1 /*$Id: mmdense.c,v 1.39 2001/08/07 03:02:46 balay Exp bsmith $*/ 2 3 /* 4 Support for the parallel dense matrix vector multiply 5 */ 6 #include "src/mat/impls/dense/mpi/mpidense.h" 7 #include "src/vec/vecimpl.h" 8 9 #undef __FUNCT__ 10 #define __FUNCT__ "MatSetUpMultiply_MPIDense" 11 int MatSetUpMultiply_MPIDense(Mat mat) 12 { 13 Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data; 14 int ierr; 15 IS from,to; 16 Vec gvec; 17 18 PetscFunctionBegin; 19 /* Create local vector that is used to scatter into */ 20 ierr = VecCreateSeq(PETSC_COMM_SELF,mat->N,&mdn->lvec);CHKERRQ(ierr); 21 22 /* Create temporary index set for building scatter gather */ 23 ierr = ISCreateStride(mat->comm,mat->N,0,1,&from);CHKERRQ(ierr); 24 ierr = ISCreateStride(PETSC_COMM_SELF,mat->N,0,1,&to);CHKERRQ(ierr); 25 26 /* Create temporary global vector to generate scatter context */ 27 /* n = mdn->cowners[mdn->rank+1] - mdn->cowners[mdn->rank]; */ 28 29 ierr = VecCreateMPI(mat->comm,mdn->nvec,mat->N,&gvec);CHKERRQ(ierr); 30 31 /* Generate the scatter context */ 32 ierr = VecScatterCreate(gvec,from,mdn->lvec,to,&mdn->Mvctx);CHKERRQ(ierr); 33 PetscLogObjectParent(mat,mdn->Mvctx); 34 PetscLogObjectParent(mat,mdn->lvec); 35 PetscLogObjectParent(mat,from); 36 PetscLogObjectParent(mat,to); 37 PetscLogObjectParent(mat,gvec); 38 39 ierr = ISDestroy(to);CHKERRQ(ierr); 40 ierr = ISDestroy(from);CHKERRQ(ierr); 41 ierr = VecDestroy(gvec);CHKERRQ(ierr); 42 PetscFunctionReturn(0); 43 } 44 45 EXTERN int MatGetSubMatrices_MPIDense_Local(Mat,int,IS*,IS*,MatReuse,Mat*); 46 #undef __FUNCT__ 47 #define __FUNCT__ "MatGetSubMatrices_MPIDense" 48 int MatGetSubMatrices_MPIDense(Mat C,int ismax,IS *isrow,IS *iscol,MatReuse scall,Mat **submat) 49 { 50 int nmax,nstages_local,nstages,i,pos,max_no,ierr; 51 52 PetscFunctionBegin; 53 /* Allocate memory to hold all the submatrices */ 54 if (scall != MAT_REUSE_MATRIX) { 55 ierr = PetscMalloc((ismax+1)*sizeof(Mat),submat);CHKERRQ(ierr); 56 } 57 /* Determine the number of stages through which submatrices are done */ 58 nmax = 20*1000000 / (C->N * sizeof(int)); 59 if (!nmax) nmax = 1; 60 nstages_local = ismax/nmax + ((ismax % nmax)?1:0); 61 62 /* Make sure every processor loops through the nstages */ 63 ierr = MPI_Allreduce(&nstages_local,&nstages,1,MPI_INT,MPI_MAX,C->comm);CHKERRQ(ierr); 64 65 66 for (i=0,pos=0; i<nstages; i++) { 67 if (pos+nmax <= ismax) max_no = nmax; 68 else if (pos == ismax) max_no = 0; 69 else max_no = ismax-pos; 70 ierr = MatGetSubMatrices_MPIDense_Local(C,max_no,isrow+pos,iscol+pos,scall,*submat+pos);CHKERRQ(ierr); 71 pos += max_no; 72 } 73 PetscFunctionReturn(0); 74 } 75 /* -------------------------------------------------------------------------*/ 76 #undef __FUNCT__ 77 #define __FUNCT__ "MatGetSubMatrices_MPIDense_Local" 78 int MatGetSubMatrices_MPIDense_Local(Mat C,int ismax,IS *isrow,IS *iscol,MatReuse scall,Mat *submats) 79 { 80 Mat_MPIDense *c = (Mat_MPIDense*)C->data; 81 Mat A = c->A; 82 Mat_SeqDense *a = (Mat_SeqDense*)A->data,*mat; 83 int N = C->N,rstart = c->rstart,count; 84 int **irow,**icol,*nrow,*ncol,*w1,*w2,*w3,*w4,*rtable,start,end,size; 85 int **sbuf1,rank,m,i,j,k,l,ct1,ierr,**rbuf1,row,proc; 86 int nrqs,msz,**ptr,index,*ctr,*pa,*tmp,bsz,nrqr; 87 int is_no,jmax,*irow_i,**rmap,*rmap_i; 88 int len,ctr_j,*sbuf1_j,*rbuf1_i; 89 int tag0,tag1; 90 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2; 91 MPI_Status *r_status1,*r_status2,*s_status1,*s_status2; 92 MPI_Comm comm; 93 PetscScalar **rbuf2,**sbuf2; 94 95 PetscFunctionBegin; 96 comm = C->comm; 97 tag0 = C->tag; 98 size = c->size; 99 rank = c->rank; 100 m = C->M; 101 102 /* Get some new tags to keep the communication clean */ 103 ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 104 105 /* Check if the col indices are sorted */ 106 for (i=0; i<ismax; i++) { 107 ierr = ISSorted(isrow[i],(PetscTruth*)&j);CHKERRQ(ierr); 108 if (!j) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"ISrow is not sorted"); 109 ierr = ISSorted(iscol[i],(PetscTruth*)&j);CHKERRQ(ierr); 110 if (!j) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"IScol is not sorted"); 111 } 112 113 len = 2*ismax*(sizeof(int *)+sizeof(int)) + (m+1)*sizeof(int); 114 ierr = PetscMalloc(len,&irow);CHKERRQ(ierr); 115 icol = irow + ismax; 116 nrow = (int*)(icol + ismax); 117 ncol = nrow + ismax; 118 rtable = ncol + ismax; 119 120 for (i=0; i<ismax; i++) { 121 ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr); 122 ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr); 123 ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr); 124 ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr); 125 } 126 127 /* Create hash table for the mapping :row -> proc*/ 128 for (i=0,j=0; i<size; i++) { 129 jmax = c->rowners[i+1]; 130 for (; j<jmax; j++) { 131 rtable[j] = i; 132 } 133 } 134 135 /* evaluate communication - mesg to who,length of mesg, and buffer space 136 required. Based on this, buffers are allocated, and data copied into them*/ 137 ierr = PetscMalloc(size*4*sizeof(int),&w1);CHKERRQ(ierr); /* mesg size */ 138 w2 = w1 + size; /* if w2[i] marked, then a message to proc i*/ 139 w3 = w2 + size; /* no of IS that needs to be sent to proc i */ 140 w4 = w3 + size; /* temp work space used in determining w1, w2, w3 */ 141 ierr = PetscMemzero(w1,size*3*sizeof(int));CHKERRQ(ierr); /* initialize work vector*/ 142 for (i=0; i<ismax; i++) { 143 ierr = PetscMemzero(w4,size*sizeof(int));CHKERRQ(ierr); /* initialize work vector*/ 144 jmax = nrow[i]; 145 irow_i = irow[i]; 146 for (j=0; j<jmax; j++) { 147 row = irow_i[j]; 148 proc = rtable[row]; 149 w4[proc]++; 150 } 151 for (j=0; j<size; j++) { 152 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 153 } 154 } 155 156 nrqs = 0; /* no of outgoing messages */ 157 msz = 0; /* total mesg length (for all procs) */ 158 w1[rank] = 0; /* no mesg sent to self */ 159 w3[rank] = 0; 160 for (i=0; i<size; i++) { 161 if (w1[i]) { w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 162 } 163 ierr = PetscMalloc((nrqs+1)*sizeof(int),&pa);CHKERRQ(ierr); /*(proc -array)*/ 164 for (i=0,j=0; i<size; i++) { 165 if (w1[i]) { pa[j] = i; j++; } 166 } 167 168 /* Each message