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 #undef __FUNCT__ 13 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ" 14 PetscErrorCode MatPtAPSymbolic_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 15 { 16 PetscErrorCode ierr; 17 18 PetscFunctionBegin; 19 if (!P->ops->ptapsymbolic_mpiaij) { 20 SETERRQ2(PETSC_ERR_SUP,"Not implemented for A=%s and P=%s",A->type_name,P->type_name); 21 } 22 ierr = (*P->ops->ptapsymbolic_mpiaij)(A,P,fill,C);CHKERRQ(ierr); 23 PetscFunctionReturn(0); 24 } 25 26 #undef __FUNCT__ 27 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ" 28 PetscErrorCode MatPtAPNumeric_MPIAIJ(Mat A,Mat P,Mat C) 29 { 30 PetscErrorCode ierr; 31 32 PetscFunctionBegin; 33 if (!P->ops->ptapnumeric_mpiaij) { 34 SETERRQ2(PETSC_ERR_SUP,"Not implemented for A=%s and P=%s",A->type_name,P->type_name); 35 } 36 ierr = (*P->ops->ptapnumeric_mpiaij)(A,P,C);CHKERRQ(ierr); 37 PetscFunctionReturn(0); 38 } 39 40 #undef __FUNCT__ 41 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 42 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 43 { 44 PetscErrorCode ierr; 45 Mat B_mpi; 46 Mat_MatMatMultMPI *ap; 47 PetscObjectContainer container; 48 FreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 49 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data; 50 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 51 Mat_SeqAIJ *p_loc,*p_oth; 52 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 53 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,nnz; 54 PetscInt nlnk,*lnk,*owners_co,*coi,*coj,i,k,pnz,row; 55 PetscInt am=A->m,pN=P->N,pn=P->n; 56 PetscBT lnkbt; 57 MPI_Comm comm=A->comm; 58 PetscMPIInt size,rank,tag,*len_si,*len_s,*len_ri; 59 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 60 PetscInt len,proc,*dnz,*onz,*owners; 61 PetscInt nzi,*bi,*bj; 62 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 63 MPI_Request *swaits,*rwaits; 64 MPI_Status *sstatus,rstatus; 65 Mat_Merge_SeqsToMPI *merge; 66 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon; 67 PetscMPIInt j; 68 69 PetscFunctionBegin; 70 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 71 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 72 73 /* destroy the container 'Mat_MatMatMultMPI' in case that P is attached to it */ 74 ierr = PetscObjectQuery((PetscObject)P,"Mat_MatMatMultMPI",(PetscObject *)&container);CHKERRQ(ierr); 75 if (container) { 76 ierr = PetscObjectContainerDestroy(container);CHKERRQ(ierr); 77 ierr = PetscObjectCompose((PetscObject)P,"Mat_MatMatMultMPI",0);CHKERRQ(ierr); 78 } 79 80 /* create the container 'Mat_MatMatMultMPI' and attach it to P */ 81 ierr = PetscNew(Mat_MatMatMultMPI,&ap);CHKERRQ(ierr); 82 ap->abi=PETSC_NULL; ap->abj=PETSC_NULL; 83 ap->abnz_max = 0; 84 85 ierr = PetscObjectContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 86 ierr = PetscObjectContainerSetPointer(container,ap);CHKERRQ(ierr); 87 ierr = PetscObjectCompose((PetscObject)P,"Mat_MatMatMultMPI",(PetscObject)container);CHKERRQ(ierr); 88 ap->MatDestroy = P->ops->destroy; 89 P->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 90 ap->reuse = MAT_INITIAL_MATRIX; 91 ierr = PetscObjectContainerSetUserDestroy(container,PetscObjectContainerDestroy_Mat_MatMatMultMPI);CHKERRQ(ierr); 92 93 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 94 ierr = MatGetBrowsOfAoCols(A,P,MAT_INITIAL_MATRIX,&ap->startsj,&ap->bufa,&ap->B_oth);CHKERRQ(ierr); 95 /* get P_loc by taking all local rows of P */ 96 ierr = MatGetLocalMat(P,MAT_INITIAL_MATRIX,&ap->B_loc);CHKERRQ(ierr); 97 98 p_loc = (Mat_SeqAIJ*)(ap->B_loc)->data; 99 p_oth = (Mat_SeqAIJ*)(ap->B_oth)->data; 100 pi_loc = p_loc->i; pj_loc = p_loc->j; 101 pi_oth = p_oth->i; pj_oth = p_oth->j; 102 103 /* first, compute symbolic AP = A_loc*P */ 104 /*--------------------------------------*/ 105 ierr = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr); 106 ap->abi = api; 107 api[0] = 0; 108 109 /* create and initialize a linked list */ 110 nlnk = pN+1; 111 ierr = PetscLLCreate(pN,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 112 113 /* Initial FreeSpace size is fill*nnz(A) */ 114 ierr = GetMoreSpace((PetscInt)(fill*(adi[am]+aoi[am])),&free_space);CHKERRQ(ierr); 115 current_space = free_space; 116 117 for (i=0;i<am;i++) { 118 apnz = 0; 119 /* diagonal portion of A */ 120 nzi = adi[i+1] - adi[i]; 121 for (j=0; j<nzi; j++){ 122 row = *adj++; 123 pnz = pi_loc[row+1] - pi_loc[row]; 124 Jptr = pj_loc + pi_loc[row]; 125 /* add non-zero cols of P into the sorted linked list lnk */ 126 ierr = PetscLLAdd(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 127 apnz += nlnk; 128 } 129 /* off-diagonal portion of A */ 130 nzi = aoi[i+1] - aoi[i]; 131 for (j=0; j<nzi; j++){ 132 row = *aoj++; 133 pnz = pi_oth[row+1] - pi_oth[row]; 134 Jptr = pj_oth + pi_oth[row]; 135 ierr = PetscLLAdd(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 136 apnz += nlnk; 137 } 138 139 api[i+1] = api[i] + apnz; 140 if (ap->abnz_max < apnz) ap->abnz_max = apnz; 141 142 /* if free space is not available, double the total space in the list */ 143 if (current_space->local_remaining<apnz) { 144 ierr = GetMoreSpace(current_space->total_array_size,¤t_space);CHKERRQ(ierr); 145 } 146 147 /* Copy data into free space, then initialize lnk */ 148 ierr = PetscLLClean(pN,pN,apnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 149 current_space->array += apnz; 150 current_space->local_used += apnz; 151 current_space->local_remaining -= apnz; 152 } 153 /* Allocate space for apj, initialize apj, and */ 154 /* destroy list of free space and other temporary array(s) */ 155 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ap->abj);CHKERRQ(ierr); 156 apj = ap->abj; 157 ierr = MakeSpaceContiguous(&free_space,ap->abj);CHKERRQ(ierr); 158 159 /* determine symbolic Co=(p->B)^T*AP - send to others */ 160 /*----------------------------------------------------*/ 161 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 162 163 /* then, compute symbolic Co = (p->B)^T*AP */ 164 pon = (p->B)->n; /* total num of rows to be sent to other processors 165 >= (num of nonzero rows of C_seq) - pn */ 166 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 167 coi[0] = 0; 168 169 /* set initial free space to be 3*pon*max( nnz(AP) per row) */ 170 nnz = 3*pon*ap->abnz_max + 1; 171 ierr = GetMoreSpace(nnz,&free_space); 172 current_space = free_space; 173 174 for (i=0; i<pon; i++) { 175 nnz = 0; 176 pnz = poti[i+1] - poti[i]; 177 j = pnz; 178 ptJ = potj + poti[i+1]; 179 while (j){/* assume cols are almost in increasing order, starting from its end saves computation */ 180 j--; ptJ--; 181 row = *ptJ; /* row of AP == col of Pot */ 182 apnz = api[row+1] - api[row]; 183 Jptr = apj + api[row]; 184 /* add non-zero cols of AP into the sorted linked list lnk */ 185 ierr = PetscLLAdd(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 186 nnz += nlnk; 187 } 188 189 /* If free space is not available, double the total space in the list */ 190 if (current_space->local_remaining<nnz) { 191 ierr = GetMoreSpace(current_space->total_array_size,¤t_space);CHKERRQ(ierr); 192 } 193 194 /* Copy data into free space, and zero out denserows */ 195 ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 196 current_space->array += nnz; 197 current_space->local_used += nnz; 198 current_space->local_remaining -= nnz; 199 coi[i+1] = coi[i] + nnz; 200 } 201 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 202 ierr = MakeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 203 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 204 205 /* send j-array (coj) of Co to other processors */ 206 /*----------------------------------------------*/ 