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 PetscErrorCode MatView_MPIAIJ_PtAP(Mat A,PetscViewer viewer) 16 { 17 PetscErrorCode ierr; 18 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 19 Mat_APMPI *ptap=a->ap; 20 PetscBool iascii; 21 PetscViewerFormat format; 22 23 PetscFunctionBegin; 24 if (!ptap) { 25 /* hack: MatDuplicate() sets oldmat->ops->view to newmat which is a base mat class with null ptpa! */ 26 A->ops->view = MatView_MPIAIJ; 27 ierr = (A->ops->view)(A,viewer);CHKERRQ(ierr); 28 PetscFunctionReturn(0); 29 } 30 31 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 32 if (iascii) { 33 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 34 if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 35 if (ptap->algType == 0) { 36 ierr = PetscViewerASCIIPrintf(viewer,"using scalable MatPtAP() implementation\n");CHKERRQ(ierr); 37 } else if (ptap->algType == 1) { 38 ierr = PetscViewerASCIIPrintf(viewer,"using nonscalable MatPtAP() implementation\n");CHKERRQ(ierr); 39 } 40 } 41 } 42 ierr = (ptap->view)(A,viewer);CHKERRQ(ierr); 43 PetscFunctionReturn(0); 44 } 45 46 PetscErrorCode MatFreeIntermediateDataStructures_MPIAIJ_AP(Mat A) 47 { 48 PetscErrorCode ierr; 49 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 50 Mat_APMPI *ptap=a->ap; 51 Mat_Merge_SeqsToMPI *merge; 52 53 PetscFunctionBegin; 54 if (!ptap) PetscFunctionReturn(0); 55 56 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 57 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 58 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 59 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 60 ierr = MatDestroy(&ptap->A_loc);CHKERRQ(ierr); /* used by MatTransposeMatMult() */ 61 ierr = MatDestroy(&ptap->Rd);CHKERRQ(ierr); 62 ierr = MatDestroy(&ptap->Ro);CHKERRQ(ierr); 63 if (ptap->AP_loc) { /* used by alg_rap */ 64 Mat_SeqAIJ *ap = (Mat_SeqAIJ*)(ptap->AP_loc)->data; 65 ierr = PetscFree(ap->i);CHKERRQ(ierr); 66 ierr = PetscFree2(ap->j,ap->a);CHKERRQ(ierr); 67 ierr = MatDestroy(&ptap->AP_loc);CHKERRQ(ierr); 68 } else { /* used by alg_ptap */ 69 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 70 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 71 } 72 ierr = MatDestroy(&ptap->C_loc);CHKERRQ(ierr); 73 ierr = MatDestroy(&ptap->C_oth);CHKERRQ(ierr); 74 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 75 76 ierr = MatDestroy(&ptap->Pt);CHKERRQ(ierr); 77 78 merge=ptap->merge; 79 if (merge) { /* used by alg_ptap */ 80 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 81 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 82 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 83 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 84 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 85 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 86 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 87 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 88 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 89 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 90 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 91 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 92 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 93 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 94 } 95 ierr = ISLocalToGlobalMappingDestroy(&ptap->ltog);CHKERRQ(ierr); 96 97 A->ops->destroy = ptap->destroy; 98 ierr = PetscFree(a->ap);CHKERRQ(ierr); 99 PetscFunctionReturn(0); 100 } 101 102 PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat A) 103 { 104 PetscErrorCode ierr; 105 106 PetscFunctionBegin; 107 ierr = (*A->ops->freeintermediatedatastructures)(A);CHKERRQ(ierr); 108 ierr = (*A->ops->destroy)(A);CHKERRQ(ierr); 109 PetscFunctionReturn(0); 110 } 111 112 PETSC_INTERN PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 113 { 114 PetscErrorCode ierr; 115 PetscBool flg; 116 MPI_Comm comm; 117 #if !defined(PETSC_HAVE_HYPRE) 118 const char *algTypes[2] = {"scalable","nonscalable"}; 119 PetscInt nalg=2; 120 #else 121 const char *algTypes[3] = {"scalable","nonscalable","hypre"}; 122 PetscInt nalg=3; 123 #endif 124 PetscInt pN=P->cmap->N,alg=1; /* set default algorithm */ 125 126 PetscFunctionBegin; 127 /* check if matrix local sizes are compatible */ 128 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 129 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) 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); 130 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) 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); 131 132 if (scall == MAT_INITIAL_MATRIX) { 133 /* pick an algorithm */ 134 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MatPtAP","Mat");CHKERRQ(ierr); 135 ierr = PetscOptionsEList("-matptap_via","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[alg],&alg,&flg);CHKERRQ(ierr); 136 ierr = PetscOptionsEnd();CHKERRQ(ierr); 137 138 if (!flg && pN > 100000) { /* may switch to scalable algorithm as default */ 139 MatInfo Ainfo,Pinfo; 140 PetscInt nz_local; 141 PetscBool alg_scalable_loc=PETSC_FALSE,alg_scalable; 142 143 ierr = MatGetInfo(A,MAT_LOCAL,&Ainfo);CHKERRQ(ierr); 144 ierr = MatGetInfo(P,MAT_LOCAL,&Pinfo);CHKERRQ(ierr); 145 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); 146 147 if (pN > fill*nz_local) alg_scalable_loc = PETSC_TRUE; 148 ierr = MPIU_Allreduce(&alg_scalable_loc,&alg_scalable,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 149 150 if (alg_scalable) { 151 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 152 } 153 } 154 155 switch (alg) { 156 case 1: 157 /* do R=P^T locally, then C=R*A*P -- nonscalable */ 158 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 159 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 160 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 