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 12 /* #define PTAP_PROFILE */ 13 14 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 15 #undef __FUNCT__ 16 #define __FUNCT__ "MatDestroy_MPIAIJ_PtAP" 17 PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat A) 18 { 19 PetscErrorCode ierr; 20 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 21 Mat_PtAPMPI *ptap=a->ptap; 22 23 PetscFunctionBegin; 24 if (ptap) { 25 Mat_Merge_SeqsToMPI *merge=ptap->merge; 26 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 27 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 28 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 29 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 30 ierr = MatDestroy(&ptap->A_loc);CHKERRQ(ierr); /* used by MatTransposeMatMult() */ 31 if (ptap->api) {ierr = PetscFree(ptap->api);CHKERRQ(ierr);} 32 if (ptap->apj) {ierr = PetscFree(ptap->apj);CHKERRQ(ierr);} 33 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 34 if (merge) { 35 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 36 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 37 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 38 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 39 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 40 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 41 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 42 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 43 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 44 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 45 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 46 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 47 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 48 ierr = merge->destroy(A);CHKERRQ(ierr); 49 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 50 } 51 ierr = PetscFree(ptap);CHKERRQ(ierr); 52 } 53 PetscFunctionReturn(0); 54 } 55 56 #undef __FUNCT__ 57 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP" 58 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 59 { 60 PetscErrorCode ierr; 61 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 62 Mat_PtAPMPI *ptap = a->ptap; 63 Mat_Merge_SeqsToMPI *merge = ptap->merge; 64 65 PetscFunctionBegin; 66 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 67 68 (*M)->ops->destroy = merge->destroy; 69 (*M)->ops->duplicate = merge->duplicate; 70 PetscFunctionReturn(0); 71 } 72 73 #undef __FUNCT__ 74 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ" 75 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 76 { 77 PetscErrorCode ierr; 78 79 PetscFunctionBegin; 80 if (scall == MAT_INITIAL_MATRIX) { 81 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 82 } 83 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr); 84 PetscFunctionReturn(0); 85 } 86 87 #undef __FUNCT__ 88 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 89 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 90 { 91 PetscErrorCode ierr; 92 Mat Cmpi; 93 Mat_PtAPMPI *ptap; 94 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 95 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 96 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 97 Mat_SeqAIJ *p_loc,*p_oth; 98 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 99 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 100 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 101 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 102 PetscBT lnkbt; 103 MPI_Comm comm=((PetscObject)A)->comm; 104 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 105 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 106 PetscInt len,proc,*dnz,*onz,*owners; 107 PetscInt nzi,*pti,*ptj; 108 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 109 MPI_Request *swaits,*rwaits; 110 MPI_Status *sstatus,rstatus; 111 Mat_Merge_SeqsToMPI *merge; 112 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 113 PetscReal afill=1.0,afill_tmp; 114 PetscInt rmax; 115 #if defined(PTAP_PROFILE) 116 PetscLogDouble t0,t1,t2,t3,t4; 117 #endif 118 119 PetscFunctionBegin; 120 #if defined(PTAP_PROFILE) 121 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 122 #endif 123 124 /* check if matrix local sizes are compatible */ 125 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 126 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); 127 } 128 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 129 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); 130 } 131 132 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 133 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 134 135 /* create struct Mat_PtAPMPI and attached it to C later */ 136 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 137 ptap->reuse = MAT_INITIAL_MATRIX; 138 139 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 