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