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 Mat_Merge_SeqsToMPI *merge=ptap->merge; 28 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 29 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 30 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 31 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 32 ierr = MatDestroy(&ptap->A_loc);CHKERRQ(ierr); /* used by MatTransposeMatMult() */ 33 if (ptap->api){ierr = PetscFree(ptap->api);CHKERRQ(ierr);} 34 if (ptap->apj){ierr = PetscFree(ptap->apj);CHKERRQ(ierr);} 35 if (ptap->apa){ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 36 if (merge) { 37 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 38 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 39 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 40 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 41 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 42 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 43 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 44 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 45 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 46 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 47 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 48 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 49 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 50 ierr = merge->destroy(A);CHKERRQ(ierr); 51 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 52 } 53 ierr = PetscFree(ptap);CHKERRQ(ierr); 54 } 55 56 PetscFunctionReturn(0); 57 } 58 59 #undef __FUNCT__ 60 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP" 61 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 62 { 63 PetscErrorCode ierr; 64 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 65 Mat_PtAPMPI *ptap = a->ptap; 66 Mat_Merge_SeqsToMPI *merge = ptap->merge; 67 68 PetscFunctionBegin; 69 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 70 (*M)->ops->destroy = merge->destroy; 71 (*M)->ops->duplicate = merge->duplicate; 72 PetscFunctionReturn(0); 73 } 74 75 #undef __FUNCT__ 76 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ" 77 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 78 { 79 PetscErrorCode ierr; 80 81 PetscFunctionBegin; 82 if (scall == MAT_INITIAL_MATRIX){ 83 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 84 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 85 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 86 } 87 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 88 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr); 89 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 90 PetscFunctionReturn(0); 91 } 92 93 #undef __FUNCT__ 94 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 95 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 96 { 97 PetscErrorCode ierr; 98 Mat Cmpi; 99 Mat_PtAPMPI *ptap; 100 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 101 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 102 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 103 Mat_SeqAIJ *p_loc,*p_oth; 104 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 105 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 106 PetscInt nlnk,*lnk,*owners_co,*coi,*coj,i,k,pnz,row; 107 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 108 PetscBT lnkbt; 109 MPI_Comm comm=((PetscObject)A)->comm; 110 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 111 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 112 PetscInt len,proc,*dnz,*onz,*owners; 113 PetscInt nzi,*bi,*bj; 114 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 115 MPI_Request *swaits,*rwaits; 116 MPI_Status *sstatus,rstatus; 117 Mat_Merge_SeqsToMPI *merge; 118 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 119 PetscReal afill=1.0,afill_tmp; 120 PetscInt rstart = P->cmap->rstart,rmax; 121 PetscScalar *vals; 122 123 PetscFunctionBegin; 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; 138 139 140 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 141 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 142 #if defined(DEBUG_MATPTAP) 143 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_oth is done...\n",rank); 144 #endif 145 146 /* get P_loc by taking all local rows of P */ 147 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 148 #if defined(DEBUG_MATPTAP) 149 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 150 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_loc is done...