xref: /petsc/src/mat/impls/dense/mpi/mmdense.c (revision 6849ba73f22fecb8f92ef896a42e4e8bd4cd6965)
1 /*
2    Support for the parallel dense matrix vector multiply
3 */
4 #include "src/mat/impls/dense/mpi/mpidense.h"
5 
6 #undef __FUNCT__
7 #define __FUNCT__ "MatSetUpMultiply_MPIDense"
8 PetscErrorCode MatSetUpMultiply_MPIDense(Mat mat)
9 {
10   Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
11   PetscErrorCode ierr;
12   IS           from,to;
13   Vec          gvec;
14 
15   PetscFunctionBegin;
16   /* Create local vector that is used to scatter into */
17   ierr = VecCreateSeq(PETSC_COMM_SELF,mat->N,&mdn->lvec);CHKERRQ(ierr);
18 
19   /* Create temporary index set for building scatter gather */
20   ierr = ISCreateStride(mat->comm,mat->N,0,1,&from);CHKERRQ(ierr);
21   ierr = ISCreateStride(PETSC_COMM_SELF,mat->N,0,1,&to);CHKERRQ(ierr);
22 
23   /* Create temporary global vector to generate scatter context */
24   /* n    = mdn->cowners[mdn->rank+1] - mdn->cowners[mdn->rank]; */
25 
26   ierr = VecCreateMPI(mat->comm,mdn->nvec,mat->N,&gvec);CHKERRQ(ierr);
27 
28   /* Generate the scatter context */
29   ierr = VecScatterCreate(gvec,from,mdn->lvec,to,&mdn->Mvctx);CHKERRQ(ierr);
30   PetscLogObjectParent(mat,mdn->Mvctx);
31   PetscLogObjectParent(mat,mdn->lvec);
32   PetscLogObjectParent(mat,from);
33   PetscLogObjectParent(mat,to);
34   PetscLogObjectParent(mat,gvec);
35 
36   ierr = ISDestroy(to);CHKERRQ(ierr);
37   ierr = ISDestroy(from);CHKERRQ(ierr);
38   ierr = VecDestroy(gvec);CHKERRQ(ierr);
39   PetscFunctionReturn(0);
40 }
41 
42 EXTERN PetscErrorCode MatGetSubMatrices_MPIDense_Local(Mat,int,const IS[],const IS[],MatReuse,Mat*);
43 #undef __FUNCT__
44 #define __FUNCT__ "MatGetSubMatrices_MPIDense"
45 PetscErrorCode MatGetSubMatrices_MPIDense(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[])
46 {
47   PetscErrorCode ierr;
48   int           nmax,nstages_local,nstages,i,pos,max_no;
49 
50   PetscFunctionBegin;
51   /* Allocate memory to hold all the submatrices */
52   if (scall != MAT_REUSE_MATRIX) {
53     ierr = PetscMalloc((ismax+1)*sizeof(Mat),submat);CHKERRQ(ierr);
54   }
55   /* Determine the number of stages through which submatrices are done */
56   nmax          = 20*1000000 / (C->N * sizeof(int));
57   if (!nmax) nmax = 1;
58   nstages_local = ismax/nmax + ((ismax % nmax)?1:0);
59 
60   /* Make sure every processor loops through the nstages */
61   ierr = MPI_Allreduce(&nstages_local,&nstages,1,MPI_INT,MPI_MAX,C->comm);CHKERRQ(ierr);
62 
63 
64   for (i=0,pos=0; i<nstages; i++) {
65     if (pos+nmax <= ismax) max_no = nmax;
66     else if (pos == ismax) max_no = 0;
67     else                   max_no = ismax-pos;
68     ierr = MatGetSubMatrices_MPIDense_Local(C,max_no,isrow+pos,iscol+pos,scall,*submat+pos);CHKERRQ(ierr);
69     pos += max_no;
70   }
71   PetscFunctionReturn(0);
72 }
73 /* -------------------------------------------------------------------------*/
74 #undef __FUNCT__
75 #define __FUNCT__ "MatGetSubMatrices_MPIDense_Local"
76 PetscErrorCode MatGetSubMatrices_MPIDense_Local(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submats)
77 {
78   Mat_MPIDense  *c = (Mat_MPIDense*)C->data;
