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 12 #undef __FUNCT__ 13 #define __FUNCT__ "MatMatMult_MPIAIJ_MPIAIJ" 14 PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill, Mat *C) 15 { 16 PetscErrorCode ierr; 17 18 PetscFunctionBegin; 19 if (scall == MAT_INITIAL_MATRIX){ 20 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 21 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 22 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 23 } 24 25 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 26 ierr = (*(*C)->ops->matmultnumeric)(A,B,*C);CHKERRQ(ierr); 27 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 28 PetscFunctionReturn(0); 29 } 30 31 #undef __FUNCT__ 32 #define __FUNCT__ "MatDestroy_MPIAIJ_MatMatMult" 33 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat A) 34 { 35 PetscErrorCode ierr; 36 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 37 Mat_PtAPMPI *ptap=a->ptap; 38 39 PetscFunctionBegin; 40 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 41 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 42 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 43 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 44 ierr = MatDestroy(&ptap->Pt);CHKERRQ(ierr); 45 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 46 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 47 ierr = PetscFree(ptap->apa);CHKERRQ(ierr); 48 ierr = ptap->destroy(A);CHKERRQ(ierr); 49 ierr = PetscFree(ptap);CHKERRQ(ierr); 50 PetscFunctionReturn(0); 51 } 52 53 #undef __FUNCT__ 54 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatMatMult" 55 PetscErrorCode MatDuplicate_MPIAIJ_MatMatMult(Mat A, MatDuplicateOption op, Mat *M) 56 { 57 PetscErrorCode ierr; 58 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 59 Mat_PtAPMPI *ptap=a->ptap; 60 61 PetscFunctionBegin; 62 ierr = (*ptap->duplicate)(A,op,M);CHKERRQ(ierr); 63 (*M)->ops->destroy = ptap->destroy; /* = MatDestroy_MPIAIJ, *M doesn't duplicate A's special structure! */ 64 (*M)->ops->duplicate = ptap->duplicate; /* = MatDuplicate_MPIAIJ */ 65 PetscFunctionReturn(0); 66 } 67 68 #undef __FUNCT__ 69 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIAIJ" 70 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 71 { 72 PetscErrorCode ierr; 73 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 74 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 75 Mat_SeqAIJ *cd=(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 76 PetscInt *adi=ad->i,*adj,*aoi=ao->i,*aoj; 77 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 78 Mat_SeqAIJ *p_loc,*p_oth; 79 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 80 PetscScalar *pa_loc,*pa_oth,*pa,*apa,valtmp,*ca; 81 PetscInt cm=C->rmap->n,anz,pnz; 82 Mat_PtAPMPI *ptap=c->ptap; 83 PetscInt *api,*apj,*apJ,i,j,k,row; 84 PetscInt cstart=C->cmap->rstart; 85 PetscInt cdnz,conz,k0,k1; 86 87 PetscFunctionBegin; 88 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 89 /*-----------------------------------------------------*/ 90 /* update numerical values of P_oth and P_loc */ 91 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 92 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 93 94 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 95 /*----------------------------------------------------------*/ 96 /* get data from symbolic products */ 97 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 98 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 99 pi_loc=p_loc->i; pj_loc=p_loc->j; pa_loc=p_loc->a; 100 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 101 102 /* get apa for storing dense row A[i,:]*P */ 103 apa = ptap->apa; 104 105 api = ptap->api; 106 apj = ptap->apj; 107 for (i=0; i<cm; i++) { 108 /* diagonal portion of A */ 109 anz = adi[i+1] - adi[i]; 110 adj = ad->j + adi[i]; 111 ada = ad->a + adi[i]; 112 for (j=0; j<anz; j++) { 113 row = adj[j]; 114 pnz = pi_loc[row+1] - pi_loc[row]; 115 pj = pj_loc + pi_loc[row]; 116 pa = pa_loc + pi_loc[row]; 117 118 /* perform dense axpy */ 119 valtmp = ada[j]; 120 for (k=0; k<pnz; k++){ 121 apa[pj[k]] += valtmp*pa[k]; 122 } 123 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 124 } 125 126 /* off-diagonal portion of A */ 127 anz = aoi[i+1] - aoi[i]; 128 aoj = ao->j + aoi[i]; 129 aoa = ao->a + aoi[i]; 130 for (j=0; j<anz; j++) { 131 row = aoj[j]; 132 pnz = pi_oth[row+1] - pi_oth[row]; 133 pj = pj_oth + pi_oth[row]; 134 pa = pa_oth + pi_oth[row]; 135 136 /* perform dense axpy */ 137 valtmp = aoa[j]; 138 for (k=0; k<pnz; k++){ 139 apa[pj[k]] += valtmp*pa[k]; 140 } 141 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 142 } 143 144 /* set values in C */ 145 apJ = apj + api[i]; 146 cdnz = cd->i[i+1] - cd->i[i]; 147 conz = co->i[i+1] - co->i[i]; 148 149 /* 1st off-diagoanl part of C */ 150 ca = coa + co->i[i]; 151 k = 0; 152 for (k0=0; k0<conz; k0++){ 153 if (apJ[k] >= cstart) break; 154 ca[k0] = apa[apJ[k]]; 155 apa[apJ[k]] = 0.0; 156 k++; 157 } 158 159 /* diagonal part of C */ 160 ca = cda + cd->i[i]; 161 for (k1=0; k1<cdnz; k1++){ 162 ca[k1] = apa[apJ[k]]; 163 apa[apJ[k]] = 0.0; 164 k++; 165 } 166 167 /* 2nd off-diagoanl part of C */ 168 ca = coa + co->i[i]; 169 for (; k0<conz; k0++){ 170 ca[k0] = apa[apJ[k]]; 171 apa[apJ[k]] = 0.0; 172 k++; 173 } 174 } 175 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 176 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 177 PetscFunctionReturn(0); 178 } 179 180 #undef __FUNCT__ 181 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ" 182 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 183 { 184 PetscErrorCode ierr; 185 MPI_Comm comm=((PetscObject)A)->comm; 186 Mat Cmpi; 187 Mat_PtAPMPI *ptap; 188 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 189 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c; 190 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 191 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 192 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 193 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 194 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 195 PetscBT lnkbt; 196 PetscScalar *apa; 197 PetscReal afill; 198 PetscBool scalable=PETSC_FALSE; 199 PetscInt nlnk_max,armax,prmax; 200 201 PetscFunctionBegin; 202 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend){ 203 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); 204 } 205 206 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 207 ierr = PetscOptionsBool("-matmatmult_scalable","Use a scalable but slower C=A*B","",scalable,&scalable,PETSC_NULL);CHKERRQ(ierr); 208 if (scalable){ 209 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(A,P,fill,C);;CHKERRQ(ierr); 210 PetscFunctionReturn(0); 211 } 212 ierr = PetscOptionsEnd();CHKERRQ(ierr); 213 214 /* create struct Mat_PtAPMPI and attached it to C later */ 215 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 216 217 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 218 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 219 220 /* get P_loc by taking all local rows of P */ 221 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 222 223 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 224 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 225 pi_loc = p_loc->i; pj_loc = p_loc->j; 226 pi_oth = p_oth->i; pj_oth = p_oth->j; 227 228 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 229 /*-------------------------------------------------------------------*/ 230 ierr = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr); 231 ptap->api = api; 232 api[0] = 0; 233 234 /* create and initialize a linked list */ 235 armax = ad->rmax+ao->rmax; 236 prmax = PetscMax(p_loc->rmax,p_oth->rmax); 237 nlnk_max = armax*prmax; 238 if (!nlnk_max || nlnk_max > pN) nlnk_max = pN; 239 ierr = PetscLLCondensedCreate(nlnk_max,pN,&lnk,&lnkbt);CHKERRQ(ierr); 240 241 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 242 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 243 current_space = free_space; 244 245 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 246 for (i=0; i<am; i++) { 247 apnz = 0; 248 /* diagonal portion of A */ 249 nzi = adi[i+1] - adi[i]; 250 for (j=0; j<nzi; j++){ 251 row = *adj++; 252 pnz = pi_loc[row+1] - pi_loc[row]; 253 Jptr = pj_loc + pi_loc[row]; 254 /* add non-zero cols of P into the sorted linked list lnk */ 255 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 256 } 257 /* off-diagonal portion of A */ 258 nzi = aoi[i+1] - aoi[i]; 259 for (j=0; j<nzi; j++){ 260 row = *aoj++; 261 pnz = pi_oth[row+1] - pi_oth[row]; 262 Jptr = pj_oth + pi_oth[row]; 263 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 264 } 265 266 apnz = lnk[0]; 267 api[i+1] = api[i] + apnz; 268 269 /* if free space is not available, double the total space in the list */ 270 if (current_space->local_remaining<apnz) { 271 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 272 nspacedouble++; 273 } 274 275 /* Copy data into free space, then initialize lnk */ 276 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 277 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 278 current_space->array += apnz; 279 current_space->local_used += apnz; 280 current_space->local_remaining -= apnz; 281 } 282 283 /* Allocate space for apj, initialize apj, and */ 284 /* destroy list of free space and other temporary array(s) */ 285 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ptap->apj);CHKERRQ(ierr); 286 apj = ptap->apj; 287 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 288 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 289 290 /* malloc apa to store dense row A[i,:]*P */ 291 ierr = PetscMalloc(pN*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 292 ierr = PetscMemzero(apa,pN*sizeof(PetscScalar));CHKERRQ(ierr); 293 ptap->apa = apa; 294 295 /* create and assemble symbolic parallel matrix Cmpi */ 296 /*----------------------------------------------------*/ 297 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 298 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 299 ierr = MatSetBlockSizes(Cmpi,A->rmap->bs,P->cmap->bs);CHKERRQ(ierr); 300 301 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 302 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 303 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 304 for (i=0; i<am; i++){ 305 row = i + rstart; 306 apnz = api[i+1] - api[i]; 307 ierr = MatSetValues(Cmpi,1,&row,apnz,apj,apa,INSERT_VALUES);CHKERRQ(ierr); 308 apj += apnz; 309 } 310 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 311 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 312 313 ptap->destroy = Cmpi->ops->destroy; 314 ptap->duplicate = Cmpi->ops->duplicate; 315 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 316 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 317 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 318 319 /* attach the supporting struct to Cmpi for reuse */ 320 c = (Mat_MPIAIJ*)Cmpi->data; 321 c->ptap = ptap; 322 323 *C = Cmpi; 324 325 /* set MatInfo */ 326 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 327 if (afill < 1.0) afill = 1.0; 328 Cmpi->info.mallocs = nspacedouble; 329 Cmpi->info.fill_ratio_given = fill; 330 Cmpi->info.fill_ratio_needed = afill; 331 332 #if defined(PETSC_USE_INFO) 333 if (api[am]) { 334 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 335 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 336 } else { 337 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 338 } 339 #endif 340 PetscFunctionReturn(0); 341 } 342 343 #undef __FUNCT__ 344 #define __FUNCT__ "MatMatMult_MPIAIJ_MPIDense" 345 PetscErrorCode MatMatMult_MPIAIJ_MPIDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 346 { 347 PetscErrorCode ierr; 348 349 PetscFunctionBegin; 350 if (scall == MAT_INITIAL_MATRIX){ 351 ierr = MatMatMultSymbolic_MPIAIJ_MPIDense(A,B,fill,C);CHKERRQ(ierr); 352 } 353 ierr = MatMatMultNumeric_MPIAIJ_MPIDense(A,B,*C);CHKERRQ(ierr); 354 PetscFunctionReturn(0); 355 } 356 357 typedef struct { 358 Mat workB; 359 PetscScalar *rvalues,*svalues; 360 MPI_Request *rwaits,*swaits; 361 } MPIAIJ_MPIDense; 362 363 #undef __FUNCT__ 364 #define __FUNCT__ "MatMPIAIJ_MPIDenseDestroy" 365 PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 366 { 367 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*) ctx; 368 PetscErrorCode ierr; 369 370 PetscFunctionBegin; 371 ierr = MatDestroy(&contents->workB);CHKERRQ(ierr); 372 ierr = PetscFree4(contents->rvalues,contents->svalues,contents->rwaits,contents->swaits);CHKERRQ(ierr); 373 ierr = PetscFree(contents);CHKERRQ(ierr); 374 PetscFunctionReturn(0); 375 } 376 377 #undef __FUNCT__ 378 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIDense" 379 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C) 380 { 381 PetscErrorCode ierr; 382 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 383 PetscInt nz = aij->B->cmap->n; 384 PetscContainer container; 385 MPIAIJ_MPIDense *contents; 386 VecScatter ctx = aij->Mvctx; 387 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 388 VecScatter_MPI_General *to = ( VecScatter_MPI_General*) ctx->todata; 389 PetscInt m=A->rmap->n,n=B->cmap->n; 390 391 PetscFunctionBegin; 392 ierr = MatCreate(((PetscObject)B)->comm,C);CHKERRQ(ierr); 393 ierr = MatSetSizes(*C,m,n,A->rmap->N,B->cmap->N);CHKERRQ(ierr); 394 ierr = MatSetBlockSizes(*C,A->rmap->bs,B->cmap->bs);CHKERRQ(ierr); 395 ierr = MatSetType(*C,MATMPIDENSE);CHKERRQ(ierr); 396 ierr = MatMPIDenseSetPreallocation(*C,PETSC_NULL);CHKERRQ(ierr); 397 ierr = MatAssemblyBegin(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 398 ierr = MatAssemblyEnd(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 399 (*C)->ops->matmult = MatMatMult_MPIAIJ_MPIDense; 400 401 ierr = PetscNew(MPIAIJ_MPIDense,&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,PETSC_NULL,&contents->workB);CHKERRQ(ierr); 404 /* Create work arrays needed */ 405 ierr = PetscMalloc4(B->cmap->N*from->starts[from->n],PetscScalar,&contents->rvalues, 406 B->cmap->N*to->starts[to->n],PetscScalar,&contents->svalues, 407 from->n,MPI_Request,&contents->rwaits, 408 to->n,MPI_Request,&contents->swaits);CHKERRQ(ierr); 409 410 ierr = PetscContainerCreate(((PetscObject)A)->comm,&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 PetscFunctionReturn(0); 416 } 417 418 #undef __FUNCT__ 419 #define __FUNCT__ "MatMPIDenseScatter" 420 /* 421 Performs an efficient scatter on the rows of B needed by this process; this is 422 a modification of the VecScatterBegin_() routines. 423 */ 424 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,Mat C,Mat *outworkB) 425 { 426 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 427 PetscErrorCode ierr; 428 PetscScalar *b,*w,*svalues,*rvalues; 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 i,j,k; 433 PetscInt *sindices,*sstarts,*rindices,*rstarts; 434 PetscMPIInt *sprocs,*rprocs,nrecvs; 435 MPI_Request *swaits,*rwaits; 436 MPI_Comm comm = ((PetscObject)A)->comm; 437 PetscMPIInt tag = ((PetscObject)ctx)->tag,ncols = B->cmap->N, nrows = aij->B->cmap->n,imdex,nrowsB = B->rmap->n; 438 MPI_Status status; 439 MPIAIJ_MPIDense *contents; 440 PetscContainer container; 441 Mat workB; 442 443 PetscFunctionBegin; 444 ierr = PetscObjectQuery((PetscObject)C,"workB",(PetscObject*)&container);CHKERRQ(ierr); 445 if (!container) SETERRQ(comm,PETSC_ERR_PLIB,"Container does not exist"); 446 ierr = PetscContainerGetPointer(container,(void**)&contents);CHKERRQ(ierr); 447 448 workB = *outworkB = contents->workB; 449 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); 450 sindices = to->indices; 451 sstarts = to->starts; 452 sprocs = to->procs; 453 swaits = contents->swaits; 454 svalues = contents->svalues; 455 456 rindices = from->indices; 457 rstarts = from->starts; 458 rprocs = from->procs; 459 rwaits = contents->rwaits; 460 rvalues = contents->rvalues; 461 462 ierr = MatDenseGetArray(B,&b);CHKERRQ(ierr); 463 ierr = MatDenseGetArray(workB,&w);CHKERRQ(ierr); 464 465 for (i=0; i<from->n; i++) { 466 ierr = MPI_Irecv(rvalues+ncols*rstarts[i],ncols*(rstarts[i+1]-rstarts[i]),MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 467 } 468 469 for (i=0; i<to->n; i++) { 470 /* pack a message at a time */ 471 CHKMEMQ; 472 for (j=0; j<sstarts[i+1]-sstarts[i]; j++){ 473 for (k=0; k<ncols; k++) { 474 svalues[ncols*(sstarts[i] + j) + k] = b[sindices[sstarts[i]+j] + nrowsB*k]; 475 } 476 } 477 CHKMEMQ; 478 ierr = MPI_Isend(svalues+ncols*sstarts[i],ncols*(sstarts[i+1]-sstarts[i]),MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 479 } 480 481 nrecvs = from->n; 482 while (nrecvs) { 483 ierr = MPI_Waitany(from->n,rwaits,&imdex,&status);CHKERRQ(ierr); 484 nrecvs--; 485 /* unpack