\n",rank); 151 #endif 152 153 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 154 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 155 pi_loc = p_loc->i; pj_loc = p_loc->j; 156 pi_oth = p_oth->i; pj_oth = p_oth->j; 157 158 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 159 /*-------------------------------------------------------------------*/ 160 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&api);CHKERRQ(ierr); 161 api[0] = 0; 162 163 /* create and initialize a linked list */ 164 nlnk = pN+1; 165 #if defined(DEBUG_MATPTAP) 166 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 167 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]); 168 #endif 169 ierr = PetscLLCreate(pN,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 170 #if defined(DEBUG_MATPTAP) 171 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 172 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call PetscLLCreate() is done\n",rank); 173 #endif 174 175 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) */ 176 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 177 current_space = free_space; 178 179 for (i=0; i<am; i++) { 180 apnz = 0; 181 /* diagonal portion of A */ 182 nzi = adi[i+1] - adi[i]; 183 aj = ad->j + adi[i]; 184 for (j=0; j<nzi; j++){ 185 row = aj[j]; 186 pnz = pi_loc[row+1] - pi_loc[row]; 187 Jptr = pj_loc + pi_loc[row]; 188 /* add non-zero cols of P into the sorted linked list lnk */ 189 ierr = PetscLLAddSorted(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 190 apnz += nlnk; 191 } 192 /* off-diagonal portion of A */ 193 nzi = aoi[i+1] - aoi[i]; 194 aj = ao->j + aoi[i]; 195 for (j=0; j<nzi; j++){ 196 row = aj[j]; 197 pnz = pi_oth[row+1] - pi_oth[row]; 198 Jptr = pj_oth + pi_oth[row]; 199 ierr = PetscLLAddSorted(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 200 apnz += nlnk; 201 } 202 203 api[i+1] = api[i] + apnz; 204 if (ap_rmax < apnz) ap_rmax = apnz; 205 206 /* if free space is not available, double the total space in the list */ 207 if (current_space->local_remaining<apnz) { 208 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 209 nspacedouble++; 210 } 211 212 /* Copy data into free space, then initialize lnk */ 213 ierr = PetscLLClean(pN,pN,apnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 214 current_space->array += apnz; 215 current_space->local_used += apnz; 216 current_space->local_remaining -= apnz; 217 } 218 /* Allocate space for apj, initialize apj, and */ 219 /* destroy list of free space and other temporary array(s) */ 220 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&apj);CHKERRQ(ierr); 221 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 222 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 223 if (afill_tmp > afill) afill = afill_tmp; 224 225 /* determine symbolic Co=(p->B)^T*AP - send to others */ 226 /*----------------------------------------------------*/ 227 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 228 229 /* then, compute symbolic Co = (p->B)^T*AP */ 230 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 231 >= (num of nonzero rows of C_seq) - pn */ 232 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 233 coi[0] = 0; 234 235 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 236 nnz = fill*(poti[pon] + api[am]); 237 ierr = PetscFreeSpaceGet(nnz,&free_space); 238 current_space = free_space; 239 240 for (i=0; i<pon; i++) { 241 nnz = 0; 242 pnz = poti[i+1] - poti[i]; 243 ptJ = potj + poti[i]; 244 for (j=0; j<pnz; j++){ 245 row = ptJ[j]; /* row of AP == col of Pot */ 246 apnz = api[row+1] - api[row]; 247 Jptr = apj + api[row]; 248 /* add non-zero cols of AP into the sorted linked list lnk */ 249 ierr = PetscLLAddSorted(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 250 nnz += nlnk; 251 } 252 253 /* If free space is not available, double the total space in the list */ 254 if (current_space->local_remaining<nnz) { 255 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 256 nspacedouble++; 257 } 258 259 /* Copy data into free space, and zero out denserows */ 260 ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 261 current_space->array += nnz; 262 current_space->local_used += nnz; 263 current_space->local_remaining -= nnz; 264 coi[i+1] = coi[i] + nnz; 265 } 266 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 267 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 268 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 269 if (afill_tmp > afill) afill = afill_tmp; 270 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 271 272 #if defined(DEBUG_MATPTAP) 273 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 274 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] local Co is done\n",rank); 275 #endif 276 277 /* send j-array (coj) of Co to other processors */ 278 /*----------------------------------------------*/ 279 /* determine row ownership */ 280 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 281 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 282 merge->rowmap->n = pn; 283 merge->rowmap->bs = 1; 