCode ierr; 1157 const char *algTypes[3] = {"scalable","nonscalable","matmatmult"}; 1158 PetscInt aN=A->cmap->N,alg=1; /* set default algorithm */ 1159 PetscBool flg; 1160 1161 PetscFunctionBegin; 1162 if (scall == MAT_INITIAL_MATRIX) { 1163 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MatTransposeMatMult","Mat");CHKERRQ(ierr); 1164 ierr = PetscOptionsEList("-mattransposematmult_via","Algorithmic approach","MatTransposeMatMult",algTypes,3,algTypes[1],&alg,&flg);CHKERRQ(ierr); 1165 ierr = PetscOptionsEnd();CHKERRQ(ierr); 1166 1167 ierr = PetscLogEventBegin(MAT_TransposeMatMultSymbolic,P,A,0,0);CHKERRQ(ierr); 1168 switch (alg) { 1169 case 1: 1170 if (!flg && aN > 100000) { /* may switch to scalable algorithm as default */ 1171 MatInfo Ainfo,Pinfo; 1172 PetscInt nz_local; 1173 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 1174 MPI_Comm comm; 1175 1176 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 1177 ierr = MatGetInfo(P,MAT_LOCAL,&Pinfo);CHKERRQ(ierr); 1178 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); /* estimated local nonzero entries */ 1179 1180 if (aN > fill*nz_local) alg_scalable_loc = PETSC_TRUE; 1181 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1182 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 1183 1184 if (alg_scalable) { 1185 alg = 0; /* scalable algorithm would slower than nonscalable algorithm */ 1186 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 1187 break; 1188 } 1189 } 1190 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(P,A,fill,C);CHKERRQ(ierr); 1191 break; 1192 case 2: 1193 { 1194 Mat Pt; 1195 Mat_APMPI *ptap; 1196 Mat_MPIAIJ *c; 1197 ierr = MatTranspose(P,MAT_INITIAL_MATRIX,&Pt);CHKERRQ(ierr); 1198 ierr = MatMatMult(Pt,A,MAT_INITIAL_MATRIX,fill,C);CHKERRQ(ierr); 1199 c = (Mat_MPIAIJ*)(*C)->data; 1200 ptap = c->ap; 1201 if (ptap) { 1202 ptap->Pt = Pt; 1203 (*C)->ops->freeintermediatedatastructures = MatFreeIntermediateDataStructures_MPIAIJ_AP; 1204 } 1205 (*C)->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 1206 PetscFunctionReturn(0); 1207 } 1208 break; 1209 default: /* scalable algorithm */ 1210 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 1211 break; 1212 } 1213 ierr = PetscLogEventEnd(MAT_TransposeMatMultSymbolic,P,A,0,0);CHKERRQ(ierr); 1214 1215 { 1216 Mat_MPIAIJ *c = (Mat_MPIAIJ*)(*C)->data; 1217 Mat_APMPI *ap = c->ap; 1218 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)(*C)),((PetscObject)(*C))->prefix,"MatFreeIntermediateDataStructures","Mat");CHKERRQ(ierr); 1219 ap->freestruct = PETSC_FALSE; 1220 ierr = PetscOptionsBool("-mat_freeintermediatedatastructures","Free intermediate data structures", "MatFreeIntermediateDataStructures",ap->freestruct,&ap->freestruct, NULL);CHKERRQ(ierr); 1221 ierr = PetscOptionsEnd();CHKERRQ(ierr); 1222 } 1223 } 1224 1225 ierr = PetscLogEventBegin(MAT_TransposeMatMultNumeric,P,A,0,0);CHKERRQ(ierr); 1226 ierr = (*(*C)->ops->mattransposemultnumeric)(P,A,*C);CHKERRQ(ierr); 1227 ierr = PetscLogEventEnd(MAT_TransposeMatMultNumeric,P,A,0,0);CHKERRQ(ierr); 1228 PetscFunctionReturn(0); 1229 } 1230 1231 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1232 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P,Mat A,Mat C) 1233 { 1234 PetscErrorCode ierr; 1235 Mat_MPIAIJ *c=(Mat_MPIAIJ*)C->data; 1236 Mat_APMPI *ptap= c->ap; 1237 Mat Pt; 1238 1239 PetscFunctionBegin; 1240 if (!ptap) { 1241 MPI_Comm comm; 1242 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1243 SETERRQ(comm,PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatFreeIntermediateDataStructures() or use '-mat_freeintermediatedatastructures'"); 1244 } 1245 1246 Pt=ptap->Pt; 1247 ierr = MatTranspose(P,MAT_REUSE_MATRIX,&Pt);CHKERRQ(ierr); 1248 ierr = MatMatMultNumeric(Pt,A,C);CHKERRQ(ierr); 1249 1250 /* supporting struct ptap consumes almost same amount of memory as C=PtAP, release it if C will not be updated by A and P */ 1251 if (ptap->freestruct) { 1252 ierr = MatFreeIntermediateDataStructures(C);CHKERRQ(ierr); 1253 } 1254 PetscFunctionReturn(0); 1255 } 1256 1257 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1258 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,PetscReal fill,Mat *C) 1259 { 1260 PetscErrorCode ierr; 1261 Mat_APMPI *ptap; 1262 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c; 1263 MPI_Comm comm; 1264 PetscMPIInt size,rank; 1265 Mat Cmpi; 1266 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1267 PetscInt pn=P->cmap->n,aN=A->cmap->N,an=A->cmap->n; 