a message at a time */ 486 CHKMEMQ; 487 for (j=0; j<rstarts[imdex+1]-rstarts[imdex]; j++){ 488 for (k=0; k<ncols; k++) { 489 w[rindices[rstarts[imdex]+j] + nrows*k] = rvalues[ncols*(rstarts[imdex] + j) + k]; 490 } 491 } 492 CHKMEMQ; 493 } 494 if (to->n) {ierr = MPI_Waitall(to->n,swaits,to->sstatus);CHKERRQ(ierr);} 495 496 ierr = MatDenseRestoreArray(B,&b);CHKERRQ(ierr); 497 ierr = MatDenseRestoreArray(workB,&w);CHKERRQ(ierr); 498 ierr = MatAssemblyBegin(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 499 ierr = MatAssemblyEnd(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 500 PetscFunctionReturn(0); 501 } 502 extern PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat); 503 504 #undef __FUNCT__ 505 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIDense" 506 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 507 { 508 PetscErrorCode ierr; 509 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 510 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 511 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 512 Mat workB; 513 514 PetscFunctionBegin; 515 516 /* diagonal block of A times all local rows of B*/ 517 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(aij->A,bdense->A,cdense->A);CHKERRQ(ierr); 518 519 /* get off processor parts of B needed to complete the product */ 520 ierr = MatMPIDenseScatter(A,B,C,&workB);CHKERRQ(ierr); 521 522 /* off-diagonal block of A times nonlocal rows of B */ 523 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A);CHKERRQ(ierr); 524 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 525 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 526 PetscFunctionReturn(0); 527 } 528 529 #undef __FUNCT__ 530 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable" 531 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable(Mat A,Mat P,Mat C) 532 { 533 PetscErrorCode ierr; 534 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 535 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 536 Mat_SeqAIJ *cd=(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 537 PetscInt *adi=ad->i,*adj,*aoi=ao->i,*aoj; 538 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 539 Mat_SeqAIJ *p_loc,*p_oth; 540 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 541 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 542 PetscInt cm=C->rmap->n,anz,pnz; 543 Mat_PtAPMPI *ptap=c->ptap; 544 PetscScalar *apa_sparse=ptap->apa; 545 PetscInt *api,*apj,*apJ,i,j,k,row; 546 PetscInt cstart=C->cmap->rstart; 547 PetscInt cdnz,conz,k0,k1,nextp; 548 549 PetscFunctionBegin; 550 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 551 /*-----------------------------------------------------*/ 552 /* update numerical values of P_oth and P_loc */ 553 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 554 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 555 556 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 557 /*----------------------------------------------------------*/ 558 /* get data from symbolic products */ 559 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 560 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 561 pi_loc=p_loc->i; pj_loc=p_loc->j; pa_loc=p_loc->a; 562 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 563 564 api = ptap->api; 565 apj = ptap->apj; 566 for (i=0; i<cm; i++) { 567 apJ = apj + api[i]; 568 569 /* diagonal portion of A */ 570 anz = adi[i+1] - adi[i]; 571 adj = ad->j + adi[i]; 572 ada = ad->a + adi[i]; 573 for (j=0; j<anz; j++) { 574 row = adj[j]; 575 pnz = pi_loc[row+1] - pi_loc[row]; 576 pj = pj_loc + pi_loc[row]; 577 pa = pa_loc + pi_loc[row]; 578 /* perform sparse axpy */ 579 valtmp = ada[j]; 580 nextp = 0; 581 for (k=0; nextp<pnz; k++) { 582 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 583 apa_sparse[k] += valtmp*pa[nextp++]; 584 } 585 } 586 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 587 } 588 589 /* off-diagonal portion of A */ 590 anz = aoi[i+1] - aoi[i]; 591 aoj = ao->j + aoi[i]; 592 aoa = ao->a + aoi[i]; 593 for (j=0; j<anz; j++) { 594 row = aoj[j]; 595 pnz = pi_oth[row+1] - pi_oth[row]; 596 pj = pj_oth + pi_oth[row]; 597 pa = pa_oth + pi_oth[row]; 598 /* perform sparse axpy */ 599 valtmp = aoa[j]; 600 nextp = 0; 601 for (k=0; nextp<pnz; k++) { 602 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 603 apa_sparse[k] += valtmp*pa[nextp++]; 604 } 605 } 606 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 607 } 608 609 /* set values in C */ 610 cdnz = cd->i[i+1] - cd->i[i]; 611 conz = co->i[i+1] - co->i[i]; 612 613 /* 1st off-diagoanl part of C */ 614 ca = coa + co->i[i]; 615 k = 0; 616 for (k0=0; k0<conz; k0++){ 617 if (apJ[k] >= cstart) break; 618 ca[k0] = apa_sparse[k]; 619 apa_sparse[k] = 0.0; 620 k++; 621 } 622 623 /* diagonal part of C */ 624 ca = cda + cd->i[i]; 625 for (k1=0; k1<cdnz; k1++){ 626 ca[k1] = apa_sparse[k]; 627 apa_sparse[k] = 0.0; 628 k++; 629 } 630 631 /* 2nd off-diagoanl part of C */ 632 ca = coa + co->i[i]; 633 for (; k0<conz; k0++){ 634 ca[k0] = apa_sparse[k]; 635 apa_sparse[k] = 0.0; 636 k++; 637 } 638 } 639 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 640 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 641 PetscFunctionReturn(0); 642 } 643 644 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ(), except using LLCondensed to avoid O(BN) memory requirement */ 645 #undef __FUNCT__ 646 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable" 647 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(Mat A,Mat P,PetscReal fill,Mat *C) 648 { 649 PetscErrorCode ierr; 650 MPI_Comm comm=((PetscObject)A)->comm; 651 Mat Cmpi; 652 Mat_PtAPMPI *ptap; 653 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 654 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c; 655 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 656 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 657 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 658 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=0; 659 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 660 PetscInt nlnk_max,armax,prmax; 661 PetscReal afill; 662 PetscScalar *apa; 663 664 PetscFunctionBegin; 665 /* create struct Mat_PtAPMPI and attached it to C later */ 666 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 667 668 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 669 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 670 671 /* get P_loc by taking all local rows of P */ 672 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 673 674 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 675 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 676 pi_loc = p_loc->i; pj_loc = p_loc->j; 677 pi_oth = p_oth->i; pj_oth = p_oth->j; 678 679 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 680 /*-------------------------------------------------------------------*/ 681 ierr = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr); 682 ptap->api = api; 683 api[0] = 0; 684 685 /* create and initialize a linked list */ 686 armax = ad->rmax+ao->rmax; 687 prmax = PetscMax(p_loc->rmax,p_oth->rmax); 688 nlnk_max = armax*prmax; 689 if (!nlnk_max || nlnk_max > pN) nlnk_max = pN; 690 ierr = PetscLLCondensedCreate_Scalable(nlnk_max,&lnk);CHKERRQ(ierr); 691 692 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 693 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 694 current_space = free_space; 695 696 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 697 for (i=0; i<am; i++) { 698 apnz = 0; 699 /* diagonal portion of A */ 700 nzi = adi[i+1] - adi[i]; 701 for (j=0; j<nzi; j++){ 702 row = *adj++; 703 pnz = pi_loc[row+1] - pi_loc[row]; 704 Jptr = pj_loc + pi_loc[row]; 705 /* add non-zero cols of P into the sorted linked list lnk */ 706 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 707 } 708 /* off-diagonal portion of A */ 709 nzi = aoi[i+1] - aoi[i]; 710 for (j=0; j<nzi; j++){ 711 row = *aoj++; 712 pnz = pi_oth[row+1] - pi_oth[row]; 713 Jptr = pj_oth + pi_oth[row]; 714 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 715 } 716 717 apnz = *lnk; 718 api[i+1] = api[i] + apnz; 719 if (apnz > apnz_max) apnz_max = apnz; 720 721 /* if free space is not available, double the total space in the list */ 722 if (current_space->local_remaining<apnz) { 723 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 