161 break; 162 #if defined(PETSC_HAVE_HYPRE) 163 case 2: 164 /* Use boomerAMGBuildCoarseOperator */ 165 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 166 ierr = MatPtAPSymbolic_AIJ_AIJ_wHYPRE(A,P,fill,C);CHKERRQ(ierr); 167 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 168 break; 169 #endif 170 default: 171 /* do R=P^T locally, then C=R*A*P */ 172 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 173 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable(A,P,fill,C);CHKERRQ(ierr); 174 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 175 break; 176 } 177 178 if (alg == 0 || alg == 1) { 179 Mat_MPIAIJ *c = (Mat_MPIAIJ*)(*C)->data; 180 Mat_APMPI *ap = c->ap; 181 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)(*C)),((PetscObject)(*C))->prefix,"MatFreeIntermediateDataStructures","Mat");CHKERRQ(ierr); 182 ap->freestruct = PETSC_FALSE; 183 ierr = PetscOptionsBool("-mat_freeintermediatedatastructures","Free intermediate data structures", "MatFreeIntermediateDataStructures",ap->freestruct,&ap->freestruct, NULL);CHKERRQ(ierr); 184 ierr = PetscOptionsEnd();CHKERRQ(ierr); 185 } 186 } 187 188 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 189 ierr = (*(*C)->ops->ptapnumeric)(A,P,*C);CHKERRQ(ierr); 190 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 191 PetscFunctionReturn(0); 192 } 193 194 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_scalable(Mat A,Mat P,Mat C) 195 { 196 PetscErrorCode ierr; 197 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 198 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 199 Mat_SeqAIJ *ap,*p_loc,*p_oth,*c_seq; 200 Mat_APMPI *ptap = c->ap; 201 Mat AP_loc,C_loc,C_oth; 202 PetscInt i,rstart,rend,cm,ncols,row,*api,*apj,am = A->rmap->n,apnz,nout; 203 PetscScalar *apa; 204 const PetscInt *cols; 205 const PetscScalar *vals; 206 207 PetscFunctionBegin; 208 if (!ptap) { 209 MPI_Comm comm; 210 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 211 SETERRQ(comm,PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be reused. Do not call MatFreeIntermediateDataStructures() or use '-mat_freeintermediatedatastructures'"); 212 } 213 214 ierr = MatZeroEntries(C);CHKERRQ(ierr); 215 216 /* 1) get R = Pd^T,Ro = Po^T */ 217 if (ptap->reuse == MAT_REUSE_MATRIX) { 218 ierr = MatTranspose(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 219 ierr = MatTranspose(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 220 } 221 222 /* 2) get AP_loc */ 223 AP_loc = ptap->AP_loc; 224 ap = (Mat_SeqAIJ*)AP_loc->data; 225 226 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 227 /*-----------------------------------------------------*/ 228 if (ptap->reuse == MAT_REUSE_MATRIX) { 229 /* P_oth and P_loc are obtained in MatPtASymbolic() when reuse == MAT_INITIAL_MATRIX */ 230 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 231 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 232 } 233 234 /* 2-2) compute numeric A_loc*P - dominating part */ 235 /* ---------------------------------------------- */ 236 /* get data from symbolic products */ 237 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 238 if (ptap->P_oth) p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 239 else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"ptap->P_oth is NULL. Cannot proceed!"); 240 241 api = ap->i; 242 apj = ap->j; 243 ierr = ISLocalToGlobalMappingApply(ptap->ltog,api[AP_loc->rmap->n],apj,apj);CHKERRQ(ierr); 244 for (i=0; i<am; i++) { 245 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 246 apnz = api[i+1] - api[i]; 247 apa = ap->a + api[i]; 248 ierr = PetscMemzero(apa,sizeof(PetscScalar)*apnz);CHKERRQ(ierr); 249 AProw_scalable(i,ad,ao,p_loc,p_oth,api,apj,apa); 250 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 251 } 252 ierr = ISGlobalToLocalMappingApply(ptap->ltog,IS_GTOLM_DROP,api[AP_loc->rmap->n],apj,&nout,apj);CHKERRQ(ierr); 253 if (api[AP_loc->rmap->n] != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Incorrect mapping %D != %D\n",api[AP_loc->rmap->n],nout); 254 255 /* 3) C_loc = Rd*AP_loc, C_oth = Ro*AP_loc */ 256 /* Always use scalable version since we are in the MPI scalable version */ 257 ierr = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(ptap->Rd,AP_loc,ptap->C_loc);CHKERRQ(ierr); 258 ierr = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(ptap->Ro,AP_loc,ptap->C_oth);CHKERRQ(ierr); 259 260 C_loc = ptap->C_loc; 261 C_oth = ptap->C_oth; 262 263 /* add C_loc and Co to to C */ 264 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 265 266 /* C_loc -> C */ 267 cm = C_loc->rmap->N; 268 c_seq = (Mat_SeqAIJ*)C_loc->data; 269 cols = c_seq->j; 270 vals = c_seq->a; 271 ierr = ISLocalToGlobalMappingApply(ptap->ltog,c_seq->i[C_loc->rmap->n],c_seq->j,c_seq->j);CHKERRQ(ierr); 272 273 /* The (fast) MatSetValues_MPIAIJ_CopyFromCSRFormat function can only be used when C->was_assembled is PETSC_FALSE and */ 274 /* when there are no off-processor parts. */ 275 /* If was_assembled is true, then the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart; in MatSetValues_MPIAIJ_CopyFromCSRFormat */ 276 /* is no longer true. Then the more complex function MatSetValues_MPIAIJ() has to be used, where the column index is looked up from */ 277 /* a table, and other, more complex stuff has to be done. */ 278 if (C->assembled) { 279 C->was_assembled = PETSC_TRUE; 280 C->assembled = PETSC_FALSE; 281 } 282 if (C->was_assembled) { 283 for (i=0; i<cm; i++) { 284 ncols = c_seq->i[i+1] - c_seq->i[i]; 285 row = rstart + i; 286 ierr = MatSetValues_MPIAIJ(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 287 cols += ncols; vals += ncols; 288 } 289 } else { 290 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat(C,c_seq->j,c_seq->i,c_seq->a);CHKERRQ(ierr); 291 } 292 ierr = ISGlobalToLocalMappingApply(ptap->ltog,IS_GTOLM_DROP,c_seq->i[C_loc->rmap->n],c_seq->j,&nout,c_seq->j);CHKERRQ(ierr); 293 if (c_seq->i[C_loc->rmap->n] != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Incorrect mapping %D != %D\n",c_seq->i[C_loc->rmap->n],nout); 294 295 /* Co -> C, off-processor part */ 296 cm = C_oth->rmap->N; 297 c_seq = (Mat_SeqAIJ*)C_oth->data; 298 cols = c_seq->j; 299 vals = c_seq->a; 300 ierr = ISLocalToGlobalMappingApply(ptap->ltog,c_seq->i[C_oth->rmap->n],c_seq->j,c_seq->j);CHKERRQ(ierr); 301 for (i=0; i<cm; i++) { 302 ncols = c_seq->i[i+1] - c_seq->i[i]; 303 row = p->garray[i]; 304 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 305 cols += ncols; vals += ncols; 306 } 307 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 308 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 309 310 ptap->reuse = MAT_REUSE_MATRIX; 311 312 ierr = ISGlobalToLocalMappingApply(ptap->ltog,IS_GTOLM_DROP,c_seq->i[C_oth->rmap->n],c_seq->j,&nout,c_seq->j);CHKERRQ(ierr); 313 if (c_seq->i[C_oth->rmap->n] != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Incorrect mapping %D != %D\n",c_seq->i[C_loc->rmap->n],nout); 314 315 /* supporting struct ptap consumes almost same amount of memory as C=PtAP, release it if C will not be updated by A and P */ 316 if (ptap->freestruct) { 317 ierr = MatFreeIntermediateDataStructures(C);CHKERRQ(ierr); 318 } 319 PetscFunctionReturn(0); 320 } 321 322 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable(Mat A,Mat P,PetscReal fill,Mat *C) 323 { 324 PetscErrorCode ierr; 325 Mat_APMPI *ptap; 326 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 327 MPI_Comm comm; 328 PetscMPIInt size,rank; 329 Mat Cmpi,P_loc,P_oth; 330 PetscFreeSpaceList free_space=NULL,current_space=NULL; 331 PetscInt am=A->rmap->n,pm=P->rmap->n,pN=P->cmap->N,pn=P->cmap->n; 332 PetscInt *lnk,i,k,pnz,row,nsend; 333 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 334 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 335 PetscInt len,proc,*dnz,*onz,*owners,nzi,nspacedouble; 336 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 337 MPI_Request *swaits,*rwaits; 338 MPI_Status *sstatus,rstatus; 339 PetscLayout rowmap; 340 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 341 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 342 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,con,j,Crmax,*aj,*ai,*pi,nout; 343 Mat_SeqAIJ *p_loc,*p_oth=NULL,*ad=(Mat_SeqAIJ*)(a->A)->data,*ao=NULL,*c_loc,*c_oth; 344 PetscScalar *apv; 345 PetscTable ta; 346 MatType mtype; 347 const char *prefix; 348 #if defined(PETSC_USE_INFO) 349 PetscReal apfill; 350 #endif 351 352 PetscFunctionBegin; 353 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 354 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 355 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 356 357 if (size > 1) ao = (Mat_SeqAIJ*)(a->B)->data; 358 359 /* create symbolic parallel matrix Cmpi */ 360 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 361 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 362 ierr = MatSetType(Cmpi,mtype);CHKERRQ(ierr); 363 364 /* create struct Mat_APMPI and attached it to C later */ 365 ierr = PetscNew(&ptap);CHKERRQ(ierr); 366 ptap->reuse = MAT_INITIAL_MATRIX; 367 ptap->algType = 0; 368 369 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 370 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&P_oth);CHKERRQ(ierr); 371 /* get P_loc by taking all local rows of P */ 372 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&P_loc);CHKERRQ(ierr); 373 374 ptap->P_loc = P_loc; 375 ptap->P_oth = P_oth; 376 377 /* (0) compute Rd = Pd^T, Ro = Po^T */ 378 /* --------------------------------- */ 379 ierr = MatTranspose(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 380 ierr = MatTranspose(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 381 382 /* (1) compute symbolic AP = A_loc*P = Ad*P_loc + Ao*P_oth (api,apj) */ 383 /* ----------------------------------------------------------------- */ 384 p_loc = (Mat_SeqAIJ*)P_loc->data; 385 if (P_oth) p_oth = (Mat_SeqAIJ*)P_oth->data; 386 387 /* create and initialize a linked list */ 388 ierr = PetscTableCreate(pn,pN,&ta);CHKERRQ(ierr); /* for compute AP_loc and Cmpi */ 389 MatRowMergeMax_SeqAIJ(p_loc,P_loc->rmap->N,ta); 390 MatRowMergeMax_SeqAIJ(p_oth,P_oth->rmap->N,ta); 391 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); /* Crmax = nnz(sum of Prows) */ 392 393 ierr = PetscLLCondensedCreate_Scalable(Crmax,&lnk);CHKERRQ(ierr); 394 395 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) */ 396 if (ao) { 397 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ad->i[am],PetscIntSumTruncate(ao->i[am],p_loc->i[pm]))),&free_space);CHKERRQ(ierr); 398 } else { 399 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ad->i[am],p_loc->i[pm])),&free_space);CHKERRQ(ierr); 400 } 401 current_space = free_space; 402 nspacedouble = 0; 403 404 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 405 api[0] = 0; 406 for (i=0; i<am; i++) { 407 /* diagonal portion: Ad[i,:]*P */ 408 ai = ad->i; pi = p_loc->i; 409 nzi = ai[i+1] - ai[i]; 410 aj = ad->j + ai[i]; 411 for (j=0; j<nzi; j++) { 412 row = aj[j]; 413 pnz = pi[row+1] - pi[row]; 414 Jptr = p_loc->j + pi[row]; 415 /* add non-zero cols of P into the sorted linked list lnk */ 416 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 417 } 418 /* off-diagonal portion: Ao[i,:]*P */ 419 if (ao) { 420 ai = ao->i; pi = p_oth->i; 421 nzi = ai[i+1] - ai[i]; 422 aj = ao->j + ai[i]; 423 for (j=0; j<nzi; j++) { 424 row = aj[j]; 425 pnz = pi[row+1] - pi[row]; 426 Jptr = p_oth->j + pi[row]; 427 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 428 } 429 } 430 apnz = lnk[0]; 431 api[i+1] = api[i] + apnz; 432 433 /* if free space is not available, double the total space in the list */ 434 if (current_space->local_remaining<apnz) { 435 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 436 nspacedouble++; 437 } 438 439 /* Copy data into free space, then initialize lnk */ 440 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 441 442 current_space->array += apnz; 443 current_space->local_used += apnz; 444 current_space->local_remaining -= apnz; 445 } 446 /* Allocate space for apj and apv, initialize apj, and */ 447 /* destroy list of free space and other temporary array(s) */ 448 ierr = PetscCalloc2(api[am],&apj,api[am],&apv);CHKERRQ(ierr); 449 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 450 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 451 452 /* Create AP_loc for reuse */ 453 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,pN,api,apj,apv,&ptap->AP_loc);CHKERRQ(ierr); 454 ierr = MatSeqAIJCompactOutExtraColumns_SeqAIJ(ptap->AP_loc, &ptap->ltog);CHKERRQ(ierr); 455 456 #if defined(PETSC_USE_INFO) 457 if (ao) { 458 apfill = (PetscReal)api[am]/(ad->i[am]+ao->i[am]+p_loc->i[pm]+1); 459 } else { 460 apfill = (PetscReal)api[am]/(ad->i[am]+p_loc->i[pm]+1); 461 } 462 ptap->AP_loc->info.mallocs = nspacedouble; 463 ptap->AP_loc->info.fill_ratio_given = fill; 464 ptap->AP_loc->info.fill_ratio_needed = apfill; 465 466 if (api[am]) { 467 ierr = PetscInfo3(ptap->AP_loc,"Scalable algorithm, AP_loc reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)apfill);CHKERRQ(ierr); 468 ierr = PetscInfo1(ptap->AP_loc,"Use MatPtAP(A,B,MatReuse,%g,&C) for best AP_loc performance.;\n",(double)apfill);CHKERRQ(ierr); 469 } else { 470 ierr = PetscInfo(ptap->AP_loc,"Scalable algorithm, AP_loc is empty \n");CHKERRQ(ierr); 471 } 472 #endif 473 474 /* (2-1) compute symbolic Co = Ro*AP_loc */ 475 /* ------------------------------------ */ 476 ierr = MatGetOptionsPrefix(A,&prefix);CHKERRQ(ierr); 477 ierr = MatSetOptionsPrefix(ptap->Ro,prefix);CHKERRQ(ierr); 478 ierr = MatAppendOptionsPrefix(ptap->Ro,"inner_offdiag_");CHKERRQ(ierr); 479 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 480 481 /* (3) send coj of C_oth to other processors */ 482 /* ------------------------------------------ */ 483 /* determine row ownership */ 484 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 485 rowmap->n = pn; 486 rowmap->bs = 1; 487 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 488 owners = rowmap->range; 489 490 /* determine the number of messages to send, their lengths */ 491 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 492 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 493 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 494 495 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 496 coi = c_oth->i; coj = c_oth->j; 497 con = ptap->C_oth->rmap->n; 498 proc = 0; 499 ierr = ISLocalToGlobalMappingApply(ptap->ltog,coi[con],coj,coj);CHKERRQ(ierr); 500 for (i=0; i<con; i++) { 501 while (prmap[i] >= owners[proc+1]) proc++; 502 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 503 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 504 } 505 506 len = 0; /* max length of buf_si[], see (4) */ 507 owners_co[0] = 0; 508 nsend = 0; 509 for (proc=0; proc<size; proc++) { 510 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 511 if (len_s[proc]) { 512 nsend++; 513 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 514 len += len_si[proc]; 515 } 516 } 517 518 /* determine the number and length of messages to receive for coi and coj */ 519 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 520 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 521 522 /* post the Irecv and Isend of coj */ 523 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 524 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 525 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 526 for (proc=0, k=0; proc<size; proc++) { 527 if (!len_s[proc]) continue; 528 i = owners_co[proc]; 529 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 530 k++; 531 } 532 533 /* (2-2) compute symbolic C_loc = Rd*AP_loc */ 534 /* ---------------------------------------- */ 535 ierr = MatSetOptionsPrefix(ptap->Rd,prefix);CHKERRQ(ierr); 536 ierr = MatAppendOptionsPrefix(ptap->Rd,"inner_diag_");CHKERRQ(ierr); 537 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 538 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 539 ierr = ISLocalToGlobalMappingApply(ptap->ltog,c_loc->i[ptap->C_loc->rmap->n],c_loc->j,c_loc->j);CHKERRQ(ierr); 540 541 /* receives coj are complete */ 542 for (i=0; i<nrecv; i++) { 543 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 544 } 545 ierr = PetscFree(rwaits);CHKERRQ(ierr); 546 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 547 548 /* add received column indices into ta to update Crmax */ 549 for (k=0; k<nrecv; k++) {/* k-th received message */ 550 Jptr = buf_rj[k]; 551 for (j=0; j<len_r[k]; j++) { 552 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 553 } 554 } 555 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 556 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 557 558 /* (4) send and recv coi */ 559 /*-----------------------*/ 560 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 561 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 562 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 563 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 564 for (proc=0,k=0; proc<size; proc++) { 565 if (!len_s[proc]) continue; 566 /* form outgoing message for i-structure: 567 buf_si[0]: nrows to be sent 568 [1:nrows]: row index (global) 569 [nrows+1:2*nrows+1]: i-structure index 570 */ 571 /*-------------------------------------------*/ 572 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 573 