140 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 141 142 /* get P_loc by taking all local rows of P */ 143 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 144 145 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 146 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 147 pi_loc = p_loc->i; pj_loc = p_loc->j; 148 pi_oth = p_oth->i; pj_oth = p_oth->j; 149 #if defined(PTAP_PROFILE) 150 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 151 #endif 152 153 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 154 /*-------------------------------------------------------------------*/ 155 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&api);CHKERRQ(ierr); 156 api[0] = 0; 157 158 /* create and initialize a linked list */ 159 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 160 161 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 162 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 163 164 current_space = free_space; 165 166 for (i=0; i<am; i++) { 167 /* diagonal portion of A */ 168 nzi = adi[i+1] - adi[i]; 169 aj = ad->j + adi[i]; 170 for (j=0; j<nzi; j++) { 171 row = aj[j]; 172 pnz = pi_loc[row+1] - pi_loc[row]; 173 Jptr = pj_loc + pi_loc[row]; 174 /* add non-zero cols of P into the sorted linked list lnk */ 175 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 176 } 177 /* off-diagonal portion of A */ 178 nzi = aoi[i+1] - aoi[i]; 179 aj = ao->j + aoi[i]; 180 for (j=0; j<nzi; j++) { 181 row = aj[j]; 182 pnz = pi_oth[row+1] - pi_oth[row]; 183 Jptr = pj_oth + pi_oth[row]; 184 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 185 } 186 apnz = lnk[0]; 187 api[i+1] = api[i] + apnz; 188 if (ap_rmax < apnz) ap_rmax = apnz; 189 190 /* if free space is not available, double the total space in the list */ 191 if (current_space->local_remaining<apnz) { 192 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 193 nspacedouble++; 194 } 195 196 /* Copy data into free space, then initialize lnk */ 197 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 198 199 current_space->array += apnz; 200 current_space->local_used += apnz; 201 current_space->local_remaining -= apnz; 202 } 203 204 /* Allocate space for apj, initialize apj, and */ 205 /* destroy list of free space and other temporary array(s) */ 206 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&apj);CHKERRQ(ierr); 207 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 208 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 209 if (afill_tmp > afill) afill = afill_tmp; 210 211 #if defined(PTAP_PROFILE) 212 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 213 #endif 214 215 /* determine symbolic Co=(p->B)^T*AP - send to others */ 216 /*----------------------------------------------------*/ 217 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 218 219 /* then, compute symbolic Co = (p->B)^T*AP */ 220 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 221 >= (num of nonzero rows of C_seq) - pn */ 222 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 223 coi[0] = 0; 224 225 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 226 nnz = fill*(poti[pon] + api[am]); 227 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 228 current_space = free_space; 229 230 for (i=0; i<pon; i++) { 231 pnz = poti[i+1] - poti[i]; 232 ptJ = potj + poti[i]; 233 for (j=0; j<pnz; j++) { 234 row = ptJ[j]; /* row of AP == col of Pot */ 235 apnz = api[row+1] - api[row]; 236 Jptr = apj + api[row]; 237 /* add non-zero cols of AP into the sorted linked list lnk */ 238 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 239 } 240 nnz = lnk[0]; 241 242 /* If free space is not available, double the total space in the list */ 243 if (current_space->local_remaining<nnz) { 244 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 245 nspacedouble++; 246 } 247 248 /* Copy data into free space, and zero out denserows */ 249 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 250 251 current_space->array += nnz; 252 current_space->local_used += nnz; 253 current_space->local_remaining -= nnz; 254 255 coi[i+1] = coi[i] + nnz; 256 } 257 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 258 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 259 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 260 if (afill_tmp > afill) afill = afill_tmp; 261 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 262 263 /* send j-array (coj) of Co to other processors */ 264 /*----------------------------------------------*/ 265 /* determine row ownership */ 266 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 267 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 268 269 merge->rowmap->n = pn; 270 merge->rowmap->bs = 1; 271 272 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 273 owners = merge->rowmap->range; 274 275 /* determine