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 ierr = MatDestroy(&ptap->Rd);CHKERRQ(ierr); 33 ierr = MatDestroy(&ptap->Ro);CHKERRQ(ierr); 34 if (ptap->AP_loc) { 35 Mat_SeqAIJ *ap = (Mat_SeqAIJ*)(ptap->AP_loc)->data; 36 ierr = PetscFree(ap->i);CHKERRQ(ierr); 37 ierr = PetscFree2(ap->j,ap->a);CHKERRQ(ierr); 38 ierr = MatDestroy(&ptap->AP_loc);CHKERRQ(ierr); 39 } 40 ierr = MatDestroy(&ptap->C_loc);CHKERRQ(ierr); 41 ierr = MatDestroy(&ptap->C_oth);CHKERRQ(ierr); 42 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 43 if (merge) { 44 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 45 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 46 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 47 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 48 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 49 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 50 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 51 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 52 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 53 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 54 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 55 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 56 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 57 ierr = merge->destroy(A);CHKERRQ(ierr); 58 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 59 } else { 60 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 61 } 62 ierr = PetscFree(ptap);CHKERRQ(ierr); 63 } 64 PetscFunctionReturn(0); 65 } 66 67 #undef __FUNCT__ 68 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP_old" 69 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP_old(Mat A, MatDuplicateOption op, Mat *M) 70 { 71 PetscErrorCode ierr; 72 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 73 Mat_PtAPMPI *ptap = a->ptap; 74 Mat_Merge_SeqsToMPI *merge = ptap->merge; 75 76 PetscFunctionBegin; 77 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 78 79 (*M)->ops->destroy = merge->destroy; 80 (*M)->ops->duplicate = merge->duplicate; 81 PetscFunctionReturn(0); 82 } 83 84 85 #undef __FUNCT__ 86 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP" 87 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 88 { 89 PetscErrorCode ierr; 90 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 91 Mat_PtAPMPI *ptap = a->ptap; 92 93 PetscFunctionBegin; 94 ierr = (*ptap->duplicate)(A,op,M);CHKERRQ(ierr); 95 (*M)->ops->destroy = ptap->destroy; 96 (*M)->ops->duplicate = ptap->duplicate; 97 PetscFunctionReturn(0); 98 } 99 100 #undef __FUNCT__ 101 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ" 102 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 103 { 104 PetscErrorCode ierr; 105 PetscBool newalg=PETSC_TRUE; 106 MPI_Comm comm; 107 108 PetscFunctionBegin; 109 /* check if matrix local sizes are compatible */ 110 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 111 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 112 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); 113 } 114 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 115 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); 116 } 117 118 ierr = PetscOptionsGetBool(NULL,"-matptap_new",&newalg,NULL);CHKERRQ(ierr); 119 if (scall == MAT_INITIAL_MATRIX) { 120 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 121 if (newalg) { 122 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 123 } else { 124 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ_old(A,P,fill,C);CHKERRQ(ierr); 125 } 126 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 127 } 128 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 129 if (newalg) { 130 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr); 131 } else { 132 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ_old(A,P,*C);CHKERRQ(ierr); 133 } 134 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 135 PetscFunctionReturn(0); 136 } 137 138 /* requires array of size pN, thus nonscalable version */ 139 #undef __FUNCT__ 140 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 141 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 142 { 143 PetscErrorCode ierr; 144 Mat_PtAPMPI *ptap; 145 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 146 MPI_Comm comm; 147 PetscMPIInt size,rank; 148 Mat Cmpi; 149 PetscFreeSpaceList free_space=NULL,current_space=NULL; 150 PetscInt am=A->rmap->n,pm=P->rmap->n; 151 PetscInt *lnk,i,k,pnz,row,nnz; 152 PetscInt pN=P->cmap->N,pn=P->cmap->n; 153 PetscBT lnkbt; 154 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 155 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 156 PetscInt len,proc,*dnz,*onz,*owners; 157 PetscInt nzi,nspacedouble; 158 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 159 MPI_Request *swaits,*rwaits; 160 MPI_Status *sstatus,rstatus; 161 PetscLayout rowmap; 162 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 163 PetscInt nsend; 164 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 165 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,con,j,ap_rmax=0; 166 PetscInt rmax,*aj,*ai,*pi; 167 Mat_SeqAIJ *p_loc,*p_oth,*ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*c_loc,*c_oth; 168 PetscScalar *apv; 169 #if defined(PETSC_USE_INFO) 170 PetscReal apfill; 171 #endif 172 #if defined(PTAP_PROFILE) 173 PetscLogDouble t0,t1,t11,t12,t2,t3,t4; 174 #endif 175 176 PetscFunctionBegin; 177 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 178 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 179 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 180 #if defined(PTAP_PROFILE) 181 ierr = PetscTime(&t0);CHKERRQ(ierr); 182 #endif 183 184 /* create struct Mat_PtAPMPI and attached it to C later */ 185 ierr = PetscNew(&ptap);CHKERRQ(ierr); 186 ptap->reuse = MAT_INITIAL_MATRIX; 187 188 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 189 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 190 /* get P_loc by taking all local rows of P */ 191 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 192 193 /* (0) compute Rd = Pd^T, Ro = Po^T */ 194 /* --------------------------------- */ 195 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 196 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 197 #if defined(PTAP_PROFILE) 198 ierr = PetscTime(&t11);CHKERRQ(ierr); 199 #endif 200 201 /* (1) compute symbolic AP = A_loc*P = Ad*P_loc + Ao*P_oth (api,apj) */ 202 /* ----------------------------------------------------------------- */ 203 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 204 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 205 206 /* create and initialize a linked list - nonscalable version */ 207 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 208 209 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) */ 210 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ad->i[am]+ao->i[am]+p_loc->i[pm])),&free_space);CHKERRQ(ierr); 211 current_space = free_space; 212 nspacedouble = 0; 213 214 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 215 api[0] = 0; 216 for (i=0; i<am; i++) { 217 /* diagonal portion: Ad[i,:]*P */ 218 ai = ad->i; pi = p_loc->i; 219 nzi = ai[i+1] - ai[i]; 220 aj = ad->j + ai[i]; 221 for (j=0; j<nzi; j++) { 222 row = aj[j]; 223 pnz = pi[row+1] - pi[row]; 224 Jptr = p_loc->j + pi[row]; 225 /* add non-zero cols of P into the sorted linked list lnk */ 226 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 227 } 228 /* off-diagonal portion: Ao[i,:]*P */ 229 ai = ao->i; pi = p_oth->i; 230 nzi = ai[i+1] - ai[i]; 231 aj = ao->j + ai[i]; 232 for (j=0; j<nzi; j++) { 233 row = aj[j]; 234 pnz = pi[row+1] - pi[row]; 235 Jptr = p_oth->j + pi[row]; 236 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 237 } 238 apnz = lnk[0]; 239 api[i+1] = api[i] + apnz; 240 if (ap_rmax < apnz) ap_rmax = apnz; 241 242 /* if free space is not available, double the total space in the list */ 243 if (current_space->local_remaining<apnz) { 244 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 245 nspacedouble++; 246 } 247 248 /* Copy data into free space, then initialize lnk */ 249 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 250 251 current_space->array += apnz; 252 current_space->local_used += apnz; 253 current_space->local_remaining -= apnz; 254 } 255 /* Allocate space for apj and apv, initialize apj, and */ 256 /* destroy list of free space and other temporary array(s) */ 257 ierr = PetscMalloc2(api[am],&apj,api[am],&apv);CHKERRQ(ierr); 258 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 259 260 /* Create AP_loc for reuse */ 261 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,pN,api,apj,apv,&ptap->AP_loc);CHKERRQ(ierr); 262 263 #if defined(PETSC_USE_INFO) 264 apfill = (PetscReal)api[am]/(ad->i[am]+ao->i[am]+p_loc->i[pm]+1); 265 ptap->AP_loc->info.mallocs = nspacedouble; 266 ptap->AP_loc->info.fill_ratio_given = fill; 267 ptap->AP_loc->info.fill_ratio_needed = apfill; 268 269 if (api[am]) { 270 ierr = PetscInfo3(ptap->AP_loc,"AP_loc reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)apfill);CHKERRQ(ierr); 271 ierr = PetscInfo1(ptap->AP_loc,"Use MatPtAP(A,B,MatReuse,%g,&C) for best AP_loc performance.;\n",(double)apfill);CHKERRQ(ierr); 272 } else { 273 ierr = PetscInfo(ptap->AP_loc,"AP_loc is empty \n");CHKERRQ(ierr); 274 } 275 #endif 276 277 #if defined(PTAP_PROFILE) 278 ierr = PetscTime(&t12);CHKERRQ(ierr); 279 #endif 280 281 /* (2-1) compute symbolic Co = Ro*AP_loc */ 282 /* ------------------------------------ */ 283 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 284 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 285 #if defined(PTAP_PROFILE) 286 ierr = PetscTime(&t1);CHKERRQ(ierr); 287 #endif 288 289 /* (3) send coj of C_oth to other processors */ 290 /* ------------------------------------------ */ 291 /* determine row ownership */ 292 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 293 rowmap->n = pn; 294 rowmap->bs = 1; 295 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 296 owners = rowmap->range; 297 298 /* determine the number of messages to send, their lengths */ 299 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 300 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 301 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 302 303 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 304 coi = c_oth->i; coj = c_oth->j; 305 con = ptap->C_oth->rmap->n; 306 proc = 0; 307 for (i=0; i<con; i++) { 308 while (prmap[i] >= owners[proc+1]) proc++; 309 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 310 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 311 } 312 313 len = 0; /* max length of buf_si[], see (4) */ 314 owners_co[0] = 0; 315 nsend = 0; 316 for (proc=0; proc<size; proc++) { 317 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 318 if (len_s[proc]) { 319 nsend++; 320 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 321 len += len_si[proc]; 322 } 323 } 324 325 /* determine the number and length of messages to receive for coi and coj */ 326 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 327 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 328 329 /* post the Irecv and Isend of coj */ 330 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 331 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 332 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 333 for (proc=0, k=0; proc<size; proc++) { 334 if (!len_s[proc]) continue; 335 i = owners_co[proc]; 336 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 337 k++; 338 } 339 340 /* (2-2) compute symbolic C_loc = Rd*AP_loc */ 341 /* ---------------------------------------- */ 342 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 343 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 344 345 /* receives coj are complete */ 346 for (i=0; i<nrecv; i++) { 347 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 348 } 349 ierr = PetscFree(rwaits);CHKERRQ(ierr); 350 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 351 352 /* (4) send and recv coi */ 353 /*-----------------------*/ 354 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 355 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 