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);CHKERRQ(ierr); 36 ierr = MatDestroy(&ptap->AP_loc);CHKERRQ(ierr); 37 38 if (ptap->api) {ierr = PetscFree(ptap->api);CHKERRQ(ierr);} 39 if (ptap->apj) {ierr = PetscFree(ptap->apj);CHKERRQ(ierr);} 40 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 41 if (merge) { 42 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 43 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 44 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 45 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 46 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 47 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 48 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 49 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 50 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 51 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 52 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 53 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 54 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 55 ierr = merge->destroy(A);CHKERRQ(ierr); 56 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 57 } 58 ierr = PetscFree(ptap);CHKERRQ(ierr); 59 } 60 PetscFunctionReturn(0); 61 } 62 63 #undef __FUNCT__ 64 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP" 65 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 66 { 67 PetscErrorCode ierr; 68 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 69 Mat_PtAPMPI *ptap = a->ptap; 70 Mat_Merge_SeqsToMPI *merge = ptap->merge; 71 72 PetscFunctionBegin; 73 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 74 75 (*M)->ops->destroy = merge->destroy; 76 (*M)->ops->duplicate = merge->duplicate; 77 PetscFunctionReturn(0); 78 } 79 80 #undef __FUNCT__ 81 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ" 82 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 83 { 84 PetscErrorCode ierr; 85 PetscBool newalg=PETSC_FALSE; 86 87 PetscFunctionBegin; 88 ierr = PetscOptionsGetBool(NULL,"-matptap_new",&newalg,NULL);CHKERRQ(ierr); 89 if (scall == MAT_INITIAL_MATRIX) { 90 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 91 if (newalg) { 92 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ_new(A,P,fill,C);CHKERRQ(ierr); 93 } else { 94 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 95 } 96 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 97 } 98 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 99 if (newalg) { 100 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ_new(A,P,*C);CHKERRQ(ierr); 101 } else { 102 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr); 103 } 104 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 105 PetscFunctionReturn(0); 106 } 107 108 #undef __FUNCT__ 109 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ_new" 110 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_new(Mat A,Mat P,PetscReal fill,Mat *C) 111 { 112 PetscErrorCode ierr; 113 Mat_PtAPMPI *ptap; 114 Mat_MPIAIJ *c,*p=(Mat_MPIAIJ*)P->data; 115 116 PetscFunctionBegin; 117 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); //rm later !!! 118 c = (Mat_MPIAIJ*)(*C)->data; 119 ptap = c->ptap; 120 ptap->reuse = MAT_INITIAL_MATRIX; 121 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 122 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 123 124 ierr = MatMatMult(A,P,MAT_INITIAL_MATRIX,2.0,&ptap->AP);CHKERRQ(ierr); 125 ierr = MatMPIAIJGetLocalMat(ptap->AP,MAT_INITIAL_MATRIX,&ptap->AP_loc);CHKERRQ(ierr); 126 PetscFunctionReturn(0); 127 } 128 129 #undef __FUNCT__ 130 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ_new" 131 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_new(Mat A,Mat P,Mat C) 132 { 133 PetscErrorCode ierr; 134 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 135 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 136 Mat_PtAPMPI *ptap = c->ptap; 137 Mat AP_loc,C_loc,Co; 138 PetscInt i,rstart,rend,cm,ncols,row; 139 PetscMPIInt rank; 140 MPI_Comm comm; 141 const PetscInt *cols; 142 const PetscScalar *vals; 143 PetscLogDouble t0,t1,t2,t3,t4,eR,eAP,eCseq,eCmpi; 144 145 PetscFunctionBegin; 146 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 147 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 148 149 ierr = MatZeroEntries(C);CHKERRQ(ierr); 150 151 /* 1) get R = Pd^T,Ro = Po^T */ 152 ierr = PetscTime(&t0);CHKERRQ(ierr); 153 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 