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