724 nspacedouble++; 725 } 726 727 /* Copy data into free space, then initialize lnk */ 728 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 729 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 730 current_space->array += apnz; 731 current_space->local_used += apnz; 732 current_space->local_remaining -= apnz; 733 } 734 735 /* Allocate space for apj, initialize apj, and */ 736 /* destroy list of free space and other temporary array(s) */ 737 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ptap->apj);CHKERRQ(ierr); 738 apj = ptap->apj; 739 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 740 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 741 742 /* create and assemble symbolic parallel matrix Cmpi */ 743 /*----------------------------------------------------*/ 744 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 745 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 746 ierr = MatSetBlockSizes(Cmpi,A->rmap->bs,P->cmap->bs);CHKERRQ(ierr); 747 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 748 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 749 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 750 751 /* malloc apa for assembly Cmpi */ 752 ierr = PetscMalloc(apnz_max*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 753 ierr = PetscMemzero(apa,apnz_max*sizeof(PetscScalar));CHKERRQ(ierr); 754 ptap->apa = apa; 755 for (i=0; i<am; i++){ 756 row = i + rstart; 757 apnz = api[i+1] - api[i]; 758 ierr = MatSetValues(Cmpi,1,&row,apnz,apj,apa,INSERT_VALUES);CHKERRQ(ierr); 759 apj += apnz; 760 } 761 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 762 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 763 764 ptap->destroy = Cmpi->ops->destroy; 765 ptap->duplicate = Cmpi->ops->duplicate; 766 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable; 767 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 768 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 769 770 /* attach the supporting struct to Cmpi for reuse */ 771 c = (Mat_MPIAIJ*)Cmpi->data; 772 c->ptap = ptap; 773 774 *C = Cmpi; 775 776 /* set MatInfo */ 777 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1) + 1.e-5; 778 if (afill < 1.0) afill = 1.0; 779 Cmpi->info.mallocs = nspacedouble; 780 Cmpi->info.fill_ratio_given = fill; 781 Cmpi->info.fill_ratio_needed = afill; 782 783 #if defined(PETSC_USE_INFO) 784 if (api[am]) { 785 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 786 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 787 } else { 788 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 789 } 790 #endif 791 PetscFunctionReturn(0); 792 } 793 794 /*-------------------------------------------------------------------------*/ 795 #undef __FUNCT__ 796 #define __FUNCT__ "MatTransposeMatMult_MPIAIJ_MPIAIJ" 797 PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat P,Mat A,MatReuse scall,PetscReal fill,Mat *C) 798 { 799 PetscErrorCode ierr; 800 PetscBool scalable=PETSC_FALSE,viamatmatmult=PETSC_FALSE; 801 802 PetscFunctionBegin; 803 ierr = PetscOptionsBool("-mattransposematmult_viamatmatmult","Use R=Pt and C=R*A","",viamatmatmult,&viamatmatmult,PETSC_NULL);CHKERRQ(ierr); 804 if (viamatmatmult){ 805 Mat Pt; 806 Mat_PtAPMPI *ptap; 807 Mat_MPIAIJ *c; 808 if (scall == MAT_INITIAL_MATRIX){ 809 ierr = MatTranspose(P,MAT_INITIAL_MATRIX,&Pt);CHKERRQ(ierr); 810 ierr = MatMatMult(Pt,A,MAT_INITIAL_MATRIX,fill,C);CHKERRQ(ierr); 811 812 c = (Mat_MPIAIJ*)(*C)->data; 813 ptap = c->ptap; 814 ptap->Pt = Pt; 815 } else if (scall == MAT_REUSE_MATRIX){ 816 c = (Mat_MPIAIJ*)(*C)->data; 817 ptap = c->ptap; 818 Pt = ptap->Pt; 819 ierr = MatTranspose(P,scall,&Pt);CHKERRQ(ierr); 820 ierr = MatMatMult(Pt,A,scall,fill,C);CHKERRQ(ierr); 821 } else { 822 SETERRQ(((PetscObject)A)->comm,PETSC_ERR_ARG_WRONGSTATE,"Not supported"); 823 } 824 PetscFunctionReturn(0); 825 } 826 827 if (scall == MAT_INITIAL_MATRIX){ 828 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 829 ierr = PetscOptionsBool("-mattransposematmult_scalable","Use a scalable but slower C=Pt*A","",scalable,&scalable,PETSC_NULL);CHKERRQ(ierr); 830 if (scalable){ 831 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(P,A,fill,C);CHKERRQ(ierr); 832 } else { 833 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 834 } 835 ierr = PetscOptionsEnd();CHKERRQ(ierr); 836 } 837 ierr = (*(*C)->ops->mattransposemultnumeric)(P,A,*C);CHKERRQ(ierr); 838 PetscFunctionReturn(0); 839 } 840 841 #undef __FUNCT__ 842 #define __FUNCT__ "MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ" 843 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 844 { 845 PetscErrorCode ierr; 846 Mat_Merge_SeqsToMPI *merge; 847 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 848 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 849 Mat_PtAPMPI *ptap; 850 PetscInt *adj,*aJ; 851 PetscInt i,j,k,anz,pnz,row,*cj; 852 MatScalar *ada,*aval,*ca,valtmp; 853 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 854 MPI_Comm comm=((PetscObject)C)->comm; 855 PetscMPIInt size,rank,taga,*len_s; 856 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 857 PetscInt **buf_ri,**buf_rj; 858 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 859 MPI_Request *s_waits,*r_waits; 860 MPI_Status *status; 861 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 862 PetscInt *ai,*aj,*coi,*coj; 863 PetscInt *poJ=po->j,*pdJ=pd->j; 864 Mat A_loc; 865 Mat_SeqAIJ *a_loc; 866 867 PetscFunctionBegin; 868 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 869 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 870 871 ptap = c->ptap; 872 merge = ptap->merge; 873 874 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 875 /*--------------------------------------------------------------*/ 876 /* get data from symbolic products */ 877 coi = merge->coi; coj = merge->coj; 878 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 879 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 880 881 bi = merge->bi; bj = merge->bj; 882 owners = merge->rowmap->range; 883 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 884 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 885 886 /* get A_loc by taking all local rows of A */ 887 A_loc = ptap->A_loc; 888 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 889 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 890 ai = a_loc->i; 891 aj = a_loc->j; 892 893 ierr = PetscMalloc((A->cmap->N)*sizeof(PetscScalar),&aval);CHKERRQ(ierr); /* non-scalable!!! */ 894 ierr = PetscMemzero(aval,A->cmap->N*sizeof(PetscScalar));CHKERRQ(ierr); 895 896 for (i=0; i<am; i++) { 897 /* 2-a) put A[i,:] to dense array aval */ 898 anz = ai[i+1] - ai[i]; 899 adj = aj + ai[i]; 900 ada = a_loc->a + ai[i]; 901 for (j=0; j<anz; j++){ 902 aval[adj[j]] = ada[j]; 903 } 904 905 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 906 /*--------------------------------------------------------------*/ 907 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 908 pnz = po->i[i+1] - po->i[i]; 909 poJ = po->j + po->i[i]; 910 pA = po->a + po->i[i]; 911 for (j=0; j<pnz; j++){ 912 row = poJ[j]; 913 cnz = coi[row+1] - coi[row]; 914 cj = coj + coi[row]; 915 ca = coa + coi[row]; 916 /* perform dense axpy */ 917 valtmp = pA[j]; 918 for (k=0; k<cnz; k++) { 919 ca[k] += valtmp*aval[cj[k]]; 920 } 921 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 922 } 923 924 /* put the value into Cd (diagonal part) */ 925 pnz = pd->i[i+1] - pd->i[i]; 926 pdJ = pd->j + pd->i[i]; 927 pA = pd->a + pd->i[i]; 928 for (j=0; j<pnz; j++){ 929 row = pdJ[j]; 930 cnz = bi[row+1] - bi[row]; 931 cj = bj + bi[row]; 932 ca = ba + bi[row]; 933 /* perform dense axpy */ 934 valtmp = pA[j]; 935 for (k=0; k<cnz; k++) { 936 ca[k] += valtmp*aval[cj[k]]; 937 } 938 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 939 } 940 941 /* zero the current row of Pt*A */ 942 aJ = aj + ai[i]; 943 for (k=0; k<anz; k++) aval[aJ[k]] = 0.0; 944 } 945 946 /* 3) send and recv matrix values coa */ 947 /*------------------------------------*/ 948 buf_ri = merge->buf_ri; 949 buf_rj = merge->buf_rj; 950 len_s = merge->len_s; 951 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 952 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 953 954 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 955 for (proc=0,k=0; proc<size; proc++){ 956 if (!len_s[proc]) continue; 