1268 PetscInt *lnk,i,k,nsend; 1269 PetscBT lnkbt; 1270 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 1271 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1272 PetscInt len,proc,*dnz,*onz,*owners,nzi; 1273 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1274 MPI_Request *swaits,*rwaits; 1275 MPI_Status *sstatus,rstatus; 1276 PetscLayout rowmap; 1277 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1278 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 1279 PetscInt *Jptr,*prmap=p->garray,con,j,Crmax; 1280 Mat_SeqAIJ *a_loc,*c_loc,*c_oth; 1281 PetscTable ta; 1282 MatType mtype; 1283 1284 PetscFunctionBegin; 1285 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1286 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1287 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1288 1289 /* create symbolic parallel matrix Cmpi */ 1290 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1291 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 1292 ierr = MatSetType(Cmpi,mtype);CHKERRQ(ierr); 1293 1294 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1295 1296 /* create struct Mat_APMPI and attached it to C later */ 1297 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1298 ptap->reuse = MAT_INITIAL_MATRIX; 1299 1300 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1301 /* --------------------------------- */ 1302 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1303 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1304 1305 /* (1) compute symbolic A_loc */ 1306 /* ---------------------------*/ 1307 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1308 1309 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1310 /* ------------------------------------ */ 1311 ierr = MatSetOptionsPrefix(ptap->Ro, "inner_offdiag_");CHKERRQ(ierr); 1312 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->A_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 1313 1314 /* (3) send coj of C_oth to other processors */ 1315 /* ------------------------------------------ */ 1316 /* determine row ownership */ 1317 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 1318 rowmap->n = pn; 1319 rowmap->bs = 1; 1320 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 1321 owners = rowmap->range; 1322 1323 /* determine the number of messages to send, their lengths */ 1324 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 1325 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1326 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1327 1328 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 1329 coi = c_oth->i; coj = c_oth->j; 1330 con = ptap->C_oth->rmap->n; 1331 proc = 0; 1332 for (i=0; i<con; i++) { 1333 while (prmap[i] >= owners[proc+1]) proc++; 1334 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1335 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1336 } 1337 1338 len = 0; /* max length of buf_si[], see (4) */ 1339 owners_co[0] = 0; 1340 nsend = 0; 1341 for (proc=0; proc<size; proc++) { 1342 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1343 if (len_s[proc]) { 1344 nsend++; 1345 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1346 len += len_si[proc]; 1347 } 1348 } 1349 1350 /* determine the number and length of messages to receive for coi and coj */ 1351 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 1352 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 1353 1354 /* post the Irecv and Isend of coj */ 1355 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1356 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1357 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 1358 for (proc=0, k=0; proc<size; proc++) { 1359 if (!len_s[proc]) continue; 1360 i = owners_co[proc]; 1361 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1362 k++; 1363 } 1364 1365 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1366 /* ---------------------------------------- */ 1367 ierr = MatSetOptionsPrefix(ptap->Rd,"inner_diag_");CHKERRQ(ierr); 1368 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->A_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 1369 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 1370 1371 /* receives coj are complete */ 1372 for (i=0; i<nrecv; i++) { 1373 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1374 } 1375 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1376 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 1377 1378 /* add received column