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 /* 13 #define DEBUG_MATPTAP 14 */ 15 16 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 17 #undef __FUNCT__ 18 #define __FUNCT__ "MatDestroy_MPIAIJ_PtAP" 19 PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat A) 20 { 21 PetscErrorCode ierr; 22 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 23 Mat_PtAPMPI *ptap=a->ptap; 24 25 PetscFunctionBegin; 26 if (ptap){ 27 ierr = PetscFree2(ptap->startsj,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 = PetscFree(ptap->api);CHKERRQ(ierr); 32 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 33 } 34 if (ptap->merge) { 35 Mat_Merge_SeqsToMPI *merge=ptap->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 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_Merge_SeqsToMPI *merge; 62 PetscContainer container; 63 64 PetscFunctionBegin; 65 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject *)&container);CHKERRQ(ierr); 66 if (container) { 67 ierr = PetscContainerGetPointer(container,(void **)&merge);CHKERRQ(ierr); 68 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exit"); 69 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 70 (*M)->ops->destroy = merge->destroy; /* =MatDestroy_MPIAIJ, *M doesn't duplicate A's container! */ 71 (*M)->ops->duplicate = merge->duplicate; /* =MatDuplicate_ MPIAIJ */ 72 PetscFunctionReturn(0); 73 } 74 75 #undef __FUNCT__ 76 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ" 77 PetscErrorCode MatPtAPSymbolic_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 78 { 79 PetscErrorCode ierr; 80 81 PetscFunctionBegin; 82 if (!P->ops->ptapsymbolic_mpiaij) SETERRQ2(((PetscObject)A)->comm,PETSC_ERR_SUP,"Not implemented for A=%s and P=%s",((PetscObject)A)->type_name,((PetscObject)P)->type_name); 83 ierr = (*P->ops->ptapsymbolic_mpiaij)(A,P,fill,C);CHKERRQ(ierr); 84 PetscFunctionReturn(0); 85 } 86 87 #undef __FUNCT__ 88 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ" 89 PetscErrorCode MatPtAPNumeric_MPIAIJ(Mat A,Mat P,Mat C) 90 { 91 PetscErrorCode ierr; 92 93 PetscFunctionBegin; 94 if (!P->ops->ptapnumeric_mpiaij) SETERRQ2(((PetscObject)A)->comm,PETSC_ERR_SUP,"Not implemented for A=%s and P=%s",((PetscObject)A)->type_name,((PetscObject)P)->type_name); 95 ierr = (*P->ops->ptapnumeric_mpiaij)(A,P,C);CHKERRQ(ierr); 96 PetscFunctionReturn(0); 97 } 98 99 #undef __FUNCT__ 100 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 101 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 102 { 103 PetscErrorCode ierr; 104 Mat Cmpi; 105 Mat_PtAPMPI *ptap; 106 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 107 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 108 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 109 Mat_SeqAIJ *p_loc,*p_oth; 110 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 111 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 112 PetscInt nlnk,*lnk,*owners_co,*coi,*coj,i,k,pnz,row; 113 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 114 PetscBT lnkbt; 115 MPI_Comm comm=((PetscObject)A)->comm; 116 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 117 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 118 PetscInt len,proc,*dnz,*onz,*owners; 119 PetscInt nzi,*bi,*bj; 120 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 121 MPI_Request *swaits,*rwaits; 122 MPI_Status *sstatus,rstatus; 123 Mat_Merge_SeqsToMPI *merge; 124 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 125 PetscReal afill=1.0,afill_tmp; 126 PetscLogDouble t0,tf,etime=0.0,t00,tff,time_matupdate=0.0,time_Cd=0.0,time_AP=0.0,time_Co=0.0; 127 PetscInt rstart = P->cmap->rstart,rmax; 128 PetscScalar *vals; 129 130 PetscFunctionBegin; 131 ierr = PetscGetTime(&t00);CHKERRQ(ierr); 132 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 133 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 134 135 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 136 /* create struct Mat_PtAPMPI and attached it to C later */ 137 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 138 ptap->reuse = MAT_INITIAL_MATRIX; 139 140 141 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 142 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 143 #if defined(DEBUG_MATPTAP) 144 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_oth is done...