957 i = merge->owners_co[proc]; 958 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 959 k++; 960 } 961 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 962 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 963 964 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 965 ierr = PetscFree(r_waits);CHKERRQ(ierr); 966 ierr = PetscFree(coa);CHKERRQ(ierr); 967 968 /* 4) insert local Cseq and received values into Cmpi */ 969 /*----------------------------------------------------*/ 970 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 971 for (k=0; k<merge->nrecv; k++){ 972 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 973 nrows = *(buf_ri_k[k]); 974 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 975 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 976 } 977 978 for (i=0; i<cm; i++) { 979 row = owners[rank] + i; /* global row index of C_seq */ 980 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 981 ba_i = ba + bi[i]; 982 bnz = bi[i+1] - bi[i]; 983 /* add received vals into ba */ 984 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 985 /* i-th row */ 986 if (i == *nextrow[k]) { 987 cnz = *(nextci[k]+1) - *nextci[k]; 988 cj = buf_rj[k] + *(nextci[k]); 989 ca = abuf_r[k] + *(nextci[k]); 990 nextcj = 0; 991 for (j=0; nextcj<cnz; j++){ 992 if (bj_i[j] == cj[nextcj]){ /* bcol == ccol */ 993 ba_i[j] += ca[nextcj++]; 994 } 995 } 996 nextrow[k]++; nextci[k]++; 997 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 998 } 999 } 1000 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1001 } 1002 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1003 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1004 1005 ierr = PetscFree(ba);CHKERRQ(ierr); 1006 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1007 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1008 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1009 ierr = PetscFree(aval);CHKERRQ(ierr); 1010 PetscFunctionReturn(0); 1011 } 1012 1013 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1014 #undef __FUNCT__ 1015 #define __FUNCT__ "MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ" 1016 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat *C) 1017 { 1018 PetscErrorCode ierr; 1019 Mat Cmpi,A_loc,POt,PDt; 1020 Mat_PtAPMPI *ptap; 1021 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1022 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c; 1023 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1024 PetscInt nnz; 1025 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1026 PetscInt am=A->rmap->n,pn=P->cmap->n; 1027 PetscBT lnkbt; 1028 MPI_Comm comm=((PetscObject)A)->comm; 1029 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1030 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1031 PetscInt len,proc,*dnz,*onz,*owners; 1032 PetscInt nzi,*bi,*bj; 1033 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1034 MPI_Request *swaits,*rwaits; 1035 MPI_Status *sstatus,rstatus; 1036 Mat_Merge_SeqsToMPI *merge; 1037 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1038 PetscReal afill=1.0,afill_tmp; 1039 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Crmax; 1040 PetscScalar *vals; 1041 Mat_SeqAIJ *a_loc, *pdt,*pot; 1042 1043 PetscFunctionBegin; 1044 /* check if matrix local sizes are compatible */ 1045 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend){ 1046 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); 1047 } 1048 1049 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1050 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1051 1052 /* create struct Mat_PtAPMPI and attached it to C later */ 1053 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 1054 1055 /* get A_loc by taking all local rows of A */ 1056 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1057 ptap->A_loc = A_loc; 1058 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1059 ai = a_loc->i; 1060 aj = a_loc->j; 1061 1062 /* determine symbolic Co=(p->B)^T*A - send to others */ 1063 /*----------------------------------------------------*/ 1064 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1065 pdt = (Mat_SeqAIJ*)PDt->data; 1066 pdti = pdt->i; pdtj = pdt->j; 1067 1068 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1069 pot = (Mat_SeqAIJ*)POt->data; 1070 poti = pot->i; potj = pot->j; 1071 1072 /* then, compute symbolic Co = (p->B)^T*A */ 1073 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1074 >= (num of nonzero rows of C_seq) - pn */ 1075 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 1076 coi[0] = 0; 1077 1078 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1079 nnz = fill*(poti[pon] + ai[am]); 1080 ierr = PetscFreeSpaceGet(nnz,&free_space); 1081 current_space = free_space; 1082 1083 /* create and initialize a linked list */ 1084 i = PetscMax(pdt->rmax,pot->rmax); 1085 Crmax = i*a_loc->rmax*size; 1086 if (!Crmax || Crmax > aN) Crmax = aN; 1087 ierr = PetscLLCondensedCreate(Crmax,aN,&lnk,&lnkbt);CHKERRQ(ierr); 1088 1089 for (i=0; i<pon; i++) { 1090 pnz = poti[i+1] - poti[i]; 1091 ptJ = potj + poti[i]; 1092 for (j=0; j<pnz; j++){ 1093 row = ptJ[j]; /* row of A_loc == col of Pot */ 1094 anz = ai[row+1] - ai[row]; 1095 Jptr = aj + ai[row]; 1096 /* add non-zero cols of AP into the sorted linked list lnk */ 1097 ierr = PetscLLCondensedAddSorted(anz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1098 } 1099 nnz = lnk[0]; 1100 1101 /* If free space is not available, double the total space in the list */ 1102 if (current_space->local_remaining<nnz) { 1103 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1104 nspacedouble++; 1105 } 1106 1107 /* Copy data into free space, and zero out denserows */ 1108 ierr = PetscLLCondensedClean(aN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1109 current_space->array += nnz; 1110 current_space->local_used += nnz; 1111 current_space->local_remaining -= nnz; 1112 coi[i+1] = coi[i] + nnz; 1113 } 1114 1115 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 1116 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1117 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1118 if (afill_tmp > afill) afill = afill_tmp; 1119 1120 /* send j-array (coj) of Co to other processors */ 1121 /*----------------------------------------------*/ 1122 /* determine row ownership */ 1123 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 1124 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1125 merge->rowmap->n = pn; 1126 merge->rowmap->bs = 1; 1127 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1128 owners = merge->rowmap->range; 1129 1130 /* determine the number of messages to send, their lengths */ 1131 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 1132 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1133 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 1134 len_s = merge->len_s; 1135 merge->nsend = 0; 1136 1137 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 1138 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1139 1140 proc = 0; 1141 for (i=0; i<pon; i++){ 1142 while (prmap[i] >= owners[proc+1]) proc++; 1143 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1144 len_s[proc] += coi[i+1] - coi[i]; 1145 } 1146 1147 len = 0; /* max length of buf_si[] */ 1148 owners_co[0] = 0; 1149 for (proc=0; proc<size; proc++){ 1150 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1151 if (len_si[proc]){ 1152 merge->nsend++; 1153 len_si[proc] = 2*(len_si[proc] + 1); 1154 len += len_si[proc]; 1155 } 1156 } 1157 1158 /* determine the number and length of messages to receive for coi and coj */ 1159 ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1160 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1161 1162 /* post the Irecv and Isend of coj */ 1163 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1164 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1165 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 1166 for (proc=0, k=0; proc<size; proc++){ 1167 if (!len_s[proc]) continue; 1168 i = owners_co[proc]; 1169 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1170 k++; 1171 } 1172 1173 /* receives and sends of coj are complete */ 1174 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 1175 for (i=0; i<merge->nrecv; i++){ 1176 PetscMPIInt icompleted; 1177 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1178 } 1179 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1180 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1181 1182 /* send and recv coi */ 1183 /*-------------------*/ 1184 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1185 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1186 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 1187 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1188 for (proc=0,k=0; proc<size; proc++){ 1189 if (!len_s[proc]) continue; 1190 /* form outgoing message for i-structure: 1191 buf_si[0]: nrows to be sent 1192 [1:nrows]: row index (global) 1193 [nrows+1:2*nrows+1]: i-structure index 1194 */ 1195 /*-------------------------------------------*/ 1196 nrows = len_si[proc]/2 - 1; 1197 buf_si_i = buf_si + nrows+1; 1198 buf_si[0] = nrows; 1199 buf_si_i[0] = 0; 1200 nrows = 0; 1201 for (i=owners_co[proc]; i<owners_co[proc+1]; i++){ 1202 nzi = coi[i+1] - coi[i]; 1203 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1204 buf_si[nrows+1] =prmap[i] -owners[proc]; /* local row index */ 1205 nrows++; 1206 } 1207 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1208 k++; 1209 buf_si += len_si[proc]; 1210 } 1211 i = merge->nrecv; 1212 while (i--) { 1213 PetscMPIInt icompleted; 1214 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1215 } 1216 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1217 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1218 ierr = PetscFree(len_si);CHKERRQ(ierr); 1219 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1220 ierr = PetscFree(swaits);CHKERRQ(ierr); 1221 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1222 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1223 1224 /* compute the local portion of C (mpi mat) */ 1225 /*------------------------------------------*/ 1226 /* allocate bi array and free space for accumulating nonzero column info */ 1227 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1228 bi[0] = 0; 1229 1230 /* set initial free space to be fill*(nnz(P) + nnz(A)) */ 1231 nnz = fill*(pdti[pn] + poti[pon] + ai[am]); 1232 ierr = PetscFreeSpaceGet(nnz,&free_space); 1233 current_space = free_space; 1234 1235 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1236 for (k=0; k<merge->nrecv; k++){ 1237 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1238 nrows = *buf_ri_k[k]; 1239 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1240 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 1241 } 1242 1243 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1244 rmax = 0; 1245 for (i=0; i<pn; i++) { 1246 /* add pdt[i,:]*AP into lnk */ 1247 pnz = pdti[i+1] - pdti[i]; 1248 ptJ = pdtj + pdti[i]; 1249 for (j=0; j<pnz; j++){ 1250 row = ptJ[j]; /* row of AP == col of Pt */ 1251 anz = ai[row+1] - ai[row]; 1252 Jptr = aj + ai[row]; 1253 /* add non-zero cols of AP into the sorted linked list lnk */ 1254 ierr = PetscLLCondensedAddSorted(anz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1255 } 1256 1257 /* add received col data into lnk */ 1258 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 1259 if (i == *nextrow[k]) { /* i-th row */ 1260 nzi = *(nextci[k]+1) - *nextci[k]; 1261 Jptr = buf_rj[k] + *nextci[k]; 1262 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1263 nextrow[k]++; nextci[k]++; 1264 } 1265 } 1266 nnz = lnk[0]; 1267 1268 /* if free space is not available, make more free space */ 1269 if (current_space->local_remaining<nnz) { 1270 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1271 nspacedouble++; 1272 } 1273 /* copy data into free space, then initialize lnk */ 1274 ierr = PetscLLCondensedClean(aN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1275 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 1276 current_space->array += nnz; 1277 current_space->local_used += nnz; 1278 current_space->local_remaining -= nnz; 1279 bi[i+1] = bi[i] + nnz; 1280 if (nnz > rmax) rmax = nnz; 1281 } 1282 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1283 1284 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1285 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 1286 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 1287 if (afill_tmp > afill) afill = afill_tmp; 1288 ierr = PetscLLCondensedDestroy(lnk,lnkbt);CHKERRQ(ierr); 1289 ierr = MatDestroy(&POt);CHKERRQ(ierr); 1290 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 1291 1292 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 1293 /*----------------------------------------------------------------------------------*/ 1294 ierr = PetscMalloc((rmax+1)*sizeof(PetscScalar),&vals);CHKERRQ(ierr); 1295 ierr = PetscMemzero(vals,rmax*sizeof(PetscScalar));CHKERRQ(ierr); 1296 1297 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1298 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1299 ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,A->cmap->bs);CHKERRQ(ierr); 1300 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1301 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1302 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1303 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 1304 for (i=0; i<pn; i++){ 1305 row = i + rstart; 1306 nnz = bi[i+1] - bi[i]; 1307 Jptr = bj + bi[i]; 1308 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 1309 } 1310 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1311 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1312 ierr = PetscFree(vals);CHKERRQ(ierr); 1313 1314 merge->bi = bi; 1315 merge->bj = bj; 1316 merge->coi = coi; 1317 merge->coj = coj; 1318 merge->buf_ri = buf_ri; 1319 merge->buf_rj = buf_rj; 1320 merge->owners_co = owners_co; 1321 merge->destroy = Cmpi->ops->destroy; 1322 merge->duplicate = Cmpi->ops->duplicate; 1323 1324 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 1325 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1326 1327 /* attach the supporting struct to Cmpi for reuse */ 1328 c = (Mat_MPIAIJ*)Cmpi->data; 1329 c->ptap = ptap; 1330 ptap->api = PETSC_NULL; 1331 ptap->apj = PETSC_NULL; 1332 ptap->merge = merge; 1333 ptap->rmax = rmax; 1334 1335 *C = Cmpi; 1336 #if defined(PETSC_USE_INFO) 1337 if (bi[pn] != 0) { 1338 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 1339 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 1340 } else { 1341 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1342 } 1343 #endif 1344 PetscFunctionReturn(0); 1345 } 1346 1347 #undef __FUNCT__ 1348 #define __FUNCT__ "MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_Scalable" 1349 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_Scalable(Mat P,Mat A,Mat C) 1350 { 1351 PetscErrorCode ierr; 1352 Mat_Merge_SeqsToMPI *merge; 1353 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1354 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1355 Mat_PtAPMPI *ptap; 1356 PetscInt *adj; 1357 PetscInt i,j,k,anz,pnz,row,*cj,nexta; 1358 MatScalar *ada,*ca,valtmp; 1359 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1360 MPI_Comm comm=((PetscObject)C)->comm; 1361 PetscMPIInt size,rank,taga,*len_s; 1362 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1363 PetscInt **buf_ri,**buf_rj; 1364 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1365 MPI_Request *s_waits,*r_waits; 1366 MPI_Status *status; 1367 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1368 PetscInt *ai,*aj,*coi,*coj; 1369 PetscInt *poJ=po->j,*pdJ=pd->j; 1370 Mat A_loc; 1371 Mat_SeqAIJ *a_loc; 1372 1373 PetscFunctionBegin; 1374 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1375 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1376 1377 ptap = c->ptap; 1378 merge = ptap->merge; 1379 1380 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1381 /*------------------------------------------*/ 1382 /* get data from symbolic products */ 1383 coi = merge->coi; coj = merge->coj; 1384 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 1385 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 1386 bi = merge->bi; bj = merge->bj; 1387 owners = merge->rowmap->range; 1388 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 1389 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 1390 1391 /* get A_loc by taking all local rows of A */ 1392 A_loc = ptap->A_loc; 1393 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 1394 