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_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 139 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 140 Mat_SeqAIJ *p_loc,*p_oth; 141 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 142 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 143 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 144 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 145 PetscBT lnkbt; 146 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 147 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 148 PetscInt len,proc,*dnz,*onz,*owners; 149 PetscInt nzi,*pti,*ptj; 150 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 151 MPI_Request *swaits,*rwaits; 152 MPI_Status *sstatus,rstatus; 153 Mat_Merge_SeqsToMPI *merge; 154 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 155 PetscReal afill=1.0,afill_tmp; 156 PetscInt rmax; 157 158 /* create struct Mat_PtAPMPI and attached it to C later */ 159 ierr = PetscNew(&ptap);CHKERRQ(ierr); 160 ierr = PetscNew(&merge);CHKERRQ(ierr); 161 ptap->merge = merge; 162 ptap->reuse = MAT_INITIAL_MATRIX; 163 164 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 165 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 166 167 /* get P_loc by taking all local rows of P */ 168 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 169 170 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 171 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 172 pi_loc = p_loc->i; pj_loc = p_loc->j; 173 pi_oth = p_oth->i; pj_oth = p_oth->j; 174 175 /* (1) compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth (api,apj) */ 176 /*--------------------------------------------------------------------------*/ 177 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 178 api[0] = 0; 179 180 /* create and initialize a linked list */ 181 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 182 183 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 184 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 185 186 current_space = free_space; 187 188 for (i=0; i<am; i++) { 189 /* diagonal portion of A */ 190 nzi = adi[i+1] - adi[i]; 191 aj = ad->j + adi[i]; 192 for (j=0; j<nzi; j++) { 193 row = aj[j]; 194 pnz = pi_loc[row+1] - pi_loc[row]; 195 Jptr = pj_loc + pi_loc[row]; 196 /* add non-zero cols of P into the sorted linked list lnk */ 197 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 198 } 199 /* off-diagonal portion of A */ 200 nzi = aoi[i+1] - aoi[i]; 201 aj = ao->j + aoi[i]; 202 for (j=0; j<nzi; j++) { 203 row = aj[j]; 204 pnz = pi_oth[row+1] - pi_oth[row]; 205 Jptr = pj_oth + pi_oth[row]; 206 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 207 } 208 apnz = lnk[0]; 209 api[i+1] = api[i] + apnz; 210 if (ap_rmax < apnz) ap_rmax = apnz; 211 212 /* if free space is not available, double the total space in the list */ 213 if (current_space->local_remaining<apnz) { 214 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 215 nspacedouble++; 216 } 217 218 /* Copy data into free space, then initialize lnk */ 219 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 220 221 current_space->array += apnz; 222 current_space->local_used += apnz; 223 current_space->local_remaining -= apnz; 224 } 225 226 /* Allocate space for apj, initialize apj, and */ 227 /* destroy list of free space and other temporary array(s) */ 228 ierr = PetscMalloc1(api[am]+1,&apj);CHKERRQ(ierr); 229 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 230 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 231 if (afill_tmp > afill) afill = afill_tmp; 232 233 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 234 /*-----------------------------------------------------------------*/ 235 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 236 237 /* then, compute symbolic Co = (p->B)^T*AP */ 238 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 239 >= (num of nonzero rows of C_seq) - pn */ 240 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 241 coi[0] = 0; 242 243 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 244 nnz = fill*(poti[pon] + api[am]); 245 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 246 current_space = free_space; 247 248 for (i=0; i<pon; i++) { 249 pnz = poti[i+1] - poti[i]; 250 ptJ = potj + poti[i]; 251 for (j=0; j<pnz; j++) { 252 row = ptJ[j]; /* row of AP == col of Pot */ 253 apnz = api[row+1] - api[row]; 254 Jptr = apj + api[row]; 255 /* add non-zero cols of AP into the sorted linked list lnk */ 256 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 257 } 258 nnz = lnk[0]; 259 260 /* If free space is not available, double the total space in the list */ 261 if (current_space->local_remaining<nnz) { 262 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 263 nspacedouble++; 264 } 265 266 /* Copy data into free space, and zero out denserows */ 267 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 268 269 current_space->array += nnz; 270 current_space->local_used += nnz; 271 current_space->local_remaining -= nnz; 272 273 coi[i+1] = coi[i] + nnz; 274 } 275 276 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 277 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 278 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 279 if (afill_tmp > afill) afill = afill_tmp; 280 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 281 282 /* (3) send j-array (coj) of Co to other processors */ 283 /*--------------------------------------------------*/ 284 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 285 len_s = merge->len_s; 286 merge->nsend = 0; 287 288 289 /* determine row ownership */ 290 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 291 merge->rowmap->n = pn; 292 merge->rowmap->bs = 1; 293 294 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 295 owners = merge->rowmap->range; 296 297 /* determine the number of messages to send, their lengths */ 298 