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 if (ptap->api) {ierr = PetscFree(ptap->api);CHKERRQ(ierr);} 33 if (ptap->apj) {ierr = PetscFree(ptap->apj);CHKERRQ(ierr);} 34 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 35 if (merge) { 36 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 37 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 38 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 39 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 40 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 41 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 42 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 43 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 44 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 45 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 46 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 47 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 48 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 49 ierr = merge->destroy(A);CHKERRQ(ierr); 50 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 51 } 52 ierr = PetscFree(ptap);CHKERRQ(ierr); 53 } 54 PetscFunctionReturn(0); 55 } 56 57 #undef __FUNCT__ 58 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP" 59 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 60 { 61 PetscErrorCode ierr; 62 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 63 Mat_PtAPMPI *ptap = a->ptap; 64 Mat_Merge_SeqsToMPI *merge = ptap->merge; 65 66 PetscFunctionBegin; 67 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 68 69 (*M)->ops->destroy = merge->destroy; 70 (*M)->ops->duplicate = merge->duplicate; 71 PetscFunctionReturn(0); 72 } 73 74 #undef __FUNCT__ 75 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ" 76 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 77 { 78 PetscErrorCode ierr; 79 80 PetscFunctionBegin; 81 if (scall == MAT_INITIAL_MATRIX) { 82 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 83 } 84 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr); 85 PetscFunctionReturn(0); 86 } 87 88 #undef __FUNCT__ 89 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 90 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 91 { 92 PetscErrorCode ierr; 93 Mat Cmpi; 94 Mat_PtAPMPI *ptap; 95 PetscFreeSpaceList free_space=NULL,current_space=NULL; 96 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 97 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 98 Mat_SeqAIJ *p_loc,*p_oth; 99 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 100 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 101 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 102 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 103 PetscBT lnkbt; 104 MPI_Comm comm; 105 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 106 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 107 PetscInt len,proc,*dnz,*onz,*owners; 108 PetscInt nzi,*pti,*ptj; 109 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 110 MPI_Request *swaits,*rwaits; 111 MPI_Status *sstatus,rstatus; 112 Mat_Merge_SeqsToMPI *merge; 113 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 114 PetscReal afill=1.0,afill_tmp; 115 PetscInt rmax; 116 #if defined(PTAP_PROFILE) 117 PetscLogDouble t0,t1,t2,t3,t4; 118 #endif 119 120 PetscFunctionBegin; 121 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 122 #if defined(PTAP_PROFILE) 123 ierr = PetscTime(&t0);CHKERRQ(ierr); 124 #endif 125 126 /* check if matrix local sizes are compatible */ 127 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 128 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%D, %D) != Prow (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 129 } 130 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 131 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%D, %D) != Prow (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 132 } 133 134 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 135 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 136 137 /* create struct Mat_PtAPMPI and attached it to C later */ 138 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 139 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 140 ptap->merge = merge; 141 ptap->reuse = MAT_INITIAL_MATRIX; 142 143 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 144 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 145 146 /* get P_loc by taking all local rows of P */ 147 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 148 149 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 150 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 151 pi_loc = p_loc->i; pj_loc = p_loc->j; 152 pi_oth = p_oth->i; pj_oth = p_oth->j; 153 #if defined(PTAP_PROFILE) 154 ierr = PetscTime(&t1);CHKERRQ(ierr); 155 #endif 156 