207 /* determine row ownership */ 208 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 209 ierr = PetscMapCreate(comm,&merge->rowmap);CHKERRQ(ierr); 210 ierr = PetscMapSetLocalSize(merge->rowmap,pn);CHKERRQ(ierr); 211 ierr = PetscMapSetType(merge->rowmap,MAP_MPI);CHKERRQ(ierr); 212 ierr = PetscMapGetGlobalRange(merge->rowmap,&owners);CHKERRQ(ierr); 213 214 /* determine the number of messages to send, their lengths */ 215 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 216 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 217 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 218 len_s = merge->len_s; 219 merge->nsend = 0; 220 221 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 222 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 223 224 proc = 0; 225 for (i=0; i<pon; i++){ 226 while (prmap[i] >= owners[proc+1]) proc++; 227 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 228 len_s[proc] += coi[i+1] - coi[i]; 229 } 230 231 len = 0; /* max length of buf_si[] */ 232 owners_co[0] = 0; 233 for (proc=0; proc<size; proc++){ 234 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 235 if (len_si[proc]){ 236 merge->nsend++; 237 len_si[proc] = 2*(len_si[proc] + 1); 238 len += len_si[proc]; 239 } 240 } 241 242 /* determine the number and length of messages to receive for coi and coj */ 243 ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 244 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 245 246 /* post the Irecv and Isend of coj */ 247 ierr = PetscObjectGetNewTag((PetscObject)merge->rowmap,&tag);CHKERRQ(ierr); 248 ierr = PetscPostIrecvInt(comm,tag,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 249 250 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 251 252 for (proc=0, k=0; proc<size; proc++){ 253 if (!len_s[proc]) continue; 254 i = owners_co[proc]; 255 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tag,comm,swaits+k);CHKERRQ(ierr); 256 k++; 257 } 258 259 /* receives and sends of coj are complete */ 260 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 261 i = merge->nrecv; 262 while (i--) { 263 ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr); 264 } 265 ierr = PetscFree(rwaits);CHKERRQ(ierr); 266 if (merge->nsend){ 267 ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr); 268 } 269 270 /* send and recv coi */ 271 /*-------------------*/ 272 ierr = PetscPostIrecvInt(comm,tag,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 273 274 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 275 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 276 for (proc=0,k=0; proc<size; proc++){ 277 if (!len_s[proc]) continue; 278 /* form outgoing message for i-structure: 279 buf_si[0]: nrows to be sent 280 [1:nrows]: row index (global) 281 [nrows+1:2*nrows+1]: i-structure index 282 */ 283 /*-------------------------------------------*/ 284 nrows = len_si[proc]/2 - 1; 285 buf_si_i = buf_si + nrows+1; 286 buf_si[0] = nrows; 287 buf_si_i[0] = 0; 288 nrows = 0; 289 for (i=owners_co[proc]; i<owners_co[proc+1]; i++){ 290 nzi = coi[i+1] - coi[i]; 291 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 292 buf_si[nrows+1] =prmap[i] -owners[proc]; /* local row index */ 293 nrows++; 294 } 295 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tag,comm,swaits+k);CHKERRQ(ierr); 296 k++; 297 buf_si += len_si[proc]; 298 } 299 i = merge->nrecv; 300 while (i--) { 301 ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr); 302 } 303 ierr = PetscFree(rwaits);CHKERRQ(ierr); 304 if (merge->nsend){ 305 ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr); 306 } 307 308 ierr = PetscLogInfo((PetscObject)A,"MatMerge_SeqsToMPI: nsend: %d, nrecv: %d\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 309 for (i=0; i<merge->nrecv; i++){ 310 ierr = PetscLogInfo((PetscObject)A,"MatMerge_SeqsToMPI: recv