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 299 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 300 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 301 302 proc = 0; 303 for (i=0; i<pon; i++) { 304 while (prmap[i] >= owners[proc+1]) proc++; 305 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 306 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 307 } 308 309 len = 0; /* max length of buf_si[], see (4) */ 310 owners_co[0] = 0; 311 for (proc=0; proc<size; proc++) { 312 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 313 if (len_s[proc]) { 314 merge->nsend++; 315 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 316 len += len_si[proc]; 317 } 318 } 319 320 /* determine the number and length of messages to receive for coi and coj */ 321 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 322 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 323 324 /* post the Irecv and Isend of coj */ 325 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 326 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 327 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 328 for (proc=0, k=0; proc<size; proc++) { 329 if (!len_s[proc]) continue; 330 i = owners_co[proc]; 331 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 332 k++; 333 } 334 335 /* receives and sends of coj are complete */ 336 for (i=0; i<merge->nrecv; i++) { 337 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 338 } 339 ierr = PetscFree(rwaits);CHKERRQ(ierr); 340 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 341 342 /* (4) send and recv coi */ 343 /*-----------------------*/ 344 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 345 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 346 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 347 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 348 for (proc=0,k=0; proc<size; proc++) { 349 if (!len_s[proc]) continue; 350 /* form outgoing message for i-structure: 351 buf_si[0]: nrows to be sent 352 [1:nrows]: row index (global) 353 [nrows+1:2*nrows+1]: i-structure index 354 */ 355 /*-------------------------------------------*/ 356 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 357 buf_si_i = buf_si + nrows+1; 358 buf_si[0] = nrows; 359 buf_si_i[0] = 0; 360 nrows = 0; 361 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 362 nzi = coi[i+1] - coi[i]; 363 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 364 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 365 nrows++; 366 } 367 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 368 k++; 369 buf_si += len_si[proc]; 370 } 371 i = merge->nrecv; 372 while (i--) { 373 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 374 } 375 ierr = PetscFree(rwaits);CHKERRQ(ierr); 376 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 377 378 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 379 ierr = PetscFree(len_ri);CHKERRQ(ierr); 380 ierr = PetscFree(swaits);CHKERRQ(ierr); 381 ierr = PetscFree(buf_s);CHKERRQ(ierr); 382 383 /* (5) compute the local portion of C (mpi mat) */ 384 /*----------------------------------------------*/ 385 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 386 387 /* allocate pti array and free space for accumulating nonzero column info */ 388 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 389 pti[0] = 0; 390 391 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 392 nnz = fill*(pi_loc[pm] + api[am]); 393 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 394 current_space = free_space; 395 396 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 397 for (k=0; k<merge->nrecv; k++) { 398 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 399 nrows = *buf_ri_k[k]; 400 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 401 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 402 } 403 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 404 rmax = 0; 405 for (i=0; i<pn; i++) { 406 /* add pdt[i,:]*AP into lnk */ 407 pnz = pdti[i+1] - pdti[i]; 408 ptJ = pdtj + pdti[i]; 409 for (j=0; j<pnz; j++) { 410 row = ptJ[j]; /* row of AP == col of Pt */ 411 apnz = api[row+1] - api[row]; 412 Jptr = apj + api[row]; 413 /* add non-zero cols of AP into the sorted linked list lnk */ 414 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 415 } 416 417 /* add received col data into lnk */ 418 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 419 if (i == *nextrow[k]) { /* i-th row */ 420 nzi = *(nextci[k]+1) - *nextci[k]; 421 Jptr = buf_rj[k] + *nextci[k]; 422 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 423 nextrow[k]++; nextci[k]++; 424 } 425 } 426 nnz = lnk[0]; 427 428 /* if free space is not available, make more free space */ 429 if (current_space->local_remaining<nnz) { 430 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 431 nspacedouble++; 432 } 433 /* copy data into free space, then initialize lnk */ 434 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 435 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 436 437 current_space->array += nnz; 438 current_space->local_used += nnz; 439 current_space->local_remaining -= nnz; 440 441 pti[i+1] = pti[i] + nnz; 442 if (nnz > rmax) rmax = nnz; 443 } 444 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 445 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 446 447 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 448 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 449 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 450 if (afill_tmp > afill) afill = afill_tmp; 451 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 452 453 /* (6) create symbolic parallel matrix Cmpi */ 454 /*------------------------------------------*/ 455 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 456 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 457 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 458 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 459 