buf_si_i = buf_si + nrows+1; 574 buf_si[0] = nrows; 575 buf_si_i[0] = 0; 576 nrows = 0; 577 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 578 nzi = coi[i+1] - coi[i]; 579 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 580 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 581 nrows++; 582 } 583 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 584 k++; 585 buf_si += len_si[proc]; 586 } 587 for (i=0; i<nrecv; i++) { 588 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 589 } 590 ierr = PetscFree(rwaits);CHKERRQ(ierr); 591 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 592 593 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 594 ierr = PetscFree(len_ri);CHKERRQ(ierr); 595 ierr = PetscFree(swaits);CHKERRQ(ierr); 596 ierr = PetscFree(buf_s);CHKERRQ(ierr); 597 598 /* (5) compute the local portion of Cmpi */ 599 /* ------------------------------------------ */ 600 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of Cmpi */ 601 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 602 current_space = free_space; 603 604 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 605 for (k=0; k<nrecv; k++) { 606 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 607 nrows = *buf_ri_k[k]; 608 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 609 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 610 } 611 612 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 613 ierr = PetscLLCondensedCreate_Scalable(Crmax,&lnk);CHKERRQ(ierr); 614 for (i=0; i<pn; i++) { 615 /* add C_loc into Cmpi */ 616 nzi = c_loc->i[i+1] - c_loc->i[i]; 617 Jptr = c_loc->j + c_loc->i[i]; 618 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 619 620 /* add received col data into lnk */ 621 for (k=0; k<nrecv; k++) { /* k-th received message */ 622 if (i == *nextrow[k]) { /* i-th row */ 623 nzi = *(nextci[k]+1) - *nextci[k]; 624 Jptr = buf_rj[k] + *nextci[k]; 625 ierr = PetscLLCondensedAddSorted_Scalable(nzi,Jptr,lnk);CHKERRQ(ierr); 626 nextrow[k]++; nextci[k]++; 627 } 628 } 629 nzi = lnk[0]; 630 631 /* copy data into free space, then initialize lnk */ 632 ierr = PetscLLCondensedClean_Scalable(nzi,current_space->array,lnk);CHKERRQ(ierr); 633 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 634 } 635 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 636 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 637 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 638 639 /* local sizes and preallocation */ 640 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 641 if (P->cmap->bs > 0) { 642 ierr = PetscLayoutSetBlockSize(Cmpi->rmap,P->cmap->bs);CHKERRQ(ierr); 643 ierr = PetscLayoutSetBlockSize(Cmpi->cmap,P->cmap->bs);CHKERRQ(ierr); 644 } 645 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 646 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 647 648 /* members in merge */ 649 ierr = PetscFree(id_r);CHKERRQ(ierr); 650 ierr = PetscFree(len_r);CHKERRQ(ierr); 651 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 652 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 653 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 654 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 655 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 656 657 /* attach the supporting struct to Cmpi for reuse */ 658 c = (Mat_MPIAIJ*)Cmpi->data; 659 c->ap = ptap; 660 ptap->duplicate = Cmpi->ops->duplicate; 661 ptap->destroy = Cmpi->ops->destroy; 662 ptap->view = Cmpi->ops->view; 663 664 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 665 Cmpi->assembled = PETSC_FALSE; 666 Cmpi->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ_scalable; 667 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 668 Cmpi->ops->view = MatView_MPIAIJ_PtAP; 669 Cmpi->ops->freeintermediatedatastructures = MatFreeIntermediateDataStructures_MPIAIJ_AP; 670 *C = Cmpi; 671 672 nout = 0; 673 ierr = ISGlobalToLocalMappingApply(ptap->ltog,IS_GTOLM_DROP,c_oth->i[ptap->C_oth->rmap->n],c_oth->j,&nout,c_oth->j);CHKERRQ(ierr); 674 if (c_oth->i[ptap->C_oth->rmap->n] != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Incorrect mapping %D != %D\n",c_oth->i[ptap->C_oth->rmap->n],nout); 675 ierr = ISGlobalToLocalMappingApply(ptap->ltog,IS_GTOLM_DROP,c_loc->i[ptap->C_loc->rmap->n],c_loc->j,&nout,c_loc->j);CHKERRQ(ierr); 676 if (c_loc->i[ptap->C_loc->rmap->n] != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Incorrect mapping %D != %D\n",c_loc->i[ptap->C_loc->rmap->n],nout); 677 678 PetscFunctionReturn(0); 679 } 680 681 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 682 { 683 PetscErrorCode ierr; 684 Mat_APMPI *ptap; 685 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 686 MPI_Comm comm; 687 PetscMPIInt size,rank; 688 Mat Cmpi; 689 PetscFreeSpaceList free_space=NULL,current_space=NULL; 690 PetscInt am=A->rmap->n,pm=P->rmap->n,pN=P->cmap->N,pn=P->cmap->n; 691 PetscInt *lnk,i,k,pnz,row,nsend; 692 PetscBT lnkbt; 693 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 694 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 695 PetscInt len,proc,*dnz,*onz,*owners,nzi,nspacedouble; 696 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 697 MPI_Request *swaits,*rwaits; 698 MPI_Status *sstatus,rstatus; 699 PetscLayout rowmap; 700 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 701 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 702 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,con,j,ap_rmax=0,Crmax,*aj,*ai,*pi; 703 Mat_SeqAIJ *p_loc,*p_oth=NULL,*ad=(Mat_SeqAIJ*)(a->A)->data,*ao=NULL,*c_loc,*c_oth; 704 PetscScalar *apv; 705 PetscTable ta; 706 MatType mtype; 707 const char *prefix; 708 #if defined(PETSC_USE_INFO) 709 PetscReal apfill; 710 #endif 711 712 PetscFunctionBegin; 713 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 