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