\n",rank); 145 #endif 146 147 /* get P_loc by taking all local rows of P */ 148 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 149 #if defined(DEBUG_MATPTAP) 150 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 151 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_loc is done...\n",rank); 152 #endif 153 154 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 155 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 156 pi_loc = p_loc->i; pj_loc = p_loc->j; 157 pi_oth = p_oth->i; pj_oth = p_oth->j; 158 159 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 160 time_matupdate += tf-t0; 161 162 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 163 /*-------------------------------------------------------------------*/ 164 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 165 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&api);CHKERRQ(ierr); 166 api[0] = 0; 167 168 /* create and initialize a linked list */ 169 nlnk = pN+1; 170 #if defined(DEBUG_MATPTAP) 171 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 172 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call PetscLLCreate(), pN %D, P_rmax %D %D; Annz %D\n",rank,pN,p_loc->rmax,p_oth->rmax,adi[am]+aoi[am]); 173 #endif 174 ierr = PetscLLCreate(pN,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 175 #if defined(DEBUG_MATPTAP) 176 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 177 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call PetscLLCreate() is done\n",rank); 178 #endif 179 180 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) */ 181 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 182 current_space = free_space; 183 184 for (i=0; i<am; i++) { 185 apnz = 0; 186 /* diagonal portion of A */ 187 nzi = adi[i+1] - adi[i]; 188 aj = ad->j + adi[i]; 189 for (j=0; j<nzi; j++){ 190 row = aj[j]; 191 pnz = pi_loc[row+1] - pi_loc[row]; 192 Jptr = pj_loc + pi_loc[row]; 193 /* add non-zero cols of P into the sorted linked list lnk */ 194 ierr = PetscLLAddSorted(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 195 apnz += nlnk; 196 } 197 /* off-diagonal portion of A */ 198 nzi = aoi[i+1] - aoi[i]; 199 aj = ao->j + aoi[i]; 200 for (j=0; j<nzi; j++){ 201 row = aj[j]; 202 pnz = pi_oth[row+1] - pi_oth[row]; 203 Jptr = pj_oth + pi_oth[row]; 204 ierr = PetscLLAddSorted(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 205 apnz += nlnk; 206 } 207 208 api[i+1] = api[i] + apnz; 209 if (ap_rmax < apnz) ap_rmax = apnz; 210 211 /* if free space is not available, double the total space in the list */ 212 if (current_space->local_remaining<apnz) { 213 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 214 nspacedouble++; 215 } 216 217 /* Copy data into free space, then initialize lnk */ 218 ierr = PetscLLClean(pN,pN,apnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 219 current_space->array += apnz; 220 current_space->local_used += apnz; 221 current_space->local_remaining -= apnz; 222 } 223 /* Allocate space for apj, initialize apj, and */ 224 /* destroy list of free space and other temporary array(s) */ 225 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&apj);CHKERRQ(ierr); 226 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 227 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]); 228 if (afill_tmp > afill) afill = afill_tmp; 229 230 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 231 time_AP += tf-t0; 232 #if defined(DEBUG_MATPTAP) 233 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 234 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] AP is done\n",rank); 235 #endif 236 237 /* determine symbolic Co=(p->B)^T*AP - send to others */ 238 /*----------------------------------------------------*/ 239 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 240 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 241 242 /* then, compute symbolic Co = (p->B)^T*AP */ 243 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 244 >= (num of nonzero rows of C_seq) - pn */ 245 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 246 coi[0] = 0; 247 248 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 249 nnz = fill*(poti[pon] + api[am]); 250 ierr = PetscFreeSpaceGet(nnz,&free_space); 251 current_space = free_space; 252 253 for (i=0; i<pon; i++) { 254 nnz = 0; 255 pnz = poti[i+1] - poti[i]; 256 ptJ = potj + poti[i]; 257 for (j=0; j<pnz; j++){ 258 row = ptJ[j]; /* row