154 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 155 ierr = PetscTime(&t1);CHKERRQ(ierr); 156 eR = t1 - t0; 157 158 /* 2) get AP_loc */ 159 AP_loc = ptap->AP_loc; 160 Mat_SeqAIJ *ap=(Mat_SeqAIJ*)AP_loc->data; 161 162 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 163 /*-----------------------------------------------------*/ 164 if (ptap->reuse == MAT_INITIAL_MATRIX) { 165 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 166 ptap->reuse = MAT_REUSE_MATRIX; 167 } else { /* update numerical values of P_oth and P_loc */ 168 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 169 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 170 } 171 172 /* 2-2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 173 /*--------------------------------------------------------------*/ 174 /* get data from symbolic products */ 175 Mat_SeqAIJ *p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 176 Mat_SeqAIJ *p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 177 PetscInt *api,*apj,am = A->rmap->n,j,col,apnz; 178 PetscScalar *apa = ptap->apa; 179 180 api = ptap->api; apj = ptap->apj; 181 for (i=0; i<am; i++) { 182 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 183 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 184 apnz = api[i+1] - api[i]; 185 for (j=0; j<apnz; j++) { 186 col = apj[j+api[i]]; 187 //if (ap->a[j+ap->i[i]] != apa[col]) SETERRQ4(NULL,PETSC_ERR_ARG_INCOMP,"%d %d ap %g != apa %g",i,col,ap->a[j+ap->i[i]],apa[col]); 188 ap->a[j+ap->i[i]] = apa[col]; 189 apa[col] = 0.0; 190 } 191 } 192 193 ierr = PetscTime(&t2);CHKERRQ(ierr); 194 eAP = t2 - t1; 195 196 /* 3) C_loc = R*AP_loc, Co = Ro*AP_loc */ 197 ierr = MatMatMult(ptap->Rd,AP_loc,MAT_INITIAL_MATRIX,2.0,&C_loc);CHKERRQ(ierr); 198 ierr = MatMatMult(ptap->Ro,AP_loc,MAT_INITIAL_MATRIX,2.0,&Co);CHKERRQ(ierr); 199 //printf("[%d] Co %d, %d\n", rank,Co->rmap->N,Co->cmap->N); 200 ierr = PetscTime(&t3);CHKERRQ(ierr); 201 eCseq = t3 - t2; 202 203 /* add C_loc and Co to to C */ 204 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 205 206 /* C_loc -> C */ 207 cm = C_loc->rmap->N; 208 for (i=0; i<cm; i++) { 209 ierr = MatGetRow(C_loc,i,&ncols,&cols,&vals);CHKERRQ(ierr); 210 row = rstart + i; 211 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 212 ierr = MatRestoreRow(C_loc,i,&ncols,&cols,&vals);CHKERRQ(ierr); 213 } 214 215 /* Co -> C, off-processor part */ 216 //printf("[%d] p->B %d, %d\n",rank,p->B->rmap->N,p->B->cmap->N); 217 for (i=0; i<Co->rmap->N; i++) { 218 ierr = MatGetRow(Co,i,&ncols,&cols,&vals);CHKERRQ(ierr); 219 row = p->garray[i]; 220 //printf("[%d] row[%d] = %d\n",rank,i,row); 221 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 222 ierr = MatRestoreRow(Co,i,&ncols,&cols,&vals);CHKERRQ(ierr); 223 } 224 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 225 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 226 ierr = PetscTime(&t4);CHKERRQ(ierr); 227 eCmpi = t4 - t3; 228 229 ierr = MatDestroy(&C_loc);CHKERRQ(ierr); 230 ierr = MatDestroy(&Co);CHKERRQ(ierr); 231 232 if (rank==1) { 233 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); 234 } 235 PetscFunctionReturn(0); 236 } 237 238 #undef __FUNCT__ 239 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 240 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 241 { 242 PetscErrorCode ierr; 243 Mat Cmpi; 244 Mat_PtAPMPI *ptap; 245 PetscFreeSpaceList free_space=NULL,current_space=NULL; 246 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 247 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 248 Mat_SeqAIJ *p_loc,*p_oth; 249 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 250 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 251 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 252 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 253 PetscBT lnkbt; 254 MPI_Comm comm; 255 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 256 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 257 PetscInt len,proc,*dnz,*onz,*owners; 258 PetscInt nzi,*pti,*ptj; 259 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 260 MPI_Request *swaits,*rwaits; 261 MPI_Status *sstatus,rstatus; 262 Mat_Merge_SeqsToMPI *merge; 263 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 264 PetscReal afill=1.0,afill_tmp; 265 PetscInt rmax; 266 267 PetscFunctionBegin; 268 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 269 