the number of messages to send, their lengths */ 276 ierr = PetscMalloc2(size,PetscMPIInt,&len_si,size,MPI_Status,&sstatus);CHKERRQ(ierr); 277 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 278 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 279 280 len_s = merge->len_s; 281 merge->nsend = 0; 282 283 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 284 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 285 286 proc = 0; 287 for (i=0; i<pon; i++) { 288 while (prmap[i] >= owners[proc+1]) proc++; 289 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 290 len_s[proc] += coi[i+1] - coi[i]; 291 } 292 293 len = 0; /* max length of buf_si[] */ 294 owners_co[0] = 0; 295 for (proc=0; proc<size; proc++) { 296 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 297 if (len_si[proc]) { 298 merge->nsend++; 299 len_si[proc] = 2*(len_si[proc] + 1); 300 len += len_si[proc]; 301 } 302 } 303 304 /* determine the number and length of messages to receive for coi and coj */ 305 ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 306 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 307 308 /* post the Irecv and Isend of coj */ 309 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 310 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 311 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 312 for (proc=0, k=0; proc<size; proc++) { 313 if (!len_s[proc]) continue; 314 i = owners_co[proc]; 315 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 316 k++; 317 } 318 319 /* receives and sends of coj are complete */ 320 for (i=0; i<merge->nrecv; i++) { 321 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 322 } 323 ierr = PetscFree(rwaits);CHKERRQ(ierr); 324 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 325 326 /* send and recv coi */ 327 /*-------------------*/ 328 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 329 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 330 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 331 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 332 for (proc=0,k=0; proc<size; proc++) { 333 if (!len_s[proc]) continue; 334 /* form outgoing message for i-structure: 335 buf_si[0]: nrows to be sent 336 [1:nrows]: row index (global) 337 [nrows+1:2*nrows+1]: i-structure index 338 */ 339 /*-------------------------------------------*/ 340 nrows = len_si[proc]/2 - 1; 341 buf_si_i = buf_si + nrows+1; 342 buf_si[0] = nrows; 343 buf_si_i[0] = 0; 344 nrows = 0; 345 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 346 nzi = coi[i+1] - coi[i]; 347 348 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 349 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 350 nrows++; 351 } 352 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 353 k++; 354 buf_si += len_si[proc]; 355 } 356 i = merge->nrecv; 357 while (i--) { 358 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 359 } 360 ierr = PetscFree(rwaits);CHKERRQ(ierr); 361 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 362 363 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 364 ierr = PetscFree(len_ri);CHKERRQ(ierr); 365 ierr = PetscFree(swaits);CHKERRQ(ierr); 366 ierr = PetscFree(buf_s);CHKERRQ(ierr); 367 368 #if defined(PTAP_PROFILE) 369 ierr = PetscGetTime(&t3);CHKERRQ(ierr); 370 #endif 371 372 /* compute the local portion of C (mpi mat) */ 373 /*------------------------------------------*/ 374 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 375 376 /* allocate pti array and free space for accumulating nonzero column info */ 377 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&pti);CHKERRQ(ierr); 378 pti[0] = 0; 379 380 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 381 nnz = fill*(pi_loc[pm] + api[am]); 382 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 383 current_space = free_space; 384 385 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 386 for (k=0; k<merge->nrecv; k++) { 387 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 388 nrows = *buf_ri_k[k]; 389 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 390 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 391 } 392 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 393 rmax = 0; 394 for (i=0; i<pn; i++) { 395 /* add pdt[i,:]*AP into lnk */ 396 pnz = pdti[i+1] - pdti[i]; 397 ptJ = pdtj + pdti[i]; 398 for (j=0; j<pnz; j++) { 399 row = ptJ[j]; /* row of AP == col of Pt */ 400 apnz = api[row+1] - api[row]; 401 Jptr = apj + api[row]; 402 /* add non-zero cols of AP into the sorted linked list lnk */ 403 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 404 } 405 406 /* add received col data into lnk */ 407 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 408 if (i == *nextrow[k]) { /* i-th row */ 409 nzi = *(nextci[k]+1) - *nextci[k]; 410 Jptr = buf_rj[k] + *nextci[k]; 