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