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