xref: /petsc/src/mat/impls/aij/mpi/mmaij.c (revision f5eb4b8140e51d60f0c4bc35d8fe3d9a5fa43604)
1cb512458SBarry Smith #ifndef lint
2*f5eb4b81SSatish Balay static char vcid[] = "$Id: mmaij.c,v 1.25 1996/01/26 04:33:52 bsmith Exp balay $";
3cb512458SBarry Smith #endif
48c79f6d3SBarry Smith 
5d6dfbf8fSBarry Smith 
68c79f6d3SBarry Smith /*
78c79f6d3SBarry Smith    Support for the parallel AIJ matrix vector multiply
88c79f6d3SBarry Smith */
98c79f6d3SBarry Smith #include "mpiaij.h"
10*f5eb4b81SSatish Balay #include "src/vec/vecimpl.h"
118c79f6d3SBarry Smith #include "../seq/aij.h"
128c79f6d3SBarry Smith 
1344a69424SLois Curfman McInnes int MatSetUpMultiply_MPIAIJ(Mat mat)
148c79f6d3SBarry Smith {
1544a69424SLois Curfman McInnes   Mat_MPIAIJ *aij = (Mat_MPIAIJ *) mat->data;
16ec8511deSBarry Smith   Mat_SeqAIJ *B = (Mat_SeqAIJ *) (aij->B->data);
17416022c9SBarry Smith   int        N = aij->N,i,j,*indices,*aj = B->j,ierr,ec = 0,*garray;
18416022c9SBarry Smith   int        shift = B->indexshift;
191eb62cbbSBarry Smith   IS         from,to;
201eb62cbbSBarry Smith   Vec        gvec;
218c79f6d3SBarry Smith 
228c79f6d3SBarry Smith   /* For the first stab we make an array as long as the number of columns */
231eb62cbbSBarry Smith   /* mark those columns that are in aij->B */
240452661fSBarry Smith   indices = (int *) PetscMalloc( N*sizeof(int) ); CHKPTRQ(indices);
25cddf8d76SBarry Smith   PetscMemzero(indices,N*sizeof(int));
26d6dfbf8fSBarry Smith   for ( i=0; i<B->m; i++ ) {
27d6dfbf8fSBarry Smith     for ( j=0; j<B->ilen[i]; j++ ) {
28dbb450caSBarry Smith       if (!indices[aj[B->i[i] +shift + j] + shift]) ec++;
29416022c9SBarry Smith       indices[aj[B->i[i] + shift + j] + shift] = 1;
30416022c9SBarry Smith     }
311eb62cbbSBarry Smith   }
328c79f6d3SBarry Smith 
331eb62cbbSBarry Smith   /* form array of columns we need */
340452661fSBarry Smith   garray = (int *) PetscMalloc( (ec+1)*sizeof(int) ); CHKPTRQ(garray);
351eb62cbbSBarry Smith   ec = 0;
361eb62cbbSBarry Smith   for ( i=0; i<N; i++ ) {
371eb62cbbSBarry Smith     if (indices[i]) garray[ec++] = i;
381eb62cbbSBarry Smith   }
391eb62cbbSBarry Smith 
401eb62cbbSBarry Smith   /* make indices now point into garray */
411eb62cbbSBarry Smith   for ( i=0; i<ec; i++ ) {
42dbb450caSBarry Smith     indices[garray[i]] = i-shift;
431eb62cbbSBarry Smith   }
441eb62cbbSBarry Smith 
451eb62cbbSBarry Smith   /* compact out the extra columns in B */
46d6dfbf8fSBarry Smith   for ( i=0; i<B->m; i++ ) {
47d6dfbf8fSBarry Smith     for ( j=0; j<B->ilen[i]; j++ ) {
48dbb450caSBarry Smith       aj[B->i[i] + shift + j] = indices[aj[B->i[i] + shift + j]+shift];
491eb62cbbSBarry Smith     }
50d6dfbf8fSBarry Smith   }
51d6dfbf8fSBarry Smith   B->n = ec;
520452661fSBarry Smith   PetscFree(indices);
531eb62cbbSBarry Smith 
541eb62cbbSBarry Smith   /* create local vector that is used to scatter into */
55fafbff53SBarry Smith   ierr = VecCreateSeq(MPI_COMM_SELF,ec,&aij->lvec); CHKERRQ(ierr);
561eb62cbbSBarry Smith 
57d6dfbf8fSBarry Smith   /* create two temporary Index sets for build scatter gather */
58fafbff53SBarry Smith   ierr = ISCreateSeq(MPI_COMM_SELF,ec,garray,&from); CHKERRQ(ierr);
59fafbff53SBarry Smith   ierr = ISCreateStrideSeq(MPI_COMM_SELF,ec,0,1,&to); CHKERRQ(ierr);
601eb62cbbSBarry Smith 
611eb62cbbSBarry Smith   /* create temporary global vector to generate scatter context */
621eb62cbbSBarry Smith   /* this is inefficient, but otherwise we must do either
631eb62cbbSBarry Smith      1) save garray until the first actual scatter when the vector is known or
641eb62cbbSBarry Smith      2) have another way of generating a scatter context without a vector.*/
6578b31e54SBarry Smith   ierr = VecCreateMPI(mat->comm,aij->n,aij->N,&gvec); CHKERRQ(ierr);
661eb62cbbSBarry Smith 
671eb62cbbSBarry Smith   /* gnerate the scatter context */
6808480c60SBarry Smith   ierr = VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx); CHKERRQ(ierr);
69a5a9c739SBarry Smith   PLogObjectParent(mat,aij->Mvctx);
70a5a9c739SBarry Smith   PLogObjectParent(mat,aij->lvec);
71464493b3SBarry Smith   PLogObjectParent(mat,from);
72464493b3SBarry Smith   PLogObjectParent(mat,to);
739e25ed09SBarry Smith   aij->garray = garray;
74464493b3SBarry Smith   PLogObjectMemory(mat,(ec+1)*sizeof(int));
7578b31e54SBarry Smith   ierr = ISDestroy(from); CHKERRQ(ierr);
7678b31e54SBarry Smith   ierr = ISDestroy(to); CHKERRQ(ierr);
771eb62cbbSBarry Smith   ierr = VecDestroy(gvec);
788c79f6d3SBarry Smith   return 0;
798c79f6d3SBarry Smith }
809e25ed09SBarry Smith 
819e25ed09SBarry Smith 
822493cbb0SBarry Smith /*
832493cbb0SBarry Smith      Takes the local part of an already assembled MPIAIJ matrix
842493cbb0SBarry Smith    and disassembles it. This is to allow new nonzeros into the matrix
852493cbb0SBarry Smith    that require more communication in the matrix vector multiply.