356 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 357 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 358 for (proc=0,k=0; proc<size; proc++) { 359 if (!len_s[proc]) continue; 360 /* form outgoing message for i-structure: 361 buf_si[0]: nrows to be sent 362 [1:nrows]: row index (global) 363 [nrows+1:2*nrows+1]: i-structure index 364 */ 365 /*-------------------------------------------*/ 366 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 367 buf_si_i = buf_si + nrows+1; 368 buf_si[0] = nrows; 369 buf_si_i[0] = 0; 370 nrows = 0; 371 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 372 nzi = coi[i+1] - coi[i]; 373 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 374 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 375 nrows++; 376 } 377 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 378 k++; 379 buf_si += len_si[proc]; 380 } 381 for (i=0; i<nrecv; i++) { 382 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 383 } 384 ierr = PetscFree(rwaits);CHKERRQ(ierr); 385 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 386 387 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 388 ierr = PetscFree(len_ri);CHKERRQ(ierr); 389 ierr = PetscFree(swaits);CHKERRQ(ierr); 390 ierr = PetscFree(buf_s);CHKERRQ(ierr); 391 #if defined(PTAP_PROFILE) 392 ierr = PetscTime(&t2);CHKERRQ(ierr); 393 #endif 394 /* (5) compute the local portion of Cmpi */ 395 /* ------------------------------------------ */ 396 /* set initial free space to be pN */ 397 ierr = PetscFreeSpaceGet(pN,&free_space);CHKERRQ(ierr); /* non-scalable version */ 398 current_space = free_space; 399 400 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 401 for (k=0; k<nrecv; k++) { 402 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 403 nrows = *buf_ri_k[k]; 404 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 405 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 406 } 407 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 408 409 rmax = 0; 410 for (i=0; i<pn; i++) { 411 /* add C_loc into Cmpi */ 412 nzi = c_loc->i[i+1] - c_loc->i[i]; 413 Jptr = c_loc->j + c_loc->i[i]; 414 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 415 416 /* add received col data into lnk */ 417 for (k=0; k<nrecv; k++) { /* k-th received message */ 418 if (i == *nextrow[k]) { /* i-th row */ 419 nzi = *(nextci[k]+1) - *nextci[k]; 420 Jptr = buf_rj[k] + *nextci[k]; 421 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 422 nextrow[k]++; nextci[k]++; 423 } 424 } 425 nnz = lnk[0]; 426 427 /* copy data into free space, then initialize lnk */ 428 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 429 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 430 if (nnz > rmax) rmax = nnz; 431 } 432 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 433 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 434 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 435 #if defined(PTAP_PROFILE) 436 ierr = PetscTime(&t3);CHKERRQ(ierr); 437 #endif 438 439 /* (6) create symbolic parallel matrix Cmpi */ 440 /*------------------------------------------*/ 441 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 442 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 443 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 444 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 445 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 446 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 447 448 /* members in merge */ 449 ierr = PetscFree(id_r);CHKERRQ(ierr); 450 ierr = PetscFree(len_r);CHKERRQ(ierr); 451 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 452 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 453 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 454 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 455 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 456 457 /* attach the supporting struct to Cmpi for reuse */ 458 c = (Mat_MPIAIJ*)Cmpi->data; 459 c->ptap = ptap; 460 ptap->rmax = ap_rmax; 461 ptap->duplicate = Cmpi->ops->duplicate; 462 ptap->destroy = Cmpi->ops->destroy; 463 464 /* Do dense axpy in MatPtAPNumeric_MPIAIJ_MPIAIJ_new() */ 465 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 466 467 468 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 469 Cmpi->assembled = PETSC_FALSE; 470 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 471 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 472 Cmpi->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ; 473 *C = Cmpi; 474 475 #if defined(PTAP_PROFILE) 476 ierr = PetscTime(&t4);CHKERRQ(ierr); 477 if (rank == 1) { 478 printf("PtAPSym: %g + %g + %g + %g + %g + %g = %g\n",t11-t0,t1-t11,t12-t11,t2-t2,t3-t2,t4-t3,t4-t0); 479 } 480 #endif 481 PetscFunctionReturn(0); 482 } 483 484 #undef __FUNCT__ 485 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 486 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 487 { 488 PetscErrorCode ierr; 489 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 490 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 491 Mat_SeqAIJ *ap,*p_loc,*p_oth,*c_seq; 492 Mat_PtAPMPI *ptap = c->ptap; 493 Mat AP_loc,C_loc,C_oth; 494 PetscInt i,rstart,rend,cm,ncols,row; 495 PetscInt *api,*apj,am = A->rmap->n,j,col,apnz; 496 PetscScalar *apa; 497 const PetscInt *cols; 498 const PetscScalar *vals; 499 #if defined(PTAP_PROFILE) 500 PetscMPIInt rank; 501 MPI_Comm comm; 502 PetscLogDouble t0,t1,t2,t3,t4,eR,eAP,eCseq,eCmpi; 503 #endif 504 505 PetscFunctionBegin; 506 ierr = MatZeroEntries(C);CHKERRQ(ierr); 507 508 /* 1) get R = Pd^T,Ro = Po^T */ 509 #if defined(PTAP_PROFILE) 510 ierr = PetscTime(&t0);CHKERRQ(ierr); 511 #endif 512 if (ptap->reuse == MAT_REUSE_MATRIX) { 513 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 514 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 515 } 516 #if defined(PTAP_PROFILE) 517 ierr = PetscTime(&t1);CHKERRQ(ierr); 518 eR = t1 - t0; 519 #endif 520 521 /* 2) get AP_loc */ 522 AP_loc = ptap->AP_loc; 523 ap = (Mat_SeqAIJ*)AP_loc->data; 524 525 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 