indices into ta to update Crmax */ 1379 a_loc = (Mat_SeqAIJ*)(ptap->A_loc)->data; 1380 1381 /* create and initialize a linked list */ 1382 ierr = PetscTableCreate(an,aN,&ta);CHKERRQ(ierr); /* for compute Crmax */ 1383 MatRowMergeMax_SeqAIJ(a_loc,ptap->A_loc->rmap->N,ta); 1384 1385 for (k=0; k<nrecv; k++) {/* k-th received message */ 1386 Jptr = buf_rj[k]; 1387 for (j=0; j<len_r[k]; j++) { 1388 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1389 } 1390 } 1391 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 1392 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 1393 1394 /* (4) send and recv coi */ 1395 /*-----------------------*/ 1396 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1397 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1398 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1399 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1400 for (proc=0,k=0; proc<size; proc++) { 1401 if (!len_s[proc]) continue; 1402 /* form outgoing message for i-structure: 1403 buf_si[0]: nrows to be sent 1404 [1:nrows]: row index (global) 1405 [nrows+1:2*nrows+1]: i-structure index 1406 */ 1407 /*-------------------------------------------*/ 1408 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 1409 buf_si_i = buf_si + nrows+1; 1410 buf_si[0] = nrows; 1411 buf_si_i[0] = 0; 1412 nrows = 0; 1413 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1414 nzi = coi[i+1] - coi[i]; 1415 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1416 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1417 nrows++; 1418 } 1419 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1420 k++; 1421 buf_si += len_si[proc]; 1422 } 1423 for (i=0; i<nrecv; i++) { 1424 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1425 } 1426 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1427 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 1428 1429 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 1430 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1431 ierr = PetscFree(swaits);CHKERRQ(ierr); 1432 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1433 1434 /* (5) compute the local portion of Cmpi */ 1435 /* ------------------------------------------ */ 1436 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of Cmpi */ 1437 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 1438 current_space = free_space; 1439 1440 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 1441 for (k=0; k<nrecv; k++) { 1442 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1443 nrows = *buf_ri_k[k]; 1444 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1445 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1446 } 1447 1448 ierr = MatPreallocateInitialize(comm,pn,an,dnz,onz);CHKERRQ(ierr); 1449 ierr = PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt);CHKERRQ(ierr); 1450 for (i=0; i<pn; i++) { 1451 /* add C_loc into Cmpi */ 1452 nzi = c_loc->i[i+1] - c_loc->i[i]; 1453 Jptr = c_loc->j + c_loc->i[i]; 1454 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1455 1456 /* add received col data into lnk */ 1457 for (k=0; k<nrecv; k++) { /* k-th received message */ 1458 if (i == *nextrow[k]) { /* i-th row */ 1459 nzi = *(nextci[k]+1) - *nextci[k]; 1460 Jptr = buf_rj[k] + *nextci[k]; 1461 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1462 nextrow[k]++; nextci[k]++; 1463 } 1464 } 1465 nzi = lnk[0]; 1466 1467 /* copy data into free space, then initialize lnk */ 1468 ierr = PetscLLCondensedClean(aN,nzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1469 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 1470 } 1471 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1472 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1473 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 1474 1475 /* local sizes and preallocation */ 1476 ierr = MatSetSizes(Cmpi,pn,an,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1477 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 1478 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1479 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1480 1481 /* members in merge */ 1482 ierr = PetscFree(id_r);CHKERRQ(ierr); 1483 ierr = PetscFree(len_r);CHKERRQ(ierr); 1484 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 1485 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 1486 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 1487 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 1488 