len_ri=%d, len_rj=%d from [%d]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);CHKERRQ(ierr); 311 } 312 313 ierr = PetscFree(len_si);CHKERRQ(ierr); 314 ierr = PetscFree(len_ri);CHKERRQ(ierr); 315 ierr = PetscFree(swaits);CHKERRQ(ierr); 316 ierr = PetscFree(sstatus);CHKERRQ(ierr); 317 ierr = PetscFree(buf_s);CHKERRQ(ierr); 318 319 /* compute the local portion of C (mpi mat) */ 320 /*------------------------------------------*/ 321 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 322 323 /* allocate bi array and free space for accumulating nonzero column info */ 324 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 325 bi[0] = 0; 326 327 /* set initial free space to be 3*pn*max( nnz(AP) per row) */ 328 nnz = 3*pn*ap->abnz_max + 1; 329 ierr = GetMoreSpace(nnz,&free_space); 330 current_space = free_space; 331 332 ierr = PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);CHKERRQ(ierr); 333 nextrow = buf_ri_k + merge->nrecv; 334 nextci = nextrow + merge->nrecv; 335 for (k=0; k<merge->nrecv; k++){ 336 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 337 nrows = *buf_ri_k[k]; 338 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 339 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 340 } 341 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 342 for (i=0; i<pn; i++) { 343 /* add pdt[i,:]*AP into lnk */ 344 nnz = 0; 345 pnz = pdti[i+1] - pdti[i]; 346 j = pnz; 347 ptJ = pdtj + pdti[i+1]; 348 while (j){/* assume cols are almost in increasing order, starting from its end saves computation */ 349 j--; ptJ--; 350 row = *ptJ; /* row of AP == col of Pt */ 351 apnz = api[row+1] - api[row]; 352 Jptr = apj + api[row]; 353 /* add non-zero cols of AP into the sorted linked list lnk */ 354 ierr = PetscLLAdd(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 355 nnz += nlnk; 356 } 357 /* add received col data into lnk */ 358 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 359 if (i == *nextrow[k]) { /* i-th row */ 360 nzi = *(nextci[k]+1) - *nextci[k]; 361 Jptr = buf_rj[k] + *nextci[k]; 362 ierr = PetscLLAdd(nzi,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 363 nnz += nlnk; 364 nextrow[k]++; nextci[k]++; 365 } 366 } 367 368 /* if free space is not available, make more free space */ 369 if (current_space->local_remaining<nnz) { 370 ierr = GetMoreSpace(current_space->total_array_size,¤t_space);CHKERRQ(ierr); 371 } 372 /* copy data into free space, then initialize lnk */ 373 ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 374 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 375 current_space->array += nnz; 376 current_space->local_used += nnz; 377 current_space->local_remaining -= nnz; 378 bi[i+1] = bi[i] + nnz; 379 } 380 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 381 ierr = PetscFree(buf_ri_k);CHKERRQ(ierr); 382 383 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 384 ierr = MakeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 385 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 386 387 /* create symbolic parallel matrix B_mpi */ 388 /*---------------------------------------*/ 389 ierr = MatCreate(comm,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE,&B_mpi);CHKERRQ(ierr); 390 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 391 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 392 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 393 394 /* B_mpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 395 B_mpi->assembled = PETSC_FALSE; 396 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 397 merge->bi = bi; 398 merge->bj = bj; 399 merge->coi = coi; 400 merge->coj = coj; 401 merge->buf_ri = buf_ri; 402 merge->buf_rj = buf_rj; 403 merge->owners_co = owners_co; 404 405 /* attach the supporting struct to B_mpi for reuse */ 406 ierr = PetscObjectContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 