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1395 ai = a_loc->i; 1396 aj = a_loc->j; 1397 1398 for (i=0; i<am; i++) { 1399 /* 2-a) put A[i,:] to dense array aval */ 1400 anz = ai[i+1] - ai[i]; 1401 adj = aj + ai[i]; 1402 ada = a_loc->a + ai[i]; 1403 1404 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1405 /*-------------------------------------------------------------*/ 1406 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1407 pnz = po->i[i+1] - po->i[i]; 1408 poJ = po->j + po->i[i]; 1409 pA = po->a + po->i[i]; 1410 for (j=0; j<pnz; j++){ 1411 row = poJ[j]; 1412 cnz = coi[row+1] - coi[row]; 1413 cj = coj + coi[row]; 1414 ca = coa + coi[row]; 1415 /* perform sparse axpy */ 1416 nexta = 0; 1417 valtmp = pA[j]; 1418 for (k=0; nexta<anz; k++) { 1419 if (cj[k] == adj[nexta]){ 1420 ca[k] += valtmp*ada[nexta]; 1421 nexta++; 1422 } 1423 } 1424 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1425 } 1426 1427 /* put the value into Cd (diagonal part) */ 1428 pnz = pd->i[i+1] - pd->i[i]; 1429 pdJ = pd->j + pd->i[i]; 1430 pA = pd->a + pd->i[i]; 1431 for (j=0; j<pnz; j++){ 1432 row = pdJ[j]; 1433 cnz = bi[row+1] - bi[row]; 1434 cj = bj + bi[row]; 1435 ca = ba + bi[row]; 1436 /* perform sparse axpy */ 1437 nexta = 0; 1438 valtmp = pA[j]; 1439 for (k=0; nexta<anz; k++) { 1440 if (cj[k] == adj[nexta]){ 1441 ca[k] += valtmp*ada[nexta]; 1442 nexta++; 1443 } 1444 } 1445 ierr = PetscLogFlops(2.0*anz);CHKERRQ(ierr); 1446 } 1447 1448 } 1449 1450 /* 3) send and recv matrix values coa */ 1451 /*------------------------------------*/ 1452 buf_ri = merge->buf_ri; 1453 buf_rj = merge->buf_rj; 1454 len_s = merge->len_s; 1455 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1456 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1457 1458 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 1459 for (proc=0,k=0; proc<size; proc++){ 1460 if (!len_s[proc]) continue; 1461 i = merge->owners_co[proc]; 1462 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1463 k++; 1464 } 1465 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1466 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1467 1468 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1469 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1470 ierr = PetscFree(coa);CHKERRQ(ierr); 1471 1472 /* 4) insert local Cseq and received values into Cmpi */ 1473 /*----------------------------------------------------*/ 1474 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1475 for (k=0; k<merge->nrecv; k++){ 1476 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1477 nrows = *(buf_ri_k[k]); 1478 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1479 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 1480 } 1481 1482 for (i=0; i<cm; i++) { 1483 row = owners[rank] + i; /* global row index of C_seq */ 1484 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1485 ba_i = ba + bi[i]; 1486 bnz = bi[i+1] - bi[i]; 1487 /* add received vals into ba */ 1488 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 1489 /* i-th row */ 1490 if (i == *nextrow[k]) { 1491 cnz = *(nextci[k]+1) - *nextci[k]; 1492 cj = buf_rj[k] + *(nextci[k]); 1493 ca = abuf_r[k] + *(nextci[k]); 1494 nextcj = 0; 1495 for (j=0; nextcj<cnz; j++){ 1496 if (bj_i[j] == cj[nextcj]){ /* bcol == ccol */ 1497 ba_i[j] += ca[nextcj++]; 1498 } 1499 } 1500 nextrow[k]++; nextci[k]++; 1501 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1502 } 1503 } 1504 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1505 } 1506 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1507 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1508 1509 ierr = PetscFree(ba);CHKERRQ(ierr); 1510 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1511 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1512 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1513 PetscFunctionReturn(0); 1514 } 1515 1516 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1517 #undef __FUNCT__ 1518 #define __FUNCT__ "MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable" 1519 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(Mat P,Mat A,PetscReal fill,Mat *C) 1520 { 1521 PetscErrorCode ierr; 1522 Mat Cmpi,A_loc,POt,PDt; 1523 Mat_PtAPMPI *ptap; 1524 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1525 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c; 1526 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1527 PetscInt nnz; 1528 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1529 PetscInt am=A->rmap->n,pn=P->cmap->n; 1530 MPI_Comm comm=((PetscObject)A)->comm; 1531 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1532 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1533 PetscInt len,proc,*dnz,*onz,*owners; 1534 PetscInt nzi,*bi,*bj; 1535 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1536 MPI_Request *swaits,*rwaits; 1537 MPI_Status *sstatus,rstatus; 1538 Mat_Merge_SeqsToMPI *merge; 1539 PetscInt *ai,*aj,*Jptr,anz,*prmap=p->garray,pon,nspacedouble=0,j; 1540 PetscReal afill=1.0,afill_tmp; 1541 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N,Crmax; 1542 PetscScalar *vals; 1543 Mat_SeqAIJ *a_loc, *pdt,*pot; 1544 1545 PetscFunctionBegin; 1546 /* check if matrix local sizes are compatible */ 1547 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend){ 1548 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); 1549 } 1550 1551 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1552 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1553 1554 /* create struct Mat_PtAPMPI and attached it to C later */ 1555 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 1556 1557 /* get A_loc by taking all local rows of A */ 1558 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1559 ptap->A_loc = A_loc; 1560 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1561 ai = a_loc->i; 1562 aj = a_loc->j; 1563 1564 /* determine symbolic Co=(p->B)^T*A - send to others */ 1565 /*----------------------------------------------------*/ 1566 ierr = MatTransposeSymbolic_SeqAIJ(p->A,&PDt);CHKERRQ(ierr); 1567 pdt = (Mat_SeqAIJ*)PDt->data; 1568 pdti = pdt->i; pdtj = pdt->j; 1569 1570 ierr = MatTransposeSymbolic_SeqAIJ(p->B,&POt);CHKERRQ(ierr); 1571 pot = (Mat_SeqAIJ*)POt->data; 1572 poti = pot->i; potj = pot->j; 1573 1574 /* then, compute symbolic Co = (p->B)^T*A */ 1575 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1576 >= (num of nonzero rows of C_seq) - pn */ 1577 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 1578 coi[0] = 0; 1579 1580 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1581 nnz = fill*(poti[pon] + ai[am]); 1582 ierr = PetscFreeSpaceGet(nnz,&free_space); 1583 current_space = free_space; 1584 1585 /* create and initialize a linked list */ 1586 i = PetscMax(pdt->rmax,pot->rmax); 1587 Crmax = i*a_loc->rmax*size; /* non-scalable! */ 1588 if (!Crmax || Crmax > aN) Crmax = aN; 1589 ierr = PetscLLCondensedCreate_Scalable(Crmax,&lnk);CHKERRQ(ierr); 1590 1591 for (i=0; i<pon; i++) { 1592 nnz = 0; 1593 pnz = poti[i+1] - poti[i]; 1594 ptJ = potj + poti[i]; 1595 for (j=0; j<pnz; j++){ 1596 row = ptJ[j]; /* row of A_loc == col of Pot */ 1597 anz = ai[row+1] - ai[row]; 1598 Jptr = aj + ai[row]; 1599 /* add non-zero cols of AP into the sorted linked list lnk */ 1600 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1601 } 1602 nnz = lnk[0]; 1603 1604 /* If free space is not available, double the total space in the list */ 1605 if (current_space->local_remaining<nnz) { 1606 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1607 nspacedouble++; 1608 } 1609 1610 /* Copy data into free space, and zero out denserows */ 1611 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1612 current_space->array += nnz; 1613 current_space->local_used += nnz; 1614 current_space->local_remaining -= nnz; 1615 coi[i+1] = coi[i] + nnz; 1616 } 1617 1618 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 1619 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1620 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + ai[am]+1); 1621 if (afill_tmp > afill) afill = afill_tmp; 1622 1623 /* send j-array (coj) of Co to other processors */ 1624 /*----------------------------------------------*/ 1625 /* determine row ownership */ 1626 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 1627 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1628 