270 /* check if matrix local sizes are compatible */ 271 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 272 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); 273 } 274 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 275 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); 276 } 277 278 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 279 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 280 281 /* create struct Mat_PtAPMPI and attached it to C later */ 282 ierr = PetscNew(&ptap);CHKERRQ(ierr); 283 ierr = PetscNew(&merge);CHKERRQ(ierr); 284 ptap->merge = merge; 285 ptap->reuse = MAT_INITIAL_MATRIX; 286 287 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 288 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 289 290 /* get P_loc by taking all local rows of P */ 291 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 292 293 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 294 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 295 pi_loc = p_loc->i; pj_loc = p_loc->j; 296 pi_oth = p_oth->i; pj_oth = p_oth->j; 297 298 /* (1) compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth (api,apj) */ 299 /*--------------------------------------------------------------------------*/ 300 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 301 api[0] = 0; 302 303 /* create and initialize a linked list */ 304 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 305 306 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 307 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 308 309 current_space = free_space; 310 311 for (i=0; i<am; i++) { 312 /* diagonal portion of A */ 313 nzi = adi[i+1] - adi[i]; 314 aj = ad->j + adi[i]; 315 for (j=0; j<nzi; j++) { 316 row = aj[j]; 317 pnz = pi_loc[row+1] - pi_loc[row]; 318 Jptr = pj_loc + pi_loc[row]; 319 /* add non-zero cols of P into the sorted linked list lnk */ 320 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 321 } 322 /* off-diagonal portion of A */ 323 nzi = aoi[i+1] - aoi[i]; 324 aj = ao->j + aoi[i]; 325 for (j=0; j<nzi; j++) { 326 row = aj[j]; 327 pnz = pi_oth[row+1] - pi_oth[row]; 328 Jptr = pj_oth + pi_oth[row]; 329 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 330 } 331 apnz = lnk[0]; 332 api[i+1] = api[i] + apnz; 333 if (ap_rmax < apnz) ap_rmax = apnz; 334 335 /* if free space is not available, double the total space in the list */ 336 if (current_space->local_remaining<apnz) { 337 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 338 nspacedouble++; 339 } 340 341 /* Copy data into free space, then initialize lnk */ 342 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 343 344 current_space->array += apnz; 345 current_space->local_used += apnz; 346 current_space->local_remaining -= apnz; 347 } 348 349 /* Allocate space for apj, initialize apj, and */ 350 /* destroy list of free space and other temporary array(s) */ 351 ierr = PetscMalloc1(api[am]+1,&apj);CHKERRQ(ierr); 352 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 353 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 354 if (afill_tmp > afill) afill = afill_tmp; 355 356 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 357 /*-----------------------------------------------------------------*/ 358 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 359 360 /* then, compute symbolic Co = (p->B)^T*AP */ 361 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 362 >= (num of nonzero rows of C_seq) - pn */ 363 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 364 coi[0] = 0; 365 366 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 367 nnz = fill*(poti[pon] + api[am]); 368 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 369 current_space = free_space; 370 371 for (i=0; i<pon; i++) { 372 pnz = poti[i+1] - poti[i]; 373 ptJ = potj + poti[i]; 374 for (j=0; j<pnz; j++) { 375 row = ptJ[j]; /* row of AP == col of Pot */ 376 apnz = api[row+1] - api[row]; 377 Jptr = apj + api[row]; 378 /* add non-zero cols of AP into the sorted linked list lnk */ 379 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 380 } 381 nnz = lnk[0]; 382 383 /* If free space is not available, double the total space in the list */ 384 if (current_space->local_remaining<nnz) { 385 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 386 nspacedouble++; 387 } 388 389 /* Copy data into free space, and zero out denserows */ 390 