284 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 285 owners = merge->rowmap->range; 286 287 /* determine the number of messages to send, their lengths */ 288 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 289 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 290 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 291 len_s = merge->len_s; 292 merge->nsend = 0; 293 294 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 295 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 296 297 proc = 0; 298 for (i=0; i<pon; i++){ 299 while (prmap[i] >= owners[proc+1]) proc++; 300 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 301 len_s[proc] += coi[i+1] - coi[i]; 302 } 303 304 len = 0; /* max length of buf_si[] */ 305 owners_co[0] = 0; 306 for (proc=0; proc<size; proc++){ 307 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 308 if (len_si[proc]){ 309 merge->nsend++; 310 len_si[proc] = 2*(len_si[proc] + 1); 311 len += len_si[proc]; 312 } 313 } 314 315 /* determine the number and length of messages to receive for coi and coj */ 316 ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 317 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 318 319 /* post the Irecv and Isend of coj */ 320 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 321 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 322 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 323 for (proc=0, k=0; proc<size; proc++){ 324 if (!len_s[proc]) continue; 325 i = owners_co[proc]; 326 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 327 k++; 328 } 329 330 /* receives and sends of coj are complete */ 331 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 332 for (i=0; i<merge->nrecv; i++){ 333 PetscMPIInt icompleted; 334 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 335 } 336 ierr = PetscFree(rwaits);CHKERRQ(ierr); 337 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 338 339 /* send and recv coi */ 340 /*-------------------*/ 341 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 342 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 343 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 344 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 345 for (proc=0,k=0; proc<size; proc++){ 346 if (!len_s[proc]) continue; 347 /* form outgoing message for i-structure: 348 buf_si[0]: nrows to be sent 349 [1:nrows]: row index (global) 350 [nrows+1:2*nrows+1]: i-structure index 351 */ 352 /*-------------------------------------------*/ 353 nrows = len_si[proc]/2 - 1; 354 buf_si_i = buf_si + nrows+1; 355 buf_si[0] = nrows; 356 buf_si_i[0] = 0; 357 nrows = 0; 358 for (i=owners_co[proc]; i<owners_co[proc+1]; i++){ 359 nzi = coi[i+1] - coi[i]; 360 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 361 buf_si[nrows+1] =prmap[i] -owners[proc]; /* local row index */ 362 nrows++; 363 } 364 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 365 k++; 366 buf_si += len_si[proc]; 367 } 368 i = merge->nrecv; 369 while (i--) { 370 PetscMPIInt icompleted; 371 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 372 } 373 ierr = PetscFree(rwaits);CHKERRQ(ierr); 374 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 375 /* 376 ierr = PetscInfo2(A,"nsend: %d, nrecv: %d\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 377 for (i=0; i<merge->nrecv; i++){ 378 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); 379 } 380 */ 381 ierr = PetscFree(len_si);CHKERRQ(ierr); 382 ierr = PetscFree(len_ri);CHKERRQ(ierr); 383 ierr = PetscFree(swaits);CHKERRQ(ierr); 384 ierr = PetscFree(sstatus);CHKERRQ(ierr); 385 ierr = PetscFree(buf_s);CHKERRQ(ierr); 386 387 #if defined(DEBUG_MATPTAP) 388 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 389 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] Co is done\n",rank); 390 #endif 391 392 /* compute the local portion of C (mpi mat) */ 393 /*------------------------------------------*/ 394 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 395 396 /* allocate bi array and free space for accumulating nonzero column info */ 397 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 398 bi[0] = 0; 399 400 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 401 nnz = fill*(pi_loc[pm] + api[am]); 402 ierr = PetscFreeSpaceGet(nnz,&free_space); 403 current_space = free_space; 404 405 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 406 for (k=0; k<merge->nrecv; k++){ 407 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 408 nrows = *buf_ri_k[k]; 409 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 410 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 411 } 412 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 413 rmax = 0; 414 for (i=0; i<pn; i++) { 415 /* add pdt[i,:]*AP into lnk */ 416 nnz = 