526 /*-----------------------------------------------------*/ 527 if (ptap->reuse == MAT_REUSE_MATRIX) { 528 /* P_oth and P_loc are obtained in MatPtASymbolic() when reuse == MAT_INITIAL_MATRIX */ 529 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 530 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 531 } 532 533 /* 2-2) compute numeric A_loc*P - dominating part */ 534 /* ---------------------------------------------- */ 535 /* get data from symbolic products */ 536 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 537 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 538 apa = ptap->apa; 539 api = ap->i; 540 apj = ap->j; 541 for (i=0; i<am; i++) { 542 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 543 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 544 apnz = api[i+1] - api[i]; 545 for (j=0; j<apnz; j++) { 546 col = apj[j+api[i]]; 547 ap->a[j+ap->i[i]] = apa[col]; 548 apa[col] = 0.0; 549 } 550 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 551 } 552 #if defined(PTAP_PROFILE) 553 ierr = PetscTime(&t2);CHKERRQ(ierr); 554 eAP = t2 - t1; 555 #endif 556 557 /* 3) C_loc = Rd*AP_loc, C_oth = Ro*AP_loc */ 558 ierr = ((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,AP_loc,ptap->C_loc);CHKERRQ(ierr); 559 ierr = ((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,AP_loc,ptap->C_oth);CHKERRQ(ierr); 560 C_loc = ptap->C_loc; 561 C_oth = ptap->C_oth; 562 #if defined(PTAP_PROFILE) 563 ierr = PetscTime(&t3);CHKERRQ(ierr); 564 eCseq = t3 - t2; 565 #endif 566 567 /* add C_loc and Co to to C */ 568 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 569 570 /* C_loc -> C */ 571 cm = C_loc->rmap->N; 572 c_seq = (Mat_SeqAIJ*)C_loc->data; 573 cols = c_seq->j; 574 vals = c_seq->a; 575 for (i=0; i<cm; i++) { 576 ncols = c_seq->i[i+1] - c_seq->i[i]; 577 row = rstart + i; 578 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 579 cols += ncols; vals += ncols; 580 } 581 582 /* Co -> C, off-processor part */ 583 cm = C_oth->rmap->N; 584 c_seq = (Mat_SeqAIJ*)C_oth->data; 585 cols = c_seq->j; 586 vals = c_seq->a; 587 for (i=0; i<cm; i++) { 588 ncols = c_seq->i[i+1] - c_seq->i[i]; 589 row = p->garray[i]; 590 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 591 cols += ncols; vals += ncols; 592 } 593 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 594 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 595 #if defined(PTAP_PROFILE) 596 ierr = PetscTime(&t4);CHKERRQ(ierr); 597 eCmpi = t4 - t3; 598 599 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 600 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 601 if (rank==1) { 602 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); 603 } 604 #endif 605 ptap->reuse = MAT_REUSE_MATRIX; 606 PetscFunctionReturn(0); 607 } 608 609 #undef __FUNCT__ 610 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ_old" 611 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_old(Mat A,Mat P,PetscReal fill,Mat *C) 612 { 613 PetscErrorCode ierr; 614 Mat Cmpi; 615 Mat_PtAPMPI *ptap; 616 PetscFreeSpaceList free_space=NULL,current_space=NULL; 617 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 618 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 619 Mat_SeqAIJ *p_loc,*p_oth; 620 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 621 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 622 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 623 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 624 PetscBT lnkbt; 625 MPI_Comm comm; 626 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 627 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 628 PetscInt len,proc,*dnz,*onz,*owners; 629 PetscInt nzi,*pti,*ptj; 630 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 631 MPI_Request *swaits,*rwaits; 632 MPI_Status *sstatus,rstatus; 633 Mat_Merge_SeqsToMPI *merge; 634 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 635 PetscReal afill=1.0,afill_tmp; 636 PetscInt rmax; 637 638 PetscFunctionBegin; 639 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 640 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 641 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 642 643 /* create struct Mat_PtAPMPI and attached it to C later */ 644 ierr = PetscNew(&ptap);CHKERRQ(ierr); 645 ierr = PetscNew(&merge);CHKERRQ(ierr); 646 ptap->merge = merge; 647 ptap->reuse = MAT_INITIAL_MATRIX; 648 649 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 650 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 651 652 /* get P_loc by taking all local rows of P */ 653 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 654 655 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 656 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 657 pi_loc = p_loc->i; pj_loc = p_loc->j; 658 pi_oth = p_oth->i; pj_oth = p_oth->j; 659 660 /* (1) compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth (api,apj) */ 661 /*--------------------------------------------------------------------------*/ 662 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 663 api[0] = 0; 664 665 /* create and initialize a linked list */ 666 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 667 668 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 669 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 670 current_space = free_space; 671 672 for (i=0; i<am; i++) { 673 /* diagonal portion of A */ 674 nzi = adi[i+1] - adi[i]; 675 aj = ad->j + adi[i]; 676 for (j=0; j<nzi; j++) { 677 row = aj[j]; 678 pnz = pi_loc[row+1] - pi_loc[row]; 679 Jptr = pj_loc + pi_loc[row]; 680 /* add non-zero cols of P into the sorted linked list lnk */ 681 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 682 } 683 /* off-diagonal portion of A */ 684 nzi = aoi[i+1] - aoi[i]; 685 aj = ao->j + aoi[i]; 686 for (j=0; j<nzi; j++) { 687 row = aj[j]; 688 pnz = pi_oth[row+1] - pi_oth[row]; 689 Jptr = pj_oth + pi_oth[row]; 690 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 691 } 692 apnz = lnk[0]; 693 api[i+1] = api[i] + apnz; 694 if (ap_rmax < apnz) ap_rmax = apnz; 