1 2 /* 3 Defines projective product routines where A is a MPIAIJ matrix 4 C = P^T * A * P 5 */ 6 7 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 8 #include <../src/mat/utils/freespace.h> 9 #include <../src/mat/impls/aij/mpi/mpiaij.h> 10 #include <petscbt.h> 11 #include <petsctime.h> 12 13 #define PTAP_PROFILE 14 15 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 16 #undef __FUNCT__ 17 #define __FUNCT__ "MatDestroy_MPIAIJ_PtAP" 18 PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat A) 19 { 20 PetscErrorCode ierr; 21 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 22 Mat_PtAPMPI *ptap=a->ptap; 23 24 PetscFunctionBegin; 25 if (ptap) { 26 Mat_Merge_SeqsToMPI *merge=ptap->merge; 27 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 28 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 29 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 30 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 31 ierr = MatDestroy(&ptap->A_loc);CHKERRQ(ierr); /* used by MatTransposeMatMult() */ 32 33 ierr = MatDestroy(&ptap->Rd);CHKERRQ(ierr); 34 ierr = MatDestroy(&ptap->Ro);CHKERRQ(ierr); 35 ierr = MatDestroy(&ptap->AP_loc);CHKERRQ(ierr); 36 ierr = MatDestroy(&ptap->C_loc);CHKERRQ(ierr); 37 ierr = MatDestroy(&ptap->C_oth);CHKERRQ(ierr); 38 39 if (ptap->api) {ierr = PetscFree(ptap->api);CHKERRQ(ierr);} 40 if (ptap->apj) {ierr = PetscFree(ptap->apj);CHKERRQ(ierr);} 41 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 42 if (merge) { 43 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 44 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 45 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 46 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 47 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 48 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 49 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 50 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 51 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 52 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 53 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 54 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 55 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 56 ierr = merge->destroy(A);CHKERRQ(ierr); 57 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 58 } 59 ierr = PetscFree(ptap);CHKERRQ(ierr); 60 } 61 PetscFunctionReturn(0); 62 } 63 64 #undef __FUNCT__ 65 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP" 66 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 67 { 68 PetscErrorCode ierr; 69 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 70 Mat_PtAPMPI *ptap = a->ptap; 71 Mat_Merge_SeqsToMPI *merge = ptap->merge; 72 73 PetscFunctionBegin; 74 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 75 76 (*M)->ops->destroy = merge->destroy; 77 (*M)->ops->duplicate = merge->duplicate; 78 PetscFunctionReturn(0); 79 } 80 81 #undef __FUNCT__ 82 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ" 83 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 84 { 85 PetscErrorCode ierr; 86 PetscBool newalg=PETSC_FALSE; 87 88 PetscFunctionBegin; 89 ierr = PetscOptionsGetBool(NULL,"-matptap_new",&newalg,NULL);CHKERRQ(ierr); 90 if (scall == MAT_INITIAL_MATRIX) { 91 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 92 if (newalg) { 93 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ_new(A,P,fill,C);CHKERRQ(ierr); 94 } else { 95 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 96 } 97 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 98 } 99 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 100 if (newalg) { 101 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ_new(A,P,*C);CHKERRQ(ierr); 102 } else { 103 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr); 104 } 105 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 106 PetscFunctionReturn(0); 107 } 108 109 #undef __FUNCT__ 110 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ_new" 111 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_new(Mat A,Mat P,PetscReal fill,Mat *C) 112 { 113 PetscErrorCode ierr; 114 Mat_PtAPMPI *ptap; 115 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 116 Mat AP; 117 Mat_MPIAIJ *c; 118 MPI_Comm comm; 119 PetscMPIInt size,rank; 120 PetscLogDouble t0,t1,t2,t3,t4; 121 122 PetscFunctionBegin; 123 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 124 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 125 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 126 127 /* check if matrix local sizes are compatible -- MV! */ 128 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 129 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%D, %D) != Prow (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 130 } 131 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 132 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%D, %D) != Prow (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 133 } 134 135 //ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); //rm later !!! 136 //============================================================================= 137 Mat Cmpi; 138 PetscFreeSpaceList free_space=NULL,current_space=NULL; 139 Mat_SeqAIJ *p_loc; 140 PetscInt *pi_loc; 141 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ,nnz; 142 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 143 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 144 PetscBT lnkbt; 145 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 146 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 147 PetscInt len,proc,*dnz,*onz,*owners; 148 PetscInt nzi,*pti,*ptj; 149 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 150 MPI_Request *swaits,*rwaits; 151 MPI_Status *sstatus,rstatus; 152 Mat_Merge_SeqsToMPI *merge; 153 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 154 PetscReal afill=1.0,afill_tmp; 155 PetscInt rmax; 156 157 ierr = PetscTime(&t0);CHKERRQ(ierr); 158 /* create struct Mat_PtAPMPI and attached it to C later */ 159 ierr = PetscNew(&ptap);CHKERRQ(ierr); 160 ierr = PetscNew(&merge);CHKERRQ(ierr); 161 ptap->merge = merge; 162 ptap->reuse = MAT_INITIAL_MATRIX; 163 164 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 165 