merge->rowmap->n = pn; 1629 merge->rowmap->bs = 1; 1630 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1631 owners = merge->rowmap->range; 1632 1633 /* determine the number of messages to send, their lengths */ 1634 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 1635 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1636 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 1637 len_s = merge->len_s; 1638 merge->nsend = 0; 1639 1640 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 1641 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1642 1643 proc = 0; 1644 for (i=0; i<pon; i++){ 1645 while (prmap[i] >= owners[proc+1]) proc++; 1646 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1647 len_s[proc] += coi[i+1] - coi[i]; 1648 } 1649 1650 len = 0; /* max length of buf_si[] */ 1651 owners_co[0] = 0; 1652 for (proc=0; proc<size; proc++){ 1653 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1654 if (len_si[proc]){ 1655 merge->nsend++; 1656 len_si[proc] = 2*(len_si[proc] + 1); 1657 len += len_si[proc]; 1658 } 1659 } 1660 1661 /* determine the number and length of messages to receive for coi and coj */ 1662 ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1663 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1664 1665 /* post the Irecv and Isend of coj */ 1666 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1667 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1668 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 1669 for (proc=0, k=0; proc<size; proc++){ 1670 if (!len_s[proc]) continue; 1671 i = owners_co[proc]; 1672 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1673 k++; 1674 } 1675 1676 /* receives and sends of coj are complete */ 1677 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 1678 for (i=0; i<merge->nrecv; i++){ 1679 PetscMPIInt icompleted; 1680 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1681 } 1682 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1683 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1684 1685 /* send and recv coi */ 1686 /*-------------------*/ 1687 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1688 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1689 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 1690 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1691 for (proc=0,k=0; proc<size; proc++){ 1692 if (!len_s[proc]) continue; 1693 /* form outgoing message for i-structure: 1694 buf_si[0]: nrows to be sent 1695 [1:nrows]: row index (global) 1696 [nrows+1:2*nrows+1]: i-structure index 1697 */ 1698 /*-------------------------------------------*/ 1699 nrows = len_si[proc]/2 - 1; 1700 buf_si_i = buf_si + nrows+1; 1701 buf_si[0] = nrows; 1702 buf_si_i[0] = 0; 1703 nrows = 0; 1704 for (i=owners_co[proc]; i<owners_co[proc+1]; i++){ 1705 nzi = coi[i+1] - coi[i]; 1706 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1707 buf_si[nrows+1] =prmap[i] -owners[proc]; /* local row index */ 1708 nrows++; 1709 } 1710 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1711 k++; 1712 buf_si += len_si[proc]; 1713 } 1714 i = merge->nrecv; 1715 while (i--) { 1716 PetscMPIInt icompleted; 1717 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1718 } 1719 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1720 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1721 ierr = PetscFree(len_si);CHKERRQ(ierr); 1722 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1723 ierr = PetscFree(swaits);CHKERRQ(ierr); 1724 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1725 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1726 1727 /* compute the local portion of C (mpi mat) */ 1728 /*------------------------------------------*/ 1729 /* allocate bi array and free space for accumulating nonzero column info */ 1730 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1731 bi[0] = 0; 1732 1733 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1734 nnz = fill*(pdti[pn] + poti[pon] + ai[am]); 1735 ierr = PetscFreeSpaceGet(nnz,&free_space); 1736 current_space = free_space; 1737 1738 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1739 for (k=0; k<merge->nrecv; k++){ 1740 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1741 nrows = *buf_ri_k[k]; 1742 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1743 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 1744 } 1745 1746 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1747 rmax = 0; 1748 for (i=0; i<pn; i++) { 1749 /* add pdt[i,:]*AP into lnk */ 1750 pnz = pdti[i+1] - pdti[i]; 1751 ptJ = pdtj + pdti[i]; 1752 for (j=0; j<pnz; j++){ 1753 row = ptJ[j]; /* row of AP == col of Pt */ 1754 anz = ai[row+1] - ai[row]; 1755 Jptr = aj + ai[row]; 1756 /* add non-zero cols of AP into the sorted linked list lnk */ 1757 ierr = PetscLLCondensedAddSorted_Scalable(anz,Jptr,lnk);CHKERRQ(ierr); 1758 } 1759 1760 /* add received col data into lnk */ 1761 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 1762 if (i == *nextrow[k]) { /* i-th row */ 1763 nzi = *(nextci[k]+1) - *nextci[k]; 1764 Jptr = buf_rj[k] + *nextci[k]; 1765 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 1766 nextrow[k]++; nextci[k]++; 1767 } 1768 } 1769 nnz = lnk[0]; 1770 1771 /* if free space is not available, make more free space */ 1772 if (current_space->local_remaining<nnz) { 1773 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1774 nspacedouble++; 1775 } 1776 /* copy data into free space, then initialize lnk */ 1777 ierr = PetscLLCondensedClean_Scalable(nnz,current_space->array,lnk);CHKERRQ(ierr); 1778 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 1779 current_space->array += nnz; 1780 current_space->local_used += nnz; 1781 current_space->local_remaining -= nnz; 1782 bi[i+1] = bi[i] + nnz; 1783 if (nnz > rmax) rmax = nnz; 1784 } 1785 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1786 1787 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1788 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 1789 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + ai[am]+1); 1790 if (afill_tmp > afill) afill = afill_tmp; 1791 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 1792 ierr = MatDestroy(&POt);CHKERRQ(ierr); 1793 ierr = MatDestroy(&PDt);CHKERRQ(ierr); 1794 1795 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 1796 /*----------------------------------------------------------------------------------*/ 1797 ierr = PetscMalloc((rmax+1)*sizeof(PetscScalar),&vals);CHKERRQ(ierr); 1798 ierr = PetscMemzero(vals,rmax*sizeof(PetscScalar));CHKERRQ(ierr); 1799 1800 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1801 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1802 ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,A->cmap->bs); CHKERRQ(ierr); 1803 ierr = MatSetType(Cmpi,MATMPIAIJ); CHKERRQ(ierr); 1804 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1805 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1806 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 1807 for (i=0; i<pn; i++){ 1808 row = i + rstart; 1809 nnz = bi[i+1] - bi[i]; 1810 Jptr = bj + bi[i]; 1811 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 1812 } 1813 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1814 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1815 ierr = PetscFree(vals);CHKERRQ(ierr); 1816 1817 merge->bi = bi; 1818 merge->bj = bj; 1819 merge->coi = coi; 1820 merge->coj = coj; 1821 merge->buf_ri = buf_ri; 1822 merge->buf_rj = buf_rj; 1823 merge->owners_co = owners_co; 1824 merge->destroy = Cmpi->ops->destroy; 1825 merge->duplicate = Cmpi->ops->duplicate; 1826 1827 Cmpi->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_Scalable; 1828 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1829 1830 /* attach the supporting struct to Cmpi for reuse */ 1831 c = (Mat_MPIAIJ*)Cmpi->data; 1832 c->ptap = ptap; 1833 ptap->api = PETSC_NULL; 1834 ptap->apj = PETSC_NULL; 1835 ptap->merge = merge; 1836 ptap->rmax = rmax; 1837 ptap->apa = PETSC_NULL; 1838 1839 *C = Cmpi; 1840 #if defined(PETSC_USE_INFO) 1841 if (bi[pn] != 0) { 1842 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 1843 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 1844 } else { 1845 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1846 } 1847 #endif 1848 PetscFunctionReturn(0); 1849 } 1850