79   Mat           A = c->A;
80   Mat_SeqDense  *a = (Mat_SeqDense*)A->data,*mat;
81   PetscErrorCode ierr;
82   int           N = C->N,rstart = c->rstart,count;
83   int           **irow,**icol,*nrow,*ncol,*w1,*w3,*w4,*rtable,start,end,size;
84   int           **sbuf1,rank,m,i,j,k,l,ct1,**rbuf1,row,proc;
85   int           nrqs,msz,**ptr,idex,*ctr,*pa,*tmp,bsz,nrqr;
86   int           is_no,jmax,*irow_i,**rmap,*rmap_i;
87   int           len,ctr_j,*sbuf1_j,*rbuf1_i;
88   int           tag0,tag1;
89   MPI_Request   *s_waits1,*r_waits1,*s_waits2,*r_waits2;
90   MPI_Status    *r_status1,*r_status2,*s_status1,*s_status2;
91   MPI_Comm      comm;
92   PetscScalar   **rbuf2,**sbuf2;
93   PetscTruth    sorted;
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],&sorted);CHKERRQ(ierr);
108     if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"ISrow is not sorted");
109     ierr = ISSorted(iscol[i],&sorted);CHKERRQ(ierr);
110     if (!sorted) 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   w3     = w1 + 2*size;      /* no of IS that needs to be sent to proc i */
139   w4     = w3 + size;      /* temp work space used in determining w1,  w3 */
140   ierr = PetscMemzero(w1,size*3*sizeof(int));CHKERRQ(ierr); /* initialize work vector*/
141   for (i=0; i<ismax; i++) {
142     ierr   = PetscMemzero(w4,size*sizeof(int));CHKERRQ(ierr); /* initialize work vector*/
143     jmax   = nrow[i];
144     irow_i = irow[i];
145     for (j=0; j<jmax; j++) {
146       row  = irow_i[j];
147       proc = rtable[row];
148       w4[proc]++;
149     }
150     for (j=0; j<size; j++) {
151       if (w4[j]) { w1[2*j] += w4[j];  w3[j]++;}
152     }
153   }
154 
155   nrqs       = 0;              /* no of outgoing messages */
156   msz        = 0;              /* total mesg length (for all procs) */
157   w1[2*rank] = 0;              /* no mesg sent to self */
158   w3[rank]   = 0;
159   for (i=0; i<size; i++) {
160     if (w1[2*i])  { w1[2*i+1] = 1; nrqs++;} /* there exists a message to proc i */
161   }
162   ierr = PetscMalloc((nrqs+1)*sizeof(int),&pa);CHKERRQ(ierr); /*(proc -array)*/
163   for (i=0,j=0; i<size; i++) {
164     if (w1[2*i]) { pa[j] = i; j++; }
165   }
166 
167   /* Each message would have a header = 1 + 2*(no of IS) + data */
168   for (i=0; i<nrqs; i++) {
169     j       = pa[i];
170     w1[2*j] += w1[2*j+1] + 2* w3[j];
171     msz     += w1[2*j];
172   }
173   /* Do a global reduction to determine how many messages to expect*/
174   ierr = PetscMaxSum(comm,w1,&bsz,&nrqr);CHKERRQ(ierr);
175 
176   /* Allocate memory for recv buffers . Prob none if nrqr = 0 ???? */
177   len      = (nrqr+1)*sizeof(int*) + nrqr*bsz*sizeof(int);
178   ierr     = PetscMalloc(len,&rbuf1);CHKERRQ(ierr);
179   rbuf1[0] = (int*)(rbuf1 + nrqr);
180   for (i=1; i<nrqr; ++i) rbuf1[i] = rbuf1[i-1] + bsz;
181 
182   /* Post the receives */
183   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&r_waits1);CHKERRQ(ierr);
184   for (i=0; i<nrqr; ++i) {
185     ierr = MPI_Irecv(rbuf1[i],bsz,MPI_INT,MPI_ANY_SOURCE,tag0,comm,r_waits1+i);CHKERRQ(ierr);
186   }
187 
188   /* Allocate Memory for outgoing messages */