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 166 167 /* get P_loc by taking all local rows of P */ 168 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 169 170 ptap->reuse = MAT_INITIAL_MATRIX; 171 172 /* (1) compute symbolic AP = A*P, then get AP_loc */ 173 /*--------------------------------------------------------------------------*/ 174 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 175 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 176 177 ierr = MatMatMult(A,P,MAT_INITIAL_MATRIX,2.0,&AP);CHKERRQ(ierr); 178 ierr = MatMPIAIJGetLocalMat(AP,MAT_INITIAL_MATRIX,&ptap->AP_loc);CHKERRQ(ierr); 179 ierr = MatDestroy(&AP);CHKERRQ(ierr); 180 181 /* (2) compute C_loc=Rd*AP_loc, Co=Ro*AP_loc */ 182 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,MAT_INITIAL_MATRIX,2.0,&ptap->C_loc);CHKERRQ(ierr); 183 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,MAT_INITIAL_MATRIX,2.0,&ptap->C_oth);CHKERRQ(ierr); 184 ierr = PetscTime(&t1);CHKERRQ(ierr); 185 186 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 187 pi_loc = p_loc->i; 188 189 Mat_SeqAIJ *ap = (Mat_SeqAIJ*)ptap->AP_loc->data; 190 api = ap->i; 191 apj = ap->j; 192 193 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 194 /*-----------------------------------------------------------------*/ 195 Mat_SeqAIJ *po = (Mat_SeqAIJ *)ptap->Ro->data; 196 poti = po->i; potj = po->j; 197 198 /* then, compute symbolic Co = (p->B)^T*AP */ 199 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 200 >= (num of nonzero rows of C_seq) - pn */ 201 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 202 coi[0] = 0; 203 204 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 205 nnz = fill*(poti[pon] + api[am]); 206 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 207 current_space = free_space; 208 209 /* create and initialize a linked list */ 210 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 211 212 for (i=0; i<pon; i++) { 213 pnz = poti[i+1] - poti[i]; 214 ptJ = potj + poti[i]; 215 for (j=0; j<pnz; j++) { 216 row = ptJ[j]; /* row of AP == col of Pot */ 217 apnz = api[row+1] - api[row]; 218 Jptr = apj + api[row]; 219 /* add non-zero cols of AP into the sorted linked list lnk */ 220 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 221 } 222 nnz = lnk[0]; 223 224 /* If free space is not available, double the total space in the list */ 225 if (current_space->local_remaining<nnz) { 226 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 227 nspacedouble++; 228 } 229 230 /* Copy data into free space, and zero out denserows */ 231 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 232 233 current_space->array += nnz; 234 current_space->local_used += nnz; 235 current_space->local_remaining -= nnz; 236 237 coi[i+1] = coi[i] + nnz; 238 } 239 240 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 241 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 242 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 243 if (afill_tmp > afill) afill = afill_tmp; 244 245 /* (3) send j-array (coj) of Co to other processors */ 246 /*--------------------------------------------------*/ 247 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 248 len_s = merge->len_s; 249 merge->nsend = 0; 250 251 /* determine row ownership */ 252 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 253 merge->rowmap->n = pn; 254 merge->rowmap->bs = 1; 255 256 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 257 owners = merge->rowmap->range; 258 259 /* determine the number of messages to send, their lengths */ 260 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 261 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 262 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 263 264 proc = 0; 265 for (i=0; i<pon; i++) { 266 while (prmap[i] >= owners[proc+1]) proc++; 267 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 268 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 269 } 270 271 len = 0; /* max length of buf_si[], see (4) */ 272 owners_co[0] = 0; 273 for (proc=0; proc<size; proc++) { 274 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 275 if (len_s[proc]) { 276 merge->nsend++; 277 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 278 len += len_si[proc]; 279 } 280 } 281 282 /* determine the number and length of messages to receive for coi and coj */ 283 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 284 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 285 286 /* post the Irecv and Isend of coj */ 287 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 288 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 289 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 290 for (proc=0, k=0; proc<size; proc++) { 291 if (!len_s[proc]) continue; 292 i = owners_co[proc]; 293 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 294 k++; 295 } 296 297 /* receives and sends of coj are complete */ 298 for (i=0; i<merge->nrecv; i++) { 299 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 300 } 301 ierr = PetscFree(rwaits);CHKERRQ(ierr); 302 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 303 304 /* (4) send and recv coi */ 305 /*-----------------------*/ 306 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 307 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 308 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 309 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 310 for (proc=0,k=0; proc<size; proc++) { 311 if (!len_s[proc]) continue; 312 /* form outgoing message for i-structure: 313 buf_si[0]: nrows to be sent 314 [1:nrows]: row index (global) 315 [nrows+1:2*nrows+1]: i-structure index 316 */ 317 /*-------------------------------------------*/ 318 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 319 buf_si_i = buf_si + nrows+1; 320 buf_si[0] = nrows; 321 buf_si_i[0] = 0; 322 nrows = 0; 323 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 324 nzi = coi[i+1] - coi[i]; 325 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 326 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 327 nrows++; 328 } 329 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 330 k++; 331 buf_si += len_si[proc]; 332 } 333 i = merge->nrecv; 334 while (i--) { 335 