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 460 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 461 462 merge->bi = pti; /* Cseq->i */ 463 merge->bj = ptj; /* Cseq->j */ 464 merge->coi = coi; /* Co->i */ 465 merge->coj = coj; /* Co->j */ 466 merge->buf_ri = buf_ri; 467 merge->buf_rj = buf_rj; 468 merge->owners_co = owners_co; 469 merge->destroy = Cmpi->ops->destroy; 470 merge->duplicate = Cmpi->ops->duplicate; 471 472 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 473 Cmpi->assembled = PETSC_FALSE; 474 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 475 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 476 477 /* attach the supporting struct to Cmpi for reuse */ 478 c = (Mat_MPIAIJ*)Cmpi->data; 479 c->ptap = ptap; 480 ptap->api = api; 481 ptap->apj = apj; 482 ptap->rmax = ap_rmax; 483 *C = Cmpi; 484 485 //============================================================================== 486 c = (Mat_MPIAIJ*)(*C)->data; 487 ptap = c->ptap; 488 489 #if 0 490 /* create struct Mat_PtAPMPI and attached it to C later */ 491 ierr = PetscNew(&ptap);CHKERRQ(ierr); 492 493 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 494 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 495 496 /* get P_loc by taking all local rows of P */ 497 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 498 #endif 499 //=========================================== 500 ptap->reuse = MAT_INITIAL_MATRIX; 501 502 /* (1) compute symbolic AP = A*P, then get AP_loc */ 503 /*--------------------------------------------------------------------------*/ 504 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 505 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 506 507 ierr = MatMatMult(A,P,MAT_INITIAL_MATRIX,2.0,&AP);CHKERRQ(ierr); 508 ierr = MatMPIAIJGetLocalMat(AP,MAT_INITIAL_MATRIX,&ptap->AP_loc);CHKERRQ(ierr); 509 ierr = MatDestroy(&AP);CHKERRQ(ierr); 510 511 /* (2) compute C_loc=Rd*AP_loc, Co=Ro*AP_loc */ 512 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,MAT_INITIAL_MATRIX,2.0,&ptap->C_loc);CHKERRQ(ierr); 513 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,MAT_INITIAL_MATRIX,2.0,&ptap->C_oth);CHKERRQ(ierr); 514 515 /* (6) create symbolic parallel matrix Cmpi */ 516 /*------------------------------------------*/ 517 518 Mat C_loc=ptap->C_loc; 519 520 /* estimate dnz, onz arrays */ 521 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 522 PetscInt i; 523 Mat_SeqAIJ *c_loc = (Mat_SeqAIJ*)C_loc->data; 524 for (i=0; i<C_loc->rmap->N; i++) { 525 printf("%d \n",c_loc->ilen[i]); 526 //dnz[i] = c_loc->ilen[i]; if (c_loc->ilen[i] > pn) dnz[i] = pn; 527 //onz[i] = c_loc->ilen[i]; if (c_loc->ilen[i] > pN - pn) onz[i] = pN - pn; 528 dnz[i] = pn; onz[i] = pN - pn; 529 } 530 531 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 532 533 /* flag 'scalable' determines which implementations to be used: 534 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 535 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 536 /* set default scalable */ 537 ptap->scalable = PETSC_FALSE; //PETSC_TRUE; 538 539 ierr = PetscOptionsGetBool(((PetscObject)(*C))->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 540 if (!ptap->scalable) { /* Do dense axpy */ 541 ierr = PetscCalloc1(P->cmap->N,&ptap->apa);CHKERRQ(ierr); 542 } else { 543 //ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 544 } 545 546 //Mat_SeqAIJ *ap=(Mat_SeqAIJ*)(ptap->AP_loc)->data; 547 //ptap->api = ap->i; 548 //ptap->apj = ap->j; 549 //ptap->rmax = ap_rmax; 550 551 PetscFunctionReturn(0); 552 } 553 554 #undef __FUNCT__ 555 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ_new" 556 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_new(Mat A,Mat P,Mat C) 557 { 558 PetscErrorCode ierr; 559 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 560 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 561 Mat_PtAPMPI *ptap = c->ptap; 562 Mat AP_loc,C_loc,C_oth; 563 PetscInt i,rstart,rend,cm,ncols,row; 564 PetscMPIInt rank; 565 MPI_Comm comm; 566 const PetscInt *cols; 567 const PetscScalar *vals; 568 PetscLogDouble t0,t1,t2,t3,t4,eR,eAP,eCseq,eCmpi; 569 570 PetscFunctionBegin; 571 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 572 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 573 574 ierr = MatZeroEntries(C);CHKERRQ(ierr); 575 576 /* 1) get R = Pd^T,Ro = Po^T */ 577 ierr = PetscTime(&t0);CHKERRQ(ierr); 578 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 579 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 580 ierr = PetscTime(&t1);CHKERRQ(ierr); 581 eR = t1 - t0; 582 583 /* 2) get AP_loc */ 584 AP_loc = ptap->AP_loc; 585 Mat_SeqAIJ *ap=(Mat_SeqAIJ*)AP_loc->data; 586 587 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 588 /*-----------------------------------------------------*/ 589 if (ptap->reuse == MAT_INITIAL_MATRIX) { 590 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 591 ptap->reuse = MAT_REUSE_MATRIX; 592 } else { /* update numerical values of P_oth and P_loc */ 593 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 594 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 595 } 596 597 /* 2-2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 598 /*--------------------------------------------------------------*/ 599 /* get data from symbolic products */ 600 Mat_SeqAIJ *p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 601 Mat_SeqAIJ *p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 602 PetscInt *api,*apj,am = A->rmap->n,j,col,apnz; 603 PetscScalar *apa = ptap->apa; 604 605 api = ap->i; 606 apj = ap->j; 607 for (i=0; i<am; i++) { 608 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 609 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 610 apnz = api[i+1] - api[i]; 611 for (j=0; j<apnz; j++) { 612 col = apj[j+api[i]]; 613 ap->a[j+ap->i[i]] = apa[col]; 614 apa[col] = 0.0; 615 } 616 } 617 618 ierr = PetscTime(&t2);CHKERRQ(ierr); 619 eAP = t2 - t1; 620 621 /* 3) C_loc = R*AP_loc, Co = Ro*AP_loc */ 622 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Rd,AP_loc,MAT_REUSE_MATRIX,2.0,&ptap->C_loc);CHKERRQ(ierr); 623 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Ro,AP_loc,MAT_REUSE_MATRIX,2.0,&ptap->C_oth);CHKERRQ(ierr); 624 C_loc = ptap->C_loc; 625 C_oth = ptap->C_oth; 626 //printf("[%d] Co %d, %d\n", rank,Co->rmap->N,Co->cmap->N); 627 ierr = PetscTime(&t3);CHKERRQ(ierr); 628 eCseq = t3 - t2; 629 630 /* add C_loc and Co to to C */ 631 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 632 633 /* C_loc -> C */ 634 cm = C_loc->rmap->N; 635 Mat_SeqAIJ *c_seq; 636 c_seq = (Mat_SeqAIJ*)C_loc->data; 637 for (i=0; i<cm; i++) { 638 ncols = c_seq->i[i+1] - c_seq->i[i]; 639 row = rstart + i; 640 cols = c_seq->j + c_seq->i[i]; 641 vals = c_seq->a + c_seq->i[i]; 642 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 643 } 644 645 /* Co -> C, off-processor part */ 646 //printf("[%d] p->B %d, %d\n",rank,p->B->rmap->N,p->B->cmap->N); 647 cm = C_oth->rmap->N; 648 c_seq = (Mat_SeqAIJ*)C_oth->data; 649 for (i=0; i<cm; i++) { 650 ncols = c_seq->i[i+1] - c_seq->i[i]; 651 row = p->garray[i]; 652 cols = c_seq->j + c_seq->i[i]; 653 vals = c_seq->a + c_seq->i[i]; 654 //printf("[%d] row[%d] = %d\n",rank,i,row); 655 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 656 } 657 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 658 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 659 ierr = PetscTime(&t4);CHKERRQ(ierr); 660 eCmpi = t4 - t3; 661 662 if (rank==1) { 663 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); 664 } 665 PetscFunctionReturn(0); 666 } 667 668 #undef __FUNCT__ 669 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 670 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 671 { 672 PetscErrorCode ierr; 673 Mat Cmpi; 674 Mat_PtAPMPI *ptap; 675 PetscFreeSpaceList free_space=NULL,current_space=NULL; 676 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 677 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 678 Mat_SeqAIJ *p_loc,*p_oth; 679 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 680 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 681 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 682 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 683 PetscBT lnkbt; 684 MPI_Comm comm; 685 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 686 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 687 PetscInt len,proc,*dnz,*onz,*owners; 688 PetscInt nzi,*pti,*ptj; 689 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 690 MPI_Request *swaits,*rwaits; 691 MPI_Status *sstatus,rstatus; 692 Mat_Merge_SeqsToMPI *merge; 693 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 694 PetscReal afill=1.0,afill_tmp; 695 PetscInt rmax; 696 697 PetscFunctionBegin; 698 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 699 700 /* check if matrix local sizes are compatible */ 701 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 702 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); 703 } 704 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 705 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); 706 } 707 708 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 709 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 710 711 /* create struct Mat_PtAPMPI and attached it to C later */ 712 ierr = PetscNew(&ptap);CHKERRQ(ierr); 713 ierr = PetscNew(&merge);CHKERRQ(ierr); 714 ptap->merge = merge; 715 ptap->reuse = MAT_INITIAL_MATRIX; 716 717 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 718 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 719 720 /* get P_loc by taking all local rows of P */ 721 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 722 723 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 724 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 725 pi_loc = p_loc->i; pj_loc = p_loc->j; 726 pi_oth = p_oth->i; pj_oth = p_oth->j; 727 728 /* (1) compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth (api,apj) */ 729 /*--------------------------------------------------------------------------*/ 730 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 731 api[0] = 0; 732 733 /* create and initialize a linked list */ 734 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 735 736 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 737 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 738 739 current_space = free_space; 740 741 for (i=0; i<am; i++) { 742 /* diagonal portion of A */ 743 nzi = adi[i+1] - adi[i]; 744 aj = ad->j + adi[i]; 745 for (j=0; j<nzi; j++) { 746 row = aj[j]; 747 pnz = pi_loc[row+1] - pi_loc[row]; 748 Jptr = pj_loc + pi_loc[row]; 749 /* add non-zero cols of P into the sorted linked list lnk */ 750 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 751 } 752 /* off-diagonal portion of A */ 753 nzi = aoi[i+1] - aoi[i]; 754 aj = ao->j + aoi[i]; 755 for (j=0; j<nzi; j++) { 756 row = aj[j]; 757 pnz = pi_oth[row+1] - pi_oth[row]; 758 Jptr = pj_oth + pi_oth[row]; 759 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 760 } 761 apnz = lnk[0]; 762 api[i+1] = api[i] + apnz; 763 if (ap_rmax < apnz) ap_rmax = apnz; 764 765 /* if free space is not available, double the total space in the list */ 766 if (current_space->local_remaining<apnz) { 767 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 768 nspacedouble++; 769 } 770 771 /* Copy data into free space, then initialize lnk */ 772 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 773 774 current_space->array += apnz; 775 current_space->local_used += apnz; 776 current_space->local_remaining -= apnz; 777 } 778 779 /* Allocate space for apj, initialize apj, and */ 780 /* destroy list of free space and other temporary array(s) */ 781 ierr = PetscMalloc1(api[am]+1,&apj);CHKERRQ(ierr); 782 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 783 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 784 if (afill_tmp > afill) afill = afill_tmp; 785 786 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 787 /*-----------------------------------------------------------------*/ 788 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 789 790 /* then, compute symbolic Co = (p->B)^T*AP */ 791 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 792 >= (num of nonzero rows of C_seq) - pn */ 793 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 794 coi[0] = 0; 795 796 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 797 nnz = fill*(poti[pon] + api[am]); 798 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 799 current_space = free_space; 800 801 for (i=0; i<pon; i++) { 802 pnz = poti[i+1] - poti[i]; 803 ptJ = potj + poti[i]; 804 for (j=0; j<pnz; j++) { 805 row = ptJ[j]; /* row of AP == col of Pot */ 806 apnz = api[row+1] - api[row]; 807 Jptr = apj + api[row]; 808 /* add non-zero cols of AP into the sorted linked list lnk */ 809 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 810 } 811 nnz = lnk[0]; 812 813 /* If free space is not available, double the total space in the list */ 814 if (current_space->local_remaining<nnz) { 815 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 816 nspacedouble++; 817 } 818 819 /* Copy data into free space, and zero out denserows */ 820 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 821 822 current_space->array += nnz; 823 current_space->local_used += nnz; 824 current_space->local_remaining -= nnz; 825 826 coi[i+1] = coi[i] + nnz; 827 } 828 829 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 830 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 831 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 832 if (afill_tmp > afill) afill = afill_tmp; 833 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 834 835 /* (3) send j-array (coj) of Co to other processors */ 836 /*--------------------------------------------------*/ 837 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 838 len_s = merge->len_s; 839 merge->nsend = 0; 840 841 842 /* determine row ownership */ 843 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 844 merge->rowmap->n = pn; 845 merge->rowmap->bs = 1; 846 847 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 848 owners = merge->rowmap->range; 849 850 /* determine the number of messages to send, their lengths */ 851 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 852 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 853 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 854 855 proc = 0; 856 for (i=0; i<pon; i++) { 857 while (prmap[i] >= owners[proc+1]) proc++; 858 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 859 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 860 } 861 862 len = 0; /* max length of buf_si[], see (4) */ 863 owners_co[0] = 0; 864 for (proc=0; proc<size; proc++) { 865 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 866 if (len_s[proc]) { 867 merge->nsend++; 868 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 869 len += len_si[proc]; 870 } 871 } 872 873 /* determine the number and length of messages to receive for coi and coj */ 874 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 875 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 876 877 /* post the Irecv and Isend of coj */ 878 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 879 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 880 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 881 for (proc=0, k=0; proc<size; proc++) { 882 if (!len_s[proc]) continue; 883 i = owners_co[proc]; 884 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 885 k++; 886 } 887 888 /* receives and sends of coj are complete */ 889 for (i=0; i<merge->nrecv; i++) { 890 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 891 } 892 ierr = PetscFree(rwaits);CHKERRQ(ierr); 893 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 894 895 /* (4) send and recv coi */ 896 /*-----------------------*/ 897 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 898 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 899 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 900 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 901 for (proc=0,k=0; proc<size; proc++) { 902 if (!len_s[proc]) continue; 903 /* form outgoing message for i-structure: 904 buf_si[0]: nrows to be sent 905 [1:nrows]: row index (global) 906 [nrows+1:2*nrows+1]: i-structure index 907 */ 908 /*-------------------------------------------*/ 909 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 910 buf_si_i = buf_si + nrows+1; 911 buf_si[0] = nrows; 912 buf_si_i[0] = 0; 913 nrows = 0; 914 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 915 nzi = coi[i+1] - coi[i]; 916 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 917 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 918 nrows++; 919 } 920 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 921 k++; 922 buf_si += len_si[proc]; 923 } 924 i = merge->nrecv; 925 while (i--) { 926 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 927 } 928 ierr = PetscFree(rwaits);CHKERRQ(ierr); 929 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 930 931 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 932 ierr = PetscFree(len_ri);CHKERRQ(ierr); 933 ierr = PetscFree(swaits);CHKERRQ(ierr); 934 ierr = PetscFree(buf_s);CHKERRQ(ierr); 935 936 /* (5) compute the local portion of C (mpi mat) */ 937 /*----------------------------------------------*/ 938 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 939 940 /* allocate pti array and free space for accumulating nonzero column info */ 941 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 942 pti[0] = 0; 943 944 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 945 nnz = fill*(pi_loc[pm] + api[am]); 946 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 947 current_space = free_space; 948 949 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 