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 391 392 current_space->array += nnz; 393 current_space->local_used += nnz; 394 current_space->local_remaining -= nnz; 395 396 coi[i+1] = coi[i] + nnz; 397 } 398 399 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 400 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 401 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 402 if (afill_tmp > afill) afill = afill_tmp; 403 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 404 405 /* (3) send j-array (coj) of Co to other processors */ 406 /*--------------------------------------------------*/ 407 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 408 len_s = merge->len_s; 409 merge->nsend = 0; 410 411 412 /* determine row ownership */ 413 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 414 merge->rowmap->n = pn; 415 merge->rowmap->bs = 1; 416 417 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 418 owners = merge->rowmap->range; 419 420 /* determine the number of messages to send, their lengths */ 421 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 422 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 423 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 424 425 proc = 0; 426 for (i=0; i<pon; i++) { 427 while (prmap[i] >= owners[proc+1]) proc++; 428 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 429 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 430 } 431 432 len = 0; /* max length of buf_si[], see (4) */ 433 owners_co[0] = 0; 434 for (proc=0; proc<size; proc++) { 435 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 436 if (len_s[proc]) { 437 merge->nsend++; 438 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 439 len += len_si[proc]; 440 } 441 } 442 443 /* determine the number and length of messages to receive for coi and coj */ 444 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 445 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 446 447 /* post the Irecv and Isend of coj */ 448 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 449 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 450 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 451 for (proc=0, k=0; proc<size; proc++) { 452 if (!len_s[proc]) continue; 453 i = owners_co[proc]; 454 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 455 k++; 456 } 457 458 /* receives and sends of coj are complete */ 459 for (i=0; i<merge->nrecv; i++) { 460 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 461 } 462 ierr = PetscFree(rwaits);CHKERRQ(ierr); 463 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 464 465 /* (4) send and recv coi */ 466 /*-----------------------*/ 467 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 468 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 469 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 470 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 471 for (proc=0,k=0; proc<size; proc++) { 472 if (!len_s[proc]) continue; 473 /* form outgoing message for i-structure: 474 buf_si[0]: nrows to be sent 475 [1:nrows]: row index (global) 476 [nrows+1:2*nrows+1]: i-structure index 477 */ 478 /*-------------------------------------------*/ 479 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 480 buf_si_i = buf_si + nrows+1; 481 buf_si[0] = nrows; 482 buf_si_i[0] = 0; 483 nrows = 0; 484 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 485 nzi = coi[i+1] - coi[i]; 486 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 487 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 488 nrows++; 489 } 490 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 491 k++; 492 buf_si += len_si[proc]; 493 } 494 i = merge->nrecv; 495 while (i--) { 496 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 497 } 498 ierr = PetscFree(rwaits);CHKERRQ(ierr); 499 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 500 501 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 502 ierr = PetscFree(len_ri);CHKERRQ(ierr); 503 ierr = PetscFree(swaits);CHKERRQ(ierr); 504 ierr = PetscFree(buf_s);CHKERRQ(ierr); 505 506 /* (5) compute the local portion of C (mpi mat) */ 507 /*----------------------------------------------*/ 508 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 509 510 /* allocate pti array and free space for accumulating nonzero column info */ 511 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 512 pti[0] = 0; 513 514 