189   len   = 2*size*sizeof(int*) + 2*msz*sizeof(int)+ size*sizeof(int);
190   ierr  = PetscMalloc(len,&sbuf1);CHKERRQ(ierr);
191   ptr   = sbuf1 + size;   /* Pointers to the data in outgoing buffers */
192   ierr  = PetscMemzero(sbuf1,2*size*sizeof(int*));CHKERRQ(ierr);
193   /* allocate memory for outgoing data + buf to receive the first reply */
194   tmp   = (int*)(ptr + size);
195   ctr   = tmp + 2*msz;
196 
197   {
198     int *iptr = tmp,ict = 0;
199     for (i=0; i<nrqs; i++) {
200       j         = pa[i];
201       iptr     += ict;
202       sbuf1[j]  = iptr;
203       ict       = w1[2*j];
204     }
205   }
206 
207   /* Form the outgoing messages */
208   /* Initialize the header space */
209   for (i=0; i<nrqs; i++) {
210     j           = pa[i];
211     sbuf1[j][0] = 0;
212     ierr        = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(int));CHKERRQ(ierr);
213     ptr[j]      = sbuf1[j] + 2*w3[j] + 1;
214   }
215 
216   /* Parse the isrow and copy data into outbuf */
217   for (i=0; i<ismax; i++) {
218     ierr = PetscMemzero(ctr,size*sizeof(int));CHKERRQ(ierr);
219     irow_i = irow[i];
220     jmax   = nrow[i];
221     for (j=0; j<jmax; j++) {  /* parse the indices of each IS */
222       row  = irow_i[j];
223       proc = rtable[row];
224       if (proc != rank) { /* copy to the outgoing buf*/
225         ctr[proc]++;
226         *ptr[proc] = row;
227         ptr[proc]++;
228       }
229     }
230     /* Update the headers for the current IS */
231     for (j=0; j<size; j++) { /* Can Optimise this loop too */
232       if ((ctr_j = ctr[j])) {
233         sbuf1_j        = sbuf1[j];
234         k              = ++sbuf1_j[0];
235         sbuf1_j[2*k]   = ctr_j;
236         sbuf1_j[2*k-1] = i;
237       }
238     }
239   }
240 
241   /*  Now  post the sends */
242   ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);CHKERRQ(ierr);
243   for (i=0; i<nrqs; ++i) {
244     j = pa[i];
245     ierr = MPI_Isend(sbuf1[j],w1[2*j],MPI_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr);
246   }
247 
248   /* Post recieves to capture the row_data from other procs */
249   ierr  = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits2);CHKERRQ(ierr);
250   ierr  = PetscMalloc((nrqs+1)*sizeof(PetscScalar*),&rbuf2);CHKERRQ(ierr);
251   for (i=0; i<nrqs; i++) {
252     j        = pa[i];
253     count    = (w1[2*j] - (2*sbuf1[j][0] + 1))*N;
254     ierr     = PetscMalloc((count+1)*sizeof(PetscScalar),&rbuf2[i]);CHKERRQ(ierr);
255     ierr     = MPI_Irecv(rbuf2[i],count,MPIU_SCALAR,j,tag1,comm,r_waits2+i);CHKERRQ(ierr);
256   }
257 
258   /* Receive messages(row_nos) and then, pack and send off the rowvalues
259      to the correct processors */
260 
261   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);CHKERRQ(ierr);
262   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status1);CHKERRQ(ierr);
263   ierr = PetscMalloc((nrqr+1)*sizeof(PetscScalar*),&sbuf2);CHKERRQ(ierr);
264 
265   {
266     PetscScalar *sbuf2_i,*v_start;
267     int         s_proc;
268     for (i=0; i<nrqr; ++i) {
269       ierr = MPI_Waitany(nrqr,r_waits1,&idex,r_status1+i);CHKERRQ(ierr);
270       s_proc          = r_status1[i].MPI_SOURCE; /* send processor */