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 1489 1490 /* attach the supporting struct to Cmpi for reuse */ 1491 c = (Mat_MPIAIJ*)Cmpi->data; 1492 c->ap = ptap; 1493 ptap->destroy = Cmpi->ops->destroy; 1494 1495 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 1496 Cmpi->assembled = PETSC_FALSE; 1497 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1498 Cmpi->ops->freeintermediatedatastructures = MatFreeIntermediateDataStructures_MPIAIJ_AP; 1499 1500 *C = Cmpi; 1501 PetscFunctionReturn(0); 1502 } 1503 1504 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P,Mat A,Mat C) 1505 { 1506 PetscErrorCode ierr; 1507 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1508 Mat_SeqAIJ *c_seq; 1509 Mat_APMPI *ptap = c->ap; 1510 Mat A_loc,C_loc,C_oth; 1511 PetscInt i,rstart,rend,cm,ncols,row; 1512 const PetscInt *cols; 1513 const PetscScalar *vals; 1514 1515 PetscFunctionBegin; 1516 if (!ptap) { 1517 MPI_Comm comm; 1518 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1519 SETERRQ(comm,PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatFreeIntermediateDataStructures() or use '-mat_freeintermediatedatastructures'"); 1520 } 1521 1522 ierr = MatZeroEntries(C);CHKERRQ(ierr); 1523 1524 if (ptap->reuse == MAT_REUSE_MATRIX) { 1525 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1526 /* 1) get R = Pd^T, Ro = Po^T */ 1527 /*----------------------------*/ 1528 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1529 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1530 1531 /* 2) compute numeric A_loc */ 1532 /*--------------------------*/ 1533 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&ptap->A_loc);CHKERRQ(ierr); 1534 } 1535 1536 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1537 A_loc = ptap->A_loc; 1538 ierr = ((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,A_loc,ptap->C_loc);CHKERRQ(ierr); 1539 ierr = ((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,A_loc,ptap->C_oth);CHKERRQ(ierr); 1540 C_loc = ptap->C_loc; 1541 C_oth = ptap->C_oth; 1542 1543 /* add C_loc and Co to to C */ 1544 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 1545 1546 /* C_loc -> C */ 1547 cm = C_loc->rmap->N; 1548 c_seq = (Mat_SeqAIJ*)C_loc->data; 1549 cols = c_seq->j; 1550 vals = c_seq->a; 1551 for (i=0; i<cm; i++) { 1552 ncols = c_seq->i[i+1] - c_seq->i[i]; 1553 row = rstart + i; 1554 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1555 cols += ncols; vals += ncols; 1556 } 1557 1558 /* Co -> C, off-processor part */ 1559 cm = C_oth->rmap->N; 1560 c_seq = (Mat_SeqAIJ*)C_oth->data; 1561 cols = c_seq->j; 1562 vals = c_seq->a; 1563 for (i=0; i<cm; i++) { 1564 ncols = c_seq->i[i+1] - c_seq->i[i]; 1565 row = p->garray[i]; 1566 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1567 cols += ncols; vals += ncols; 1568 } 1569 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1570 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1571 1572 ptap->reuse = MAT_REUSE_MATRIX; 1573 1574 /* supporting struct ptap consumes almost same amount of memory as C=PtAP, release it if C will not be updated by A and P */ 1575 if (ptap->freestruct) { 1576 ierr = MatFreeIntermediateDataStructures(C);CHKERRQ(ierr); 1577 } 1578 PetscFunctionReturn(0); 1579 } 1580 1581 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 1582 { 1583 PetscErrorCode ierr; 1584 Mat_Merge_SeqsToMPI *merge; 1585 Mat_MPIAIJ *p =(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1586 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1587 Mat_APMPI *ptap; 1588 PetscInt *adj; 1589 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1590 MatScalar *ada,*ca,valtmp; 1591 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1592 MPI_Comm comm; 1593 PetscMPIInt size,rank,taga,*len_s; 1594 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1595 PetscInt **buf_ri,**buf_rj; 1596 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1597 MPI_Request *s_waits,*r_waits; 1598 MPI_Status *status; 1599 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1600 PetscInt *ai,*aj,*coi,*coj,*poJ,*pdJ; 1601 Mat A_loc; 1602 Mat_SeqAIJ *a_loc; 1603 1604 PetscFunctionBegin; 1605 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1606 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1607 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1608 1609 ptap = c->ap; 1610 if (!ptap) SETERRQ(comm,PETSC_ERR_ARG_WRONGSTATE,"PtA cannot be reused. Do not call MatFreeIntermediateDataStructures() or use '-mat_freeintermediatedatastructures'"); 1611 merge = ptap->merge; 1612 1613 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1614 /*------------------------------------------*/ 1615 /* get data from symbolic products */ 1616 coi = merge->coi; coj = merge->coj; 1617 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1618 bi = merge->bi; bj = merge->bj; 1619 owners = merge->rowmap->range; 1620 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); 1621 1622 /* get A_loc by taking all local rows of A */ 1623 A_loc = ptap->A_loc; 1624 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 1625 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1626 ai = a_loc->i; 1627 aj = a_loc->j; 1628 1629 for (i=0; i<am; i++) { 1630 anz = ai[i+1] - ai[i]; 1631 adj = aj + ai[i]; 1632 ada = a_loc->a + ai[i]; 1633 1634 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1635 /*-------------------------------------------------------------*/ 1636 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1637 pnz = po->i[i+1] - po->i[i]; 1638 poJ = po->j + po->i[i]; 1639 pA = po->a + po->i[i]; 1640 for (j=0; j<pnz; j++) { 1641 row = poJ[j]; 1642 cj = coj + coi[row]; 1643 ca = coa + coi[row]; 1644 /* perform sparse axpy */ 1645 nexta = 0; 1646 valtmp = pA[j]; 1647 for (k=0; nexta<anz; k++) { 1648 if (cj[k] == adj[nexta]) { 1649 ca[k] += valtmp*ada[nexta]; 1650 nexta++; 1651 } 1652 } 1653 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1654 } 1655 1656 /* put the value into Cd (diagonal part) */ 1657 pnz = pd->i[i+1] - pd->i[i]; 1658 pdJ = pd->j + pd->i[i]; 1659 pA = pd->a + pd->i[i]; 1660 for (j=0; j<pnz; j++) { 1661 row = pdJ[j]; 1662 cj = bj + bi[row]; 1663 ca = ba + bi[row]; 1664 /* perform sparse axpy */ 1665 nexta = 0; 1666 valtmp = pA[j]; 1667 for (k=0; nexta<anz; k++) { 1668 if (cj[k] == adj[nexta]) { 1669 ca[k] += valtmp*ada[nexta]; 1670 nexta++; 1671 } 1672 } 1673 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1674 } 1675 } 1676 1677 /* 3) send and recv matrix values coa */ 1678 /*------------------------------------*/ 1679 buf_ri = merge->buf_ri; 1680 buf_rj = merge->buf_rj; 1681 len_s = merge->len_s; 1682 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1683 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1684 1685 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1686 for (proc=0,k=0; proc<size; proc++) { 1687 if (!len_s[proc]) continue; 1688 i = merge->owners_co[proc]; 1689 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1690 k++; 1691 } 1692 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1693 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1694 1695 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1696 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1697 ierr = PetscFree(coa);CHKERRQ(ierr); 1698 1699 /* 4) insert local Cseq and received values into Cmpi */ 1700 /*----------------------------------------------------*/ 1701 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1702 for (k=0; k<merge->nrecv; k++) { 1703 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1704 nrows = *(buf_ri_k[k]); 1705 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1706 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1707 } 1708 1709 for (i=0; i<cm; i++) { 1710 row = owners[rank] + i; /* global row index of C_seq */ 1711 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1712 ba_i = ba + bi[i]; 1713 bnz = bi[i+1] - bi[i]; 1714 /* add received vals into ba */ 1715 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1716 /* i-th row */ 1717 if (i == *nextrow[k]) { 1718 cnz = *(nextci[k]+1) - *nextci[k]; 1719 cj = buf_rj[k] + *(nextci[k]); 1720 ca = abuf_r[k] + *(nextci[k]); 1721 nextcj = 0; 1722 for (j=0; nextcj<cnz; j++) { 1723 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1724 ba_i[j] += ca[nextcj++]; 1725 } 1726 } 1727 nextrow[k]++; nextci[k]++; 1728 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1729 } 1730 } 1731 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1732 } 1733 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1734 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1735 1736 ierr = PetscFree(ba);CHKERRQ(ierr); 1737 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1738 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1739 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1740 1741 if (ptap->freestruct) { 1742 ierr = MatFreeIntermediateDataStructures(C);CHKERRQ(ierr); 1743 } 1744 PetscFunctionReturn(0); 