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 515 nnz = fill*(pi_loc[pm] + api[am]); 516 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 517 current_space = free_space; 518 519 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 520 for (k=0; k<merge->nrecv; k++) { 521 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 522 nrows = *buf_ri_k[k]; 523 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 524 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 525 } 526 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 527 rmax = 0; 528 for (i=0; i<pn; i++) { 529 /* add pdt[i,:]*AP into lnk */ 530 pnz = pdti[i+1] - pdti[i]; 531 ptJ = pdtj + pdti[i]; 532 for (j=0; j<pnz; j++) { 533 row = ptJ[j]; /* row of AP == col of Pt */ 534 apnz = api[row+1] - api[row]; 535 Jptr = apj + api[row]; 536 /* add non-zero cols of AP into the sorted linked list lnk */ 537 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 538 } 539 540 /* add received col data into lnk */ 541 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 542 if (i == *nextrow[k]) { /* i-th row */ 543 nzi = *(nextci[k]+1) - *nextci[k]; 544 Jptr = buf_rj[k] + *nextci[k]; 545 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 546 nextrow[k]++; nextci[k]++; 547 } 548 } 549 nnz = lnk[0]; 550 551 /* if free space is not available, make more free space */ 552 if (current_space->local_remaining<nnz) { 553 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 554 nspacedouble++; 555 } 556 /* copy data into free space, then initialize lnk */ 557 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 558 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 559 560 current_space->array += nnz; 561 current_space->local_used += nnz; 562 current_space->local_remaining -= nnz; 563 564 pti[i+1] = pti[i] + nnz; 565 if (nnz > rmax) rmax = nnz; 566 } 567 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 568 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 569 570 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 571 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 572 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 573 if (afill_tmp > afill) afill = afill_tmp; 574 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 575 576 /* (6) create symbolic parallel matrix Cmpi */ 577 /*------------------------------------------*/ 578 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 579 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 580 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 581 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 582 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 583 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 584 585 merge->bi = pti; /* Cseq->i */ 586 merge->bj = ptj; /* Cseq->j */ 587 merge->coi = coi; /* Co->i */ 588 merge->coj = coj; /* Co->j */ 589 merge->buf_ri = buf_ri; 590 merge->buf_rj = buf_rj; 591 merge->owners_co = owners_co; 592 merge->destroy = Cmpi->ops->destroy; 593 merge->duplicate = Cmpi->ops->duplicate; 594 595 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 596 Cmpi->assembled = PETSC_FALSE; 597 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 598 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 599 600 /* attach the supporting struct to Cmpi for reuse */ 601 c = (Mat_MPIAIJ*)Cmpi->data; 602 c->ptap = ptap; 603 ptap->api = api; 604 ptap->apj = apj; 605 ptap->rmax = ap_rmax; 606 *C = Cmpi; 607 608 /* flag 'scalable' determines which implementations to be used: 609 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 610 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 611 /* set default scalable */ 612 ptap->scalable = PETSC_FALSE; //PETSC_TRUE; 613 614 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 615 if (!ptap->scalable) { /* Do dense axpy */ 616 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 617 } else { 618 ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 619 } 620 621 #if defined(PETSC_USE_INFO) 622 if (pti[pn] != 0) { 623 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 624 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 625 } else { 626 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 627 } 628 #endif 629 PetscFunctionReturn(0); 630 } 631 632 #undef __FUNCT__ 633 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 634 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 635 { 636 PetscErrorCode ierr; 637 