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