411 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 412 nextrow[k]++; nextci[k]++; 413 } 414 } 415 nnz = lnk[0]; 416 417 /* if free space is not available, make more free space */ 418 if (current_space->local_remaining<nnz) { 419 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 420 nspacedouble++; 421 } 422 /* copy data into free space, then initialize lnk */ 423 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 424 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 425 426 current_space->array += nnz; 427 current_space->local_used += nnz; 428 current_space->local_remaining -= nnz; 429 430 pti[i+1] = pti[i] + nnz; 431 if (nnz > rmax) rmax = nnz; 432 } 433 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 434 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 435 436 ierr = PetscMalloc((pti[pn]+1)*sizeof(PetscInt),&ptj);CHKERRQ(ierr); 437 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 438 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 439 if (afill_tmp > afill) afill = afill_tmp; 440 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 441 442 /* create symbolic parallel matrix Cmpi */ 443 /*--------------------------------------*/ 444 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 445 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 446 ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,P->cmap->bs);CHKERRQ(ierr); 447 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 448 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 449 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 450 451 merge->bi = pti; /* Cseq->i */ 452 merge->bj = ptj; /* Cseq->j */ 453 merge->coi = coi; /* Co->i */ 454 merge->coj = coj; /* Co->j */ 455 merge->buf_ri = buf_ri; 456 merge->buf_rj = buf_rj; 457 merge->owners_co = owners_co; 458 merge->destroy = Cmpi->ops->destroy; 459 merge->duplicate = Cmpi->ops->duplicate; 460 461 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 462 Cmpi->assembled = PETSC_FALSE; 463 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 464 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 465 466 /* attach the supporting struct to Cmpi for reuse */ 467 c = (Mat_MPIAIJ*)Cmpi->data; 468 c->ptap = ptap; 469 ptap->api = api; 470 ptap->apj = apj; 471 ptap->merge = merge; 472 ptap->rmax = ap_rmax; 473 *C = Cmpi; 474 475 /* flag 'scalable' determines which implementations to be used: 476 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 477 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 478 /* set default scalable */ 479 ptap->scalable = PETSC_TRUE; 480 481 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,PETSC_NULL);CHKERRQ(ierr); 482 if (!ptap->scalable) { /* Do dense axpy */ 483 ierr = PetscMalloc(pN*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr); 484 ierr = PetscMemzero(ptap->apa,pN*sizeof(PetscScalar));CHKERRQ(ierr); 485 } else { 486 ierr = PetscMalloc((ap_rmax+1)*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr); 487 ierr = PetscMemzero(ptap->apa,ap_rmax*sizeof(PetscScalar));CHKERRQ(ierr); 488 } 489 490 #if defined(PTAP_PROFILE) 491 ierr = PetscGetTime(&t4);CHKERRQ(ierr); 492 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); 493 #endif 494 495 #if defined(PETSC_USE_INFO) 496 if (pti[pn] != 0) { 497 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 498 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 499 } else { 500 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 501 } 502 #endif 503 PetscFunctionReturn(0); 504 } 505 506 #undef __FUNCT__ 507 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 508 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 509 { 510 PetscErrorCode ierr; 511 Mat_Merge_SeqsToMPI *merge; 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 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 518 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 519 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 520 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 521 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 522 MPI_Comm comm=((PetscObject)C)->comm; 523 PetscMPIInt size,rank,taga,*len_s; 524 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 525 PetscInt **buf_ri,**buf_rj; 526 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 527 MPI_Request *s_waits,*r_waits; 528 MPI_Status *status; 529 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 530 PetscInt *api,*apj,*coi,*coj; 531 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 532 PetscBool scalable; 533 #if defined(PTAP_PROFILE) 534 PetscLogDouble t0,t1,t2,t3,t4,et2_AP=0.0,et2_PtAP=0.0,t2_0,t2_1,t2_2; 535 #endif 536 537 PetscFunctionBegin; 538 #if defined(PTAP_PROFILE) 539 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 540 #endif 541 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 