xref: /petsc/src/mat/impls/dense/mpi/mmdense.c (revision fd705b320d8d44969be9ca25a36dbdd35fbe8e12)
1 #define PETSCMAT_DLL
2 
3 /*
4    Support for the parallel dense matrix vector multiply
5 */
6 #include "src/mat/impls/dense/mpi/mpidense.h"
7 
8 #undef __FUNCT__
9 #define __FUNCT__ "MatSetUpMultiply_MPIDense"
10 PetscErrorCode MatSetUpMultiply_MPIDense(Mat mat)
11 {
12   Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
13   PetscErrorCode ierr;
14   IS           from,to;
15   Vec          gvec;
16 
17   PetscFunctionBegin;
18   /* Create local vector that is used to scatter into */
19   ierr = VecCreateSeq(PETSC_COMM_SELF,mat->cmap.N,&mdn->lvec);CHKERRQ(ierr);
20 
21   /* Create temporary index set for building scatter gather */
22   ierr = ISCreateStride(((PetscObject)mat)->comm,mat->cmap.N,0,1,&from);CHKERRQ(ierr);
23   ierr = ISCreateStride(PETSC_COMM_SELF,mat->cmap.N,0,1,&to);CHKERRQ(ierr);
24 
25   /* Create temporary global vector to generate scatter context */
26   /* n    = mdn->cowners[mdn->rank+1] - mdn->cowners[mdn->rank]; */
27 
28   ierr = VecCreateMPI(((PetscObject)mat)->comm,mdn->nvec,mat->cmap.N,&gvec);CHKERRQ(ierr);
29 
30   /* Generate the scatter context */
31   ierr = VecScatterCreate(gvec,from,mdn->lvec,to,&mdn->Mvctx);CHKERRQ(ierr);
32   ierr = PetscLogObjectParent(mat,mdn->Mvctx);CHKERRQ(ierr);
33   ierr = PetscLogObjectParent(mat,mdn->lvec);CHKERRQ(ierr);
34   ierr = PetscLogObjectParent(mat,from);CHKERRQ(ierr);
35   ierr = PetscLogObjectParent(mat,to);CHKERRQ(ierr);
36   ierr = PetscLogObjectParent(mat,gvec);CHKERRQ(ierr);
37 
38   ierr = ISDestroy(to);CHKERRQ(ierr);
39   ierr = ISDestroy(from);CHKERRQ(ierr);
40   ierr = VecDestroy(gvec);CHKERRQ(ierr);
41   PetscFunctionReturn(0);
42 }
43 
44 EXTERN PetscErrorCode MatGetSubMatrices_MPIDense_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,Mat*);
45 #undef __FUNCT__
46 #define __FUNCT__ "MatGetSubMatrices_MPIDense"
47 PetscErrorCode MatGetSubMatrices_MPIDense(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[])
48 {
49   PetscErrorCode ierr;
50   PetscInt           nmax,nstages_local,nstages,i,pos,max_no;
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->cmap.N * sizeof(PetscInt));
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,MPIU_INT,MPI_MAX,((PetscObject)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 PetscErrorCode MatGetSubMatrices_MPIDense_Local(Mat C,PetscInt ismax,const IS isrow[],const 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   PetscErrorCode ierr;
84   PetscMPIInt    rank,size,tag0,tag1,idex,end,i;
85   PetscInt       N = C->cmap.N,rstart = C->rmap.rstart,count;
86   PetscInt       **irow,**icol,*nrow,*ncol,*w1,*w3,*w4,*rtable,start;
87   PetscInt       **sbuf1,m,j,k,l,ct1,**rbuf1,row,proc;
88   PetscInt       nrqs,msz,**ptr,*ctr,*pa,*tmp,bsz,nrqr;
89   PetscInt       is_no,jmax,*irow_i,**rmap,*rmap_i;
90   PetscInt       len,ctr_j,*sbuf1_j,*rbuf1_i;
91   MPI_Request    *s_waits1,*r_waits1,*s_waits2,*r_waits2;
92   MPI_Status     *r_status1,*r_status2,*s_status1,*s_status2;
93   MPI_Comm       comm;
94   PetscScalar    **rbuf2,**sbuf2;
95   PetscTruth     sorted;
96 
97   PetscFunctionBegin;
98   comm   = ((PetscObject)C)->comm;
99   tag0   = ((PetscObject)C)->tag;
100   size   = c->size;
101   rank   = c->rank;
102   m      = C->rmap.N;
103 