1745 } 1746 1747 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat *C) 1748 { 1749 PetscErrorCode ierr; 1750 Mat Cmpi,A_loc,POt,PDt; 1751 Mat_APMPI *ptap; 1752 PetscFreeSpaceList free_space=NULL,current_space=NULL; 1753 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*a=(Mat_MPIAIJ*)A->data,*c; 1754 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1755 PetscInt nnz; 1756 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1757 PetscInt am =A->rmap->n,pn=P->cmap->n; 1758 MPI_Comm comm; 1759 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1760 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1761 PetscInt len,proc,*dnz,*onz,*owners; 1762 PetscInt nzi,*bi,*bj; 1763 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1764 MPI_Request *swaits,*rwaits; 1765 MPI_Status *sstatus,rstatus; 1766 Mat_Merge_SeqsToMPI *merge; 1767 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1768 PetscReal afill =1.0,afill_tmp; 1769 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Armax; 1770 PetscScalar *vals; 1771 Mat_SeqAIJ *a_loc,*pdt,*pot; 1772 PetscTable ta; 1773 MatType mtype; 1774 1775 PetscFunctionBegin; 1776 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1777 /* check if matrix local sizes are compatible */ 1778 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); 1779 1780 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1781 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1782 1783 /* create struct Mat_APMPI and attached it to C later */ 1784 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1785 1786 /* get A_loc by taking all local rows of A */ 1787 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1788 1789 ptap->A_loc = A_loc; 1790 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1791 ai = a_loc->i; 1792 aj = a_loc->j; 1793 1794 /* determine symbolic Co=(p->B)^T*A - send to others */ 1795 /*----------------------------------------------------*/ 1796 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1797 pdt = (Mat_SeqAIJ*)PDt->data; 1798 pdti = pdt->i; pdtj = pdt->j; 1799 1800 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1801 pot = (Mat_SeqAIJ*)POt->data; 1802 poti = pot->i; potj = pot->j; 1803 1804 /* then, compute symbolic Co = (p->B)^T*A */ 1805 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1806 >= (num of nonzero rows of C_seq) - pn */ 1807 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 1808 coi[0] = 0; 1809 1810 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1811 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(poti[pon],ai[am])); 1812 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1813 current_space = free_space; 1814 1815 /* create and initialize a linked list */ 1816 ierr = PetscTableCreate(A->cmap->n + a->B->cmap->N,aN,&ta);CHKERRQ(ierr); 1817 MatRowMergeMax_SeqAIJ(a_loc,am,ta); 1818 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1819 1820 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1821 1822 for (i=0; i<pon; i++) { 1823 pnz = poti[i+1] - poti[i]; 1824 ptJ = potj + poti[i]; 1825 for (j=0; j<pnz; j++) { 1826 row = ptJ[j]; /* row of A_loc == col of Pot */ 1827 anz = ai[row+1] - ai[row]; 1828 Jptr = aj + ai[row]; 1829 /* add non-zero cols of AP into the sorted linked list lnk */ 1830 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1831 } 1832 nnz = lnk[0]; 1833 1834 /* If free space is not available, double the total space in the list */ 1835 if (current_space->local_remaining<nnz) { 1836 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1837 nspacedouble++; 1838 } 1839 1840 /* Copy data into free space, and zero out denserows */ 1841 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1842 1843 current_space->array += nnz; 1844 current_space->local_used += nnz; 1845 current_space->local_remaining -= nnz; 1846 1847 coi[i+1] = coi[i] + nnz; 1848 } 1849 1850 ierr = PetscMalloc1(coi[pon]+1,&coj);CHKERRQ(ierr); 1851 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1852 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); /* must destroy to get a new one for C */ 1853 1854 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1855 if (afill_tmp > afill) afill = afill_tmp; 1856 1857 /* send j-array (coj) of Co to other processors */ 1858 /*----------------------------------------------*/ 1859 /* determine row ownership */ 1860 ierr = PetscNew(&merge);CHKERRQ(ierr); 1861 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1862 1863 