271       rbuf1_i         = rbuf1[idex]; /* Actual message from s_proc */
272       /* no of rows = end - start; since start is array idex[], 0idex, whel end
273          is length of the buffer - which is 1idex */
274       start           = 2*rbuf1_i[0] + 1;
275       ierr            = MPI_Get_count(r_status1+i,MPI_INT,&end);CHKERRQ(ierr);
276       /* allocate memory sufficinet to hold all the row values */
277       ierr = PetscMalloc((end-start)*N*sizeof(PetscScalar),&sbuf2[idex]);CHKERRQ(ierr);
278       sbuf2_i      = sbuf2[idex];
279       /* Now pack the data */
280       for (j=start; j<end; j++) {
281         row = rbuf1_i[j] - rstart;
282         v_start = a->v + row;
283         for (k=0; k<N; k++) {
284           sbuf2_i[0] = v_start[0];
285           sbuf2_i++; v_start += C->m;
286         }
287       }
288       /* Now send off the data */
289       ierr = MPI_Isend(sbuf2[idex],(end-start)*N,MPIU_SCALAR,s_proc,tag1,comm,s_waits2+i);CHKERRQ(ierr);
290     }
291   }
292   /* End Send-Recv of IS + row_numbers */
293   ierr = PetscFree(r_status1);CHKERRQ(ierr);
294   ierr = PetscFree(r_waits1);CHKERRQ(ierr);
295   ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status1);CHKERRQ(ierr);
296   ierr      = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr);
297   ierr = PetscFree(s_status1);CHKERRQ(ierr);
298   ierr = PetscFree(s_waits1);CHKERRQ(ierr);
299 
300   /* Create the submatrices */
301   if (scall == MAT_REUSE_MATRIX) {
302     for (i=0; i<ismax; i++) {
303       mat = (Mat_SeqDense *)(submats[i]->data);
304       if ((submats[i]->m != nrow[i]) || (submats[i]->n != ncol[i])) {
305         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
306       }
307       ierr = PetscMemzero(mat->v,submats[i]->m*submats[i]->n*sizeof(PetscScalar));CHKERRQ(ierr);
308       submats[i]->factor = C->factor;
309     }
310   } else {
311     for (i=0; i<ismax; i++) {
312       ierr = MatCreate(PETSC_COMM_SELF,nrow[i],ncol[i],nrow[i],ncol[i],submats+i);CHKERRQ(ierr);
313       ierr = MatSetType(submats[i],A->type_name);CHKERRQ(ierr);
314       ierr = MatSeqDenseSetPreallocation(submats[i],PETSC_NULL);CHKERRQ(ierr);
315     }
316   }
317 
318   /* Assemble the matrices */
319   {
320     int         col;
321     PetscScalar *imat_v,*mat_v,*imat_vi,*mat_vi;
322 
323     for (i=0; i<ismax; i++) {
324       mat       = (Mat_SeqDense*)submats[i]->data;
325       mat_v     = a->v;
326       imat_v    = mat->v;
327       irow_i    = irow[i];
328       m         = nrow[i];
329       for (j=0; j<m; j++) {
330         row      = irow_i[j] ;
331         proc     = rtable[row];
332         if (proc == rank) {
333           row      = row - rstart;
334           mat_vi   = mat_v + row;
335           imat_vi  = imat_v + j;
336           for (k=0; k<ncol[i]; k++) {
337             col = icol[i][k];
338             imat_vi[k*m] = mat_vi[col*C->m];
339           }
340         }
341       }
342     }
343   }
344 
345   /* Create row map. This maps c->row to submat->row for each submat*/
346   /* this is a very expensive operation wrt memory usage */
347   len     = (1+ismax)*sizeof(int*)+ ismax*C->M*sizeof(int);