714 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 715 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 716 717 if (size > 1) ao = (Mat_SeqAIJ*)(a->B)->data; 718 719 /* create symbolic parallel matrix Cmpi */ 720 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 721 ierr = MatGetType(A,&mtype);CHKERRQ(ierr); 722 ierr = MatSetType(Cmpi,mtype);CHKERRQ(ierr); 723 724 /* Do dense axpy in MatPtAPNumeric_MPIAIJ_MPIAIJ() */ 725 Cmpi->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ; 726 727 /* create struct Mat_APMPI and attached it to C later */ 728 ierr = PetscNew(&ptap);CHKERRQ(ierr); 729 ptap->reuse = MAT_INITIAL_MATRIX; 730 ptap->algType = 1; 731 732 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 733 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 734 /* get P_loc by taking all local rows of P */ 735 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 736 737 /* (0) compute Rd = Pd^T, Ro = Po^T */ 738 /* --------------------------------- */ 739 ierr = MatTranspose(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 740 ierr = MatTranspose(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 741 742 /* (1) compute symbolic AP = A_loc*P = Ad*P_loc + Ao*P_oth (api,apj) */ 743 /* ----------------------------------------------------------------- */ 744 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 745 if (ptap->P_oth) p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 746 747 /* create and initialize a linked list */ 748 ierr = PetscTableCreate(pn,pN,&ta);CHKERRQ(ierr); /* for compute AP_loc and Cmpi */ 749 MatRowMergeMax_SeqAIJ(p_loc,ptap->P_loc->rmap->N,ta); 750 MatRowMergeMax_SeqAIJ(p_oth,ptap->P_oth->rmap->N,ta); 751 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); /* Crmax = nnz(sum of Prows) */ 752 /* printf("[%d] est %d, Crmax %d; pN %d\n",rank,5*(p_loc->rmax+p_oth->rmax + (PetscInt)(1.e-2*pN)),Crmax,pN); */ 753 754 ierr = PetscLLCondensedCreate(Crmax,pN,&lnk,&lnkbt);CHKERRQ(ierr); 755 756 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) */ 757 if (ao) { 758 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ad->i[am],PetscIntSumTruncate(ao->i[am],p_loc->i[pm]))),&free_space);CHKERRQ(ierr); 759 } else { 760 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ad->i[am],p_loc->i[pm])),&free_space);CHKERRQ(ierr); 761 } 762 current_space = free_space; 763 nspacedouble = 0; 764 765 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 766 api[0] = 0; 767 for (i=0; i<am; i++) { 768 /* diagonal portion: Ad[i,:]*P */ 769 ai = ad->i; pi = p_loc->i; 770 nzi = ai[i+1] - ai[i]; 771 aj = ad->j + ai[i]; 772 for (j=0; j<nzi; j++) { 773 row = aj[j]; 774 pnz = pi[row+1] - pi[row]; 775 Jptr = p_loc->j + pi[row]; 776 /* add non-zero cols of P into the sorted linked list lnk */ 777 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 778 } 779 /* off-diagonal portion: Ao[i,:]*P */ 780 if (ao) { 781 ai = ao->i; pi = p_oth->i; 782 nzi = ai[i+1] - ai[i]; 783 aj = ao->j + ai[i]; 784 for (j=0; j<nzi; j++) { 785 row = aj[j]; 786 pnz = pi[row+1] - pi[row]; 787 Jptr = p_oth->j + pi[row]; 788 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 789 } 790 } 791 apnz = lnk[0]; 792 api[i+1] = api[i] + apnz; 793 if (ap_rmax < apnz) ap_rmax = apnz; 794 795 /* if free space is not available, double the total space in the list */ 796 if (current_space->local_remaining<apnz) { 797 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 798 nspacedouble++; 799 } 800 801 /* Copy data into free space, then initialize lnk */ 802 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 803 804 current_space->array += apnz; 805 current_space->local_used += apnz; 806 current_space->local_remaining -= apnz; 807 } 808 /* Allocate space for apj and apv, initialize apj, and */ 809 /* destroy list of free space and other temporary array(s) */ 810 ierr = PetscMalloc2(api[am],&apj,api[am],&apv);CHKERRQ(ierr); 811 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 812 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 813 814 /* Create AP_loc for reuse */ 815 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,pN,api,apj,apv,&ptap->AP_loc);CHKERRQ(ierr); 816 817 #if defined(PETSC_USE_INFO) 818 if (ao) { 819 apfill = (PetscReal)api[am]/(ad->i[am]+ao->i[am]+p_loc->i[pm]+1); 820 } else { 821 apfill = (PetscReal)api[am]/(ad->i[am]+p_loc->i[pm]+1); 822 } 823 ptap->AP_loc->info.mallocs = nspacedouble; 824 ptap->AP_loc->info.fill_ratio_given = fill; 825 ptap->AP_loc->info.fill_ratio_needed = apfill; 826 827 if (api[am]) { 828 ierr = PetscInfo3(ptap->AP_loc,"Nonscalable algorithm, AP_loc reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)apfill);CHKERRQ(ierr); 829 ierr = PetscInfo1(ptap->AP_loc,"Use MatPtAP(A,B,MatReuse,%g,&C) for best AP_loc performance.;\n",(double)apfill);CHKERRQ(ierr); 830 } else { 831 ierr = PetscInfo(ptap->AP_loc,"Nonscalable algorithm, AP_loc is empty \n");CHKERRQ(ierr); 832 } 833 #endif 834 835 /* (2-1) compute symbolic Co = Ro*AP_loc */ 836 /* ------------------------------------ */ 837 ierr = MatGetOptionsPrefix(A,&prefix);CHKERRQ(ierr); 838 ierr = MatSetOptionsPrefix(ptap->Ro,prefix);CHKERRQ(ierr); 839 ierr = MatAppendOptionsPrefix(ptap->Ro,"inner_offdiag_");CHKERRQ(ierr); 840 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 841 842 /* (3) send coj of C_oth to other processors */ 843 /* ------------------------------------------ */ 844 /* determine row ownership */ 845 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 846 rowmap->n = pn; 847 rowmap->bs = 1; 848 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 849 owners = rowmap->range; 850 851 /* determine the number of messages to send, their lengths */ 852 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 853 