merge->rowmap->n = pn; 1864 merge->rowmap->bs = 1; 1865 1866 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1867 owners = merge->rowmap->range; 1868 1869 /* determine the number of messages to send, their lengths */ 1870 ierr = PetscCalloc1(size,&len_si);CHKERRQ(ierr); 1871 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 1872 1873 len_s = merge->len_s; 1874 merge->nsend = 0; 1875 1876 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 1877 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1878 1879 proc = 0; 1880 for (i=0; i<pon; i++) { 1881 while (prmap[i] >= owners[proc+1]) proc++; 1882 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1883 len_s[proc] += coi[i+1] - coi[i]; 1884 } 1885 1886 len = 0; /* max length of buf_si[] */ 1887 owners_co[0] = 0; 1888 for (proc=0; proc<size; proc++) { 1889 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1890 if (len_si[proc]) { 1891 merge->nsend++; 1892 len_si[proc] = 2*(len_si[proc] + 1); 1893 len += len_si[proc]; 1894 } 1895 } 1896 1897 /* determine the number and length of messages to receive for coi and coj */ 1898 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1899 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1900 1901 /* post the Irecv and Isend of coj */ 1902 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1903 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1904 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 1905 for (proc=0, k=0; proc<size; proc++) { 1906 if (!len_s[proc]) continue; 1907 i = owners_co[proc]; 1908 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1909 k++; 1910 } 1911 1912 /* receives and sends of coj are complete */ 1913 ierr = PetscMalloc1(size,&sstatus);CHKERRQ(ierr); 1914 for (i=0; i<merge->nrecv; i++) { 1915 PetscMPIInt icompleted; 1916 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1917 } 1918 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1919 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1920 1921 /* add received column indices into table to update Armax */ 1922 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/examples/tutorials/ex56.c! */ 1923 for (k=0; k<merge->nrecv; k++) {/* k-th received message */ 1924 Jptr = buf_rj[k]; 1925 for (j=0; j<merge->len_r[k]; j++) { 1926 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 1927 } 1928 } 1929 ierr = PetscTableGetCount(ta,&Armax);CHKERRQ(ierr); 1930 /* 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); */ 1931 1932 /* send and recv coi */ 1933 /*-------------------*/ 1934 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1935 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1936 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1937 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1938 for (proc=0,k=0; proc<size; proc++) { 1939 if (!len_s[proc]) continue; 1940 /* form outgoing message for i-structure: 1941 buf_si[0]: nrows to be sent 1942 [1:nrows]: row index (global) 1943 [nrows+1:2*nrows+1]: i-structure index 1944 */ 1945 /*-------------------------------------------*/ 1946 nrows = len_si[proc]/2 - 1; 1947 buf_si_i = buf_si + nrows+1; 1948 buf_si[0] = nrows; 1949 buf_si_i[0] = 0; 1950 nrows = 0; 1951 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1952 nzi = coi[i+1] - coi[i]; 1953 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1954 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1955 nrows++; 1956 } 1957 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1958 k++; 1959 buf_si += len_si[proc]; 1960 } 1961 i = merge->nrecv; 1962 while (i--) { 1963 PetscMPIInt icompleted; 1964 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1965 } 1966 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1967 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1968 ierr = PetscFree(len_si);CHKERRQ(ierr); 1969 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1970 ierr = PetscFree(swaits);CHKERRQ(ierr); 1971 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1972 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1973 1974 /* compute the local portion of C (mpi mat) */ 1975 /*------------------------------------------*/ 1976 /* allocate bi array and free space for accumulating nonzero column info */ 1977 ierr = PetscMalloc1(pn+1,&bi);CHKERRQ(ierr); 1978 bi[0] = 0; 1979 1980 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1981 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(pdti[pn],PetscIntSumTruncate(poti[pon],ai[am]))); 1982 