104   /* Get some new tags to keep the communication clean */
105   ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr);
106 
107     /* Check if the col indices are sorted */
108   for (i=0; i<ismax; i++) {
109     ierr = ISSorted(isrow[i],&sorted);CHKERRQ(ierr);
110     if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"ISrow is not sorted");
111     ierr = ISSorted(iscol[i],&sorted);CHKERRQ(ierr);
112     if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"IScol is not sorted");
113   }
114 
115   len    =  2*ismax*(sizeof(PetscInt*)+sizeof(PetscInt)) + (m+1)*sizeof(PetscInt);
116   ierr   = PetscMalloc(len,&irow);CHKERRQ(ierr);
117   icol   = irow + ismax;
118   nrow   = (PetscInt*)(icol + ismax);
119   ncol   = nrow + ismax;
120   rtable = ncol + ismax;
121 
122   for (i=0; i<ismax; i++) {
123     ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr);
124     ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr);
125     ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr);
126     ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr);
127   }
128 
129   /* Create hash table for the mapping :row -> proc*/
130   for (i=0,j=0; i<size; i++) {
131     jmax = C->rmap.range[i+1];
132     for (; j<jmax; j++) {
133       rtable[j] = i;
134     }
135   }
136 
137   /* evaluate communication - mesg to who,length of mesg, and buffer space
138      required. Based on this, buffers are allocated, and data copied into them*/
139   ierr   = PetscMalloc(size*4*sizeof(PetscInt),&w1);CHKERRQ(ierr); /* mesg size */
140   w3     = w1 + 2*size;      /* no of IS that needs to be sent to proc i */
141   w4     = w3 + size;      /* temp work space used in determining w1,  w3 */
142   ierr = PetscMemzero(w1,size*3*sizeof(PetscInt));CHKERRQ(ierr); /* initialize work vector*/
143   for (i=0; i<ismax; i++) {
144     ierr   = PetscMemzero(w4,size*sizeof(PetscInt));CHKERRQ(ierr); /* initialize work vector*/
145     jmax   = nrow[i];
146     irow_i = irow[i];
147     for (j=0; j<jmax; j++) {
148       row  = irow_i[j];
149       proc = rtable[row];
150       w4[proc]++;
151     }
152     for (j=0; j<size; j++) {
153       if (w4[j]) { w1[2*j] += w4[j];  w3[j]++;}
154     }
155   }
156 
157   nrqs       = 0;              /* no of outgoing messages */
158   msz        = 0;              /* total mesg length (for all procs) */
159   w1[2*rank] = 0;              /* no mesg sent to self */
160   w3[rank]   = 0;
161   for (i=0; i<size; i++) {
162     if (w1[2*i])  { w1[2*i+1] = 1; nrqs++;} /* there exists a message to proc i */
163   }
164   ierr = PetscMalloc((nrqs+1)*sizeof(PetscInt),&pa);CHKERRQ(ierr); /*(proc -array)*/
165   for (i=0,j=0; i<size; i++) {
166     if (w1[2*i]) { pa[j] = i; j++; }
167   }
168 
169   /* Each message would have a header = 1 + 2*(no of IS) + data */
170   for (i=0; i<nrqs; i++) {
171     j       = pa[i];
172     w1[2*j] += w1[2*j+1] + 2* w3[j];
173     msz     += w1[2*j];
174   }
175   /* Do a global reduction to determine how many messages to expect*/
176   ierr = PetscMaxSum(comm,w1,&bsz,&nrqr);CHKERRQ(ierr);
177 
178   /* Allocate memory for recv buffers . Prob none if nrqr = 0 ???? */
179   len      = (nrqr+1)*sizeof(PetscInt*) + nrqr*bsz*sizeof(PetscInt);
180   ierr     = PetscMalloc(len,&rbuf1);CHKERRQ(ierr);
181   rbuf1[0] = (PetscInt*)(rbuf1 + nrqr);
182   for (i=1; i<nrqr; ++i) rbuf1[i] = rbuf1[i-1] + bsz;