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 854 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 855 856 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 857 coi = c_oth->i; coj = c_oth->j; 858 con = ptap->C_oth->rmap->n; 859 proc = 0; 860 for (i=0; i<con; i++) { 861 while (prmap[i] >= owners[proc+1]) proc++; 862 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 863 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 864 } 865 866 len = 0; /* max length of buf_si[], see (4) */ 867 owners_co[0] = 0; 868 nsend = 0; 869 for (proc=0; proc<size; proc++) { 870 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 871 if (len_s[proc]) { 872 nsend++; 873 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 874 len += len_si[proc]; 875 } 876 } 877 878 /* determine the number and length of messages to receive for coi and coj */ 879 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 880 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 881 882 /* post the Irecv and Isend of coj */ 883 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 884 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 885 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 886 for (proc=0, k=0; proc<size; proc++) { 887 if (!len_s[proc]) continue; 888 i = owners_co[proc]; 889 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 890 k++; 891 } 892 893 /* (2-2) compute symbolic C_loc = Rd*AP_loc */ 894 /* ---------------------------------------- */ 895 ierr = MatSetOptionsPrefix(ptap->Rd,prefix);CHKERRQ(ierr); 896 ierr = MatAppendOptionsPrefix(ptap->Rd,"inner_diag_");CHKERRQ(ierr); 897 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 898 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 899 900 /* receives coj are complete */ 901 for (i=0; i<nrecv; i++) { 902 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 903 } 904 ierr = PetscFree(rwaits);CHKERRQ(ierr); 905 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 906 907 /* add received column indices into ta to update Crmax */ 908 for (k=0; k<nrecv; k++) {/* k-th received message */ 909 Jptr = buf_rj[k]; 910 for (j=0; j<len_r[k]; j++) { 911 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 912 } 913 } 914 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 915 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 916 917 /* (4) send and recv coi */ 918 /*-----------------------*/ 919 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 920 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 921 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 922 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 923 for (proc=0,k=0; proc<size; proc++) { 924 if (!len_s[proc]) continue; 925 /* form outgoing message for i-structure: 926 buf_si[0]: nrows to be sent 927 [1:nrows]: row index (global) 928 [nrows+1:2*nrows+1]: i-structure index 929 */ 930 /*-------------------------------------------*/ 931 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 932 buf_si_i = buf_si + nrows+1; 933 buf_si[0] = nrows; 934 buf_si_i[0] = 0; 935 nrows = 0; 936 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 937 nzi = coi[i+1] - coi[i]; 938 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 939 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 940 nrows++; 941 } 942 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 943 k++; 944 buf_si += len_si[proc]; 945 } 946 for (i=0; i<nrecv; i++) { 947 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 948 } 949 ierr = PetscFree(rwaits);CHKERRQ(ierr); 950 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 951 952 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 953 ierr = PetscFree(len_ri);CHKERRQ(ierr); 954 ierr = PetscFree(swaits);CHKERRQ(ierr); 955 ierr = PetscFree(buf_s);CHKERRQ(ierr); 956 957 /* (5) compute the local portion of Cmpi */ 958 /* ------------------------------------------ */ 959 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of Cmpi */ 960 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 961 current_space = free_space; 962 963 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 964 for (k=0; k<nrecv; k++) { 965 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 966 nrows = *buf_ri_k[k]; 967 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 968 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 969 } 970 971 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 972 ierr = PetscLLCondensedCreate(Crmax,pN,&lnk,&lnkbt);CHKERRQ(ierr); 973 for (i=0; i<pn; i++) { 974 /* add C_loc into Cmpi */ 975 nzi = c_loc->i[i+1] - c_loc->i[i]; 976 Jptr = c_loc->j + c_loc->i[i]; 977 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 978 979 /* add received col data into lnk */ 980 for (k=0; k<nrecv; k++) { /* k-th received message */ 981 if (i == *nextrow[k]) { /* i-th row */ 982 nzi = *(nextci[k]+1) - *nextci[k]; 983 Jptr = buf_rj[k] + *nextci[k]; 984 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 985 nextrow[k]++; nextci[k]++; 986 } 987 } 988 nzi = lnk[0]; 989 990 /* copy data into free space, then initialize lnk */ 991 ierr = PetscLLCondensedClean(pN,nzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 992 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 993 } 994 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 995 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 996 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 997 998 /* local sizes and preallocation */ 999 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1000 if (P->cmap->bs > 0) { 1001 ierr = PetscLayoutSetBlockSize(Cmpi->rmap,P->cmap->bs);CHKERRQ(ierr); 1002 ierr = PetscLayoutSetBlockSize(Cmpi->cmap,P->cmap->bs);CHKERRQ(ierr); 1003 } 1004 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1005 