would have a header = 1 + 2*(no of IS) + data */ 169 for (i=0; i<nrqs; i++) { 170 j = pa[i]; 171 w1[j] += w2[j] + 2* w3[j]; 172 msz += w1[j]; 173 } 174 /* Do a global reduction to determine how many messages to expect*/ 175 { 176 int *rw1; 177 ierr = PetscMalloc(2*size*sizeof(int),&rw1);CHKERRQ(ierr); 178 ierr = MPI_Allreduce(w1,rw1,2*size,MPI_INT,PetscMaxSum_Op,comm);CHKERRQ(ierr); 179 bsz = rw1[rank]; 180 nrqr = rw1[size+rank]; 181 ierr = PetscFree(rw1);CHKERRQ(ierr); 182 } 183 184 /* Allocate memory for recv buffers . Prob none if nrqr = 0 ???? */ 185 len = (nrqr+1)*sizeof(int*) + nrqr*bsz*sizeof(int); 186 ierr = PetscMalloc(len,&rbuf1);CHKERRQ(ierr); 187 rbuf1[0] = (int*)(rbuf1 + nrqr); 188 for (i=1; i<nrqr; ++i) rbuf1[i] = rbuf1[i-1] + bsz; 189 190 /* Post the receives */ 191 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&r_waits1);CHKERRQ(ierr); 192 for (i=0; i<nrqr; ++i) { 193 ierr = MPI_Irecv(rbuf1[i],bsz,MPI_INT,MPI_ANY_SOURCE,tag0,comm,r_waits1+i);CHKERRQ(ierr); 194 } 195 196 /* Allocate Memory for outgoing messages */ 197 len = 2*size*sizeof(int*) + 2*msz*sizeof(int)+ size*sizeof(int); 198 ierr = PetscMalloc(len,&sbuf1);CHKERRQ(ierr); 199 ptr = sbuf1 + size; /* Pointers to the data in outgoing buffers */ 200 ierr = PetscMemzero(sbuf1,2*size*sizeof(int*));CHKERRQ(ierr); 201 /* allocate memory for outgoing data + buf to receive the first reply */ 202 tmp = (int*)(ptr + size); 203 ctr = tmp + 2*msz; 204 205 { 206 int *iptr = tmp,ict = 0; 207 for (i=0; i<nrqs; i++) { 208 j = pa[i]; 209 iptr += ict; 210 sbuf1[j] = iptr; 211 ict = w1[j]; 212 } 213 } 214 215 /* Form the outgoing messages */ 216 /* Initialize the header space */ 217 for (i=0; i<nrqs; i++) { 218 j = pa[i]; 219 sbuf1[j][0] = 0; 220 ierr = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(int));CHKERRQ(ierr); 221 ptr[j] = sbuf1[j] + 2*w3[j] + 1; 222 } 223 224 /* Parse the isrow and copy data into outbuf */ 225 for (i=0; i<ismax; i++) { 226 ierr = PetscMemzero(ctr,size*sizeof(int));CHKERRQ(ierr); 227 irow_i = irow[i]; 228 jmax = nrow[i]; 229 for (j=0; j<jmax; j++) { /* parse the indices of each IS */ 230 row = irow_i[j]; 231 proc = rtable[row]; 232 if (proc != rank) { /* copy to the outgoing buf*/ 233 ctr[proc]++; 234 *ptr[proc] = row; 235 ptr[proc]++; 236 } 237 } 238 /* Update the headers for the current IS */ 239 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 240 if ((ctr_j = ctr[j])) { 241 sbuf1_j = sbuf1[j]; 242 k = ++sbuf1_j[0]; 243 sbuf1_j[2*k] = ctr_j; 244 sbuf1_j[2*k-1] = i; 245 } 246 } 247 } 248 249 /* Now post the sends */ 250 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);CHKERRQ(ierr); 251 for (i=0; i<nrqs; ++i) { 252 j = pa[i]; 253 ierr = MPI_Isend(sbuf1[j],w1[j],MPI_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr); 254 } 255 256 /* Post recieves to capture the row_data from other procs */ 257 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits2);CHKERRQ(ierr); 258 ierr = PetscMalloc((nrqs+1)*sizeof(PetscScalar*),&rbuf2);CHKERRQ(ierr); 259 for (i=0; i<nrqs; i++) { 260 j = pa[i]; 261 count = (w1[j] - (2*sbuf1[j][0] + 1))*N; 262 ierr = PetscMalloc((count+1)*sizeof(PetscScalar),&rbuf2[i]);CHKERRQ(ierr); 263 ierr = MPI_Irecv(rbuf2[i],count,MPIU_SCALAR,j,tag1,comm,r_waits2+i);CHKERRQ(ierr); 264 } 265 266 /* Receive messages(row_nos) and then, pack and send off the rowvalues 267 to the correct processors */ 268 269 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);CHKERRQ(ierr); 270 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status1);CHKERRQ(ierr); 271 ierr = PetscMalloc((nrqr+1)*sizeof(PetscScalar*),&sbuf2);CHKERRQ(ierr); 272 273 { 274 PetscScalar *sbuf2_i,*v_start; 275 int s_proc; 276 for (i=0; i<nrqr; ++i) { 277 ierr = MPI_Waitany(nrqr,r_waits1,&index,r_status1+i);CHKERRQ(ierr); 278 s_proc = r_status1[i].MPI_SOURCE; /* send processor */ 279 rbuf1_i = rbuf1[index]; /* Actual message from s_proc */ 280 /* no of rows = end - start; since start is array index[], 0index, whel end 281 is length of the buffer - which is 1index */ 282 start = 2*rbuf1_i[0] + 1; 283 ierr = MPI_Get_count(r_status1+i,MPI_INT,&end);CHKERRQ(ierr); 284 /* allocate memory sufficinet to hold all the row values */ 285 ierr = PetscMalloc((end-start)*N*sizeof(PetscScalar),&sbuf2[index]);CHKERRQ(ierr); 286 sbuf2_i = sbuf2[index]; 287 /* Now pack the data */ 288 for (j=start; j<end; j++) { 289 row = rbuf1_i[j] - rstart; 290 v_start = a->v + row; 291 for (k=0; k<N; k++) { 292 sbuf2_i[0] = v_start[0]; 293 sbuf2_i++; v_start += C->m; 294 } 295 } 296 /* Now send off the data */ 297 ierr = MPI_Isend(sbuf2[index],(end-start)*N,MPIU_SCALAR,s_proc,tag1,comm,s_waits2+i);CHKERRQ(ierr); 298 } 299 } 300 /* End Send-Recv of IS + row_numbers */ 301 ierr = PetscFree(r_status1);CHKERRQ(ierr); 302 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 303 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status1);CHKERRQ(ierr); 304 ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr); 305 ierr = PetscFree(s_status1);CHKERRQ(ierr); 306 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 307 308 /* Create the submatrices */ 309 if (scall == MAT_REUSE_MATRIX) { 310 for (i=0; i<ismax; i++) { 311 mat = (Mat_SeqDense *)(submats[i]->data); 312 if ((submats[i]->m != nrow[i]) || (submats[i]->n != ncol[i])) { 313 SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 314 } 315 ierr = PetscMemzero(mat->v,submats[i]->m*submats[i]->n*sizeof(PetscScalar));CHKERRQ(ierr); 316 submats[i]->factor = C->factor; 317 } 318 } else { 319 for (i=0; i<ismax; i++) { 320 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nrow[i],ncol[i],PETSC_NULL,submats+i);CHKERRQ(ierr); 321 } 322 } 323 324 /* Assemble the matrices */ 325 { 326 int col; 327 PetscScalar *imat_v,*mat_v,*imat_vi,*mat_vi; 328 329 for (i=0; i<ismax; i++) { 330 mat = (Mat_SeqDense*)submats[i]->data; 331 mat_v = a->v; 332 imat_v = mat->v; 333 irow_i = irow[i]; 334 m = nrow[i]; 335 for (j=0; j<m; j++) { 336 row = irow_i[j] ; 337 proc = rtable[row]; 338 if (proc == rank) { 339 row = row - rstart; 340 mat_vi = mat_v + row; 341 imat_vi = imat_v + j; 342 for (k=0; k<ncol[i]; k++) { 343 col = icol[i][k]; 344 