542 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 543 544 ptap = c->ptap; 545 if (!ptap) SETERRQ(((PetscObject)C)->comm,PETSC_ERR_ARG_INCOMP,"MatPtAP() has not been called to create matrix C yet, cannot use MAT_REUSE_MATRIX"); 546 merge = ptap->merge; 547 apa = ptap->apa; 548 scalable = ptap->scalable; 549 550 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 551 /*--------------------------------------------------*/ 552 if (ptap->reuse == MAT_INITIAL_MATRIX) { 553 ptap->reuse = MAT_REUSE_MATRIX; 554 } else { /* update numerical values of P_oth and P_loc */ 555 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 556 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 557 } 558 #if defined(PTAP_PROFILE) 559 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 560 #endif 561 562 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 563 /*--------------------------------------------------------------*/ 564 /* get data from symbolic products */ 565 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 566 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 567 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 568 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 569 570 coi = merge->coi; coj = merge->coj; 571 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 572 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 573 574 bi = merge->bi; bj = merge->bj; 575 owners = merge->rowmap->range; 576 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 577 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 578 579 api = ptap->api; apj = ptap->apj; 580 581 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but fast */ 582 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 583 /*-----------------------------------------------------------------------------------------------------*/ 584 for (i=0; i<am; i++) { 585 #if defined(PTAP_PROFILE) 586 ierr = PetscGetTime(&t2_0);CHKERRQ(ierr); 587 #endif 588 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 589 /*------------------------------------------------------------*/ 590 apJ = apj + api[i]; 591 592 /* diagonal portion of A */ 593 anz = adi[i+1] - adi[i]; 594 adj = ad->j + adi[i]; 595 ada = ad->a + adi[i]; 596 for (j=0; j<anz; j++) { 597 row = adj[j]; 598 pnz = pi_loc[row+1] - pi_loc[row]; 599 pj = pj_loc + pi_loc[row]; 600 pa = pa_loc + pi_loc[row]; 601 602 /* perform dense axpy */ 603 valtmp = ada[j]; 604 for (k=0; k<pnz; k++) { 605 apa[pj[k]] += valtmp*pa[k]; 606 } 607 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 608 } 609 610 /* off-diagonal portion of A */ 611 anz = aoi[i+1] - aoi[i]; 612 aoj = ao->j + aoi[i]; 613 aoa = ao->a + aoi[i]; 614 for (j=0; j<anz; j++) { 615 row = aoj[j]; 616 pnz = pi_oth[row+1] - pi_oth[row]; 617 pj = pj_oth + pi_oth[row]; 618 pa = pa_oth + pi_oth[row]; 619 620 /* perform dense axpy */ 621 valtmp = aoa[j]; 622 for (k=0; k<pnz; k++) { 623 apa[pj[k]] += valtmp*pa[k]; 624 } 625 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 626 } 627 #if defined(PTAP_PROFILE) 628 ierr = PetscGetTime(&t2_1);CHKERRQ(ierr); 629 et2_AP += t2_1 - t2_0; 630 #endif 631 632 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 633 /*--------------------------------------------------------------*/ 634 apnz = api[i+1] - api[i]; 635 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 636 pnz = po->i[i+1] - po->i[i]; 637 poJ = po->j + po->i[i]; 638 pA = po->a + po->i[i]; 639 for (j=0; j<pnz; j++) { 640 row = poJ[j]; 641 cnz = coi[row+1] - coi[row]; 642 cj = coj + coi[row]; 643 ca = coa + coi[row]; 644 /* perform dense axpy */ 645 valtmp = pA[j]; 646 for (k=0; k<cnz; k++) { 647 ca[k] += valtmp*apa[cj[k]]; 648 } 649 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 650 } 651 652 /* put the value into Cd (diagonal part) */ 653 pnz = pd->i[i+1] - pd->i[i]; 654 pdJ = pd->j + pd->i[i]; 655 pA = pd->a + pd->i[i]; 656 for (j=0; j<pnz; j++) { 657 row = pdJ[j]; 658 cnz = bi[row+1] - bi[row]; 659 cj = bj + bi[row]; 660 ca = ba + bi[row]; 661 /* perform dense axpy */ 662 valtmp = pA[j]; 663 for (k=0; k<cnz; k++) { 664 ca[k] += valtmp*apa[cj[k]]; 665 } 666 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 667 } 668 669 /* zero the current row of A*P */ 670 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 671 #if defined(PTAP_PROFILE) 672 ierr = PetscGetTime(&t2_2);CHKERRQ(ierr); 673 et2_PtAP += t2_2 - t2_1; 674 #endif 675 } 676 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 677 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 678 /*-----------------------------------------------------------------------------------------*/ 679 pA=pa_loc; 680 for (i=0; i<am; i++) { 681 #if defined(PTAP_PROFILE) 682 ierr = PetscGetTime(&t2_0);CHKERRQ(ierr); 683 #endif 684 /* form i-th sparse row of A*P */ 685 apnz = api[i+1] - api[i]; 686 apJ = apj + api[i]; 687 /* diagonal portion of A */ 688 anz = adi[i+1] - adi[i]; 689 adj = ad->j + adi[i]; 690 ada = ad->a + adi[i]; 691 for (j=0; j<anz; j++) { 692 row = adj[j]; 693 pnz = pi_loc[row+1] - pi_loc[row]; 694 pj = pj_loc + pi_loc[row]; 695 pa = pa_loc + pi_loc[row]; 696 valtmp = ada[j]; 697 nextp = 0; 698 for (k=0; nextp<pnz; k++) { 699 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 700 apa[k] += valtmp*pa[nextp++]; 701 } 702 } 703 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 704 } 705 /* off-diagonal portion of A */ 706 anz = aoi[i+1] - aoi[i]; 707 aoj = ao->j + aoi[i]; 708 aoa = ao->a + aoi[i]; 709 for (j=0; j<anz; j++) { 710 row = aoj[j]; 711 pnz = pi_oth[row+1] - pi_oth[row]; 712 pj = pj_oth + pi_oth[row]; 713 pa = pa_oth + pi_oth[row]; 714 valtmp = aoa[j]; 715 nextp = 0; 716 for (k=0; nextp<pnz; k++) { 717 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 718 apa[k] += valtmp*pa[nextp++]; 719 } 720 } 721 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 722 } 723 #if defined(PTAP_PROFILE) 724 ierr = PetscGetTime(&t2_1);CHKERRQ(ierr); 725 et2_AP += t2_1 - t2_0; 726 #endif 727 728 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 729 /*--------------------------------------------------------------*/ 730 pnz = pi_loc[i+1] - pi_loc[i]; 731 pJ = pj_loc + pi_loc[i]; 732 for (j=0; j<pnz; j++) { 733 nextap = 0; 734 row = pJ[j]; /* global index */ 735 if (row < pcstart || row >=pcend) { /* put the value into Co */ 736 row = *poJ; 737 cj = coj + coi[row]; 738 ca = coa + coi[row]; poJ++; 739 } else { /* put the value into Cd */ 740 row = *pdJ; 741 cj = bj + bi[row]; 742 ca = ba + bi[row]; pdJ++; 743 } 744 valtmp = pA[j]; 745 for (k=0; nextap<apnz; k++) { 746 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 747 } 748 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 749 } 750 pA += pnz; 751 /* zero the current row info for A*P */ 752 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 753 #if defined(PTAP_PROFILE) 754 ierr = PetscGetTime(&t2_2);CHKERRQ(ierr); 755 et2_PtAP += t2_2 - t2_1; 756 #endif 757 } 758 } 759 #if defined(PTAP_PROFILE) 760 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 761 #endif 762 763 /* 3) send and recv matrix values coa */ 764 /*------------------------------------*/ 765 buf_ri = merge->buf_ri; 766 buf_rj = merge->buf_rj; 767 len_s = merge->len_s; 768 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 769 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 770 771 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 772 for (proc=0,k=0; proc<size; proc++) { 773 if (!len_s[proc]) continue; 774 i = merge->owners_co[proc]; 775 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 776 k++; 777 } 778 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 779 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 780 781 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 782 ierr = PetscFree(r_waits);CHKERRQ(ierr); 783 ierr = PetscFree(coa);CHKERRQ(ierr); 784 #if defined(PTAP_PROFILE) 785 ierr = PetscGetTime(&t3);CHKERRQ(ierr); 786 #endif 787 788 /* 4) insert local Cseq and received values into Cmpi */ 789 /*------------------------------------------------------*/ 790 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 791 for (k=0; k<merge->nrecv; k++) { 792 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 793 nrows = *(buf_ri_k[k]); 794 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 795 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 796 } 797 798 for (i=0; i<cm; i++) { 799 row = owners[rank] + i; /* global row index of C_seq */ 800 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 801 ba_i = ba + bi[i]; 802 bnz = bi[i+1] - bi[i]; 803 /* add received vals into ba */ 804 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 805 /* i-th row */ 806 if (i == *nextrow[k]) { 807 cnz = *(nextci[k]+1) - *nextci[k]; 808 cj = buf_rj[k] + *(nextci[k]); 809 ca = abuf_r[k] + *(nextci[k]); 810 nextcj = 0; 811 for (j=0; nextcj<cnz; j++) { 812 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 813 ba_i[j] += ca[nextcj++]; 814 } 815 } 816 nextrow[k]++; nextci[k]++; 817 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 818 } 819 } 820 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 821 } 822 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 823 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 824 825 ierr = PetscFree(ba);CHKERRQ(ierr); 826 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 827 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 828 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 829 #if defined(PTAP_PROFILE) 830 ierr = PetscGetTime(&t4);CHKERRQ(ierr); 831 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); 832 #endif 833 PetscFunctionReturn(0); 834 } 835