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 336 } 337 ierr = PetscFree(rwaits);CHKERRQ(ierr); 338 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 339 340 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 341 ierr = PetscFree(len_ri);CHKERRQ(ierr); 342 ierr = PetscFree(swaits);CHKERRQ(ierr); 343 ierr = PetscFree(buf_s);CHKERRQ(ierr); 344 345 ierr = PetscTime(&t2);CHKERRQ(ierr); 346 347 /* (5) compute the local portion of C (mpi mat) */ 348 /*----------------------------------------------*/ 349 Mat_SeqAIJ *pd = (Mat_SeqAIJ *)ptap->Rd->data; 350 pdti = pd->i; pdtj = pd->j; 351 352 /* allocate pti array and free space for accumulating nonzero column info */ 353 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 354 pti[0] = 0; 355 356 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 357 nnz = fill*(pi_loc[pm] + api[am]); 358 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 359 current_space = free_space; 360 361 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 362 for (k=0; k<merge->nrecv; k++) { 363 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 364 nrows = *buf_ri_k[k]; 365 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 366 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 367 } 368 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 369 rmax = 0; 370 for (i=0; i<pn; i++) { 371 /* add pdt[i,:]*AP into lnk */ 372 pnz = pdti[i+1] - pdti[i]; 373 ptJ = pdtj + pdti[i]; 374 for (j=0; j<pnz; j++) { 375 row = ptJ[j]; /* row of AP == col of Pt */ 376 apnz = api[row+1] - api[row]; 377 Jptr = apj + api[row]; 378 /* add non-zero cols of AP into the sorted linked list lnk */ 379 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 380 } 381 382 /* add received col data into lnk */ 383 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 384 if (i == *nextrow[k]) { /* i-th row */ 385 nzi = *(nextci[k]+1) - *nextci[k]; 386 Jptr = buf_rj[k] + *nextci[k]; 387 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 388 nextrow[k]++; nextci[k]++; 389 } 390 } 391 nnz = lnk[0]; 392 393 /* if free space is not available, make more free space */ 394 if (current_space->local_remaining<nnz) { 395 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 396 nspacedouble++; 397 } 398 /* copy data into free space, then initialize lnk */ 399 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 400 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 401 402 current_space->array += nnz; 403 current_space->local_used += nnz; 404 current_space->local_remaining -= nnz; 405 406 pti[i+1] = pti[i] + nnz; 407 if (nnz > rmax) rmax = nnz; 408 } 409 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 410 411 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 412 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 413 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 414 if (afill_tmp > afill) afill = afill_tmp; 415 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 416 417 ierr = PetscTime(&t3);CHKERRQ(ierr); 418 419 /* (6) create symbolic parallel matrix Cmpi */ 420 /*------------------------------------------*/ 421 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 422 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 423 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 424 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 425 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 426 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 427 428 merge->bi = pti; /* Cseq->i */ 429 merge->bj = ptj; /* Cseq->j */ 430 merge->coi = coi; /* Co->i */ 431 merge->coj = coj; /* Co->j */ 432 merge->buf_ri = buf_ri; 433 merge->buf_rj = buf_rj; 434 merge->owners_co = owners_co; 435 merge->destroy = Cmpi->ops->destroy; 436 merge->duplicate = Cmpi->ops->duplicate; 437 438 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 439 Cmpi->assembled = PETSC_FALSE; 440 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 441 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 442 443 /* attach the supporting struct to Cmpi for reuse */ 444 c = (Mat_MPIAIJ*)Cmpi->data; 445 c->ptap = ptap; 446 ptap->api = NULL; 447 ptap->apj = NULL; 448 ptap->rmax = ap_rmax; 449 *C = Cmpi; 450 451 /* flag 'scalable' determines which implementations to be used: 452 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 453 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 454 /* set default scalable */ 455 ptap->scalable = PETSC_FALSE; //PETSC_TRUE; 456 457 ierr = PetscOptionsGetBool(((PetscObject)(*C))->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 458 if (!ptap->scalable) { /* Do dense axpy */ 459 ierr = PetscCalloc1(P->cmap->N,&ptap->apa);CHKERRQ(ierr); 460 } else { 461 //ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 462 } 463 464 ierr = PetscTime(&t4);CHKERRQ(ierr); 465 if (rank == 1) printf("PtAPSym: %g + %g + %g = %g\n",t1-t0,t2-t1,t3-t2,t4-t3); 466 467 PetscFunctionReturn(0); 468 } 469 470 #undef __FUNCT__ 471 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ_new" 472 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_new(Mat A,Mat P,Mat C) 473 { 474 PetscErrorCode ierr; 475 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 476 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 477 Mat_PtAPMPI *ptap = c->ptap; 478 Mat AP_loc,C_loc,C_oth; 479 PetscInt i,rstart,rend,cm,ncols,row; 480 PetscMPIInt rank; 481 MPI_Comm comm; 482 const PetscInt *cols; 483 const PetscScalar *vals; 484 PetscLogDouble t0,t1,t2,t3,t4,eR,eAP,eCseq,eCmpi; 485 486 PetscFunctionBegin; 487 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 488 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 489 490 ierr = MatZeroEntries(C);CHKERRQ(ierr); 491 492 /* 1) get R = Pd^T,Ro = Po^T */ 493 ierr = PetscTime(&t0);CHKERRQ(ierr); 494 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 495 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 496 ierr = PetscTime(&t1);CHKERRQ(ierr); 497 eR = t1 - t0; 498 499 /* 2) get AP_loc */ 500 AP_loc = ptap->AP_loc; 501 Mat_SeqAIJ *ap=(Mat_SeqAIJ*)AP_loc->data; 502 503 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 504 /*-----------------------------------------------------*/ 505 if (ptap->reuse == MAT_INITIAL_MATRIX) { 506 