695 696 /* if free space is not available, double the total space in the list */ 697 if (current_space->local_remaining<apnz) { 698 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 699 nspacedouble++; 700 } 701 702 /* Copy data into free space, then initialize lnk */ 703 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 704 705 current_space->array += apnz; 706 current_space->local_used += apnz; 707 current_space->local_remaining -= apnz; 708 } 709 710 /* Allocate space for apj, initialize apj, and */ 711 /* destroy list of free space and other temporary array(s) */ 712 ierr = PetscMalloc1(api[am]+1,&apj);CHKERRQ(ierr); 713 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 714 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 715 if (afill_tmp > afill) afill = afill_tmp; 716 717 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 718 /*-----------------------------------------------------------------*/ 719 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 720 721 /* then, compute symbolic Co = (p->B)^T*AP */ 722 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 723 >= (num of nonzero rows of C_seq) - pn */ 724 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 725 coi[0] = 0; 726 727 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 728 nnz = fill*(poti[pon] + api[am]); 729 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 730 current_space = free_space; 731 732 for (i=0; i<pon; i++) { 733 pnz = poti[i+1] - poti[i]; 734 ptJ = potj + poti[i]; 735 for (j=0; j<pnz; j++) { 736 row = ptJ[j]; /* row of AP == col of Pot */ 737 apnz = api[row+1] - api[row]; 738 Jptr = apj + api[row]; 739 /* add non-zero cols of AP into the sorted linked list lnk */ 740 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 741 } 742 nnz = lnk[0]; 743 744 /* If free space is not available, double the total space in the list */ 745 if (current_space->local_remaining<nnz) { 746 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 747 nspacedouble++; 748 } 749 750 /* Copy data into free space, and zero out denserows */ 751 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 752 753 current_space->array += nnz; 754 current_space->local_used += nnz; 755 current_space->local_remaining -= nnz; 756 757 coi[i+1] = coi[i] + nnz; 758 } 759 760 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 761 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 762 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 763 if (afill_tmp > afill) afill = afill_tmp; 764 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 765 766 /* (3) send j-array (coj) of Co to other processors */ 767 /*--------------------------------------------------*/ 768 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 769 len_s = merge->len_s; 770 merge->nsend = 0; 771 772 773 /* determine row ownership */ 774 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 775 merge->rowmap->n = pn; 776 merge->rowmap->bs = 1; 777 778 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 779 owners = merge->rowmap->range; 780 781 /* determine the number of messages to send, their lengths */ 782 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 783 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 784 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 785 786 proc = 0; 787 for (i=0; i<pon; i++) { 788 while (prmap[i] >= owners[proc+1]) proc++; 789 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 790 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 791 } 792 793 len = 0; /* max length of buf_si[], see (4) */ 794 owners_co[0] = 0; 795 for (proc=0; proc<size; proc++) { 796 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 797 if (len_s[proc]) { 798 merge->nsend++; 799 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 800 len += len_si[proc]; 801 } 802 } 803 804 /* determine the number and length of messages to receive for coi and coj */ 805 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 806 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 807 808 /* post the Irecv and Isend of coj */ 809 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 810 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 811 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 812 for (proc=0, k=0; proc<size; proc++) { 813 if (!len_s[proc]) continue; 814 i = owners_co[proc]; 815 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 816 k++; 817 } 818 819 /* receives and sends of coj are complete */ 820 for (i=0; i<merge->nrecv; i++) { 821 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 822 } 823 ierr = PetscFree(rwaits);CHKERRQ(ierr); 824 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 825 826 /* (4) send and recv coi */ 827 /*-----------------------*/ 828 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 829 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 830 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 831 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 832 for (proc=0,k=0; proc<size; proc++) { 833 if (!len_s[proc]) continue; 834 /* form outgoing message for i-structure: 835 buf_si[0]: nrows to be sent 836 [1:nrows]: row index (global) 837 [nrows+1:2*nrows+1]: i-structure index 838 */ 839 /*-------------------------------------------*/ 840 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 841 buf_si_i = buf_si + nrows+1; 842 buf_si[0] = nrows; 843 buf_si_i[0] = 0; 844 nrows = 0; 845 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 846 nzi = coi[i+1] - coi[i]; 847 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 848 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 849 nrows++; 850 } 851 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 852 k++; 853 buf_si += len_si[proc]; 854 } 855 i = merge->nrecv; 856 while (i--) { 857 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 858 } 859 ierr = PetscFree(rwaits);CHKERRQ(ierr); 860 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 861 862 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 863 ierr = PetscFree(len_ri);CHKERRQ(ierr); 864 