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1006 1007 /* members in merge */ 1008 ierr = PetscFree(id_r);CHKERRQ(ierr); 1009 ierr = PetscFree(len_r);CHKERRQ(ierr); 1010 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 1011 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 1012 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 1013 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 1014 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 1015 1016 /* attach the supporting struct to Cmpi for reuse */ 1017 c = (Mat_MPIAIJ*)Cmpi->data; 1018 c->ap = ptap; 1019 ptap->duplicate = Cmpi->ops->duplicate; 1020 ptap->destroy = Cmpi->ops->destroy; 1021 ptap->view = Cmpi->ops->view; 1022 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 1023 1024 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 1025 Cmpi->assembled = PETSC_FALSE; 1026 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1027 Cmpi->ops->view = MatView_MPIAIJ_PtAP; 1028 Cmpi->ops->freeintermediatedatastructures = MatFreeIntermediateDataStructures_MPIAIJ_AP; 1029 *C = Cmpi; 1030 PetscFunctionReturn(0); 1031 } 1032 1033 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 1034 { 1035 PetscErrorCode ierr; 1036 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1037 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 1038 Mat_SeqAIJ *ap,*p_loc,*p_oth=NULL,*c_seq; 1039 Mat_APMPI *ptap = c->ap; 1040 Mat AP_loc,C_loc,C_oth; 1041 PetscInt i,rstart,rend,cm,ncols,row; 1042 PetscInt *api,*apj,am = A->rmap->n,j,col,apnz; 1043 PetscScalar *apa; 1044 const PetscInt *cols; 1045 const PetscScalar *vals; 1046 1047 PetscFunctionBegin; 1048 if (!ptap) { 1049 MPI_Comm comm; 1050 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1051 SETERRQ(comm,PETSC_ERR_ARG_WRONGSTATE,"PtAP cannot be reused. Do not call MatFreeIntermediateDataStructures() or use '-mat_freeintermediatedatastructures'"); 1052 } 1053 1054 ierr = MatZeroEntries(C);CHKERRQ(ierr); 1055 /* 1) get R = Pd^T,Ro = Po^T */ 1056 if (ptap->reuse == MAT_REUSE_MATRIX) { 1057 ierr = MatTranspose(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 1058 ierr = MatTranspose(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 1059 } 1060 1061 /* 2) get AP_loc */ 1062 AP_loc = ptap->AP_loc; 1063 ap = (Mat_SeqAIJ*)AP_loc->data; 1064 1065 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 1066 /*-----------------------------------------------------*/ 1067 if (ptap->reuse == MAT_REUSE_MATRIX) { 1068 /* P_oth and P_loc are obtained in MatPtASymbolic() when reuse == MAT_INITIAL_MATRIX */ 1069 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1070 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1071 } 1072 1073 /* 2-2) compute numeric A_loc*P - dominating part */ 1074 /* ---------------------------------------------- */ 1075 /* get data from symbolic products */ 1076 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1077 if (ptap->P_oth) { 1078 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 1079 } 1080 apa = ptap->apa; 1081 api = ap->i; 1082 apj = ap->j; 1083 for (i=0; i<am; i++) { 1084 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 1085 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 1086 apnz = api[i+1] - api[i]; 1087 for (j=0; j<apnz; j++) { 1088 col = apj[j+api[i]]; 1089 ap->a[j+ap->i[i]] = apa[col]; 1090 apa[col] = 0.0; 1091 } 1092 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 1093 } 1094 1095 /* 3) C_loc = Rd*AP_loc, C_oth = Ro*AP_loc */ 1096 ierr = ((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,AP_loc,ptap->C_loc);CHKERRQ(ierr); 1097 ierr = ((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,AP_loc,ptap->C_oth);CHKERRQ(ierr); 1098 C_loc = ptap->C_loc; 1099 C_oth = ptap->C_oth; 1100 1101 /* add C_loc and Co to to C */ 1102 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 1103 1104 /* C_loc -> C */ 1105 cm = C_loc->rmap->N; 1106 c_seq = (Mat_SeqAIJ*)C_loc->data; 1107 cols = c_seq->j; 1108 vals = c_seq->a; 1109 1110 1111 /* The (fast) MatSetValues_MPIAIJ_CopyFromCSRFormat function can only be used when C->was_assembled is PETSC_FALSE and */ 1112 /* when there are no off-processor parts. */ 1113 /* If was_assembled is true, then the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart; in MatSetValues_MPIAIJ_CopyFromCSRFormat */ 1114 /* is no longer true. Then the more complex function MatSetValues_MPIAIJ() has to be used, where the column index is looked up from */ 1115 /* a table, and other, more complex stuff has to be done. */ 1116 if (C->assembled) { 1117 C->was_assembled = PETSC_TRUE; 1118 C->assembled = PETSC_FALSE; 1119 } 1120 if (C->was_assembled) { 1121 for (i=0; i<cm; i++) { 1122 ncols = c_seq->i[i+1] - c_seq->i[i]; 1123 row = rstart + i; 1124 ierr = MatSetValues_MPIAIJ(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1125 cols += ncols; vals += ncols; 1126 } 1127 } else { 1128 ierr = MatSetValues_MPIAIJ_CopyFromCSRFormat(C,c_seq->j,c_seq->i,c_seq->a);CHKERRQ(ierr); 1129 } 1130 1131 /* Co -> C, off-processor part */ 1132 cm = C_oth->rmap->N; 1133 c_seq = (Mat_SeqAIJ*)C_oth->data; 1134 cols = c_seq->j; 1135 vals = c_seq->a; 1136 for (i=0; i<cm; i++) { 1137 ncols = c_seq->i[i+1] - c_seq->i[i]; 1138 row = p->garray[i]; 1139 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 1140 cols += ncols; vals += ncols; 1141 } 1142 1143 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1144 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1145 1146 ptap->reuse = MAT_REUSE_MATRIX; 1147 1148 /* supporting struct ptap consumes almost same amount of memory as C=PtAP, release it if C will not be updated by A and P */ 1149 if (ptap->freestruct) { 1150 ierr = MatFreeIntermediateDataStructures(C);CHKERRQ(ierr); 1151 } 1152 PetscFunctionReturn(0); 1153 } 1154