157 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 158 /*-------------------------------------------------------------------*/ 159 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&api);CHKERRQ(ierr); 160 api[0] = 0; 161 162 /* create and initialize a linked list */ 163 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 164 165 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 166 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 167 168 current_space = free_space; 169 170 for (i=0; i<am; i++) { 171 /* diagonal portion of A */ 172 nzi = adi[i+1] - adi[i]; 173 aj = ad->j + adi[i]; 174 for (j=0; j<nzi; j++) { 175 row = aj[j]; 176 pnz = pi_loc[row+1] - pi_loc[row]; 177 Jptr = pj_loc + pi_loc[row]; 178 /* add non-zero cols of P into the sorted linked list lnk */ 179 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 180 } 181 /* off-diagonal portion of A */ 182 nzi = aoi[i+1] - aoi[i]; 183 aj = ao->j + aoi[i]; 184 for (j=0; j<nzi; j++) { 185 row = aj[j]; 186 pnz = pi_oth[row+1] - pi_oth[row]; 187 Jptr = pj_oth + pi_oth[row]; 188 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 189 } 190 apnz = lnk[0]; 191 api[i+1] = api[i] + apnz; 192 if (ap_rmax < apnz) ap_rmax = apnz; 193 194 /* if free space is not available, double the total space in the list */ 195 if (current_space->local_remaining<apnz) { 196 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 197 nspacedouble++; 198 } 199 200 /* Copy data into free space, then initialize lnk */ 201 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 202 203 current_space->array += apnz; 204 current_space->local_used += apnz; 205 current_space->local_remaining -= apnz; 206 } 207 208 /* Allocate space for apj, initialize apj, and */ 209 /* destroy list of free space and other temporary array(s) */ 210 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&apj);CHKERRQ(ierr); 211 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 212 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 213 if (afill_tmp > afill) afill = afill_tmp; 214 215 #if defined(PTAP_PROFILE) 216 ierr = PetscTime(&t2);CHKERRQ(ierr); 217 #endif 218 219 /* determine symbolic Co=(p->B)^T*AP - send to others */ 220 /*----------------------------------------------------*/ 221 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 222 223 /* then, compute symbolic Co = (p->B)^T*AP */ 224 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 225 >= (num of nonzero rows of C_seq) - pn */ 226 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 227 coi[0] = 0; 228 229 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 230 nnz = fill*(poti[pon] + api[am]); 231 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 232 current_space = free_space; 233 234 for (i=0; i<pon; i++) { 235 pnz = poti[i+1] - poti[i]; 236 ptJ = potj + poti[i]; 237 for (j=0; j<pnz; j++) { 238 row = ptJ[j]; /* row of AP == col of Pot */ 239 apnz = api[row+1] - api[row]; 240 Jptr = apj + api[row]; 241 /* add non-zero cols of AP into the sorted linked list lnk */ 242 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 243 } 244 nnz = lnk[0]; 245 246 /* If free space is not available, double the total space in the list */ 247 if (current_space->local_remaining<nnz) { 248 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 249 nspacedouble++; 250 } 251 252 /* Copy data into free space, and zero out denserows */ 253 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 254 255 current_space->array += nnz; 256 current_space->local_used += nnz; 257 current_space->local_remaining -= nnz; 258 259 coi[i+1] = coi[i] + nnz; 260 } 261 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 262 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 263 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 264 if (afill_tmp > afill) afill = afill_tmp; 265 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 266 267 /* send j-array (coj) of Co to other processors */ 268 /*----------------------------------------------*/ 269 /* determine row ownership */ 270 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 271 merge->rowmap->n = pn; 272 merge->rowmap->bs = 1; 273 274 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 275 owners = merge->rowmap->range; 276 277 /* determine the number of messages to send, their lengths */ 278 ierr = PetscMalloc2(size,PetscMPIInt,&len_si,size,MPI_Status,&sstatus);CHKERRQ(ierr); 279 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 280 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 281 282 len_s = merge->len_s; 283 merge->nsend = 0; 284 285 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 286 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 287 288 proc = 0; 289 for (i=0; i<pon; i++) { 290 while (prmap[i] >= owners[proc+1]) proc++; 