950 for (k=0; k<merge->nrecv; k++) { 951 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 952 nrows = *buf_ri_k[k]; 953 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 954 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 955 } 956 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 957 rmax = 0; 958 for (i=0; i<pn; i++) { 959 /* add pdt[i,:]*AP into lnk */ 960 pnz = pdti[i+1] - pdti[i]; 961 ptJ = pdtj + pdti[i]; 962 for (j=0; j<pnz; j++) { 963 row = ptJ[j]; /* row of AP == col of Pt */ 964 apnz = api[row+1] - api[row]; 965 Jptr = apj + api[row]; 966 /* add non-zero cols of AP into the sorted linked list lnk */ 967 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 968 } 969 970 /* add received col data into lnk */ 971 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 972 if (i == *nextrow[k]) { /* i-th row */ 973 nzi = *(nextci[k]+1) - *nextci[k]; 974 Jptr = buf_rj[k] + *nextci[k]; 975 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 976 nextrow[k]++; nextci[k]++; 977 } 978 } 979 nnz = lnk[0]; 980 981 /* if free space is not available, make more free space */ 982 if (current_space->local_remaining<nnz) { 983 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 984 nspacedouble++; 985 } 986 /* copy data into free space, then initialize lnk */ 987 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 988 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 989 990 current_space->array += nnz; 991 current_space->local_used += nnz; 992 current_space->local_remaining -= nnz; 993 994 pti[i+1] = pti[i] + nnz; 995 if (nnz > rmax) rmax = nnz; 996 } 997 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 998 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 999 1000 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 1001 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 1002 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 1003 if (afill_tmp > afill) afill = afill_tmp; 1004 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1005 1006 /* (6) create symbolic parallel matrix Cmpi */ 1007 /*------------------------------------------*/ 1008 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1009 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1010 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 1011 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1012 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1013 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1014 1015 merge->bi = pti; /* Cseq->i */ 1016 merge->bj = ptj; /* Cseq->j */ 1017 merge->coi = coi; /* Co->i */ 1018 merge->coj = coj; /* Co->j */ 1019 merge->buf_ri = buf_ri; 1020 merge->buf_rj = buf_rj; 1021 merge->owners_co = owners_co; 1022 merge->destroy = Cmpi->ops->destroy; 1023 merge->duplicate = Cmpi->ops->duplicate; 1024 1025 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 1026 Cmpi->assembled = PETSC_FALSE; 1027 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1028 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 1029 1030 /* attach the supporting struct to Cmpi for reuse */ 1031 c = (Mat_MPIAIJ*)Cmpi->data; 1032 c->ptap = ptap; 1033 ptap->api = api; 1034 ptap->apj = apj; 1035 ptap->rmax = ap_rmax; 1036 *C = Cmpi; 1037 1038 /* flag 'scalable' determines which implementations to be used: 1039 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 1040 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 1041 /* set default scalable */ 1042 ptap->scalable = PETSC_FALSE; //PETSC_TRUE; 1043 1044 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 1045 if (!ptap->scalable) { /* Do dense axpy */ 1046 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 1047 } else { 1048 ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 1049 } 1050 1051 #if defined(PETSC_USE_INFO) 1052 if (pti[pn] != 0) { 1053 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 1054 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 1055 } else { 1056 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1057 } 1058 #endif 1059 PetscFunctionReturn(0); 1060 } 1061 1062 #undef __FUNCT__ 1063 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 1064 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 1065 { 1066 PetscErrorCode ierr; 1067 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1068 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 1069 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1070 Mat_SeqAIJ *p_loc,*p_oth; 1071 Mat_PtAPMPI *ptap; 1072 Mat_Merge_SeqsToMPI *merge; 1073 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 1074 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 1075 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 1076 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 1077 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1078 MPI_Comm comm; 1079 PetscMPIInt size,rank,taga,*len_s; 1080 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1081 PetscInt **buf_ri,**buf_rj; 1082 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1083 MPI_Request *s_waits,*r_waits; 1084 MPI_Status *status; 1085 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1086 PetscInt *api,*apj,*coi,*coj; 1087 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 1088 PetscBool scalable; 1089 #if defined(PTAP_PROFILE) 1090 PetscLogDouble t0,t1,t2,eP,t3,t4,et2_AP=0.0,ePtAP=0.0,t2_0,t2_1,t2_2; 1091 #endif 1092 1093 PetscFunctionBegin; 1094 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1095 #if defined(PTAP_PROFILE) 1096 ierr = PetscTime(&t0);CHKERRQ(ierr); 1097 #endif 1098 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1099 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1100 1101 ptap = c->ptap; 1102 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"); 1103 merge = ptap->merge; 1104 apa = ptap->apa; 1105 scalable = ptap->scalable; 1106 1107 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 1108 /*-----------------------------------------------------*/ 