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 507 ptap->reuse = MAT_REUSE_MATRIX; 508 } else { /* update numerical values of P_oth and P_loc */ 509 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 510 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 511 } 512 513 /* 2-2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 514 /*--------------------------------------------------------------*/ 515 /* get data from symbolic products */ 516 Mat_SeqAIJ *p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 517 Mat_SeqAIJ *p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 518 PetscInt *api,*apj,am = A->rmap->n,j,col,apnz; 519 PetscScalar *apa = ptap->apa; 520 521 api = ap->i; 522 apj = ap->j; 523 for (i=0; i<am; i++) { 524 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 525 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 526 apnz = api[i+1] - api[i]; 527 for (j=0; j<apnz; j++) { 528 col = apj[j+api[i]]; 529 ap->a[j+ap->i[i]] = apa[col]; 530 apa[col] = 0.0; 531 } 532 } 533 534 ierr = PetscTime(&t2);CHKERRQ(ierr); 535 eAP = t2 - t1; 536 537 /* 3) C_loc = R*AP_loc, Co = Ro*AP_loc */ 538 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Rd,AP_loc,MAT_REUSE_MATRIX,2.0,&ptap->C_loc);CHKERRQ(ierr); 539 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Ro,AP_loc,MAT_REUSE_MATRIX,2.0,&ptap->C_oth);CHKERRQ(ierr); 540 C_loc = ptap->C_loc; 541 C_oth = ptap->C_oth; 542 //printf("[%d] Co %d, %d\n", rank,Co->rmap->N,Co->cmap->N); 543 ierr = PetscTime(&t3);CHKERRQ(ierr); 544 eCseq = t3 - t2; 545 546 /* add C_loc and Co to to C */ 547 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 548 549 /* C_loc -> C */ 550 cm = C_loc->rmap->N; 551 Mat_SeqAIJ *c_seq; 552 c_seq = (Mat_SeqAIJ*)C_loc->data; 553 for (i=0; i<cm; i++) { 554 ncols = c_seq->i[i+1] - c_seq->i[i]; 555 row = rstart + i; 556 cols = c_seq->j + c_seq->i[i]; 557 vals = c_seq->a + c_seq->i[i]; 558 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 559 } 560 561 /* Co -> C, off-processor part */ 562 //printf("[%d] p->B %d, %d\n",rank,p->B->rmap->N,p->B->cmap->N); 563 cm = C_oth->rmap->N; 564 c_seq = (Mat_SeqAIJ*)C_oth->data; 565 for (i=0; i<cm; i++) { 566 ncols = c_seq->i[i+1] - c_seq->i[i]; 567 row = p->garray[i]; 568 cols = c_seq->j + c_seq->i[i]; 569 vals = c_seq->a + c_seq->i[i]; 570 //printf("[%d] row[%d] = %d\n",rank,i,row); 571 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 572 } 573 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 574 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 575 ierr = PetscTime(&t4);CHKERRQ(ierr); 576 eCmpi = t4 - t3; 577 578 if (rank==1) { 579 ierr = PetscPrintf(MPI_COMM_SELF," R %g, AP %g, Cseq %g, Cmpi %g = %g\n", eR,eAP,eCseq,eCmpi,eR+eAP+eCseq+eCmpi);CHKERRQ(ierr); 580 } 581 PetscFunctionReturn(0); 582 } 583 584 #undef __FUNCT__ 585 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 586 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 587 { 588 PetscErrorCode ierr; 589 Mat Cmpi; 590 Mat_PtAPMPI *ptap; 591 PetscFreeSpaceList free_space=NULL,current_space=NULL; 592 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 593 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 594 Mat_SeqAIJ *p_loc,*p_oth; 595 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 596 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 597 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 598 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 599 PetscBT lnkbt; 600 MPI_Comm comm; 601 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 602 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 603 PetscInt len,proc,*dnz,*onz,*owners; 604 PetscInt nzi,*pti,*ptj; 605 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 606 MPI_Request *swaits,*rwaits; 607 MPI_Status *sstatus,rstatus; 608 Mat_Merge_SeqsToMPI *merge; 609 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 610 PetscReal afill=1.0,afill_tmp; 611 PetscInt rmax; 612 613 PetscFunctionBegin; 614 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 615 616 /* check if matrix local sizes are compatible */ 617 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 618 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%D, %D) != Prow (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 619 } 620 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 621 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%D, %D) != Prow (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 622 } 623 624 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 625 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 626 627 /* create struct Mat_PtAPMPI and attached it to C later */ 628 ierr = PetscNew(&ptap);CHKERRQ(ierr); 629 ierr = PetscNew(&merge);CHKERRQ(ierr); 630 ptap->merge = merge; 631 ptap->reuse = MAT_INITIAL_MATRIX; 632 633 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 634 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 635 636 /* get P_loc by taking all local rows of P */ 637 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 638 639 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 640 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 641 pi_loc = p_loc->i; pj_loc = p_loc->j; 642 pi_oth = p_oth->i; pj_oth = p_oth->j; 643 644 /* (1) compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth (api,apj) */ 645 /*--------------------------------------------------------------------------*/ 646 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 647 api[0] = 0; 648 649 /* create and initialize a linked list */ 650 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 651 652 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 653 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 654 655 current_space = free_space; 656 657 for (i=0; i<am; i++) { 658 /* diagonal portion of A */ 659 nzi = adi[i+1] - adi[i]; 660 aj = ad->j + adi[i]; 661 for (j=0; j<nzi; j++) { 662 row = aj[j]; 663 pnz = pi_loc[row+1] - pi_loc[row]; 664 Jptr = pj_loc + pi_loc[row]; 665 /* add non-zero cols of P into the sorted linked list lnk */ 666 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 667 } 668 /* off-diagonal portion of A */ 669 nzi = aoi[i+1] - aoi[i]; 670 aj = ao->j + aoi[i]; 671 for (j=0; j<nzi; j++) { 672 row = aj[j]; 673 pnz = pi_oth[row+1] - pi_oth[row]; 674 Jptr = pj_oth + pi_oth[row]; 675 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 676 } 677 apnz = lnk[0]; 678 api[i+1] = api[i] + apnz; 679 if (ap_rmax < apnz) ap_rmax = apnz; 680 681 /* if free space is not available, double the total space in the list */ 682 if (current_space->local_remaining<apnz) { 683 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 684 nspacedouble++; 685 } 686 687 /* Copy data into free space, then initialize lnk */ 688 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 689 690 current_space->array += apnz; 691 current_space->local_used += apnz; 692 current_space->local_remaining -= apnz; 693 } 694 695 /* Allocate space for apj, initialize apj, and */ 696 /* destroy list of free space and other temporary array(s) */ 697 ierr = PetscMalloc1(api[am]+1,&apj);CHKERRQ(ierr); 698 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 699 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 700 if (afill_tmp > afill) afill = afill_tmp; 701 702 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 703 /*-----------------------------------------------------------------*/ 704 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 705 706 /* then, compute symbolic Co = (p->B)^T*AP */ 707 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 708 >= (num of nonzero rows of C_seq) - pn */ 709 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 710 coi[0] = 0; 711 712 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 713 nnz = fill*(poti[pon] + api[am]); 714 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 715 current_space = free_space; 716 717 for (i=0; i<pon; i++) { 718 pnz = poti[i+1] - poti[i]; 719 ptJ = potj + poti[i]; 720 for (j=0; j<pnz; j++) { 721 row = ptJ[j]; /* row of AP == col of Pot */ 722 apnz = api[row+1] - api[row]; 723 Jptr = apj + api[row]; 724 /* add non-zero cols of AP into the sorted linked list lnk */ 725 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 726 } 727 nnz = lnk[0]; 728 729 /* If free space is not available, double the total space in the list */ 730 if (current_space->local_remaining<nnz) { 731 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 732 nspacedouble++; 733 } 734 735 /* Copy data into free space, and zero out denserows */ 736 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 737 738 current_space->array += nnz; 739 current_space->local_used += nnz; 740 current_space->local_remaining -= nnz; 741 742 coi[i+1] = coi[i] + nnz; 743 } 744 745 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 746 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 747 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 748 if (afill_tmp > afill) afill = afill_tmp; 749 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 750 751 /* (3) send j-array (coj) of Co to other processors */ 752 /*--------------------------------------------------*/ 753 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 754 len_s = merge->len_s; 755 merge->nsend = 0; 756 757 758 /* determine row ownership */ 759 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 760 merge->rowmap->n = pn; 761 merge->rowmap->bs = 1; 762 763 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 764 owners = merge->rowmap->range; 765 766 /* determine the number of messages to send, their lengths */ 767 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 768 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 769 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 770 771 proc = 0; 772 for (i=0; i<pon; i++) { 773 while (prmap[i] >= owners[proc+1]) proc++; 774 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 775 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 776 } 777 778 len = 0; /* max length of buf_si[], see (4) */ 779 owners_co[0] = 0; 780 for (proc=0; proc<size; proc++) { 781 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 782 if (len_s[proc]) { 783 merge->nsend++; 784 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 785 len += len_si[proc]; 786 } 787 } 788 789 /* determine the number and length of messages to receive for coi and coj */ 790 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 791 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 792 793 /* post the Irecv and Isend of coj */ 794 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 795 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 796 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 797 for (proc=0, k=0; proc<size; proc++) { 798 if (!len_s[proc]) continue; 799 i = owners_co[proc]; 800 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 801 k++; 802 } 803 804 /* receives and sends of coj are complete */ 805 for (i=0; i<merge->nrecv; i++) { 806 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 807 } 808 ierr = PetscFree(rwaits);CHKERRQ(ierr); 809 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 810 811 /* (4) send and recv coi */ 812 /*-----------------------*/ 813 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 814 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 815 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 816 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 817 for (proc=0,k=0; proc<size; proc++) { 818 if (!len_s[proc]) continue; 819 /* form outgoing message for i-structure: 820 buf_si[0]: nrows to be sent 821 [1:nrows]: row index (global) 822 [nrows+1:2*nrows+1]: i-structure index 823 */ 824 /*-------------------------------------------*/ 825 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 826 buf_si_i = buf_si + nrows+1; 827 buf_si[0] = nrows; 828 buf_si_i[0] = 0; 829 nrows = 0; 830 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 831 nzi = coi[i+1] - coi[i]; 832 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 833 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 834 nrows++; 835 } 836 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 837 k++; 838 buf_si += len_si[proc]; 839 } 840 i = merge->nrecv; 841 while (i--) { 842 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 843 } 844 ierr = PetscFree(rwaits);CHKERRQ(ierr); 845 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 846 847 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 