291 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 292 len_s[proc] += coi[i+1] - coi[i]; 293 } 294 295 len = 0; /* max length of buf_si[] */ 296 owners_co[0] = 0; 297 for (proc=0; proc<size; proc++) { 298 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 299 if (len_si[proc]) { 300 merge->nsend++; 301 len_si[proc] = 2*(len_si[proc] + 1); 302 len += len_si[proc]; 303 } 304 } 305 306 /* determine the number and length of messages to receive for coi and coj */ 307 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 308 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 309 310 /* post the Irecv and Isend of coj */ 311 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 312 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 313 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 314 for (proc=0, k=0; proc<size; proc++) { 315 if (!len_s[proc]) continue; 316 i = owners_co[proc]; 317 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 318 k++; 319 } 320 321 /* receives and sends of coj are complete */ 322 for (i=0; i<merge->nrecv; i++) { 323 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 324 } 325 ierr = PetscFree(rwaits);CHKERRQ(ierr); 326 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 327 328 /* send and recv coi */ 329 /*-------------------*/ 330 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 331 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 332 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 333 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 334 for (proc=0,k=0; proc<size; proc++) { 335 if (!len_s[proc]) continue; 336 /* form outgoing message for i-structure: 337 buf_si[0]: nrows to be sent 338 [1:nrows]: row index (global) 339 [nrows+1:2*nrows+1]: i-structure index 340 */ 341 /*-------------------------------------------*/ 342 nrows = len_si[proc]/2 - 1; 343 buf_si_i = buf_si + nrows+1; 344 buf_si[0] = nrows; 345 buf_si_i[0] = 0; 346 nrows = 0; 347 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 348 nzi = coi[i+1] - coi[i]; 349 350 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 351 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 352 nrows++; 353 } 354 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 355 k++; 356 buf_si += len_si[proc]; 357 } 358 i = merge->nrecv; 359 while (i--) { 360 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 361 } 362 ierr = PetscFree(rwaits);CHKERRQ(ierr); 363 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 364 365 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 366 ierr = PetscFree(len_ri);CHKERRQ(ierr); 367 ierr = PetscFree(swaits);CHKERRQ(ierr); 368 ierr = PetscFree(buf_s);CHKERRQ(ierr); 369 370 #if defined(PTAP_PROFILE) 371 ierr = PetscTime(&t3);CHKERRQ(ierr); 372 #endif 373 374 /* compute the local portion of C (mpi mat) */ 375 /*------------------------------------------*/ 376 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 377 378 /* allocate pti array and free space for accumulating nonzero column info */ 379 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&pti);CHKERRQ(ierr); 380 pti[0] = 0; 381 382 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 383 nnz = fill*(pi_loc[pm] + api[am]); 384 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 385 current_space = free_space; 386 387 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 388 for (k=0; k<merge->nrecv; k++) { 389 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 390 nrows = *buf_ri_k[k]; 391 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 392 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 393 } 394 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 395 rmax = 0; 396 for (i=0; i<pn; i++) { 397 /* add pdt[i,:]*AP into lnk */ 398 pnz = pdti[i+1] - pdti[i]; 399 ptJ = pdtj + pdti[i]; 400 for (j=0; j<pnz; j++) { 401 row = ptJ[j]; /* row of AP == col of Pt */ 402 apnz = api[row+1] - api[row]; 403 Jptr = apj + api[row]; 404 /* add non-zero cols of AP into the sorted linked list lnk */ 405 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 406 } 407 408 /* add received col data into lnk */ 409 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 410 if (i == *nextrow[k]) { /* i-th row */ 411 nzi = *(nextci[k]+1) - *nextci[k]; 412 Jptr = buf_rj[k] + *nextci[k]; 413 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 414 nextrow[k]++; nextci[k]++; 415 } 416 } 417 nnz = lnk[0]; 418 419 /* if free space is not available, make more free space */ 420 if (current_space->local_remaining<nnz) { 421 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 422 nspacedouble++; 423 } 424 /* copy data into free space, then initialize lnk */ 425 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 426 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 