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 638 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 639 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 640 Mat_SeqAIJ *p_loc,*p_oth; 641 Mat_PtAPMPI *ptap; 642 Mat_Merge_SeqsToMPI *merge; 643 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 644 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 645 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 646 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 647 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 648 MPI_Comm comm; 649 PetscMPIInt size,rank,taga,*len_s; 650 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 651 PetscInt **buf_ri,**buf_rj; 652 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 653 MPI_Request *s_waits,*r_waits; 654 MPI_Status *status; 655 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 656 PetscInt *api,*apj,*coi,*coj; 657 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 658 PetscBool scalable; 659 #if defined(PTAP_PROFILE) 660 PetscLogDouble t0,t1,t2,eP,t3,t4,et2_AP=0.0,ePtAP=0.0,t2_0,t2_1,t2_2; 661 #endif 662 663 PetscFunctionBegin; 664 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 665 #if defined(PTAP_PROFILE) 666 ierr = PetscTime(&t0);CHKERRQ(ierr); 667 #endif 668 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 669 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 670 671 ptap = c->ptap; 672 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"); 673 merge = ptap->merge; 674 apa = ptap->apa; 675 scalable = ptap->scalable; 676 677 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 678 /*-----------------------------------------------------*/ 679 if (ptap->reuse == MAT_INITIAL_MATRIX) { 680 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 681 ptap->reuse = MAT_REUSE_MATRIX; 682 } else { /* update numerical values of P_oth and P_loc */ 683 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 684 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 685 } 686 #if defined(PTAP_PROFILE) 687 ierr = PetscTime(&t1);CHKERRQ(ierr); 688 eP = t1-t0; 689 #endif 690 /* 691 printf("[%d] Ad: %d, %d; Ao: %d, %d; P_loc: %d, %d; P_oth %d, %d;\n",rank, 692 a->A->rmap->N,a->A->cmap->N,a->B->rmap->N,a->B->cmap->N, 693 ptap->P_loc->rmap->N,ptap->P_loc->cmap->N, 694 ptap->P_oth->rmap->N,ptap->P_oth->cmap->N); 695 */ 696 697 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 698 /*--------------------------------------------------------------*/ 699 /* get data from symbolic products */ 700 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 701 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 702 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 703 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 704 705 coi = merge->coi; coj = merge->coj; 706 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 707 708 bi = merge->bi; bj = merge->bj; 709 owners = merge->rowmap->range; 710 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 711 712 api = ptap->api; apj = ptap->apj; 713 714 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but faster (could take 1/3 scalable time) */ 715 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 716 #if 0 717 /* ------ 10x slower -------------- */ 718 /*==================================*/ 719 Mat R = ptap->R; 720 Mat_SeqAIJ *r = (Mat_SeqAIJ*)R->data; 721 PetscInt *ri=r->i,*rj=r->j,rnz,arow,l,prow,pcol,pN=P->cmap->N; 722 PetscScalar *ra=r->a,tmp,cdense[pN]; 723 724 ierr = PetscMemzero(cdense,pN*sizeof(PetscScalar));CHKERRQ(ierr); 725 for (i=0; i<cm; i++) { /* each row of C or R */ 726 rnz = ri[i+1] - ri[i]; 727 728 for (j=0; j<rnz; j++) { /* each nz of R */ 729 arow = rj[ri[i] + j]; 730 731 /* diagonal portion of A */ 732 anz = ad->i[arow+1] - ad->i[arow]; 733 for (k=0; k<anz; k++) { /* each nz of Ad */ 734 tmp = ra[ri[i] + j]*ad->a[ad->i[arow] + k]; 735 prow = ad->j[ad->i[arow] + k]; 736 pnz = pi_loc[prow+1] - pi_loc[prow]; 737 738 for (l=0; l<pnz; l++) { /* each nz of P_loc */ 739 pcol = pj_loc[pi_loc[prow] + l]; 740 cdense[pcol] += tmp*pa_loc[pi_loc[prow] + l]; 741 } 742 } 743 744 /* off-diagonal portion of A */ 745 anz = ao->i[arow+1] - ao->i[arow]; 746 for (k=0; k<anz; k++) { /* each nz of Ao */ 747 tmp = ra[ri[i] + j]*ao->a[ao->i[arow] + k]; 748 prow = ao->j[ao->i[arow] + k]; 749 pnz = pi_oth[prow+1] - pi_oth[prow]; 750 751 for (l=0; l<pnz; l++) { /* each nz of P_oth */ 752 pcol = pj_oth[pi_oth[prow] + l]; 753 cdense[pcol] += tmp*pa_oth[pi_oth[prow] + l]; 754 } 755 } 756 757 } //for (j=0; j<rnz; j++) 758 759 /* copy cdense[] into ca; zero cdense[] */ 760 cnz = bi[i+1] - bi[i]; 761 cj = bj + bi[i]; 762 ca = ba + bi[i]; 763 for (j=0; j<cnz; j++) { 764 ca[j] += cdense[cj[j]]; 765 cdense[cj[j]] = 0.0; 766 } 767 #if 0 768 if (rank == 0) { 769 printf("[%d] row %d: ",rank,i); 770 for (j=0; j<pN; j++) printf(" %g,",cdense[j]); 771 printf("\n"); 772 } 773 for (j=0; j<pN; j++) cdense[j]=0.0; // zero cdnese[] 774 #endif 775 } //for (i=0; i<cm; i++) { 776 #endif 777 778 //========================================== 779 780 ierr = PetscTime(&t1);CHKERRQ(ierr); 781 for (i=0; i<am; i++) { 782 #if defined(PTAP_PROFILE) 783 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 784 #endif 785 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 786 /*------------------------------------------------------------*/ 787 apJ = apj + api[i]; 788 789 /* diagonal portion of A */ 790 anz = adi[i+1] - adi[i]; 791 adj = ad->j + adi[i]; 792 ada = ad->a + adi[i]; 793 for (j=0; j<anz; j++) { 794 row = adj[j]; 795 pnz = pi_loc[row+1] - pi_loc[row]; 796 pj = pj_loc + pi_loc[row]; 797 pa = pa_loc + pi_loc[row]; 798 799 /* perform dense axpy */ 800 valtmp = ada[j]; 801 for (k=0; k<pnz; k++) { 802 apa[pj[k]] += valtmp*pa[k]; 803 } 804 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 805 } 806 807 /* off-diagonal portion of A */ 808 anz = aoi[i+1] - aoi[i]; 809 aoj = ao->j + aoi[i]; 810 aoa = ao->a + aoi[i]; 811 for (j=0; j<anz; j++) { 812 row = aoj[j]; 813 pnz = pi_oth[row+1] - pi_oth[row]; 814 pj = pj_oth + pi_oth[row]; 815 pa = pa_oth + pi_oth[row]; 816 817 /* perform dense axpy */ 818 valtmp = aoa[j]; 819 for (k=0; k<pnz; k++) { 820 apa[pj[k]] += valtmp*pa[k]; 821 } 822 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 823 } 824 #if defined(PTAP_PROFILE) 825 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 826 et2_AP += t2_1 - t2_0; 827 #endif 828 829 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 830 /*--------------------------------------------------------------*/ 831 apnz = api[i+1] - api[i]; 832 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 833 pnz = po->i[i+1] - po->i[i]; 834 poJ = po->j + po->i[i]; 835 pA = po->a + po->i[i]; 836 for (j=0; j<pnz; j++) { 837 row = poJ[j]; 838 cnz = coi[row+1] - coi[row]; 839 cj = coj + coi[row]; 840 ca = coa + coi[row]; 841 /* perform dense axpy */ 842 valtmp = pA[j]; 843 for (k=0; k<cnz; k++) { 844 ca[k] += valtmp*apa[cj[k]]; 845 } 846 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 847 } 848 #if 1 849 /* put the value into Cd (diagonal part) */ 850 pnz = pd->i[i+1] - pd->i[i]; 851 pdJ = pd->j + pd->i[i]; 852 pA = pd->a + pd->i[i]; 853 for (j=0; j<pnz; j++) { 854 row = pdJ[j]; 855 cnz = bi[row+1] - bi[row]; 856 cj = bj + bi[row]; 857 ca = ba + bi[row]; 858 /* perform dense axpy */ 859 valtmp = pA[j]; 860 for (k=0; k<cnz; k++) { 861 ca[k] += valtmp*apa[cj[k]]; 862 } 863 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 864 } 865 #endif 866 /* zero the current row of A*P */ 867 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 868 #if defined(PTAP_PROFILE) 869 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 870 ePtAP += t2_2 - t2_1; 871 #endif 872 } 873 874 if (rank == 100) { 875 for (row=0; row<cm; row++) { 876 printf("[%d] row %d: ",rank,row); 877 cnz = bi[row+1] - bi[row]; 878 for (j=0; j<cnz; j++) printf(" %g,",ba[bi[row]+j]); 879 printf("\n"); 880 } 881 } 882 883 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 884 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 885 /*-----------------------------------------------------------------------------------------*/ 886 pA=pa_loc; 887 for (i=0; i<am; i++) { 888 #if defined(PTAP_PROFILE) 889 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 890 #endif 891 /* form i-th sparse row of A*P */ 892 apnz = api[i+1] - api[i]; 893 apJ = apj + api[i]; 894 /* diagonal portion of A */ 895 anz = adi[i+1] - adi[i]; 896 adj = ad->j + adi[i]; 897 ada = ad->a + adi[i]; 898 for (j=0; j<anz; j++) { 899 row = adj[j]; 900 pnz = pi_loc[row+1] - pi_loc[row]; 901 pj = pj_loc + pi_loc[row]; 902 pa = pa_loc + pi_loc[row]; 903 valtmp = ada[j]; 904 nextp = 0; 905 for (k=0; nextp<pnz; k++) { 906 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 907 apa[k] += valtmp*pa[nextp++]; 908 } 