848 ierr = PetscFree(len_ri);CHKERRQ(ierr); 849 ierr = PetscFree(swaits);CHKERRQ(ierr); 850 ierr = PetscFree(buf_s);CHKERRQ(ierr); 851 852 /* (5) compute the local portion of C (mpi mat) */ 853 /*----------------------------------------------*/ 854 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 855 856 /* allocate pti array and free space for accumulating nonzero column info */ 857 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 858 pti[0] = 0; 859 860 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 861 nnz = fill*(pi_loc[pm] + api[am]); 862 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 863 current_space = free_space; 864 865 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 866 for (k=0; k<merge->nrecv; k++) { 867 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 868 nrows = *buf_ri_k[k]; 869 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 870 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 871 } 872 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 873 rmax = 0; 874 for (i=0; i<pn; i++) { 875 /* add pdt[i,:]*AP into lnk */ 876 pnz = pdti[i+1] - pdti[i]; 877 ptJ = pdtj + pdti[i]; 878 for (j=0; j<pnz; j++) { 879 row = ptJ[j]; /* row of AP == col of Pt */ 880 apnz = api[row+1] - api[row]; 881 Jptr = apj + api[row]; 882 /* add non-zero cols of AP into the sorted linked list lnk */ 883 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 884 } 885 886 /* add received col data into lnk */ 887 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 888 if (i == *nextrow[k]) { /* i-th row */ 889 nzi = *(nextci[k]+1) - *nextci[k]; 890 Jptr = buf_rj[k] + *nextci[k]; 891 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 892 nextrow[k]++; nextci[k]++; 893 } 894 } 895 nnz = lnk[0]; 896 897 /* if free space is not available, make more free space */ 898 if (current_space->local_remaining<nnz) { 899 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 900 nspacedouble++; 901 } 902 /* copy data into free space, then initialize lnk */ 903 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 904 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 905 906 current_space->array += nnz; 907 current_space->local_used += nnz; 908 current_space->local_remaining -= nnz; 909 910 pti[i+1] = pti[i] + nnz; 911 if (nnz > rmax) rmax = nnz; 912 } 913 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 914 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 915 916 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 917 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 918 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 919 if (afill_tmp > afill) afill = afill_tmp; 920 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 921 922 /* (6) create symbolic parallel matrix Cmpi */ 923 /*------------------------------------------*/ 924 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 925 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 926 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 927 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 928 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 929 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 930 931 merge->bi = pti; /* Cseq->i */ 932 merge->bj = ptj; /* Cseq->j */ 933 merge->coi = coi; /* Co->i */ 934 merge->coj = coj; /* Co->j */ 935 merge->buf_ri = buf_ri; 936 merge->buf_rj = buf_rj; 937 merge->owners_co = owners_co; 938 merge->destroy = Cmpi->ops->destroy; 939 merge->duplicate = Cmpi->ops->duplicate; 940 941 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 942 Cmpi->assembled = PETSC_FALSE; 943 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 944 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 945 946 /* attach the supporting struct to Cmpi for reuse */ 947 c = (Mat_MPIAIJ*)Cmpi->data; 948 c->ptap = ptap; 949 ptap->api = api; 950 ptap->apj = apj; 951 ptap->rmax = ap_rmax; 952 *C = Cmpi; 953 954 /* flag 'scalable' determines which implementations to be used: 955 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 956 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 957 /* set default scalable */ 958 ptap->scalable = PETSC_FALSE; //PETSC_TRUE; 959 960 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 961 if (!ptap->scalable) { /* Do dense axpy */ 962 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 963 } else { 964 ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 965 } 966 967 #if defined(PETSC_USE_INFO) 968 if (pti[pn] != 0) { 969 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 970 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 971 } else { 972 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 973 } 974 #endif 975 PetscFunctionReturn(0); 976 } 977 978 #undef __FUNCT__ 979 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 980 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 981 { 982 PetscErrorCode ierr; 983 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 984 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 985 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 986 Mat_SeqAIJ *p_loc,*p_oth; 987 Mat_PtAPMPI *ptap; 988 Mat_Merge_SeqsToMPI *merge; 989 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 990 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 991 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 992 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 993 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 994 MPI_Comm comm; 995 PetscMPIInt size,rank,taga,*len_s; 996 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 997 PetscInt **buf_ri,**buf_rj; 998 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 999 MPI_Request *s_waits,*r_waits; 1000 MPI_Status *status; 1001 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1002 PetscInt *api,*apj,*coi,*coj; 1003 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 1004 PetscBool scalable; 1005 #if defined(PTAP_PROFILE) 1006 PetscLogDouble t0,t1,t2,eP,t3,t4,et2_AP=0.0,ePtAP=0.0,t2_0,t2_1,t2_2; 1007 #endif 1008 1009 PetscFunctionBegin; 1010 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1011 #if defined(PTAP_PROFILE) 1012 ierr = PetscTime(&t0);CHKERRQ(ierr); 1013 #endif 1014 