1109 if (ptap->reuse == MAT_INITIAL_MATRIX) { 1110 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 1111 ptap->reuse = MAT_REUSE_MATRIX; 1112 } else { /* update numerical values of P_oth and P_loc */ 1113 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1114 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1115 } 1116 #if defined(PTAP_PROFILE) 1117 ierr = PetscTime(&t1);CHKERRQ(ierr); 1118 eP = t1-t0; 1119 #endif 1120 /* 1121 printf("[%d] Ad: %d, %d; Ao: %d, %d; P_loc: %d, %d; P_oth %d, %d;\n",rank, 1122 a->A->rmap->N,a->A->cmap->N,a->B->rmap->N,a->B->cmap->N, 1123 ptap->P_loc->rmap->N,ptap->P_loc->cmap->N, 1124 ptap->P_oth->rmap->N,ptap->P_oth->cmap->N); 1125 */ 1126 1127 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 1128 /*--------------------------------------------------------------*/ 1129 /* get data from symbolic products */ 1130 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1131 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 1132 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 1133 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 1134 1135 coi = merge->coi; coj = merge->coj; 1136 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1137 1138 bi = merge->bi; bj = merge->bj; 1139 owners = merge->rowmap->range; 1140 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 1141 1142 api = ptap->api; apj = ptap->apj; 1143 1144 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but faster (could take 1/3 scalable time) */ 1145 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 1146 #if 0 1147 /* ------ 10x slower -------------- */ 1148 /*==================================*/ 1149 Mat R = ptap->R; 1150 Mat_SeqAIJ *r = (Mat_SeqAIJ*)R->data; 1151 PetscInt *ri=r->i,*rj=r->j,rnz,arow,l,prow,pcol,pN=P->cmap->N; 1152 PetscScalar *ra=r->a,tmp,cdense[pN]; 1153 1154 ierr = PetscMemzero(cdense,pN*sizeof(PetscScalar));CHKERRQ(ierr); 1155 for (i=0; i<cm; i++) { /* each row of C or R */ 1156 rnz = ri[i+1] - ri[i]; 1157 1158 for (j=0; j<rnz; j++) { /* each nz of R */ 1159 arow = rj[ri[i] + j]; 1160 1161 /* diagonal portion of A */ 1162 anz = ad->i[arow+1] - ad->i[arow]; 1163 for (k=0; k<anz; k++) { /* each nz of Ad */ 1164 tmp = ra[ri[i] + j]*ad->a[ad->i[arow] + k]; 1165 prow = ad->j[ad->i[arow] + k]; 1166 pnz = pi_loc[prow+1] - pi_loc[prow]; 1167 1168 for (l=0; l<pnz; l++) { /* each nz of P_loc */ 1169 pcol = pj_loc[pi_loc[prow] + l]; 1170 cdense[pcol] += tmp*pa_loc[pi_loc[prow] + l]; 1171 } 1172 } 1173 1174 /* off-diagonal portion of A */ 1175 anz = ao->i[arow+1] - ao->i[arow]; 1176 for (k=0; k<anz; k++) { /* each nz of Ao */ 1177 tmp = ra[ri[i] + j]*ao->a[ao->i[arow] + k]; 1178 prow = ao->j[ao->i[arow] + k]; 1179 pnz = pi_oth[prow+1] - pi_oth[prow]; 1180 1181 for (l=0; l<pnz; l++) { /* each nz of P_oth */ 1182 pcol = pj_oth[pi_oth[prow] + l]; 1183 cdense[pcol] += tmp*pa_oth[pi_oth[prow] + l]; 1184 } 1185 } 1186 1187 } //for (j=0; j<rnz; j++) 1188 1189 /* copy cdense[] into ca; zero cdense[] */ 1190 cnz = bi[i+1] - bi[i]; 1191 cj = bj + bi[i]; 1192 ca = ba + bi[i]; 1193 for (j=0; j<cnz; j++) { 1194 ca[j] += cdense[cj[j]]; 1195 cdense[cj[j]] = 0.0; 1196 } 1197 #if 0 1198 if (rank == 0) { 1199 printf("[%d] row %d: ",rank,i); 1200 for (j=0; j<pN; j++) printf(" %g,",cdense[j]); 1201 printf("\n"); 1202 } 1203 for (j=0; j<pN; j++) cdense[j]=0.0; // zero cdnese[] 1204 #endif 1205 } //for (i=0; i<cm; i++) { 1206 #endif 1207 1208 //========================================== 1209 1210 ierr = PetscTime(&t1);CHKERRQ(ierr); 1211 for (i=0; i<am; i++) { 1212 #if defined(PTAP_PROFILE) 1213 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1214 #endif 1215 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 1216 /*------------------------------------------------------------*/ 1217 apJ = apj + api[i]; 1218 1219 /* diagonal portion of A */ 1220 anz = adi[i+1] - adi[i]; 1221 adj = ad->j + adi[i]; 1222 ada = ad->a + adi[i]; 1223 for (j=0; j<anz; j++) { 1224 row = adj[j]; 1225 pnz = pi_loc[row+1] - pi_loc[row]; 1226 pj = pj_loc + pi_loc[row]; 1227 pa = pa_loc + pi_loc[row]; 1228 1229 /* perform dense axpy */ 1230 valtmp = ada[j]; 1231 for (k=0; k<pnz; k++) { 1232 apa[pj[k]] += valtmp*pa[k]; 1233 } 1234 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1235 } 1236 1237 /* off-diagonal portion of A */ 1238 anz = aoi[i+1] - aoi[i]; 1239 aoj = ao->j + aoi[i]; 1240 aoa = ao->a + aoi[i]; 1241 for (j=0; j<anz; j++) { 1242 row = aoj[j]; 1243 pnz = pi_oth[row+1] - pi_oth[row]; 1244 pj = pj_oth + pi_oth[row]; 1245 pa = pa_oth + pi_oth[row]; 1246 1247 /* perform dense axpy */ 1248 valtmp = aoa[j]; 1249 for (k=0; k<pnz; k++) { 1250 apa[pj[k]] += valtmp*pa[k]; 1251 } 1252 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1253 } 1254 #if defined(PTAP_PROFILE) 1255 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1256 et2_AP += t2_1 - t2_0; 1257 #endif 1258 1259 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1260 /*--------------------------------------------------------------*/ 1261 apnz = api[i+1] - api[i]; 1262 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 1263 pnz = po->i[i+1] - po->i[i]; 1264 poJ = po->j + po->i[i]; 1265 pA = po->a + po->i[i]; 1266 for (j=0; j<pnz; j++) { 1267 row = poJ[j]; 1268 cnz = coi[row+1] - coi[row]; 1269 cj = coj + coi[row]; 1270 ca = coa + coi[row]; 1271 /* perform dense axpy */ 1272 valtmp = pA[j]; 1273 for (k=0; k<cnz; k++) { 1274 ca[k] += valtmp*apa[cj[k]]; 1275 } 1276 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1277 } 1278 #if 1 1279 /* put the value into Cd (diagonal part) */ 1280 pnz = pd->i[i+1] - pd->i[i]; 1281 pdJ = pd->j + pd->i[i]; 1282 pA = pd->a + pd->i[i]; 1283 for (j=0; j<pnz; j++) { 1284 row = pdJ[j]; 1285 cnz = bi[row+1] - bi[row]; 1286 cj = bj + bi[row]; 1287 ca = ba + bi[row]; 1288 /* perform dense axpy */ 1289 valtmp = pA[j]; 1290 for (k=0; k<cnz; k++) { 1291 ca[k] += valtmp*apa[cj[k]]; 1292 } 1293 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1294 } 1295 #endif 1296 /* zero the current row of A*P */ 1297 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 1298 #if defined(PTAP_PROFILE) 1299 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1300 ePtAP += t2_2 - t2_1; 1301 #endif 1302 } 1303 1304 if (rank == 100) { 1305 for (row=0; row<cm; row++) { 1306 printf("[%d] row %d: ",rank,row); 1307 cnz = bi[row+1] - bi[row]; 1308 for (j=0; j<cnz; j++) printf(" %g,",ba[bi[row]+j]); 1309 printf("\n"); 1310 } 1311 } 1312 1313 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 1314 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 1315 /*-----------------------------------------------------------------------------------------*/ 1316 pA=pa_loc; 1317 for (i=0; i<am; i++) { 1318 #if defined(PTAP_PROFILE) 1319 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1320 #endif 1321 /* form i-th sparse row of A*P */ 1322 apnz = api[i+1] - api[i]; 1323 apJ = apj + api[i]; 1324 /* diagonal portion of A */ 1325 anz = adi[i+1] - adi[i]; 1326 adj = ad->j + adi[i]; 1327 ada = ad->a + adi[i]; 1328 for (j=0; j<anz; j++) { 1329 row = adj[j]; 1330 pnz = pi_loc[row+1] - pi_loc[row]; 1331 pj = pj_loc + pi_loc[row]; 1332 pa = pa_loc + pi_loc[row]; 1333 valtmp = ada[j]; 1334 nextp = 0; 1335 for (k=0; nextp<pnz; k++) { 1336 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1337 apa[k] += valtmp*pa[nextp++]; 1338 } 1339 } 1340 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1341 } 1342 /* off-diagonal portion of A */ 1343 anz = aoi[i+1] - aoi[i]; 1344 aoj = ao->j + aoi[i]; 1345 aoa = ao->a + aoi[i]; 1346 for (j=0; j<anz; j++) { 1347 row = aoj[j]; 1348 pnz = pi_oth[row+1] - pi_oth[row]; 1349 pj = pj_oth + pi_oth[row]; 1350 pa = pa_oth + pi_oth[row]; 1351 valtmp = aoa[j]; 1352 nextp = 0; 1353 for (k=0; nextp<pnz; k++) { 1354 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1355 apa[k] += valtmp*pa[nextp++]; 1356 } 1357 } 1358 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1359 } 1360 #if defined(PTAP_PROFILE) 1361 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1362 et2_AP += t2_1 - t2_0; 1363 #endif 1364 1365 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1366 /*--------------------------------------------------------------*/ 1367 pnz = pi_loc[i+1] - pi_loc[i]; 1368 pJ = pj_loc + pi_loc[i]; 1369 for (j=0; j<pnz; j++) { 1370 nextap = 0; 1371 row = pJ[j]; /* global index */ 1372 if (row < pcstart || row >=pcend) { /* put the value into Co */ 1373 row = *poJ; 1374 cj = coj + coi[row]; 1375 ca = coa + coi[row]; poJ++; 1376 } else { /* put the value into Cd */ 1377 row = *pdJ; 1378 cj = bj + bi[row]; 1379 ca = ba + bi[row]; pdJ++; 1380 } 1381 valtmp = pA[j]; 1382 for (k=0; nextap<apnz; k++) { 1383 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 1384 } 1385 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 1386 } 1387 pA += pnz; 1388 /* zero the current row info for A*P */ 1389 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 1390 #if defined(PTAP_PROFILE) 1391 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1392 ePtAP += t2_2 - t2_1; 1393 #endif 1394 } 1395 } 1396 #if defined(PTAP_PROFILE) 1397 ierr = PetscTime(&t2);CHKERRQ(ierr); 1398 #endif 1399 1400 /* 3) send and recv matrix values coa */ 1401 /*------------------------------------*/ 1402 buf_ri = merge->buf_ri; 1403 buf_rj = merge->buf_rj; 1404 len_s = merge->len_s; 1405 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1406 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1407 1408 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1409 for (proc=0,k=0; proc<size; proc++) { 1410 if (!len_s[proc]) continue; 1411 i = merge->owners_co[proc]; 1412 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1413 k++; 1414 } 1415 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1416 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1417 1418 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1419 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1420 ierr = PetscFree(coa);CHKERRQ(ierr); 1421 #if defined(PTAP_PROFILE) 1422 ierr = PetscTime(&t3);CHKERRQ(ierr); 1423 #endif 1424 1425 /* 4) insert local Cseq and received values into Cmpi */ 1426 /*------------------------------------------------------*/ 1427 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1428 for (k=0; k<merge->nrecv; k++) { 1429 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1430 nrows = *(buf_ri_k[k]); 1431 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1432 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1433 } 1434 1435 for (i=0; i<cm; i++) { 1436 row = owners[rank] + i; /* global row index of C_seq */ 1437 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1438 ba_i = ba + bi[i]; 1439 bnz = bi[i+1] - bi[i]; 1440 /* add received vals into ba */ 1441 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1442 /* i-th row */ 1443 if (i == *nextrow[k]) { 1444 cnz = *(nextci[k]+1) - *nextci[k]; 1445 cj = buf_rj[k] + *(nextci[k]); 1446 ca = abuf_r[k] + *(nextci[k]); 1447 nextcj = 0; 1448 for (j=0; nextcj<cnz; j++) { 1449 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1450 ba_i[j] += ca[nextcj++]; 1451 } 1452 } 1453 nextrow[k]++; nextci[k]++; 1454 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1455 } 1456 } 1457 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1458 } 1459 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1460 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1461 1462 ierr = PetscFree(ba);CHKERRQ(ierr); 1463 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1464 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1465 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1466 #if defined(PTAP_PROFILE) 1467 ierr = PetscTime(&t4);CHKERRQ(ierr); 1468 if (rank==1) { 1469 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); 1470 } 1471 #endif 1472 PetscFunctionReturn(0); 1473 } 1474