of AP == col of Pot */ 259 apnz = api[row+1] - api[row]; 260 Jptr = apj + api[row]; 261 /* add non-zero cols of AP into the sorted linked list lnk */ 262 ierr = PetscLLAddSorted(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 263 nnz += nlnk; 264 } 265 266 /* If free space is not available, double the total space in the list */ 267 if (current_space->local_remaining<nnz) { 268 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 269 nspacedouble++; 270 } 271 272 /* Copy data into free space, and zero out denserows */ 273 ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 274 current_space->array += nnz; 275 current_space->local_used += nnz; 276 current_space->local_remaining -= nnz; 277 coi[i+1] = coi[i] + nnz; 278 } 279 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 280 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 281 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]); 282 if (afill_tmp > afill) afill = afill_tmp; 283 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 284 285 #if defined(DEBUG_MATPTAP) 286 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 287 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] local Co is done\n",rank); 288 #endif 289 290 /* send j-array (coj) of Co to other processors */ 291 /*----------------------------------------------*/ 292 /* determine row ownership */ 293 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 294 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 295 merge->rowmap->n = pn; 296 merge->rowmap->bs = 1; 297 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 298 owners = merge->rowmap->range; 299 300 /* determine the number of messages to send, their lengths */ 301 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 302 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 303 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 304 len_s = merge->len_s; 305 merge->nsend = 0; 306 307 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 308 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 309 310 proc = 0; 311 for (i=0; i<pon; i++){ 312 while (prmap[i] >= owners[proc+1]) proc++; 313 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 314 len_s[proc] += coi[i+1] - coi[i]; 315 } 316 317 len = 0; /* max length of buf_si[] */ 318 owners_co[0] = 0; 319 for (proc=0; proc<size; proc++){ 320 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 321 if (len_si[proc]){ 322 merge->nsend++; 323 len_si[proc] = 2*(len_si[proc] + 1); 324 len += len_si[proc]; 325 } 326 } 327 328 /* determine the number and length of messages to receive for coi and coj */ 329 ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 330 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 331 332 /* post the Irecv and Isend of coj */ 333 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 334 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 335 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 336 for (proc=0, k=0; proc<size; proc++){ 337 if (!len_s[proc]) continue; 338 i = owners_co[proc]; 339 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 340 k++; 341 } 342 343 /* receives and sends of coj are complete */ 344 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 345 for (i=0; i<merge->nrecv; i++){ 346 ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr); 347 } 348 ierr = PetscFree(rwaits);CHKERRQ(ierr); 349 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 350 351 /* send and recv coi */ 352 /*-------------------*/ 353 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 354 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 355 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 356 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 357 for (proc=0,k=0; proc<size; proc++){ 358 if (!len_s[proc]) continue; 359 /* form outgoing message for i-structure: 360 buf_si[0]: nrows to be sent 361 [1:nrows]: row index (global) 362 [nrows+1:2*nrows+1]: i-structure index 363 */ 364 /*-------------------------------------------*/ 365 nrows = len_si[proc]/2 - 1; 366 buf_si_i = buf_si + nrows+1; 367 buf_si[0] = nrows; 368 buf_si_i[0] = 0; 369 nrows = 0; 370 for (i=owners_co[proc]; i<owners_co[proc+1]; i++){ 371 nzi = coi[i+1] - coi[i]; 372 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 373 buf_si[nrows+1] =prmap[i] -owners[proc]; /* local row index */ 374 