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1015 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1016 1017 ptap = c->ptap; 1018 if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_INCOMP,"MatPtAP() has not been called to create matrix C yet, cannot use MAT_REUSE_MATRIX"); 1019 merge = ptap->merge; 1020 apa = ptap->apa; 1021 scalable = ptap->scalable; 1022 1023 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 1024 /*-----------------------------------------------------*/ 1025 if (ptap->reuse == MAT_INITIAL_MATRIX) { 1026 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 1027 ptap->reuse = MAT_REUSE_MATRIX; 1028 } else { /* update numerical values of P_oth and P_loc */ 1029 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1030 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1031 } 1032 #if defined(PTAP_PROFILE) 1033 ierr = PetscTime(&t1);CHKERRQ(ierr); 1034 eP = t1-t0; 1035 #endif 1036 /* 1037 printf("[%d] Ad: %d, %d; Ao: %d, %d; P_loc: %d, %d; P_oth %d, %d;\n",rank, 1038 a->A->rmap->N,a->A->cmap->N,a->B->rmap->N,a->B->cmap->N, 1039 ptap->P_loc->rmap->N,ptap->P_loc->cmap->N, 1040 ptap->P_oth->rmap->N,ptap->P_oth->cmap->N); 1041 */ 1042 1043 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 1044 /*--------------------------------------------------------------*/ 1045 /* get data from symbolic products */ 1046 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1047 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 1048 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 1049 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 1050 1051 coi = merge->coi; coj = merge->coj; 1052 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1053 1054 bi = merge->bi; bj = merge->bj; 1055 owners = merge->rowmap->range; 1056 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 1057 1058 api = ptap->api; apj = ptap->apj; 1059 1060 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but faster (could take 1/3 scalable time) */ 1061 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 1062 #if 0 1063 /* ------ 10x slower -------------- */ 1064 /*==================================*/ 1065 Mat R = ptap->R; 1066 Mat_SeqAIJ *r = (Mat_SeqAIJ*)R->data; 1067 PetscInt *ri=r->i,*rj=r->j,rnz,arow,l,prow,pcol,pN=P->cmap->N; 1068 PetscScalar *ra=r->a,tmp,cdense[pN]; 1069 1070 ierr = PetscMemzero(cdense,pN*sizeof(PetscScalar));CHKERRQ(ierr); 1071 for (i=0; i<cm; i++) { /* each row of C or R */ 1072 rnz = ri[i+1] - ri[i]; 1073 1074 for (j=0; j<rnz; j++) { /* each nz of R */ 1075 arow = rj[ri[i] + j]; 1076 1077 /* diagonal portion of A */ 1078 anz = ad->i[arow+1] - ad->i[arow]; 1079 for (k=0; k<anz; k++) { /* each nz of Ad */ 1080 tmp = ra[ri[i] + j]*ad->a[ad->i[arow] + k]; 1081 prow = ad->j[ad->i[arow] + k]; 1082 pnz = pi_loc[prow+1] - pi_loc[prow]; 1083 1084 for (l=0; l<pnz; l++) { /* each nz of P_loc */ 1085 pcol = pj_loc[pi_loc[prow] + l]; 1086 cdense[pcol] += tmp*pa_loc[pi_loc[prow] + l]; 1087 } 1088 } 1089 1090 /* off-diagonal portion of A */ 1091 anz = ao->i[arow+1] - ao->i[arow]; 1092 for (k=0; k<anz; k++) { /* each nz of Ao */ 1093 tmp = ra[ri[i] + j]*ao->a[ao->i[arow] + k]; 1094 prow = ao->j[ao->i[arow] + k]; 1095 pnz = pi_oth[prow+1] - pi_oth[prow]; 1096 1097 for (l=0; l<pnz; l++) { /* each nz of P_oth */ 1098 pcol = pj_oth[pi_oth[prow] + l]; 1099 cdense[pcol] += tmp*pa_oth[pi_oth[prow] + l]; 1100 } 1101 } 1102 1103 } //for (j=0; j<rnz; j++) 1104 1105 /* copy cdense[] into ca; zero cdense[] */ 1106 cnz = bi[i+1] - bi[i]; 1107 cj = bj + bi[i]; 1108 ca = ba + bi[i]; 1109 for (j=0; j<cnz; j++) { 1110 ca[j] += cdense[cj[j]]; 1111 cdense[cj[j]] = 0.0; 1112 } 1113 #if 0 1114 if (rank == 0) { 1115 printf("[%d] row %d: ",rank,i); 1116 for (j=0; j<pN; j++) printf(" %g,",cdense[j]); 1117 printf("\n"); 1118 } 1119 for (j=0; j<pN; j++) cdense[j]=0.0; // zero cdnese[] 1120 #endif 1121 } //for (i=0; i<cm; i++) { 1122 #endif 1123 1124 //========================================== 1125 1126 ierr = PetscTime(&t1);CHKERRQ(ierr); 1127 for (i=0; i<am; i++) { 1128 #if defined(PTAP_PROFILE) 1129 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1130 #endif 1131 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 1132 /*------------------------------------------------------------*/ 1133 apJ = apj + api[i]; 1134 1135 /* diagonal portion of A */ 1136 anz = adi[i+1] - adi[i]; 1137 adj = ad->j + adi[i]; 1138 ada = ad->a + adi[i]; 1139 for (j=0; j<anz; j++) { 1140 row = adj[j]; 1141 pnz = pi_loc[row+1] - pi_loc[row]; 1142 pj = pj_loc + pi_loc[row]; 1143 pa = pa_loc + pi_loc[row]; 1144 1145 /* perform dense axpy */ 1146 valtmp = ada[j]; 1147 for (k=0; k<pnz; k++) { 1148 apa[pj[k]] += valtmp*pa[k]; 1149 } 1150 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1151 } 1152 1153 /* off-diagonal portion of A */ 1154 anz = aoi[i+1] - aoi[i]; 1155 aoj = ao->j + aoi[i]; 1156 aoa = ao->a + aoi[i]; 1157 for (j=0; j<anz; j++) { 1158 row = aoj[j]; 1159 pnz = pi_oth[row+1] - pi_oth[row]; 1160 pj = pj_oth + pi_oth[row]; 1161 pa = pa_oth + pi_oth[row]; 1162 1163 /* perform dense axpy */ 1164 valtmp = aoa[j]; 1165 for (k=0; k<pnz; k++) { 1166 apa[pj[k]] += valtmp*pa[k]; 1167 } 1168 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1169 } 1170 #if defined(PTAP_PROFILE) 1171 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1172 et2_AP += t2_1 - t2_0; 1173 #endif 1174 1175 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1176 /*--------------------------------------------------------------*/ 1177 apnz = api[i+1] - api[i]; 1178 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 1179 pnz = po->i[i+1] - po->i[i]; 1180 poJ = po->j + po->i[i]; 1181 pA = po->a + po->i[i]; 1182 for (j=0; j<pnz; j++) { 1183 row = poJ[j]; 1184 cnz = coi[row+1] - coi[row]; 1185 cj = coj + coi[row]; 1186 ca = coa + coi[row]; 1187 /* perform dense axpy */ 1188 valtmp = pA[j]; 1189 for (k=0; k<cnz; k++) { 1190 ca[k] += valtmp*apa[cj[k]]; 1191 } 1192 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1193 } 1194 #if 1 1195 /* put the value into Cd (diagonal part) */ 1196 pnz = pd->i[i+1] - pd->i[i]; 1197 pdJ = pd->j + pd->i[i]; 1198 pA = pd->a + pd->i[i]; 1199 for (j=0; j<pnz; j++) { 1200 row = pdJ[j]; 1201 cnz = bi[row+1] - bi[row]; 