imat_vi[k*m] = mat_vi[col*C->m]; 345 } 346 } 347 } 348 } 349 } 350 351 /* Create row map. This maps c->row to submat->row for each submat*/ 352 /* this is a very expensive operation wrt memory usage */ 353 len = (1+ismax)*sizeof(int*)+ ismax*C->M*sizeof(int); 354 ierr = PetscMalloc(len,&rmap);CHKERRQ(ierr); 355 rmap[0] = (int *)(rmap + ismax); 356 ierr = PetscMemzero(rmap[0],ismax*C->M*sizeof(int));CHKERRQ(ierr); 357 for (i=1; i<ismax; i++) { rmap[i] = rmap[i-1] + C->M;} 358 for (i=0; i<ismax; i++) { 359 rmap_i = rmap[i]; 360 irow_i = irow[i]; 361 jmax = nrow[i]; 362 for (j=0; j<jmax; j++) { 363 rmap_i[irow_i[j]] = j; 364 } 365 } 366 367 /* Now Receive the row_values and assemble the rest of the matrix */ 368 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status2);CHKERRQ(ierr); 369 370 { 371 int is_max,tmp1,col,*sbuf1_i,is_sz; 372 PetscScalar *rbuf2_i,*imat_v,*imat_vi; 373 374 for (tmp1=0; tmp1<nrqs; tmp1++) { /* For each message */ 375 ierr = MPI_Waitany(nrqs,r_waits2,&i,r_status2+tmp1);CHKERRQ(ierr); 376 /* Now dig out the corresponding sbuf1, which contains the IS data_structure */ 377 sbuf1_i = sbuf1[pa[i]]; 378 is_max = sbuf1_i[0]; 379 ct1 = 2*is_max+1; 380 rbuf2_i = rbuf2[i]; 381 for (j=1; j<=is_max; j++) { /* For each IS belonging to the message */ 382 is_no = sbuf1_i[2*j-1]; 383 is_sz = sbuf1_i[2*j]; 384 mat = (Mat_SeqDense*)submats[is_no]->data; 385 imat_v = mat->v; 386 rmap_i = rmap[is_no]; 387 m = nrow[is_no]; 388 for (k=0; k<is_sz; k++,rbuf2_i+=N) { /* For each row */ 389 row = sbuf1_i[ct1]; ct1++; 390 row = rmap_i[row]; 391 imat_vi = imat_v + row; 392 for (l=0; l<ncol[is_no]; l++) { /* For each col */ 393 col = icol[is_no][l]; 394 imat_vi[l*m] = rbuf2_i[col]; 395 } 396 } 397 } 398 } 399 } 400 /* End Send-Recv of row_values */ 401 ierr = PetscFree(r_status2);CHKERRQ(ierr); 402 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 403 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status2);CHKERRQ(ierr); 404 ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr); 405 ierr = PetscFree(s_status2);CHKERRQ(ierr); 406 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 407 408 /* Restore the indices */ 409 for (i=0; i<ismax; i++) { 410 ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr); 411 ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr); 412 } 413 414 /* Destroy allocated memory */ 415 ierr = PetscFree(irow);CHKERRQ(ierr); 416 ierr = PetscFree(w1);CHKERRQ(ierr); 417 ierr = PetscFree(pa);CHKERRQ(ierr); 418 419 420 for (i=0; i<nrqs; ++i) { 421 ierr = PetscFree(rbuf2[i]);CHKERRQ(ierr); 422 } 423 ierr = PetscFree(rbuf2);CHKERRQ(ierr); 424 ierr = PetscFree(sbuf1);CHKERRQ(ierr); 425 ierr = PetscFree(rbuf1);CHKERRQ(ierr); 426 427 for (i=0; i<nrqr; ++i) { 428 ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr); 429 } 430 431 ierr = PetscFree(sbuf2);CHKERRQ(ierr); 432 ierr = PetscFree(rmap);CHKERRQ(ierr); 433 434 for (i=0; i<ismax; i++) { 435 ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 436 ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 437 } 438 439 PetscFunctionReturn(0); 440 } 441 442