nrows++; 375 } 376 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 377 k++; 378 buf_si += len_si[proc]; 379 } 380 i = merge->nrecv; 381 while (i--) { 382 ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr); 383 } 384 ierr = PetscFree(rwaits);CHKERRQ(ierr); 385 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 386 /* 387 ierr = PetscInfo2(A,"nsend: %d, nrecv: %d\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 388 for (i=0; i<merge->nrecv; i++){ 389 ierr = PetscInfo3(A,"recv len_ri=%d, len_rj=%d from [%d]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);CHKERRQ(ierr); 390 } 391 */ 392 ierr = PetscFree(len_si);CHKERRQ(ierr); 393 ierr = PetscFree(len_ri);CHKERRQ(ierr); 394 ierr = PetscFree(swaits);CHKERRQ(ierr); 395 ierr = PetscFree(sstatus);CHKERRQ(ierr); 396 ierr = PetscFree(buf_s);CHKERRQ(ierr); 397 398 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 399 time_Co += tf-t0; 400 401 #if defined(DEBUG_MATPTAP) 402 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 403 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] Co is done\n",rank); 404 #endif 405 406 /* compute the local portion of C (mpi mat) */ 407 /*------------------------------------------*/ 408 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 409 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 410 411 /* allocate bi array and free space for accumulating nonzero column info */ 412 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 413 bi[0] = 0; 414 415 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 416 nnz = fill*(pi_loc[pm] + api[am]); 417 ierr = PetscFreeSpaceGet(nnz,&free_space); 418 current_space = free_space; 419 420 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 421 for (k=0; k<merge->nrecv; k++){ 422 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 423 nrows = *buf_ri_k[k]; 424 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 425 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 426 } 427 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 428 rmax = 0; 429 for (i=0; i<pn; i++) { 430 /* add pdt[i,:]*AP into lnk */ 431 nnz = 0; 432 pnz = pdti[i+1] - pdti[i]; 433 ptJ = pdtj + pdti[i]; 434 for (j=0; j<pnz; j++){ 435 row = ptJ[j]; /* row of AP == col of Pt */ 436 apnz = api[row+1] - api[row]; 437 Jptr = apj + api[row]; 438 /* add non-zero cols of AP into the sorted linked list lnk */ 439 ierr = PetscLLAddSorted(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 440 nnz += nlnk; 441 } 442 /* add received col data into lnk */ 443 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 444 if (i == *nextrow[k]) { /* i-th row */ 445 nzi = *(nextci[k]+1) - *nextci[k]; 446 Jptr = buf_rj[k] + *nextci[k]; 447 ierr = PetscLLAddSorted(nzi,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 448 nnz += nlnk; 449 nextrow[k]++; nextci[k]++; 450 } 451 } 452 453 /* if free space is not available, make more free space */ 454 if (current_space->local_remaining<nnz) { 455 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 456 nspacedouble++; 457 } 458 /* copy data into free space, then initialize lnk */ 459 ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 460 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 461 current_space->array += nnz; 462 current_space->local_used += nnz; 463 current_space->local_remaining -= nnz; 464 bi[i+1] = bi[i] + nnz; 465 if (nnz > rmax) rmax = nnz; 466 } 467 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 468 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 469 470 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 471 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 472 afill_tmp = (PetscReal)bi[pn]/(pi_loc[pm] + api[am]); 473 if (afill_tmp > afill) afill = afill_tmp; 474 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 475 476 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 477 time_Cd += tf-t0; 478 479 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part - check runex55_SA ? */ 480 /*------------------------------------------------------------------------------------------------------*/ 481 ierr = PetscMalloc((rmax+1)*sizeof(PetscScalar),&vals);CHKERRQ(ierr); 482 ierr = PetscMemzero(vals,rmax*sizeof(PetscScalar));CHKERRQ(ierr); 483 /* 484 ierr = PetscSynchronizedPrintf(comm, "[%d] pn %D, pN %D, cstart %D, rmax %D\n", rank,pn,pN,rstart,rmax);CHKERRQ(ierr); 