183 
184   /* Post the receives */
185   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&r_waits1);CHKERRQ(ierr);
186   for (i=0; i<nrqr; ++i) {
187     ierr = MPI_Irecv(rbuf1[i],bsz,MPIU_INT,MPI_ANY_SOURCE,tag0,comm,r_waits1+i);CHKERRQ(ierr);
188   }
189 
190   /* Allocate Memory for outgoing messages */
191   len   = 2*size*sizeof(PetscInt*) + 2*msz*sizeof(PetscInt)+ size*sizeof(PetscInt);
192   ierr  = PetscMalloc(len,&sbuf1);CHKERRQ(ierr);
193   ptr   = sbuf1 + size;   /* Pointers to the data in outgoing buffers */
194   ierr  = PetscMemzero(sbuf1,2*size*sizeof(PetscInt*));CHKERRQ(ierr);
195   /* allocate memory for outgoing data + buf to receive the first reply */
196   tmp   = (PetscInt*)(ptr + size);
197   ctr   = tmp + 2*msz;
198 
199   {
200     PetscInt *iptr = tmp,ict = 0;
201     for (i=0; i<nrqs; i++) {
202       j         = pa[i];
203       iptr     += ict;
204       sbuf1[j]  = iptr;
205       ict       = w1[2*j];
206     }
207   }
208 
209   /* Form the outgoing messages */
210   /* Initialize the header space */
211   for (i=0; i<nrqs; i++) {
212     j           = pa[i];
213     sbuf1[j][0] = 0;
214     ierr        = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr);
215     ptr[j]      = sbuf1[j] + 2*w3[j] + 1;
216   }
217 
218   /* Parse the isrow and copy data into outbuf */
219   for (i=0; i<ismax; i++) {
220     ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr);
221     irow_i = irow[i];
222     jmax   = nrow[i];
223     for (j=0; j<jmax; j++) {  /* parse the indices of each IS */
224       row  = irow_i[j];
225       proc = rtable[row];
226       if (proc != rank) { /* copy to the outgoing buf*/
227         ctr[proc]++;
228         *ptr[proc] = row;
229         ptr[proc]++;
230       }
231     }
232     /* Update the headers for the current IS */
233     for (j=0; j<size; j++) { /* Can Optimise this loop too */
234       if ((ctr_j = ctr[j])) {
235         sbuf1_j        = sbuf1[j];
236         k              = ++sbuf1_j[0];
237         sbuf1_j[2*k]   = ctr_j;
238         sbuf1_j[2*k-1] = i;
239       }
240     }
241   }
242 
243   /*  Now  post the sends */
244   ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);CHKERRQ(ierr);
245   for (i=0; i<nrqs; ++i) {
246     j = pa[i];
247     ierr = MPI_Isend(sbuf1[j],w1[2*j],MPIU_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr);
248   }
249 
250   /* Post recieves to capture the row_data from other procs */
251   ierr  = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits2);CHKERRQ(ierr);
252   ierr  = PetscMalloc((nrqs+1)*sizeof(PetscScalar*),&rbuf2);CHKERRQ(ierr);
253   for (i=0; i<nrqs; i++) {
254     j        = pa[i];
255     count    = (w1[2*j] - (2*sbuf1[j][0] + 1))*N;
256     ierr     = PetscMalloc((count+1)*sizeof(PetscScalar),&rbuf2[i]);CHKERRQ(ierr);
257     ierr     = MPI_Irecv(rbuf2[i],count,MPIU_SCALAR,j,tag1,comm,r_waits2+i);CHKERRQ(ierr);
258   }
259 
260   /* Receive messages(row_nos) and then, pack and send off the rowvalues
261      to the correct processors */
262 
263   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);CHKERRQ(ierr);
264   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status1);CHKERRQ(ierr);
265   ierr = PetscMalloc((nrqr+1)*sizeof(PetscScalar*),&sbuf2);CHKERRQ(ierr);
266 
267   {
268     PetscScalar *sbuf2_i,*v_start;
