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