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1983 current_space = free_space; 1984 1985 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1986 for (k=0; k<merge->nrecv; k++) { 1987 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1988 nrows = *buf_ri_k[k]; 1989 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1990 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recieved i-structure */ 1991 } 1992 1993 ierr = PetscLLCondensedCreate_Scalable(Armax,&lnk);CHKERRQ(ierr); 1994 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1995 rmax = 0; 1996 for (i=0; i<pn; i++) { 1997 /* add pdt[i,:]*AP into lnk */ 1998 pnz = pdti[i+1] - pdti[i]; 1999 ptJ = pdtj + pdti[i]; 2000 for (j=0; j<pnz; j++) { 2001 row = ptJ[j]; /* row of AP == col of Pt */ 2002 anz = ai[row+1] - ai[row]; 2003 Jptr = aj + ai[row]; 2004 /* add non-zero cols of AP into the sorted linked list lnk */ 2005 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 2006 } 2007 2008 /* add received col data into lnk */ 2009 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 2010 if (i == *nextrow[k]) { /* i-th row */ 2011 nzi = *(nextci[k]+1) - *nextci[k]; 2012 Jptr = buf_rj[k] + *nextci[k]; 2013 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 2014 nextrow[k]++; nextci[k]++; 2015 } 2016 } 2017 nnz = lnk[0]; 2018 2019 /* if free space is not available, make more free space */ 2020 if (current_space->local_remaining<nnz) { 2021 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 2022 nspacedouble++; 2023 } 2024 /* copy data into free space, then initialize lnk */ 2025 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 2026 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 2027 2028 current_space->array += nnz; 2029 current_space->local_used += nnz; 2030 current_space->local_remaining -= nnz; 2031 2032 bi[i+1] = bi[i] + nnz; 2033 if (nnz > rmax) rmax = nnz; 2034 } 2035 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 2036 2037 ierr = PetscMalloc1(bi[pn]+1,&bj);CHKERRQ(ierr); 2038 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 2039 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 2040 if (afill_tmp > afill) afill = afill_tmp; 2041 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 2042 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 2043 2044 ierr = MatDestroy(&POt);CHKERRQ(ierr); 2045 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 2046 2047 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 2048 /*----------------------------------------------------------------------------------*/ 2049 ierr = PetscCalloc1(rmax+1,&vals);CHKERRQ(ierr); 2050 2051 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 2052 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2053 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(A->cmap->bs));CHKERRQ(ierr); 2054 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 2055 ierr = MatSetType(Cmpi,mtype);CHKERRQ(ierr); 2056 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 2057 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 2058 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 2059 for (i=0; i<pn; i++) { 2060 row = i + rstart; 2061 nnz = bi[i+1] - bi[i]; 2062 Jptr = bj + bi[i]; 2063 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 2064 } 2065 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2066 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2067 ierr = PetscFree(vals);CHKERRQ(ierr); 2068 2069 merge->bi = bi; 2070 merge->bj = bj; 2071 merge->coi = coi; 2072 merge->coj = coj; 2073 merge->buf_ri = buf_ri; 2074 merge->buf_rj = buf_rj; 2075 merge->owners_co = owners_co; 2076 2077 /* attach the supporting struct to Cmpi for reuse */ 2078 c = (Mat_MPIAIJ*)Cmpi->data; 2079 2080 c->ap = ptap; 2081 ptap->api = NULL; 2082 ptap->apj = NULL; 2083 ptap->merge = merge; 2084 ptap->apa = NULL; 2085 ptap->destroy = Cmpi->ops->destroy; 2086 ptap->duplicate = Cmpi->ops->duplicate; 2087 2088 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2089 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 2090 Cmpi->ops->freeintermediatedatastructures = MatFreeIntermediateDataStructures_MPIAIJ_AP; 2091 2092 *C = Cmpi; 2093 #if defined(PETSC_USE_INFO) 2094 if (bi[pn] != 0) { 2095 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 2096 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 2097 } else { 2098 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 2099 } 2100 #endif 2101 PetscFunctionReturn(0); 2102 } 2103