407 ierr = PetscObjectContainerSetPointer(container,merge);CHKERRQ(ierr); 408 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 409 *C = B_mpi; 410 PetscFunctionReturn(0); 411 } 412 413 #undef __FUNCT__ 414 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 415 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 416 { 417 PetscErrorCode ierr; 418 Mat_Merge_SeqsToMPI *merge; 419 Mat_MatMatMultMPI *ap; 420 PetscObjectContainer cont_merge,cont_ptap; 421 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data; 422 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 423 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 424 Mat_SeqAIJ *p_loc,*p_oth; 425 PetscInt *adi=ad->i,*aoi=ao->i,*adj=ad->j,*aoj=ao->j,*apJ,nextp,flops=0; 426 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 427 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 428 MatScalar *ada=ad->a,*aoa=ao->a,*apa,*pa,*ca,*pa_loc,*pa_oth; 429 PetscInt am=A->m,cm=C->m,pon=(p->B)->n; 430 MPI_Comm comm=C->comm; 431 PetscMPIInt size,rank,taga,*len_s; 432 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 433 PetscInt **buf_ri,**buf_rj; 434 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 435 MPI_Request *s_waits,*r_waits; 436 MPI_Status *status; 437 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 438 PetscInt *api,*apj,*coi,*coj; 439 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=p->cstart,pcend=p->cend; 440 441 PetscFunctionBegin; 442 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 443 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 444 445 ierr = PetscObjectQuery((PetscObject)C,"MatMergeSeqsToMPI",(PetscObject *)&cont_merge);CHKERRQ(ierr); 446 if (cont_merge) { 447 ierr = PetscObjectContainerGetPointer(cont_merge,(void **)&merge);CHKERRQ(ierr); 448 } else { 449 SETERRQ(PETSC_ERR_ARG_WRONGSTATE, "Matrix C does not posses an object container"); 450 } 451 452 ierr = PetscObjectQuery((PetscObject)P,"Mat_MatMatMultMPI",(PetscObject *)&cont_ptap);CHKERRQ(ierr); 453 if (cont_ptap) { 454 ierr = PetscObjectContainerGetPointer(cont_ptap,(void **)&ap);CHKERRQ(ierr); 455 if (ap->reuse == MAT_INITIAL_MATRIX){ 456 ap->reuse = MAT_REUSE_MATRIX; 457 } else { /* update numerical values of P_oth and P_loc */ 458 ierr = MatGetBrowsOfAoCols(A,P,MAT_REUSE_MATRIX,&ap->startsj,&ap->bufa,&ap->B_oth);CHKERRQ(ierr); 459 ierr = MatGetLocalMat(P,MAT_REUSE_MATRIX,&ap->B_loc);CHKERRQ(ierr); 460 } 461 } else { 462 SETERRQ(PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 463 } 464 465 /* get data from symbolic products */ 466 p_loc = (Mat_SeqAIJ*)(ap->B_loc)->data; 467 p_oth = (Mat_SeqAIJ*)(ap->B_oth)->data; 468 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a,pA=pa_loc; 469 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 470 471 coi = merge->coi; coj = merge->coj; 472 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 473 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 474 475 bi = merge->bi; bj = merge->bj; 476 ierr = PetscMapGetGlobalRange(merge->rowmap,&owners);CHKERRQ(ierr); 477 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 478 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 479 480 /* get data from symbolic A*P */ 481 ierr = PetscMalloc((ap->abnz_max+1)*sizeof(MatScalar),&apa);CHKERRQ(ierr); 482 ierr = PetscMemzero(apa,ap->abnz_max*sizeof(MatScalar));CHKERRQ(ierr); 483 484 /* compute numeric C_seq=P_loc^T*A_loc*P */ 485 api = ap->abi; apj = ap->abj; 486 for (i=0;i<am;i++) { 487 /* form i-th sparse row of A*P */ 488 apnz = api[i+1] - api[i]; 489 apJ = apj + api[i]; 490 /* diagonal portion of A */ 491 anz = adi[i+1] - adi[i]; 492 for (j=0;j<anz;j++) { 493 row = *adj++; 494 pnz = pi_loc[row+1] - pi_loc[row]; 495 pj = pj_loc + pi_loc[row]; 496 pa = pa_loc + pi_loc[row]; 497 nextp = 0; 498 for (k=0; nextp<pnz; k++) { 