485 ierr = PetscSynchronizedFlush(comm);CHKERRQ(ierr); 486 */ 487 488 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 489 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 490 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 491 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 492 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 493 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 494 for (i=0; i<pn; i++){ 495 row = i + rstart; 496 nnz = bi[i+1] - bi[i]; 497 Jptr = bj + bi[i]; 498 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 499 } 500 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 501 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 502 ierr = PetscFree(vals);CHKERRQ(ierr); 503 504 merge->bi = bi; 505 merge->bj = bj; 506 merge->coi = coi; 507 merge->coj = coj; 508 merge->buf_ri = buf_ri; 509 merge->buf_rj = buf_rj; 510 merge->owners_co = owners_co; 511 merge->destroy = Cmpi->ops->destroy; 512 merge->duplicate = Cmpi->ops->duplicate; 513 514 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 515 /* Cmpi->assembled = PETSC_FALSE; */ 516 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 517 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 518 519 /* attach the supporting struct to Cmpi for reuse */ 520 c = (Mat_MPIAIJ*)Cmpi->data; 521 c->ptap = ptap; 522 ptap->api = api; 523 ptap->apj = apj; 524 ptap->merge = merge; 525 ptap->abnz_max = ap_rmax; 526 527 *C = Cmpi; 528 #if defined(PETSC_USE_INFO) 529 if (bi[pn] != 0) { 530 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 531 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 532 } else { 533 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 534 } 535 #endif 536 ierr = PetscGetTime(&tff);CHKERRQ(ierr); 537 etime += tff - t00; 538 /* 539 PetscInt prid=0; 540 if (rank == prid){ 541 ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] PtAPSym time %g = matsetup %g + AP %g + Co %g + Cd %g\n",rank,etime,time_matupdate,time_AP,time_Co,time_Cd); 542 } 543 */ 544 #if defined(DEBUG_MATPTAP) 545 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 546 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] exit MatPtAPSymbolic_MPIAIJ_MPIAIJ\n",rank); 547 #endif 548 PetscFunctionReturn(0); 549 } 550 551 #undef __FUNCT__ 552 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 553 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 554 { 555 PetscErrorCode ierr; 556 Mat_Merge_SeqsToMPI *merge; 557 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 558 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 559 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 560 Mat_SeqAIJ *p_loc,*p_oth; 561 Mat_PtAPMPI *ptap; 562 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 563 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 564 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 565 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 566 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 567 MPI_Comm comm=((PetscObject)C)->comm; 568 PetscMPIInt size,rank,taga,*len_s; 569 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 570 PetscInt **buf_ri,**buf_rj; 571 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 572 MPI_Request *s_waits,*r_waits; 573 MPI_Status *status; 574 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 575 PetscInt *api,*apj,*coi,*coj; 576 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 577 PetscInt sparse_axpy; 578 PetscLogDouble t0,tf,etime=0.0,t00,tff,time_matupdate=0.0,time_malloc=0.0,time_Cseq0=0.0,time_Cseq1=0.0,time_setvals=0.0; 579 580 PetscFunctionBegin; 581 /* ierr = MPI_Barrier(comm);CHKERRQ(ierr); */ 582 ierr = PetscGetTime(&t00);CHKERRQ(ierr); 583 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 584 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 585 586 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 587 ptap = c->ptap; 588 merge = ptap->merge; 589 590 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 591 /*--------------------------------------------------*/ 592 if (ptap->reuse == MAT_INITIAL_MATRIX){ 593 ptap->reuse = MAT_REUSE_MATRIX; 594 } else { /* update numerical values of P_oth and P_loc */ 595 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 596 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 597 } 598 599 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 