0; 417 pnz = pdti[i+1] - pdti[i]; 418 ptJ = pdtj + pdti[i]; 419 for (j=0; j<pnz; j++){ 420 row = ptJ[j]; /* row of AP == col of Pt */ 421 apnz = api[row+1] - api[row]; 422 Jptr = apj + api[row]; 423 /* add non-zero cols of AP into the sorted linked list lnk */ 424 ierr = PetscLLAddSorted(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 425 nnz += nlnk; 426 } 427 /* add received col data into lnk */ 428 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 429 if (i == *nextrow[k]) { /* i-th row */ 430 nzi = *(nextci[k]+1) - *nextci[k]; 431 Jptr = buf_rj[k] + *nextci[k]; 432 ierr = PetscLLAddSorted(nzi,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 433 nnz += nlnk; 434 nextrow[k]++; nextci[k]++; 435 } 436 } 437 438 /* if free space is not available, make more free space */ 439 if (current_space->local_remaining<nnz) { 440 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 441 nspacedouble++; 442 } 443 /* copy data into free space, then initialize lnk */ 444 ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 445 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 446 current_space->array += nnz; 447 current_space->local_used += nnz; 448 current_space->local_remaining -= nnz; 449 bi[i+1] = bi[i] + nnz; 450 if (nnz > rmax) rmax = nnz; 451 } 452 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 453 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 454 455 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 456 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 457 afill_tmp = (PetscReal)bi[pn]/(pi_loc[pm] + api[am]+1); 458 if (afill_tmp > afill) afill = afill_tmp; 459 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 460 461 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part - check runex55_SA ? */ 462 /*------------------------------------------------------------------------------------------------------*/ 463 ierr = PetscMalloc((rmax+1)*sizeof(PetscScalar),&vals);CHKERRQ(ierr); 464 ierr = PetscMemzero(vals,rmax*sizeof(PetscScalar));CHKERRQ(ierr); 465 466 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 467 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 468 ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,P->cmap->bs);CHKERRQ(ierr); 469 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 470 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 471 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 472 473 for (i=0; i<pn; i++){ 474 row = i + rstart; 475 nnz = bi[i+1] - bi[i]; 476 Jptr = bj + bi[i]; 477 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 478 } 479 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 480 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 481 ierr = PetscFree(vals);CHKERRQ(ierr); 482 483 merge->bi = bi; 484 merge->bj = bj; 485 merge->coi = coi; 486 merge->coj = coj; 487 merge->buf_ri = buf_ri; 488 merge->buf_rj = buf_rj; 489 merge->owners_co = owners_co; 490 merge->destroy = Cmpi->ops->destroy; 491 merge->duplicate = Cmpi->ops->duplicate; 492 493 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 494 /* Cmpi->assembled = PETSC_FALSE; */ 495 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 496 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 497 498 /* attach the supporting struct to Cmpi for reuse */ 499 c = (Mat_MPIAIJ*)Cmpi->data; 500 c->ptap = ptap; 501 ptap->api = api; 502 ptap->apj = apj; 503 ptap->merge = merge; 504 ptap->rmax = ap_rmax; 505 506 *C = Cmpi; 507 508 /* flag 'scalable' determines which implementations to be used: 509 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 510 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 511 /* set default scalable */ 512 ptap->scalable = PETSC_TRUE; 513 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,PETSC_NULL);CHKERRQ(ierr); 514 if (!ptap->scalable){ /* Do dense axpy */ 515 ierr = PetscMalloc(pN*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr); 516 } else { 517 ierr = PetscMalloc((ap_rmax+1)*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr); 518 } 519 520 #if defined(PETSC_USE_INFO) 521 if (bi[pn] != 0) { 522 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 523 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 524 } else { 525 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 526 } 527 #endif 528 PetscFunctionReturn(0); 529 } 530 531 #undef __FUNCT__ 532 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 533 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 534 { 535 PetscErrorCode ierr; 536 Mat_Merge_SeqsToMPI *merge; 537 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 538 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 539 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 540 Mat_SeqAIJ *p_loc,*p_oth; 541 Mat_PtAPMPI *ptap; 542 