909 } 910 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 911 } 912 /* off-diagonal portion of A */ 913 anz = aoi[i+1] - aoi[i]; 914 aoj = ao->j + aoi[i]; 915 aoa = ao->a + aoi[i]; 916 for (j=0; j<anz; j++) { 917 row = aoj[j]; 918 pnz = pi_oth[row+1] - pi_oth[row]; 919 pj = pj_oth + pi_oth[row]; 920 pa = pa_oth + pi_oth[row]; 921 valtmp = aoa[j]; 922 nextp = 0; 923 for (k=0; nextp<pnz; k++) { 924 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 925 apa[k] += valtmp*pa[nextp++]; 926 } 927 } 928 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 929 } 930 #if defined(PTAP_PROFILE) 931 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 932 et2_AP += t2_1 - t2_0; 933 #endif 934 935 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 936 /*--------------------------------------------------------------*/ 937 pnz = pi_loc[i+1] - pi_loc[i]; 938 pJ = pj_loc + pi_loc[i]; 939 for (j=0; j<pnz; j++) { 940 nextap = 0; 941 row = pJ[j]; /* global index */ 942 if (row < pcstart || row >=pcend) { /* put the value into Co */ 943 row = *poJ; 944 cj = coj + coi[row]; 945 ca = coa + coi[row]; poJ++; 946 } else { /* put the value into Cd */ 947 row = *pdJ; 948 cj = bj + bi[row]; 949 ca = ba + bi[row]; pdJ++; 950 } 951 valtmp = pA[j]; 952 for (k=0; nextap<apnz; k++) { 953 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 954 } 955 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 956 } 957 pA += pnz; 958 /* zero the current row info for A*P */ 959 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 960 #if defined(PTAP_PROFILE) 961 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 962 ePtAP += t2_2 - t2_1; 963 #endif 964 } 965 } 966 #if defined(PTAP_PROFILE) 967 ierr = PetscTime(&t2);CHKERRQ(ierr); 968 #endif 969 970 /* 3) send and recv matrix values coa */ 971 /*------------------------------------*/ 972 buf_ri = merge->buf_ri; 973 buf_rj = merge->buf_rj; 974 len_s = merge->len_s; 975 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 976 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 977 978 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 979 for (proc=0,k=0; proc<size; proc++) { 980 if (!len_s[proc]) continue; 981 i = merge->owners_co[proc]; 982 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 983 k++; 984 } 985 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 986 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 987 988 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 989 ierr = PetscFree(r_waits);CHKERRQ(ierr); 990 ierr = PetscFree(coa);CHKERRQ(ierr); 991 #if defined(PTAP_PROFILE) 992 ierr = PetscTime(&t3);CHKERRQ(ierr); 993 #endif 994 995 /* 4) insert local Cseq and received values into Cmpi */ 996 /*------------------------------------------------------*/ 997 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 998 for (k=0; k<merge->nrecv; k++) { 999 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1000 nrows = *(buf_ri_k[k]); 1001 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1002 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1003 } 1004 1005 for (i=0; i<cm; i++) { 1006 row = owners[rank] + i; /* global row index of C_seq */ 1007 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1008 ba_i = ba + bi[i]; 1009 bnz = bi[i+1] - bi[i]; 1010 /* add received vals into ba */ 1011 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1012 /* i-th row */ 1013 if (i == *nextrow[k]) { 1014 cnz = *(nextci[k]+1) - *nextci[k]; 1015 cj = buf_rj[k] + *(nextci[k]); 1016 ca = abuf_r[k] + *(nextci[k]); 1017 nextcj = 0; 1018 for (j=0; nextcj<cnz; j++) { 1019 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1020 ba_i[j] += ca[nextcj++]; 1021 } 1022 } 1023 nextrow[k]++; nextci[k]++; 1024 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1025 } 1026 } 1027 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1028 } 1029 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1030 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1031 1032 ierr = PetscFree(ba);CHKERRQ(ierr); 1033 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1034 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1035 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1036 #if defined(PTAP_PROFILE) 1037 ierr = PetscTime(&t4);CHKERRQ(ierr); 1038 if (rank==1) { 1039 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); 1040 } 1041 #endif 1042 PetscFunctionReturn(0); 1043 } 1044