600 time_matupdate += tf-t0; 601 602 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 603 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 604 /*--------------------------------------------------------------*/ 605 /* get data from symbolic products */ 606 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 607 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 608 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 609 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 610 611 coi = merge->coi; coj = merge->coj; 612 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 613 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 614 615 bi = merge->bi; bj = merge->bj; 616 owners = merge->rowmap->range; 617 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 618 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 619 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 620 time_malloc += tf-t0; 621 622 api = ptap->api; apj = ptap->apj; 623 /* flag 'sparse_axpy' determines which implementations to be used: 624 0: do dense axpy in MatPtAPNumeric() - fastest, but requires storage of a dense array apa; (default) 625 1: do one sparse axpy - uses same memory as sparse_axpy=0 and might execute less flops 626 (apnz vs. cnz in the outerproduct), slower than case '0' when cnz is not too large than apnz; 627 2: do two sparse axpy in MatPtAPNumeric() - slowest, uses a sparse array apa */ 628 /* set default sparse_axpy */ 629 sparse_axpy = 0; 630 ierr = PetscOptionsGetInt(PETSC_NULL,"-matptap_sparseaxpy",&sparse_axpy,PETSC_NULL);CHKERRQ(ierr); 631 if (sparse_axpy == 0){ /* Do not perform sparse axpy */ 632 /*--------------------------------------------------*/ 633 /* malloc apa to store dense row A[i,:]*P */ 634 ierr = PetscMalloc((P->cmap->N)*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 635 ierr = PetscMemzero(apa,P->cmap->N*sizeof(PetscScalar));CHKERRQ(ierr); 636 637 for (i=0; i<am; i++) { 638 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 639 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 640 /*------------------------------------------------------------*/ 641 apJ = apj + api[i]; 642 643 /* diagonal portion of A */ 644 anz = adi[i+1] - adi[i]; 645 adj = ad->j + adi[i]; 646 ada = ad->a + adi[i]; 647 for (j=0; j<anz; j++) { 648 row = adj[j]; 649 pnz = pi_loc[row+1] - pi_loc[row]; 650 pj = pj_loc + pi_loc[row]; 651 pa = pa_loc + pi_loc[row]; 652 653 /* perform dense axpy */ 654 valtmp = ada[j]; 655 for (k=0; k<pnz; k++){ 656 apa[pj[k]] += valtmp*pa[k]; 657 } 658 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 659 } 660 661 /* off-diagonal portion of A */ 662 anz = aoi[i+1] - aoi[i]; 663 aoj = ao->j + aoi[i]; 664 aoa = ao->a + aoi[i]; 665 for (j=0; j<anz; j++) { 666 row = aoj[j]; 667 pnz = pi_oth[row+1] - pi_oth[row]; 668 pj = pj_oth + pi_oth[row]; 669 pa = pa_oth + pi_oth[row]; 670 671 /* perform dense axpy */ 672 valtmp = aoa[j]; 673 for (k=0; k<pnz; k++){ 674 apa[pj[k]] += valtmp*pa[k]; 675 } 676 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 677 } 678 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 679 time_Cseq0 += tf - t0; 680 681 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 682 /*--------------------------------------------------------------*/ 683 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 684 apnz = api[i+1] - api[i]; 685 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 686 pnz = po->i[i+1] - po->i[i]; 687 poJ = po->j + po->i[i]; 688 pA = po->a + po->i[i]; 689 for (j=0; j<pnz; j++){ 690 row = poJ[j]; 691 cnz = coi[row+1] - coi[row]; 692 cj = coj + coi[row]; 693 ca = coa + coi[row]; 694 /* perform dense axpy */ 695 valtmp = pA[j]; 696 for (k=0; k<cnz; k++) { 697 ca[k] += valtmp*apa[cj[k]]; 698 } 699 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 700 } 701 702 /* put the value into Cd (diagonal part) */ 703 pnz = pd->i[i+1] - pd->i[i]; 704 pdJ = pd->j + pd->i[i]; 705 pA = pd->a + pd->i[i]; 706 for (j=0; j<pnz; j++){ 707 row = pdJ[j]; 708 cnz = bi[row+1] - bi[row]; 709 cj = bj + bi[row]; 710 ca = ba + bi[row]; 711 /* perform dense axpy */ 712 valtmp = pA[j]; 713 for (k=0; k<cnz; k++) { 714 ca[k] += valtmp*apa[cj[k]]; 715 } 716 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 717 } 718 719 /* zero the current row of A*P */ 720 