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 543 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 544 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 545 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 546 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 547 MPI_Comm comm=((PetscObject)C)->comm; 548 PetscMPIInt size,rank,taga,*len_s; 549 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 550 PetscInt **buf_ri,**buf_rj; 551 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 552 MPI_Request *s_waits,*r_waits; 553 MPI_Status *status; 554 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 555 PetscInt *api,*apj,*coi,*coj; 556 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 557 PetscBool scalable; 558 559 PetscFunctionBegin; 560 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 561 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 562 563 ptap = c->ptap; 564 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"); 565 merge = ptap->merge; 566 apa = ptap->apa; 567 scalable = ptap->scalable; 568 569 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 570 /*--------------------------------------------------*/ 571 if (ptap->reuse == MAT_INITIAL_MATRIX){ 572 ptap->reuse = MAT_REUSE_MATRIX; 573 } else { /* update numerical values of P_oth and P_loc */ 574 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 575 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 576 } 577 578 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 579 /*--------------------------------------------------------------*/ 580 /* get data from symbolic products */ 581 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 582 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 583 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 584 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 585 586 coi = merge->coi; coj = merge->coj; 587 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 588 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 589 590 bi = merge->bi; bj = merge->bj; 591 owners = merge->rowmap->range; 592 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 593 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 594 595 api = ptap->api; apj = ptap->apj; 596 597 if (!scalable){ /* Do dense axpy */ 598 /*--------------------------------------------------*/ 599 /* apa (length of pN) stores dense row A[i,:]*P - nonscalable! */ 600 ierr = PetscMemzero(apa,P->cmap->N*sizeof(PetscScalar));CHKERRQ(ierr); 601 602 for (i=0; i<am; i++) { 603 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 604 /*------------------------------------------------------------*/ 605 apJ = apj + api[i]; 606 607 /* diagonal portion of A */ 608 anz = adi[i+1] - adi[i]; 609 adj = ad->j + adi[i]; 610 ada = ad->a + adi[i]; 611 for (j=0; j<anz; j++) { 612 row = adj[j]; 613 pnz = pi_loc[row+1] - pi_loc[row]; 614 pj = pj_loc + pi_loc[row]; 615 pa = pa_loc + pi_loc[row]; 616 617 /* perform dense axpy */ 618 valtmp = ada[j]; 619 for (k=0; k<pnz; k++){ 620 apa[pj[k]] += valtmp*pa[k]; 621 } 622 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 623 } 624 625 /* off-diagonal portion of A */ 626 anz = aoi[i+1] - aoi[i]; 627 aoj = ao->j + aoi[i]; 628 aoa = ao->a + aoi[i]; 629 for (j=0; j<anz; j++) { 630 row = aoj[j]; 631 pnz = pi_oth[row+1] - pi_oth[row]; 632 pj = pj_oth + pi_oth[row]; 633 pa = pa_oth + pi_oth[row]; 634 635 /* perform dense axpy */ 636 valtmp = aoa[j]; 637 for (k=0; k<pnz; k++){ 638 apa[pj[k]] += valtmp*pa[k]; 639 } 640 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 641 } 642 643 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 644 /*--------------------------------------------------------------*/ 645 apnz = api[i+1] - api[i]; 646 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 647 pnz = po->i[i+1] - po->i[i]; 648 poJ = po->j + po->i[i]; 649 pA = po->a + po->i[i]; 650 for (j=0; j<pnz; j++){ 651 row = poJ[j]; 652 cnz = coi[row+1] - coi[row]; 653 cj = coj + coi[row]; 654 ca = coa + coi[row]; 655 /* perform dense axpy */ 656 valtmp = pA[j]; 657 for (k=0; k<cnz; k++) { 658 ca[k] += valtmp*apa[cj[k]]; 659 } 660 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 661 } 662 663 /* put the value into Cd (diagonal part) */ 664 pnz = pd->i[i+1] - pd->i[i]; 665 pdJ = pd->j + pd->i[i]; 666 pA = pd->a + pd->i[i]; 667 for (j=0; j<pnz; j++){ 668 row = pdJ[j]; 669 cnz = bi[row+1] - bi[row]; 670 cj = bj + bi[row]; 671 ca = ba + bi[row]; 672 /* perform dense axpy */ 673 valtmp = pA[j]; 674 for (k=0; k<cnz; k++) { 675 ca[k] += valtmp*apa[cj[k]]; 676 } 677 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 678 } 679 680 /* zero the current row of A*P */ 681 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 682 } 683 } else {/* Perform sparse