499 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 500 apa[k] += (*ada)*pa[nextp++]; 501 } 502 } 503 flops += 2*pnz; 504 ada++; 505 } 506 /* off-diagonal portion of A */ 507 anz = aoi[i+1] - aoi[i]; 508 for (j=0; j<anz; j++) { 509 row = *aoj++; 510 pnz = pi_oth[row+1] - pi_oth[row]; 511 pj = pj_oth + pi_oth[row]; 512 pa = pa_oth + pi_oth[row]; 513 nextp = 0; 514 for (k=0; nextp<pnz; k++) { 515 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 516 apa[k] += (*aoa)*pa[nextp++]; 517 } 518 } 519 flops += 2*pnz; 520 aoa++; 521 } 522 523 /* Compute P_loc[i,:]^T*AP[i,:] using outer product */ 524 pnz = pi_loc[i+1] - pi_loc[i]; 525 for (j=0; j<pnz; j++) { 526 nextap = 0; 527 row = *pJ++; /* global index */ 528 if (row < pcstart || row >=pcend) { /* put the value into Co */ 529 cj = coj + coi[*poJ]; 530 ca = coa + coi[*poJ++]; 531 } else { /* put the value into Cd */ 532 cj = bj + bi[*pdJ]; 533 ca = ba + bi[*pdJ++]; 534 } 535 for (k=0; nextap<apnz; k++) { 536 if (cj[k]==apJ[nextap]) ca[k] += (*pA)*apa[nextap++]; 537 } 538 flops += 2*apnz; 539 pA++; 540 } 541 542 /* zero the current row info for A*P */ 543 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 544 } 545 ierr = PetscFree(apa);CHKERRQ(ierr); 546 547 /* send and recv matrix values */ 548 /*-----------------------------*/ 549 buf_ri = merge->buf_ri; 550 buf_rj = merge->buf_rj; 551 len_s = merge->len_s; 552 ierr = PetscObjectGetNewTag((PetscObject)merge->rowmap,&taga);CHKERRQ(ierr); 553 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 554 555 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&s_waits);CHKERRQ(ierr); 556 for (proc=0,k=0; proc<size; proc++){ 557 if (!len_s[proc]) continue; 558 i = merge->owners_co[proc]; 559 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 560 k++; 561 } 562 ierr = PetscMalloc(size*sizeof(MPI_Status),&status);CHKERRQ(ierr); 563 ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr); 564 ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr); 565 ierr = PetscFree(status);CHKERRQ(ierr); 566 567 ierr = PetscFree(s_waits);CHKERRQ(ierr); 568 ierr = PetscFree(r_waits);CHKERRQ(ierr); 569 ierr = PetscFree(coa);CHKERRQ(ierr); 570 571 /* insert local and received values into C */ 572 /*-----------------------------------------*/ 573 ierr = PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);CHKERRQ(ierr); 574 nextrow = buf_ri_k + merge->nrecv; 575 nextci = nextrow + merge->nrecv; 576 577 for (k=0; k<merge->nrecv; k++){ 578 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 579 nrows = *(buf_ri_k[k]); 580 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 581 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 582 } 583 584 for (i=0; i<cm; i++) { 585 row = owners[rank] + i; /* global row index of C_seq */ 586 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 587 ba_i = ba + bi[i]; 588 bnz = bi[i+1] - bi[i]; 589 /* add received vals into ba */ 590 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 591 /* i-th row */ 592 if (i == *nextrow[k]) { 593 cnz = *(nextci[k]+1) - *nextci[k]; 594 cj = buf_rj[k] + *(nextci[k]); 595 ca = abuf_r[k] + *(nextci[k]); 596 nextcj = 0; 597 for (j=0; nextcj<cnz; j++){ 598 if (bj_i[j] == cj[nextcj]){ /* bcol == ccol */ 599 ba_i[j] += ca[nextcj++]; 600 } 601 } 602 nextrow[k]++; nextci[k]++; 603 } 604 } 605 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 606 flops += 2*cnz; 607 } 608 ierr = MatSetOption(C,MAT_COLUMNS_SORTED);CHKERRQ(ierr); /* -cause delay? */ 609 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 610 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 611 612 ierr = PetscFree(ba);CHKERRQ(ierr); 613 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 614 ierr = PetscFree(buf_ri_k);CHKERRQ(ierr); 615 ierr = PetscLogFlops(flops);CHKERRQ(ierr); 616 PetscFunctionReturn(0); 617 } 618