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 721 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 722 time_Cseq1 += tf - t0; 723 } 724 } else if (sparse_axpy == 1){ /* Perform one sparse axpy */ 725 /*------------------------------------------------------*/ 726 /* malloc apa to store dense row A[i,:]*P */ 727 ierr = PetscMalloc((P->cmap->N)*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 728 ierr = PetscMemzero(apa,P->cmap->N*sizeof(PetscScalar));CHKERRQ(ierr); 729 730 for (i=0; i<am; i++) { 731 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 732 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 733 /*------------------------------------------------------------*/ 734 apJ = apj + api[i]; 735 736 /* diagonal portion of A */ 737 anz = adi[i+1] - adi[i]; 738 adj = ad->j + adi[i]; 739 ada = ad->a + adi[i]; 740 for (j=0; j<anz; j++) { 741 row = adj[j]; 742 pnz = pi_loc[row+1] - pi_loc[row]; 743 pj = pj_loc + pi_loc[row]; 744 pa = pa_loc + pi_loc[row]; 745 746 /* perform dense axpy */ 747 valtmp = ada[j]; 748 for (k=0; k<pnz; k++){ 749 apa[pj[k]] += valtmp*pa[k]; 750 } 751 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 752 } 753 754 /* off-diagonal portion of A */ 755 anz = aoi[i+1] - aoi[i]; 756 aoj = ao->j + aoi[i]; 757 aoa = ao->a + aoi[i]; 758 for (j=0; j<anz; j++) { 759 row = aoj[j]; 760 pnz = pi_oth[row+1] - pi_oth[row]; 761 pj = pj_oth + pi_oth[row]; 762 pa = pa_oth + pi_oth[row]; 763 764 /* perform dense axpy */ 765 valtmp = aoa[j]; 766 for (k=0; k<pnz; k++){ 767 apa[pj[k]] += valtmp*pa[k]; 768 } 769 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 770 } 771 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 772 time_Cseq0 += tf - t0; 773 774 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 775 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 776 /*--------------------------------------------------------------*/ 777 apnz = api[i+1] - api[i]; 778 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 779 pnz = po->i[i+1] - po->i[i]; 780 poJ = po->j + po->i[i]; 781 pA = po->a + po->i[i]; 782 for (j=0; j<pnz; j++){ 783 row = poJ[j]; 784 cj = coj + coi[row]; 785 ca = coa + coi[row]; 786 valtmp = pA[j]; 787 /* perform sparse axpy */ 788 nextap = 0; 789 for (k=0; nextap<apnz; k++) { 790 if (cj[k]==apJ[nextap]) { /* global column index */ 791 ca[k] += valtmp*apa[cj[k]]; nextap++; 792 } 793 } 794 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 795 } 796 797 /* put the value into Cd (diagonal part) */ 798 pnz = pd->i[i+1] - pd->i[i]; 799 pdJ = pd->j + pd->i[i]; 800 pA = pd->a + pd->i[i]; 801 for (j=0; j<pnz; j++){ 802 row = pdJ[j]; 803 cj = bj + bi[row]; 804 ca = ba + bi[row]; 805 valtmp = pA[j]; 806 /* perform sparse axpy */ 807 nextap = 0; 808 for (k=0; nextap<apnz; k++) { 809 if (cj[k]==apJ[nextap]) { /* global column index */ 810 ca[k] += valtmp*apa[cj[k]]; 811 nextap++; 812 } 813 } 814 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 815 } 816 817 /* zero the current row of A*P */ 818 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 819 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 820 time_Cseq1 += tf - t0; 821 } 822 } else if (sparse_axpy == 2){/* Perform two sparse axpy */ 823 /*----------------------------------------------------*/ 824 /* malloc apa to store sparse row A[i,:]*P */ 825 ierr = PetscMalloc((ptap->abnz_max+1)*sizeof(MatScalar),&apa);CHKERRQ(ierr); 826 ierr = PetscMemzero(apa,ptap->abnz_max*sizeof(MatScalar));CHKERRQ(ierr); 827 828 pA=pa_loc; 829 for (i=0; i<am; i++) { 830 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 831 /* form i-th sparse row of A*P */ 832 apnz = api[i+1] - api[i]; 833 apJ = apj + api[i]; 834 /* diagonal portion of A */ 835 anz = adi[i+1] - adi[i]; 836 adj = ad->j + adi[i]; 837 ada = ad->a + adi[i]; 838 for (j=0; j<anz; j++) { 839 row = adj[j]; 840 pnz = pi_loc[row+1] - pi_loc[row]; 841 pj = pj_loc + pi_loc[row]; 842 pa = pa_loc + pi_loc[row]; 843 valtmp = ada[j]; 844 nextp = 0; 845 for (k=0; nextp<pnz; k++) { 846 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 847 apa[k] += valtmp*pa[nextp++]; 848 } 849 } 850 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 851 } 852 /* off-diagonal portion of A */ 853 anz = aoi[i+1] - aoi[i]; 854 aoj = ao->j + aoi[i]; 855 aoa = ao->a + aoi[i]; 856 for (j=0; j<anz; j++) { 857 row = aoj[j]; 858 pnz = pi_oth[row+1] - pi_oth[row]; 