ierr = PetscFree(swaits);CHKERRQ(ierr); 865 ierr = PetscFree(buf_s);CHKERRQ(ierr); 866 867 /* (5) compute the local portion of C (mpi mat) */ 868 /*----------------------------------------------*/ 869 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 870 871 /* allocate pti array and free space for accumulating nonzero column info */ 872 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 873 pti[0] = 0; 874 875 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 876 nnz = fill*(pi_loc[pm] + api[am]); 877 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 878 current_space = free_space; 879 880 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 881 for (k=0; k<merge->nrecv; k++) { 882 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 883 nrows = *buf_ri_k[k]; 884 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 885 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 886 } 887 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 888 rmax = 0; 889 for (i=0; i<pn; i++) { 890 /* add pdt[i,:]*AP into lnk */ 891 pnz = pdti[i+1] - pdti[i]; 892 ptJ = pdtj + pdti[i]; 893 for (j=0; j<pnz; j++) { 894 row = ptJ[j]; /* row of AP == col of Pt */ 895 apnz = api[row+1] - api[row]; 896 Jptr = apj + api[row]; 897 /* add non-zero cols of AP into the sorted linked list lnk */ 898 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 899 } 900 901 /* add received col data into lnk */ 902 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 903 if (i == *nextrow[k]) { /* i-th row */ 904 nzi = *(nextci[k]+1) - *nextci[k]; 905 Jptr = buf_rj[k] + *nextci[k]; 906 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 907 nextrow[k]++; nextci[k]++; 908 } 909 } 910 nnz = lnk[0]; 911 912 /* if free space is not available, make more free space */ 913 if (current_space->local_remaining<nnz) { 914 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 915 nspacedouble++; 916 } 917 /* copy data into free space, then initialize lnk */ 918 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 919 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 920 921 current_space->array += nnz; 922 current_space->local_used += nnz; 923 current_space->local_remaining -= nnz; 924 925 pti[i+1] = pti[i] + nnz; 926 if (nnz > rmax) rmax = nnz; 927 } 928 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 929 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 930 931 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 932 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 933 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 934 if (afill_tmp > afill) afill = afill_tmp; 935 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 936 937 /* (6) create symbolic parallel matrix Cmpi */ 938 /*------------------------------------------*/ 939 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 940 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 941 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 942 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 943 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 944 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 945 946 merge->bi = pti; /* Cseq->i */ 947 merge->bj = ptj; /* Cseq->j */ 948 merge->coi = coi; /* Co->i */ 949 merge->coj = coj; /* Co->j */ 950 merge->buf_ri = buf_ri; 951 merge->buf_rj = buf_rj; 952 merge->owners_co = owners_co; 953 merge->destroy = Cmpi->ops->destroy; 954 merge->duplicate = Cmpi->ops->duplicate; 955 956 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 957 Cmpi->assembled = PETSC_FALSE; 958 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 959 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP_old; 960 Cmpi->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ_old; 961 962 /* attach the supporting struct to Cmpi for reuse */ 963 c = (Mat_MPIAIJ*)Cmpi->data; 964 c->ptap = ptap; 965 ptap->api = api; 966 ptap->apj = apj; 967 ptap->rmax = ap_rmax; 968 *C = Cmpi; 969 970 /* flag 'scalable' determines which implementations to be used: 971 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 972 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 973 /* set default scalable */ 974 ptap->scalable = PETSC_FALSE; /* PETSC_TRUE; */ 975 976 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 977 if (!ptap->scalable) { /* Do dense axpy */ 978 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 979 } else { 980 ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 981 } 982 983 #if defined(PETSC_USE_INFO) 984 if (pti[pn] != 0) { 985 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 986 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 987 } else { 988 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 989 } 990 #endif 991 PetscFunctionReturn(0); 992 } 993 994 #undef __FUNCT__ 995 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ_old" 996 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_old(Mat A,Mat P,Mat C) 997 { 998 PetscErrorCode ierr; 999 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1000 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 1001 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1002 Mat_SeqAIJ *p_loc,*p_oth; 1003 Mat_PtAPMPI *ptap; 1004 Mat_Merge_SeqsToMPI *merge; 1005 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 1006 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 1007 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 1008 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 1009 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1010 MPI_Comm comm; 1011 PetscMPIInt size,rank,taga,*len_s; 1012 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1013 PetscInt **buf_ri,**buf_rj; 1014 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1015 MPI_Request *s_waits,*r_waits; 1016 MPI_Status *status; 1017 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1018 PetscInt *api,*apj,*coi,*coj; 1019 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 1020 PetscBool scalable; 1021 #if defined(PTAP_PROFILE) 1022 PetscLogDouble t0,t1,t2,eP,t3,t4,et2_AP=0.0,ePtAP=0.0,t2_0,t2_1,t2_2; 