862493cbb0SBarry Smith    Thus certain data-structures must be rebuilt.
872493cbb0SBarry Smith 
882493cbb0SBarry Smith    Kind of slow! But that's what application programmers get when
892493cbb0SBarry Smith    they are sloppy.
902493cbb0SBarry Smith */
912493cbb0SBarry Smith int DisAssemble_MPIAIJ(Mat A)
922493cbb0SBarry Smith {
932493cbb0SBarry Smith   Mat_MPIAIJ *aij = (Mat_MPIAIJ *) A->data;
942493cbb0SBarry Smith   Mat        B = aij->B,Bnew;
95ec8511deSBarry Smith   Mat_SeqAIJ *Baij = (Mat_SeqAIJ*)B->data;
962493cbb0SBarry Smith   int        ierr,i,j,m=Baij->m,n = aij->N,col,ct = 0,*garray = aij->garray;
97416022c9SBarry Smith   int        *nz,ec,shift = Baij->indexshift;
982493cbb0SBarry Smith   Scalar     v;
992493cbb0SBarry Smith 
1002493cbb0SBarry Smith   /* free stuff related to matrix-vec multiply */
101464493b3SBarry Smith   ierr = VecGetSize(aij->lvec,&ec); /* needed for PLogObjectMemory below */
1022493cbb0SBarry Smith   ierr = VecDestroy(aij->lvec); CHKERRQ(ierr); aij->lvec = 0;
10308480c60SBarry Smith   ierr = VecScatterDestroy(aij->Mvctx); CHKERRQ(ierr); aij->Mvctx = 0;
104464493b3SBarry Smith   if (aij->colmap) {
1050452661fSBarry Smith     PetscFree(aij->colmap); aij->colmap = 0;
106464493b3SBarry Smith     PLogObjectMemory(A,-Baij->n*sizeof(int));
107464493b3SBarry Smith   }
1082493cbb0SBarry Smith 
1092493cbb0SBarry Smith   /* make sure that B is assembled so we can access its values */
1102493cbb0SBarry Smith   ierr = MatAssemblyBegin(B,FINAL_ASSEMBLY); CHKERRQ(ierr);
1112493cbb0SBarry Smith   MatAssemblyEnd(B,FINAL_ASSEMBLY); CHKERRQ(ierr);
1122493cbb0SBarry Smith 
1132493cbb0SBarry Smith   /* invent new B and copy stuff over */
1140452661fSBarry Smith   nz = (int *) PetscMalloc( m*sizeof(int) ); CHKPTRQ(nz);
11548b35521SBarry Smith   for ( i=0; i<m; i++ ) {
11648b35521SBarry Smith     nz[i] = Baij->i[i+1]-Baij->i[i];
11748b35521SBarry Smith   }
118fafbff53SBarry Smith   ierr = MatCreateSeqAIJ(MPI_COMM_SELF,m,n,0,nz,&Bnew); CHKERRQ(ierr);
1190452661fSBarry Smith   PetscFree(nz);
1202493cbb0SBarry Smith   for ( i=0; i<m; i++ ) {
121dbb450caSBarry Smith     for ( j=Baij->i[i]+shift; j<Baij->i[i+1]+shift; j++ ) {
122dbb450caSBarry Smith       col = garray[Baij->j[ct]+shift];
1232493cbb0SBarry Smith       v = Baij->a[ct++];
124dbb450caSBarry Smith       ierr = MatSetValues(Bnew,1,&i,1,&col,&v,INSERT_VALUES); CHKERRQ(ierr);
1252493cbb0SBarry Smith     }
1262493cbb0SBarry Smith   }
1270452661fSBarry Smith   PetscFree(aij->garray); aij->garray = 0;
128464493b3SBarry Smith   PLogObjectMemory(A,-ec*sizeof(int));
1292493cbb0SBarry Smith   ierr = MatDestroy(B); CHKERRQ(ierr);
130464493b3SBarry Smith   PLogObjectParent(A,Bnew);
1312493cbb0SBarry Smith   aij->B = Bnew;
132227d817aSBarry Smith   A->was_assembled = PETSC_FALSE;
1332493cbb0SBarry Smith   return 0;
1342493cbb0SBarry Smith }
1352493cbb0SBarry Smith 
13648b35521SBarry Smith 
137