859 pj = pj_oth + pi_oth[row]; 860 pa = pa_oth + pi_oth[row]; 861 valtmp = aoa[j]; 862 nextp = 0; 863 for (k=0; nextp<pnz; k++) { 864 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 865 apa[k] += valtmp*pa[nextp++]; 866 } 867 } 868 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 869 } 870 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 871 time_Cseq0 += tf - t0; 872 873 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 874 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 875 /*--------------------------------------------------------------*/ 876 pnz = pi_loc[i+1] - pi_loc[i]; 877 pJ = pj_loc + pi_loc[i]; 878 for (j=0; j<pnz; j++) { 879 nextap = 0; 880 row = pJ[j]; /* global index */ 881 if (row < pcstart || row >=pcend) { /* put the value into Co */ 882 row = *poJ; 883 cj = coj + coi[row]; 884 ca = coa + coi[row]; poJ++; 885 } else { /* put the value into Cd */ 886 row = *pdJ; 887 cj = bj + bi[row]; 888 ca = ba + bi[row]; pdJ++; 889 } 890 valtmp = pA[j]; 891 for (k=0; nextap<apnz; k++) { 892 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 893 } 894 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 895 } 896 pA += pnz; 897 /* zero the current row info for A*P */ 898 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 899 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 900 time_Cseq1 += tf - t0; 901 } 902 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"sparse_axpy only takes values 0, 1 and 2"); 903 ierr = PetscFree(apa);CHKERRQ(ierr); 904 905 /* 3) send and recv matrix values coa */ 906 /*------------------------------------*/ 907 buf_ri = merge->buf_ri; 908 buf_rj = merge->buf_rj; 909 len_s = merge->len_s; 910 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 911 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 912 913 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 914 for (proc=0,k=0; proc<size; proc++){ 915 if (!len_s[proc]) continue; 916 i = merge->owners_co[proc]; 917 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 918 k++; 919 } 920 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 921 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 922 923 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 924 ierr = PetscFree(r_waits);CHKERRQ(ierr); 925 ierr = PetscFree(coa);CHKERRQ(ierr); 926 927 /* 4) insert local Cseq and received values into Cmpi */ 928 /*------------------------------------------------------*/ 929 ierr = PetscGetTime(&t0);CHKERRQ(ierr); 930 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 931 932 for (k=0; k<merge->nrecv; k++){ 933 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 934 nrows = *(buf_ri_k[k]); 935 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 936 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 937 } 938 939 for (i=0; i<cm; i++) { 940 row = owners[rank] + i; /* global row index of C_seq */ 941 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 942 ba_i = ba + bi[i]; 943 bnz = bi[i+1] - bi[i]; 944 /* add received vals into ba */ 945 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 946 /* i-th row */ 947 if (i == *nextrow[k]) { 948 cnz = *(nextci[k]+1) - *nextci[k]; 949 cj = buf_rj[k] + *(nextci[k]); 950 ca = abuf_r[k] + *(nextci[k]); 951 nextcj = 0; 952 for (j=0; nextcj<cnz; j++){ 953 if (bj_i[j] == cj[nextcj]){ /* bcol == ccol */ 954 ba_i[j] += ca[nextcj++]; 955 } 956 } 957 nextrow[k]++; nextci[k]++; 958 } 959 } 960 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 961 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 962 } 963 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 964 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 965 ierr = PetscGetTime(&tf);CHKERRQ(ierr); 966 time_setvals += tf-t0; 967 968 ierr = PetscFree(ba);CHKERRQ(ierr); 969 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 970 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 971 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 972 973 ierr = PetscGetTime(&tff);CHKERRQ(ierr); 974 etime += tff - t00; 975 /* 976 PetscInt prid=0; 977 if (rank == prid){ 978 ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] PtAPNum time %g = matupdate %g + malloc %g + Cseq %g + %g + setvals %g\n",rank,etime,time_matupdate,time_malloc,time_Cseq0,time_Cseq1,time_setvals); 979 } 980 */ 981 PetscFunctionReturn(0); 982 } 983