427 428 current_space->array += nnz; 429 current_space->local_used += nnz; 430 current_space->local_remaining -= nnz; 431 432 pti[i+1] = pti[i] + nnz; 433 if (nnz > rmax) rmax = nnz; 434 } 435 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 436 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 437 438 ierr = PetscMalloc((pti[pn]+1)*sizeof(PetscInt),&ptj);CHKERRQ(ierr); 439 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 440 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 441 if (afill_tmp > afill) afill = afill_tmp; 442 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 443 444 /* create symbolic parallel matrix Cmpi */ 445 /*--------------------------------------*/ 446 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 447 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 448 ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,P->cmap->bs);CHKERRQ(ierr); 449 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 450 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 451 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 452 453 merge->bi = pti; /* Cseq->i */ 454 merge->bj = ptj; /* Cseq->j */ 455 merge->coi = coi; /* Co->i */ 456 merge->coj = coj; /* Co->j */ 457 merge->buf_ri = buf_ri; 458 merge->buf_rj = buf_rj; 459 merge->owners_co = owners_co; 460 merge->destroy = Cmpi->ops->destroy; 461 merge->duplicate = Cmpi->ops->duplicate; 462 463 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 464 Cmpi->assembled = PETSC_FALSE; 465 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 466 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 467 468 /* attach the supporting struct to Cmpi for reuse */ 469 c = (Mat_MPIAIJ*)Cmpi->data; 470 c->ptap = ptap; 471 ptap->api = api; 472 ptap->apj = apj; 473 ptap->rmax = ap_rmax; 474 *C = Cmpi; 475 476 /* flag 'scalable' determines which implementations to be used: 477 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 478 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 479 /* set default scalable */ 480 ptap->scalable = PETSC_TRUE; 481 482 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 483 if (!ptap->scalable) { /* Do dense axpy */ 484 ierr = PetscMalloc(pN*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr); 485 ierr = PetscMemzero(ptap->apa,pN*sizeof(PetscScalar));CHKERRQ(ierr); 486 } else { 487 ierr = PetscMalloc((ap_rmax+1)*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr); 488 ierr = PetscMemzero(ptap->apa,ap_rmax*sizeof(PetscScalar));CHKERRQ(ierr); 489 } 490 491 #if defined(PTAP_PROFILE) 492 ierr = PetscTime(&t4);CHKERRQ(ierr); 493 if (rank==1) PetscPrintf(MPI_COMM_SELF," [%d] PtAPSymbolic %g/P + %g/AP + %g/comm + %g/PtAP = %g\n",rank,t1-t0,t2-t1,t3-t2,t4-t3,t4-t0);CHKERRQ(ierr); 494 #endif 495 496 #if defined(PETSC_USE_INFO) 497 if (pti[pn] != 0) { 498 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 499 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 500 } else { 501 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 502 } 503 #endif 504 PetscFunctionReturn(0); 505 } 506 507 #undef __FUNCT__ 508 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 509 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 510 { 511 PetscErrorCode ierr; 512 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 513 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 514 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 515 Mat_SeqAIJ *p_loc,*p_oth; 516 Mat_PtAPMPI *ptap; 517 Mat_Merge_SeqsToMPI *merge; 518 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 519 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 520 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 521 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 522 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 523 MPI_Comm comm; 524 PetscMPIInt size,rank,taga,*len_s; 525 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 526 PetscInt **buf_ri,**buf_rj; 527 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 528 MPI_Request *s_waits,*r_waits; 529 MPI_Status *status; 530 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 531 PetscInt *api,*apj,*coi,*coj; 532 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 533 PetscBool scalable; 534 #if defined(PTAP_PROFILE) 535 PetscLogDouble t0,t1,t2,t3,t4,et2_AP=0.0,et2_PtAP=0.0,t2_0,t2_1,t2_2; 536 #endif 537 538 PetscFunctionBegin; 539 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 540 #if defined(PTAP_PROFILE) 541 ierr = PetscTime(&t0);CHKERRQ(ierr); 542 #endif 543 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 544 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 545 546 ptap = c->ptap; 547 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"); 548 merge = ptap->merge; 549 apa = ptap->apa; 550 