269     PetscInt         s_proc;
270     for (i=0; i<nrqr; ++i) {
271       ierr = MPI_Waitany(nrqr,r_waits1,&idex,r_status1+i);CHKERRQ(ierr);
272       s_proc          = r_status1[i].MPI_SOURCE; /* send processor */
273       rbuf1_i         = rbuf1[idex]; /* Actual message from s_proc */
274       /* no of rows = end - start; since start is array idex[], 0idex, whel end
275          is length of the buffer - which is 1idex */
276       start           = 2*rbuf1_i[0] + 1;
277       ierr            = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr);
278       /* allocate memory sufficinet to hold all the row values */
279       ierr = PetscMalloc((end-start)*N*sizeof(PetscScalar),&sbuf2[idex]);CHKERRQ(ierr);
280       sbuf2_i      = sbuf2[idex];
281       /* Now pack the data */
282       for (j=start; j<end; j++) {
283         row = rbuf1_i[j] - rstart;
284         v_start = a->v + row;
285         for (k=0; k<N; k++) {
286           sbuf2_i[0] = v_start[0];
287           sbuf2_i++; v_start += C->rmap.n;
288         }
289       }
290       /* Now send off the data */
291       ierr = MPI_Isend(sbuf2[idex],(end-start)*N,MPIU_SCALAR,s_proc,tag1,comm,s_waits2+i);CHKERRQ(ierr);
292     }
293   }
294   /* End Send-Recv of IS + row_numbers */
295   ierr = PetscFree(r_status1);CHKERRQ(ierr);
296   ierr = PetscFree(r_waits1);CHKERRQ(ierr);
297   ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status1);CHKERRQ(ierr);
298   if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr);}
299   ierr = PetscFree(s_status1);CHKERRQ(ierr);
300   ierr = PetscFree(s_waits1);CHKERRQ(ierr);
301 
302   /* Create the submatrices */
303   if (scall == MAT_REUSE_MATRIX) {
304     for (i=0; i<ismax; i++) {
305       mat = (Mat_SeqDense *)(submats[i]->data);
306       if ((submats[i]->rmap.n != nrow[i]) || (submats[i]->cmap.n != ncol[i])) {
307         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
308       }
309       ierr = PetscMemzero(mat->v,submats[i]->rmap.n*submats[i]->cmap.n*sizeof(PetscScalar));CHKERRQ(ierr);
310       submats[i]->factor = C->factor;
311     }
312   } else {
313     for (i=0; i<ismax; i++) {
314       ierr = MatCreate(PETSC_COMM_SELF,submats+i);CHKERRQ(ierr);
315       ierr = MatSetSizes(submats[i],nrow[i],ncol[i],nrow[i],ncol[i]);CHKERRQ(ierr);
316       ierr = MatSetType(submats[i],((PetscObject)A)->type_name);CHKERRQ(ierr);
317       ierr = MatSeqDenseSetPreallocation(submats[i],PETSC_NULL);CHKERRQ(ierr);
318     }
319   }
320 
321   /* Assemble the matrices */
322   {
323     PetscInt         col;
324     PetscScalar *imat_v,*mat_v,*imat_vi,*mat_vi;
325 
326     for (i=0; i<ismax; i++) {
327       mat       = (Mat_SeqDense*)submats[i]->data;
328       mat_v     = a->v;
329       imat_v    = mat->v;
330       irow_i    = irow[i];
331       m         = nrow[i];
332       for (j=0; j<m; j++) {
333         row      = irow_i[j] ;
334         proc     = rtable[row];
335         if (proc == rank) {
336           row      = row - rstart;
337           mat_vi   = mat_v + row;
338           imat_vi  = imat_v + j;
339           for (k=0; k<ncol[i]; k++) {
340             col = icol[i][k];
341             imat_vi[k*m] = mat_vi[col*C->rmap.n];
342           }
343         }
344       }
345     }
346   }
347 
348   /* Create row map. This maps c->row to submat->row for each submat*/