1202 cj = bj + bi[row]; 1203 ca = ba + bi[row]; 1204 /* perform dense axpy */ 1205 valtmp = pA[j]; 1206 for (k=0; k<cnz; k++) { 1207 ca[k] += valtmp*apa[cj[k]]; 1208 } 1209 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1210 } 1211 #endif 1212 /* zero the current row of A*P */ 1213 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 1214 #if defined(PTAP_PROFILE) 1215 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1216 ePtAP += t2_2 - t2_1; 1217 #endif 1218 } 1219 1220 if (rank == 100) { 1221 for (row=0; row<cm; row++) { 1222 printf("[%d] row %d: ",rank,row); 1223 cnz = bi[row+1] - bi[row]; 1224 for (j=0; j<cnz; j++) printf(" %g,",ba[bi[row]+j]); 1225 printf("\n"); 1226 } 1227 } 1228 1229 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 1230 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 1231 /*-----------------------------------------------------------------------------------------*/ 1232 pA=pa_loc; 1233 for (i=0; i<am; i++) { 1234 #if defined(PTAP_PROFILE) 1235 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1236 #endif 1237 /* form i-th sparse row of A*P */ 1238 apnz = api[i+1] - api[i]; 1239 apJ = apj + api[i]; 1240 /* diagonal portion of A */ 1241 anz = adi[i+1] - adi[i]; 1242 adj = ad->j + adi[i]; 1243 ada = ad->a + adi[i]; 1244 for (j=0; j<anz; j++) { 1245 row = adj[j]; 1246 pnz = pi_loc[row+1] - pi_loc[row]; 1247 pj = pj_loc + pi_loc[row]; 1248 pa = pa_loc + pi_loc[row]; 1249 valtmp = ada[j]; 1250 nextp = 0; 1251 for (k=0; nextp<pnz; k++) { 1252 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1253 apa[k] += valtmp*pa[nextp++]; 1254 } 1255 } 1256 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1257 } 1258 /* off-diagonal portion of A */ 1259 anz = aoi[i+1] - aoi[i]; 1260 aoj = ao->j + aoi[i]; 1261 aoa = ao->a + aoi[i]; 1262 for (j=0; j<anz; j++) { 1263 row = aoj[j]; 1264 pnz = pi_oth[row+1] - pi_oth[row]; 1265 pj = pj_oth + pi_oth[row]; 1266 pa = pa_oth + pi_oth[row]; 1267 valtmp = aoa[j]; 1268 nextp = 0; 1269 for (k=0; nextp<pnz; k++) { 1270 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1271 apa[k] += valtmp*pa[nextp++]; 1272 } 1273 } 1274 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1275 } 1276 #if defined(PTAP_PROFILE) 1277 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1278 et2_AP += t2_1 - t2_0; 1279 #endif 1280 1281 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1282 /*--------------------------------------------------------------*/ 1283 pnz = pi_loc[i+1] - pi_loc[i]; 1284 pJ = pj_loc + pi_loc[i]; 1285 for (j=0; j<pnz; j++) { 1286 nextap = 0; 1287 row = pJ[j]; /* global index */ 1288 if (row < pcstart || row >=pcend) { /* put the value into Co */ 1289 row = *poJ; 1290 cj = coj + coi[row]; 1291 ca = coa + coi[row]; poJ++; 1292 } else { /* put the value into Cd */ 1293 row = *pdJ; 1294 cj = bj + bi[row]; 1295 ca = ba + bi[row]; pdJ++; 1296 } 1297 valtmp = pA[j]; 1298 for (k=0; nextap<apnz; k++) { 1299 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 1300 } 1301 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 1302 } 1303 pA += pnz; 1304 /* zero the current row info for A*P */ 1305 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 1306 #if defined(PTAP_PROFILE) 1307 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1308 ePtAP += t2_2 - t2_1; 1309 #endif 1310 } 1311 } 1312 #if defined(PTAP_PROFILE) 1313 ierr = PetscTime(&t2);CHKERRQ(ierr); 1314 #endif 1315 1316 /* 3) send and recv matrix values coa */ 1317 /*------------------------------------*/ 1318 buf_ri = merge->buf_ri; 1319 buf_rj = merge->buf_rj; 1320 len_s = merge->len_s; 1321 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1322 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1323 1324 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1325 for (proc=0,k=0; proc<size; proc++) { 1326 if (!len_s[proc]) continue; 1327 i = merge->owners_co[proc]; 1328 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1329 k++; 1330 } 1331 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1332 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1333 1334 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1335 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1336 ierr = PetscFree(coa);CHKERRQ(ierr); 1337 #if defined(PTAP_PROFILE) 1338 ierr = PetscTime(&t3);CHKERRQ(ierr); 1339 #endif 1340 1341 /* 4) insert local Cseq and received values into Cmpi */ 1342 /*------------------------------------------------------*/ 1343 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1344 for (k=0; k<merge->nrecv; k++) { 1345 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1346 nrows = *(buf_ri_k[k]); 1347 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1348 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1349 } 1350 1351 for (i=0; i<cm; i++) { 1352 row = owners[rank] + i; /* global row index of C_seq */ 1353 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1354 ba_i = ba + bi[i]; 1355 bnz = bi[i+1] - bi[i]; 1356 /* add received vals into ba */ 1357 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1358 /* i-th row */ 1359 if (i == *nextrow[k]) { 1360 cnz = *(nextci[k]+1) - *nextci[k]; 1361 cj = buf_rj[k] + *(nextci[k]); 1362 ca = abuf_r[k] + *(nextci[k]); 1363 nextcj = 0; 1364 for (j=0; nextcj<cnz; j++) { 1365 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1366 ba_i[j] += ca[nextcj++]; 1367 } 1368 } 1369 nextrow[k]++; nextci[k]++; 1370 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1371 } 1372 } 1373 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1374 } 1375 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1376 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1377 1378 ierr = PetscFree(ba);CHKERRQ(ierr); 1379 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1380 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1381 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1382 #if defined(PTAP_PROFILE) 1383 ierr = PetscTime(&t4);CHKERRQ(ierr); 1384 if (rank==1) { 1385 ierr = PetscPrintf(MPI_COMM_SELF," [%d] PtAPNum %g/P + %g/PtAP( %g/A*P + %g/Pt*AP ) + %g/comm + %g/Cloc = %g\n\n",rank,eP,t2-t1,et2_AP,ePtAP,t3-t2,t4-t3,t4-t0);CHKERRQ(ierr); 1386 } 1387 #endif 1388 PetscFunctionReturn(0); 1389 } 1390