1023 #endif 1024 1025 PetscFunctionBegin; 1026 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1027 #if defined(PTAP_PROFILE) 1028 ierr = PetscTime(&t0);CHKERRQ(ierr); 1029 #endif 1030 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1031 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1032 1033 ptap = c->ptap; 1034 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"); 1035 merge = ptap->merge; 1036 apa = ptap->apa; 1037 scalable = ptap->scalable; 1038 1039 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 1040 /*-----------------------------------------------------*/ 1041 if (ptap->reuse == MAT_INITIAL_MATRIX) { 1042 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 1043 ptap->reuse = MAT_REUSE_MATRIX; 1044 } else { /* update numerical values of P_oth and P_loc */ 1045 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1046 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1047 } 1048 #if defined(PTAP_PROFILE) 1049 ierr = PetscTime(&t1);CHKERRQ(ierr); 1050 eP = t1-t0; 1051 #endif 1052 /* 1053 printf("[%d] Ad: %d, %d; Ao: %d, %d; P_loc: %d, %d; P_oth %d, %d;\n",rank, 1054 a->A->rmap->N,a->A->cmap->N,a->B->rmap->N,a->B->cmap->N, 1055 ptap->P_loc->rmap->N,ptap->P_loc->cmap->N, 1056 ptap->P_oth->rmap->N,ptap->P_oth->cmap->N); 1057 */ 1058 1059 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 1060 /*--------------------------------------------------------------*/ 1061 /* get data from symbolic products */ 1062 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1063 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 1064 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 1065 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 1066 1067 coi = merge->coi; coj = merge->coj; 1068 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1069 1070 bi = merge->bi; bj = merge->bj; 1071 owners = merge->rowmap->range; 1072 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 1073 1074 api = ptap->api; apj = ptap->apj; 1075 1076 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but faster (could take 1/3 scalable time) */ 1077 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 1078 #if 0 1079 /* ------ 10x slower -------------- */ 1080 /*==================================*/ 1081 Mat R = ptap->R; 1082 Mat_SeqAIJ *r = (Mat_SeqAIJ*)R->data; 1083 PetscInt *ri=r->i,*rj=r->j,rnz,arow,l,prow,pcol,pN=P->cmap->N; 1084 PetscScalar *ra=r->a,tmp,cdense[pN]; 1085 1086 ierr = PetscMemzero(cdense,pN*sizeof(PetscScalar));CHKERRQ(ierr); 1087 for (i=0; i<cm; i++) { /* each row of C or R */ 1088 rnz = ri[i+1] - ri[i]; 1089 1090 for (j=0; j<rnz; j++) { /* each nz of R */ 1091 arow = rj[ri[i] + j]; 1092 1093 /* diagonal portion of A */ 1094 anz = ad->i[arow+1] - ad->i[arow]; 1095 for (k=0; k<anz; k++) { /* each nz of Ad */ 1096 tmp = ra[ri[i] + j]*ad->a[ad->i[arow] + k]; 1097 prow = ad->j[ad->i[arow] + k]; 1098 pnz = pi_loc[prow+1] - pi_loc[prow]; 1099 1100 for (l=0; l<pnz; l++) { /* each nz of P_loc */ 1101 pcol = pj_loc[pi_loc[prow] + l]; 1102 cdense[pcol] += tmp*pa_loc[pi_loc[prow] + l]; 1103 } 1104 } 1105 1106 /* off-diagonal portion of A */ 1107 anz = ao->i[arow+1] - ao->i[arow]; 1108 for (k=0; k<anz; k++) { /* each nz of Ao */ 1109 tmp = ra[ri[i] + j]*ao->a[ao->i[arow] + k]; 1110 prow = ao->j[ao->i[arow] + k]; 1111 pnz = pi_oth[prow+1] - pi_oth[prow]; 1112 1113 for (l=0; l<pnz; l++) { /* each nz of P_oth */ 1114 pcol = pj_oth[pi_oth[prow] + l]; 1115 cdense[pcol] += tmp*pa_oth[pi_oth[prow] + l]; 1116 } 1117 } 1118 1119 } /* for (j=0; j<rnz; j++) */ 1120 1121 /* copy cdense[] into ca; zero cdense[] */ 1122 cnz = bi[i+1] - bi[i]; 1123 cj = bj + bi[i]; 1124 ca = ba + bi[i]; 1125 for (j=0; j<cnz; j++) { 1126 ca[j] += cdense[cj[j]]; 1127 cdense[cj[j]] = 0.0; 1128 } 1129 #if 0 1130 if (rank == 0) { 1131 printf("[%d] row %d: ",rank,i); 1132 for (j=0; j<pN; j++) printf(" %g,",cdense[j]); 1133 printf("\n"); 1134 } 1135 for (j=0; j<pN; j++) cdense[j]=0.0; /* zero cdnese[] */ 1136 #endif 1137 } /* for (i=0; i<cm; i++) { */ 1138 #endif 1139 1140 /* ========================================== */ 1141 #if defined(PTAP_PROFILE) 1142 ierr = PetscTime(&t1);CHKERRQ(ierr); 1143 #endif 1144 for (i=0; i<am; i++) { 1145 #if defined(PTAP_PROFILE) 1146 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1147 #endif 1148 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 1149 /*------------------------------------------------------------*/ 1150 apJ = apj + api[i]; 1151 1152 /* diagonal portion of A */ 1153 anz = adi[i+1] - adi[i]; 1154 adj = ad->j + adi[i]; 1155 ada = ad->a + adi[i]; 1156 for (j=0; j<anz; j++) { 1157 row = adj[j]; 1158 pnz = pi_loc[row+1] - pi_loc[row]; 1159 pj = pj_loc + pi_loc[row]; 1160 pa = pa_loc + pi_loc[row]; 1161 1162 /* perform dense axpy */ 1163 valtmp = ada[j]; 1164 for (k=0; k<pnz; k++) { 1165 apa[pj[k]] += valtmp*pa[k]; 1166 } 1167 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1168 } 1169 1170 /* off-diagonal portion of A */ 1171 anz = aoi[i+1] - aoi[i]; 1172 aoj = ao->j + aoi[i]; 1173 aoa = ao->a + aoi[i]; 1174 for (j=0; j<anz; j++) { 1175 row = aoj[j]; 1176 pnz = pi_oth[row+1] - pi_oth[row]; 1177 pj = pj_oth + pi_oth[row]; 1178 pa = pa_oth + pi_oth[row]; 1179 1180 /* perform dense axpy */ 1181 valtmp = aoa[j]; 1182 for (k=0; k<pnz; k++) { 1183 apa[pj[k]] += valtmp*pa[k]; 1184 } 1185 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1186 } 1187 #if defined(PTAP_PROFILE) 1188 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1189 et2_AP += t2_1 - t2_0; 1190 #endif 1191 1192 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1193 /*--------------------------------------------------------------*/ 1194 apnz = api[i+1] - api[i]; 1195 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 1196 pnz = po->i[i+1] - po->i[i]; 1197 poJ = po->j + po->i[i]; 1198 pA = po->a + po->i[i]; 1199 for (j=0; j<pnz; j++) { 1200 row = poJ[j]; 1201 cnz = coi[row+1] - coi[row]; 1202 cj = coj + coi[row]; 1203 ca = coa + coi[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 #if 1 1212 /* put the value into Cd (diagonal part) */ 1213 pnz = pd->i[i+1] - pd->i[i]; 1214 pdJ = pd->j + pd->i[i]; 1215 pA = pd->a + pd->i[i]; 1216 for (j=0; j<pnz; j++) { 1217 row = pdJ[j]; 1218 cnz = bi[row+1] - bi[row]; 1219 cj = bj + bi[row]; 1220 ca = ba + bi[row]; 1221 /* perform dense axpy */ 1222 valtmp = pA[j]; 1223 for (k=0; k<cnz; k++) { 1224 ca[k] += valtmp*apa[cj[k]]; 1225 } 1226 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1227 } 1228 #endif 1229 /* zero the current row of A*P */ 1230 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 1231 #if defined(PTAP_PROFILE) 1232 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1233 ePtAP += t2_2 - t2_1; 1234 #endif 1235 } 1236 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 1237 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 1238 /*-----------------------------------------------------------------------------------------*/ 1239 pA=pa_loc; 1240 for (i=0; i<am; i++) { 1241 #if defined(PTAP_PROFILE) 1242 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1243 #endif 1244 /* form i-th sparse row of A*P */ 1245 apnz = api[i+1] - api[i]; 1246 apJ = apj + api[i]; 1247 /* diagonal portion of A */ 1248 anz = adi[i+1] - adi[i]; 1249 adj = ad->j + adi[i]; 1250 ada = ad->a + adi[i]; 1251 for (j=0; j<anz; j++) { 1252 row = adj[j]; 1253 pnz = pi_loc[row+1] - pi_loc[row]; 1254 pj = pj_loc + pi_loc[row]; 1255 pa = pa_loc + pi_loc[row]; 1256 valtmp = ada[j]; 1257 nextp = 0; 1258 for (k=0; nextp<pnz; k++) { 1259 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1260 apa[k] += valtmp*pa[nextp++]; 1261 } 1262 } 1263 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1264 } 1265 /* off-diagonal portion of A */ 1266 anz = aoi[i+1] - aoi[i]; 1267 aoj = ao->j + aoi[i]; 1268 aoa = ao->a + aoi[i]; 1269 for (j=0; j<anz; j++) { 1270 row = aoj[j]; 1271 pnz = pi_oth[row+1] - pi_oth[row]; 1272 pj = pj_oth + pi_oth[row]; 1273 pa = pa_oth + pi_oth[row]; 1274 valtmp = aoa[j]; 1275 nextp = 0; 1276 for (k=0; nextp<pnz; k++) { 1277 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1278 apa[k] += valtmp*pa[nextp++]; 1279 } 1280 } 1281 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1282 } 1283 #if defined(PTAP_PROFILE) 1284 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1285 et2_AP += t2_1 - t2_0; 1286 #endif 1287 1288 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1289 /*--------------------------------------------------------------*/ 1290 pnz = pi_loc[i+1] - pi_loc[i]; 1291 pJ = pj_loc + pi_loc[i]; 1292 for (j=0; j<pnz; j++) { 1293 nextap = 0; 1294 row = pJ[j]; /* global index */ 1295 if (row < pcstart || row >=pcend) { /* put the value into Co */ 1296 row = *poJ; 1297 cj = coj + coi[row]; 1298 ca = coa + coi[row]; poJ++; 1299 } else { /* put the value into Cd */ 1300 row = *pdJ; 1301 cj = bj + bi[row]; 1302 ca = ba + bi[row]; pdJ++; 1303 } 1304 valtmp = pA[j]; 1305 for (k=0; nextap<apnz; k++) { 1306 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 1307 } 1308 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 1309 } 1310 pA += pnz; 1311 /* zero the current row info for A*P */ 1312 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 1313 #if defined(PTAP_PROFILE) 1314 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1315 ePtAP += t2_2 - t2_1; 1316 #endif 1317 } 1318 } 1319 #if defined(PTAP_PROFILE) 1320 ierr = PetscTime(&t2);CHKERRQ(ierr); 1321 #endif 1322 1323 /* 3) send and recv matrix values coa */ 1324 /*------------------------------------*/ 1325 buf_ri = merge->buf_ri; 1326 buf_rj = merge->buf_rj; 1327 len_s = merge->len_s; 1328 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1329 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1330 1331 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1332 for (proc=0,k=0; proc<size; proc++) { 1333 if (!len_s[proc]) continue; 1334 i = merge->owners_co[proc]; 1335 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1336 k++; 1337 } 1338 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1339 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1340 1341 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1342 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1343 ierr = PetscFree(coa);CHKERRQ(ierr); 1344 #if defined(PTAP_PROFILE) 1345 ierr = PetscTime(&t3);CHKERRQ(ierr); 1346 #endif 1347 1348 /* 4) insert local Cseq and received values into Cmpi */ 1349 /*------------------------------------------------------*/ 1350 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1351 for (k=0; k<merge->nrecv; k++) { 1352 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1353 nrows = *(buf_ri_k[k]); 1354 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1355 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1356 } 1357 1358 for (i=0; i<cm; i++) { 1359 row = owners[rank] + i; /* global row index of C_seq */ 1360 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1361 ba_i = ba + bi[i]; 1362 bnz = bi[i+1] - bi[i]; 1363 /* add received vals into ba */ 1364 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1365 /* i-th row */ 1366 if (i == *nextrow[k]) { 1367 cnz = *(nextci[k]+1) - *nextci[k]; 1368 cj = buf_rj[k] + *(nextci[k]); 1369 ca = abuf_r[k] + *(nextci[k]); 1370 nextcj = 0; 1371 for (j=0; nextcj<cnz; j++) { 1372 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1373 ba_i[j] += ca[nextcj++]; 1374 } 1375 } 1376 nextrow[k]++; nextci[k]++; 1377 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1378 } 1379 } 1380 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1381 } 1382 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1383 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1384 1385 ierr = PetscFree(ba);CHKERRQ(ierr); 1386 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1387 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1388 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1389 #if defined(PTAP_PROFILE) 1390 ierr = PetscTime(&t4);CHKERRQ(ierr); 1391 if (rank==1) { 1392 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); 1393 } 1394 #endif 1395 PetscFunctionReturn(0); 1396 } 1397