349   /* this is a very expensive operation wrt memory usage */
350   len     = (1+ismax)*sizeof(PetscInt*)+ ismax*C->rmap.N*sizeof(PetscInt);
351   ierr    = PetscMalloc(len,&rmap);CHKERRQ(ierr);
352   rmap[0] = (PetscInt*)(rmap + ismax);
353   ierr    = PetscMemzero(rmap[0],ismax*C->rmap.N*sizeof(PetscInt));CHKERRQ(ierr);
354   for (i=1; i<ismax; i++) { rmap[i] = rmap[i-1] + C->rmap.N;}
355   for (i=0; i<ismax; i++) {
356     rmap_i = rmap[i];
357     irow_i = irow[i];
358     jmax   = nrow[i];
359     for (j=0; j<jmax; j++) {
360       rmap_i[irow_i[j]] = j;
361     }
362   }
363 
364   /* Now Receive the row_values and assemble the rest of the matrix */
365   ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status2);CHKERRQ(ierr);
366 
367   {
368     PetscInt    is_max,tmp1,col,*sbuf1_i,is_sz;
369     PetscScalar *rbuf2_i,*imat_v,*imat_vi;
370 
371     for (tmp1=0; tmp1<nrqs; tmp1++) { /* For each message */
372       ierr = MPI_Waitany(nrqs,r_waits2,&i,r_status2+tmp1);CHKERRQ(ierr);
373       /* Now dig out the corresponding sbuf1, which contains the IS data_structure */
374       sbuf1_i = sbuf1[pa[i]];
375       is_max  = sbuf1_i[0];
376       ct1     = 2*is_max+1;
377       rbuf2_i = rbuf2[i];
378       for (j=1; j<=is_max; j++) { /* For each IS belonging to the message */
379         is_no     = sbuf1_i[2*j-1];
380         is_sz     = sbuf1_i[2*j];
381         mat       = (Mat_SeqDense*)submats[is_no]->data;
382         imat_v    = mat->v;
383         rmap_i    = rmap[is_no];
384         m         = nrow[is_no];
385         for (k=0; k<is_sz; k++,rbuf2_i+=N) {  /* For each row */
386           row      = sbuf1_i[ct1]; ct1++;
387           row      = rmap_i[row];
388           imat_vi  = imat_v + row;
389           for (l=0; l<ncol[is_no]; l++) { /* For each col */
390             col = icol[is_no][l];
391             imat_vi[l*m] = rbuf2_i[col];
392           }
393         }
394       }
395     }
396   }
397   /* End Send-Recv of row_values */
398   ierr = PetscFree(r_status2);CHKERRQ(ierr);
399   ierr = PetscFree(r_waits2);CHKERRQ(ierr);
400   ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status2);CHKERRQ(ierr);
401   if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr);}
402   ierr = PetscFree(s_status2);CHKERRQ(ierr);
403   ierr = PetscFree(s_waits2);CHKERRQ(ierr);
404 
405   /* Restore the indices */
406   for (i=0; i<ismax; i++) {
407     ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr);
408     ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr);
409   }
410 
411   /* Destroy allocated memory */
412   ierr = PetscFree(irow);CHKERRQ(ierr);
413   ierr = PetscFree(w1);CHKERRQ(ierr);
414   ierr = PetscFree(pa);CHKERRQ(ierr);
415 
416 
417   for (i=0; i<nrqs; ++i) {
418     ierr = PetscFree(rbuf2[i]);CHKERRQ(ierr);
419   }
420   ierr = PetscFree(rbuf2);CHKERRQ(ierr);
421   ierr = PetscFree(sbuf1);CHKERRQ(ierr);
422   ierr = PetscFree(rbuf1);CHKERRQ(ierr);
423 
424   for (i=0; i<nrqr; ++i) {
425     ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr);
426   }
427 
428   ierr = PetscFree(sbuf2);CHKERRQ(ierr);
429   ierr = PetscFree(rmap);CHKERRQ(ierr);
430 
431   for (i=0; i<ismax; i++) {
432     ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
433     ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
434   }
435 
436   PetscFunctionReturn(0);
437 }
438 
439