axpy */ 684 /*----------------------------------------------------*/ 685 /* apa (length ap_rmax) stores sparse row A[i,:]*P */ 686 ierr = PetscMemzero(apa,ptap->rmax*sizeof(MatScalar));CHKERRQ(ierr); 687 688 pA=pa_loc; 689 for (i=0; i<am; i++) { 690 /* form i-th sparse row of A*P */ 691 apnz = api[i+1] - api[i]; 692 apJ = apj + api[i]; 693 /* diagonal portion of A */ 694 anz = adi[i+1] - adi[i]; 695 adj = ad->j + adi[i]; 696 ada = ad->a + adi[i]; 697 for (j=0; j<anz; j++) { 698 row = adj[j]; 699 pnz = pi_loc[row+1] - pi_loc[row]; 700 pj = pj_loc + pi_loc[row]; 701 pa = pa_loc + pi_loc[row]; 702 valtmp = ada[j]; 703 nextp = 0; 704 for (k=0; nextp<pnz; k++) { 705 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 706 apa[k] += valtmp*pa[nextp++]; 707 } 708 } 709 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 710 } 711 /* off-diagonal portion of A */ 712 anz = aoi[i+1] - aoi[i]; 713 aoj = ao->j + aoi[i]; 714 aoa = ao->a + aoi[i]; 715 for (j=0; j<anz; j++) { 716 row = aoj[j]; 717 pnz = pi_oth[row+1] - pi_oth[row]; 718 pj = pj_oth + pi_oth[row]; 719 pa = pa_oth + pi_oth[row]; 720 valtmp = aoa[j]; 721 nextp = 0; 722 for (k=0; nextp<pnz; k++) { 723 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 724 apa[k] += valtmp*pa[nextp++]; 725 } 726 } 727 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 728 } 729 730 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 731 /*--------------------------------------------------------------*/ 732 pnz = pi_loc[i+1] - pi_loc[i]; 733 pJ = pj_loc + pi_loc[i]; 734 for (j=0; j<pnz; j++) { 735 nextap = 0; 736 row = pJ[j]; /* global index */ 737 if (row < pcstart || row >=pcend) { /* put the value into Co */ 738 row = *poJ; 739 cj = coj + coi[row]; 740 ca = coa + coi[row]; poJ++; 741 } else { /* put the value into Cd */ 742 row = *pdJ; 743 cj = bj + bi[row]; 744 ca = ba + bi[row]; pdJ++; 745 } 746 valtmp = pA[j]; 747 for (k=0; nextap<apnz; k++) { 748 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 749 } 750 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 751 } 752 pA += pnz; 753 /* zero the current row info for A*P */ 754 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 755 } 756 } 757 758 /* 3) send and recv matrix values coa */ 759 /*------------------------------------*/ 760 buf_ri = merge->buf_ri; 761 buf_rj = merge->buf_rj; 762 len_s = merge->len_s; 763 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 764 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 765 766 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 767 for (proc=0,k=0; proc<size; proc++){ 768 if (!len_s[proc]) continue; 769 i = merge->owners_co[proc]; 770 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 771 k++; 772 } 773 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 774 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 775 776 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 777 ierr = PetscFree(r_waits);CHKERRQ(ierr); 778 ierr = PetscFree(coa);CHKERRQ(ierr); 779 780 /* 4) insert local Cseq and received values into Cmpi */ 781 /*------------------------------------------------------*/ 782 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 783 for (k=0; k<merge->nrecv; k++){ 784 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 785 nrows = *(buf_ri_k[k]); 786 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 787 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 788 } 789 790 for (i=0; i<cm; i++) { 791 row = owners[rank] + i; /* global row index of C_seq */ 792 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 793 ba_i = ba + bi[i]; 794 bnz = bi[i+1] - bi[i]; 795 /* add received vals into ba */ 796 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 797 /* i-th row */ 798 if (i == *nextrow[k]) { 799 cnz = *(nextci[k]+1) - *nextci[k]; 800 cj = buf_rj[k] + *(nextci[k]); 801 ca = abuf_r[k] + *(nextci[k]); 802 nextcj = 0; 803 for (j=0; nextcj<cnz; j++){ 804 if (bj_i[j] == cj[nextcj]){ /* bcol == ccol */ 805 ba_i[j] += ca[nextcj++]; 806 } 807 } 808 nextrow[k]++; nextci[k]++; 809 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 810 } 811 } 812 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 813 } 814 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 815 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 816 817 ierr = PetscFree(ba);CHKERRQ(ierr); 818 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 819 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 820 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 821 #if defined(PETSC_USE_INFO) 822 if (scalable){ 823 ierr = PetscInfo(C,"Use scalable sparse axpy\n");CHKERRQ(ierr); 824 } else { 825 ierr = PetscInfo(C,"Use non-scalable dense axpy\n");CHKERRQ(ierr); 826 } 827 #endif 828 PetscFunctionReturn(0); 829 } 830