scalable = ptap->scalable; 551 552 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 553 /*--------------------------------------------------*/ 554 if (ptap->reuse == MAT_INITIAL_MATRIX) { 555 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 556 ptap->reuse = MAT_REUSE_MATRIX; 557 } else { /* update numerical values of P_oth and P_loc */ 558 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 559 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 560 } 561 #if defined(PTAP_PROFILE) 562 ierr = PetscTime(&t1);CHKERRQ(ierr); 563 #endif 564 565 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 566 /*--------------------------------------------------------------*/ 567 /* get data from symbolic products */ 568 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 569 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 570 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 571 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 572 573 coi = merge->coi; coj = merge->coj; 574 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 575 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 576 577 bi = merge->bi; bj = merge->bj; 578 owners = merge->rowmap->range; 579 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 580 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 581 582 api = ptap->api; apj = ptap->apj; 583 584 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but fast */ 585 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 586 /*-----------------------------------------------------------------------------------------------------*/ 587 for (i=0; i<am; i++) { 588 #if defined(PTAP_PROFILE) 589 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 590 #endif 591 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 592 /*------------------------------------------------------------*/ 593 apJ = apj + api[i]; 594 595 /* diagonal portion of A */ 596 anz = adi[i+1] - adi[i]; 597 adj = ad->j + adi[i]; 598 ada = ad->a + adi[i]; 599 for (j=0; j<anz; j++) { 600 row = adj[j]; 601 pnz = pi_loc[row+1] - pi_loc[row]; 602 pj = pj_loc + pi_loc[row]; 603 pa = pa_loc + pi_loc[row]; 604 605 /* perform dense axpy */ 606 valtmp = ada[j]; 607 for (k=0; k<pnz; k++) { 608 apa[pj[k]] += valtmp*pa[k]; 609 } 610 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 611 } 612 613 /* off-diagonal portion of A */ 614 anz = aoi[i+1] - aoi[i]; 615 aoj = ao->j + aoi[i]; 616 aoa = ao->a + aoi[i]; 617 for (j=0; j<anz; j++) { 618 row = aoj[j]; 619 pnz = pi_oth[row+1] - pi_oth[row]; 620 pj = pj_oth + pi_oth[row]; 621 pa = pa_oth + pi_oth[row]; 622 623 /* perform dense axpy */ 624 valtmp = aoa[j]; 625 for (k=0; k<pnz; k++) { 626 apa[pj[k]] += valtmp*pa[k]; 627 } 628 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 629 } 630 #if defined(PTAP_PROFILE) 631 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 632 et2_AP += t2_1 - t2_0; 633 #endif 634 635 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 636 /*--------------------------------------------------------------*/ 637 apnz = api[i+1] - api[i]; 638 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 639 pnz = po->i[i+1] - po->i[i]; 640 poJ = po->j + po->i[i]; 641 pA = po->a + po->i[i]; 642 for (j=0; j<pnz; j++) { 643 row = poJ[j]; 644 cnz = coi[row+1] - coi[row]; 645 cj = coj + coi[row]; 646 ca = coa + coi[row]; 647 /* perform dense axpy */ 648 valtmp = pA[j]; 649 for (k=0; k<cnz; k++) { 650 ca[k] += valtmp*apa[cj[k]]; 651 } 652 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 653 } 654 655 /* put the value into Cd (diagonal part) */ 656 pnz = pd->i[i+1] - pd->i[i]; 657 pdJ = pd->j + pd->i[i]; 658 pA = pd->a + pd->i[i]; 659 for (j=0; j<pnz; j++) { 660 row = pdJ[j]; 661 cnz = bi[row+1] - bi[row]; 662 cj = bj + bi[row]; 663 ca = ba + bi[row]; 664 /* perform dense axpy */ 665 valtmp = pA[j]; 666 for (k=0; k<cnz; k++) { 667 ca[k] += valtmp*apa[cj[k]]; 668 } 669 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 670 } 671 672 /* zero the current row of A*P */ 673 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 674 #if defined(PTAP_PROFILE) 675 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 676 et2_PtAP += t2_2 - t2_1; 677 #endif 678 } 679 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 680 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 681 /*-----------------------------------------------------------------------------------------*/ 682 pA=pa_loc; 683 for (i=0; i<am; i++) { 684 #if defined(PTAP_PROFILE) 685 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 686 #endif 687 /* form i-th sparse row of A*P */ 688 apnz = api[i+1] - api[i]; 689 apJ = apj + api[i]; 690 /* diagonal portion of A */ 691 anz = adi[i+1] - adi[i]; 692 adj = ad->j + adi[i]; 693 ada = ad->a + adi[i]; 694 for (j=0; j<anz; j++) { 695 row = adj[j]; 696 pnz = pi_loc[row+1] - pi_loc[row]; 697 pj = pj_loc + pi_loc[row]; 698 pa = pa_loc + pi_loc[row]; 699 valtmp = ada[j]; 700 nextp = 0; 701 for (k=0; nextp<pnz; k++) { 702 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 703 apa[k] += valtmp*pa[nextp++]; 704 } 705 } 706 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 707 } 708 /* off-diagonal portion of A */ 709 anz = aoi[i+1] - aoi[i]; 710 aoj = ao->j + aoi[i]; 711 aoa = ao->a + aoi[i]; 712 for (j=0; j<anz; j++) { 713 row = aoj[j]; 714 pnz = pi_oth[row+1] - pi_oth[row]; 715 pj = pj_oth + pi_oth[row]; 716 pa = pa_oth + pi_oth[row]; 717 valtmp = aoa[j]; 718 nextp = 0; 719 for (k=0; nextp<pnz; k++) { 720 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 721 apa[k] += valtmp*pa[nextp++]; 722 } 723 } 724 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 725 } 726 #if defined(PTAP_PROFILE) 727 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 728 et2_AP += t2_1 - t2_0; 729 #endif 730 731 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 732 /*--------------------------------------------------------------*/ 733 pnz = pi_loc[i+1] - pi_loc[i]; 734 pJ = pj_loc + pi_loc[i]; 735 for (j=0; j<pnz; j++) { 736 nextap = 0; 737 row = pJ[j]; /* global index */ 738 if (row < pcstart || row >=pcend) { /* put the value into Co */ 739 row = *poJ; 740 cj = coj + coi[row]; 741 ca = coa + coi[row]; poJ++; 742 } else { /* put the value into Cd */ 743 row = *pdJ; 744 cj = bj + bi[row]; 745 ca = ba + bi[row]; pdJ++; 746 } 747 valtmp = pA[j]; 748 for (k=0; nextap<apnz; k++) { 749 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 750 } 751 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 752 } 753 pA += pnz; 754 /* zero the current row info for A*P */ 755 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 756 #if defined(PTAP_PROFILE) 757 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 758 et2_PtAP += t2_2 - t2_1; 759 #endif 760 } 761 } 762 #if defined(PTAP_PROFILE) 763 ierr = PetscTime(&t2);CHKERRQ(ierr); 764 #endif 765 766 /* 3) send and recv matrix values coa */ 767 /*------------------------------------*/ 768 buf_ri = merge->buf_ri; 769 buf_rj = merge->buf_rj; 770 len_s = merge->len_s; 771 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 772 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 773 774 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 775 for (proc=0,k=0; proc<size; proc++) { 776 if (!len_s[proc]) continue; 777 i = merge->owners_co[proc]; 778 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 779 k++; 780 } 781 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 782 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 783 784 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 785 ierr = PetscFree(r_waits);CHKERRQ(ierr); 786 ierr = PetscFree(coa);CHKERRQ(ierr); 787 #if defined(PTAP_PROFILE) 788 ierr = PetscTime(&t3);CHKERRQ(ierr); 789 #endif 790 791 /* 4) insert local Cseq and received values into Cmpi */ 792 /*------------------------------------------------------*/ 793 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 794 for (k=0; k<merge->nrecv; k++) { 795 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 796 nrows = *(buf_ri_k[k]); 797 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 798 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 799 } 800 801 for (i=0; i<cm; i++) { 802 row = owners[rank] + i; /* global row index of C_seq */ 803 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 804 ba_i = ba + bi[i]; 805 bnz = bi[i+1] - bi[i]; 806 /* add received vals into ba */ 807 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 808 /* i-th row */ 809 if (i == *nextrow[k]) { 810 cnz = *(nextci[k]+1) - *nextci[k]; 811 cj = buf_rj[k] + *(nextci[k]); 812 ca = abuf_r[k] + *(nextci[k]); 813 nextcj = 0; 814 for (j=0; nextcj<cnz; j++) { 815 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 816 ba_i[j] += ca[nextcj++]; 817 } 818 } 819 nextrow[k]++; nextci[k]++; 820 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 821 } 822 } 823 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 824 } 825 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 826 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 827 828 ierr = PetscFree(ba);CHKERRQ(ierr); 829 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 830 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 831 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 832 #if defined(PTAP_PROFILE) 833 ierr = PetscTime(&t4);CHKERRQ(ierr); 834 if (rank==1) PetscPrintf(MPI_COMM_SELF," [%d] PtAPNum %g/P + %g/PtAP( %g + %g ) + %g/comm + %g/Cloc = %g\n\n",rank,t1-t0,t2-t1,et2_AP,et2_PtAP,t3-t2,t4-t3,t4-t0);CHKERRQ(ierr); 835 #endif 836 PetscFunctionReturn(0); 837 } 838