348   ierr    = PetscMalloc(len,&rmap);CHKERRQ(ierr);
349   rmap[0] = (int*)(rmap + ismax);
350   ierr    = PetscMemzero(rmap[0],ismax*C->M*sizeof(int));CHKERRQ(ierr);
351   for (i=1; i<ismax; i++) { rmap[i] = rmap[i-1] + C->M;}
352   for (i=0; i<ismax; i++) {
353     rmap_i = rmap[i];
354     irow_i = irow[i];
355     jmax   = nrow[i];
356     for (j=0; j<jmax; j++) {
357       rmap_i[irow_i[j]] = j;
358     }
359   }
360 
361   /* Now Receive the row_values and assemble the rest of the matrix */
362   ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status2);CHKERRQ(ierr);
363 
364   {
365     int         is_max,tmp1,col,*sbuf1_i,is_sz;
366     PetscScalar *rbuf2_i,*imat_v,*imat_vi;
367 
368     for (tmp1=0; tmp1<nrqs; tmp1++) { /* For each message */
369       ierr = MPI_Waitany(nrqs,r_waits2,&i,r_status2+tmp1);CHKERRQ(ierr);
370       /* Now dig out the corresponding sbuf1, which contains the IS data_structure */
371       sbuf1_i = sbuf1[pa[i]];
372       is_max  = sbuf1_i[0];
373       ct1     = 2*is_max+1;
374       rbuf2_i = rbuf2[i];
375       for (j=1; j<=is_max; j++) { /* For each IS belonging to the message */
376         is_no     = sbuf1_i[2*j-1];
377         is_sz     = sbuf1_i[2*j];
378         mat       = (Mat_SeqDense*)submats[is_no]->data;
379         imat_v    = mat->v;
380         rmap_i    = rmap[is_no];
381         m         = nrow[is_no];
382         for (k=0; k<is_sz; k++,rbuf2_i+=N) {  /* For each row */
383           row      = sbuf1_i[ct1]; ct1++;
384           row      = rmap_i[row];
385           imat_vi  = imat_v + row;
386           for (l=0; l<ncol[is_no]; l++) { /* For each col */
387             col = icol[is_no][l];
388             imat_vi[l*m] = rbuf2_i[col];
389           }
390         }
391       }
392     }
393   }
394   /* End Send-Recv of row_values */
395   ierr = PetscFree(r_status2);CHKERRQ(ierr);
396   ierr = PetscFree(r_waits2);CHKERRQ(ierr);
397   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status2);CHKERRQ(ierr);
398   ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr);
399   ierr = PetscFree(s_status2);CHKERRQ(ierr);
400   ierr = PetscFree(s_waits2);CHKERRQ(ierr);
401 
402   /* Restore the indices */
403   for (i=0; i<ismax; i++) {
404     ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr);
405     ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr);
406   }
407 
408   /* Destroy allocated memory */
409   ierr = PetscFree(irow);CHKERRQ(ierr);
410   ierr = PetscFree(w1);CHKERRQ(ierr);
411   ierr = PetscFree(pa);CHKERRQ(ierr);
412 
413 
414   for (i=0; i<nrqs; ++i) {
415     ierr = PetscFree(rbuf2[i]);CHKERRQ(ierr);
416   }
417   ierr = PetscFree(rbuf2);CHKERRQ(ierr);
418   ierr = PetscFree(sbuf1);CHKERRQ(ierr);
419   ierr = PetscFree(rbuf1);CHKERRQ(ierr);
420 
421   for (i=0; i<nrqr; ++i) {
422     ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr);
423   }
424 
425   ierr = PetscFree(sbuf2);CHKERRQ(ierr);
426   ierr = PetscFree(rmap);CHKERRQ(ierr);
427 
428   for (i=0; i<ismax; i++) {
429     ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
430     ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
431   }
432 
433   PetscFunctionReturn(0);
434 }
435 
436