xref: /petsc/src/mat/impls/aij/mpi/mmaij.c (revision 2493cbb059784fbcc9a1325dd70885e9c6c6722b)
1cb512458SBarry Smith #ifndef lint
2*2493cbb0SBarry Smith static char vcid[] = "$Id: mmaij.c,v 1.11 1995/06/08 03:09:29 bsmith Exp bsmith $";
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"
108c79f6d3SBarry Smith #include "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;
1644a69424SLois Curfman McInnes   Mat_AIJ    *B = (Mat_AIJ *) (aij->B->data);
176abc6512SBarry Smith   int        N = aij->N,i,j,*indices,*aj = B->j;
181eb62cbbSBarry Smith   int        ierr,ec = 0,*garray;
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 */
2478b31e54SBarry Smith   indices = (int *) PETSCMALLOC( N*sizeof(int) ); CHKPTRQ(indices);
2578b31e54SBarry Smith   PETSCMEMSET(indices,0,N*sizeof(int));
26d6dfbf8fSBarry Smith   for ( i=0; i<B->m; i++ ) {
27d6dfbf8fSBarry Smith     for ( j=0; j<B->ilen[i]; j++ ) {
28d6dfbf8fSBarry Smith      if (!indices[aj[B->i[i] - 1 + j]-1]) ec++;
29d6dfbf8fSBarry Smith      indices[aj[B->i[i] - 1 + j]-1] = 1;}
301eb62cbbSBarry Smith   }
318c79f6d3SBarry Smith 
321eb62cbbSBarry Smith   /* form array of columns we need */
3378b31e54SBarry Smith   garray = (int *) PETSCMALLOC( (ec+1)*sizeof(int) ); CHKPTRQ(garray);
341eb62cbbSBarry Smith   ec = 0;
351eb62cbbSBarry Smith   for ( i=0; i<N; i++ ) {
361eb62cbbSBarry Smith     if (indices[i]) garray[ec++] = i;
371eb62cbbSBarry Smith   }
381eb62cbbSBarry Smith 
391eb62cbbSBarry Smith   /* make indices now point into garray */
401eb62cbbSBarry Smith   for ( i=0; i<ec; i++ ) {
411eb62cbbSBarry Smith     indices[garray[i]] = i+1;
421eb62cbbSBarry Smith   }
431eb62cbbSBarry Smith 
441eb62cbbSBarry Smith   /* compact out the extra columns in B */
45d6dfbf8fSBarry Smith   for ( i=0; i<B->m; i++ ) {
46d6dfbf8fSBarry Smith     for ( j=0; j<B->ilen[i]; j++ ) {
47d6dfbf8fSBarry Smith       aj[B->i[i] - 1 + j] = indices[aj[B->i[i] - 1 + j]-1];
481eb62cbbSBarry Smith     }
49d6dfbf8fSBarry Smith   }
50d6dfbf8fSBarry Smith   B->n = ec;
5178b31e54SBarry Smith   PETSCFREE(indices);
521eb62cbbSBarry Smith 
531eb62cbbSBarry Smith   /* create local vector that is used to scatter into */
5478b31e54SBarry Smith   ierr = VecCreateSequential(MPI_COMM_SELF,ec,&aij->lvec); CHKERRQ(ierr);
551eb62cbbSBarry Smith 
56d6dfbf8fSBarry Smith   /* create two temporary Index sets for build scatter gather */
5778b31e54SBarry Smith   ierr = ISCreateSequential(MPI_COMM_SELF,ec,garray,&from); CHKERRQ(ierr);
5878b31e54SBarry Smith   ierr = ISCreateStrideSequential(MPI_COMM_SELF,ec,0,1,&to); CHKERRQ(ierr);
591eb62cbbSBarry Smith 
601eb62cbbSBarry Smith   /* create temporary global vector to generate scatter context */
611eb62cbbSBarry Smith   /* this is inefficient, but otherwise we must do either
621eb62cbbSBarry Smith      1) save garray until the first actual scatter when the vector is known or
631eb62cbbSBarry Smith      2) have another way of generating a scatter context without a vector.*/
6478b31e54SBarry Smith   ierr = VecCreateMPI(mat->comm,aij->n,aij->N,&gvec); CHKERRQ(ierr);
651eb62cbbSBarry Smith 
661eb62cbbSBarry Smith   /* gnerate the scatter context */
6778b31e54SBarry Smith   ierr = VecScatterCtxCreate(gvec,from,aij->lvec,to,&aij->Mvctx); CHKERRQ(ierr);
68a5a9c739SBarry Smith   PLogObjectParent(mat,aij->Mvctx);
69a5a9c739SBarry Smith   PLogObjectParent(mat,aij->lvec);
709e25ed09SBarry Smith   aij->garray = garray;
7178b31e54SBarry Smith   ierr = ISDestroy(from); CHKERRQ(ierr);
7278b31e54SBarry Smith   ierr = ISDestroy(to); CHKERRQ(ierr);
731eb62cbbSBarry Smith   ierr = VecDestroy(gvec);
748c79f6d3SBarry Smith   return 0;
758c79f6d3SBarry Smith }
769e25ed09SBarry Smith 
779e25ed09SBarry Smith 
78*2493cbb0SBarry Smith /*
79*2493cbb0SBarry Smith      Takes the local part of an already assembled MPIAIJ matrix
80*2493cbb0SBarry Smith    and disassembles it. This is to allow new nonzeros into the matrix
81*2493cbb0SBarry Smith    that require more communication in the matrix vector multiply.
82*2493cbb0SBarry Smith    Thus certain data-structures must be rebuilt.
83*2493cbb0SBarry Smith 
84*2493cbb0SBarry Smith    Kind of slow! But that's what application programmers get when
85*2493cbb0SBarry Smith    they are sloppy.
86*2493cbb0SBarry Smith */
87*2493cbb0SBarry Smith int DisAssemble_MPIAIJ(Mat A)
88*2493cbb0SBarry Smith {
89*2493cbb0SBarry Smith   Mat_MPIAIJ *aij = (Mat_MPIAIJ *) A->data;
90*2493cbb0SBarry Smith   Mat        B = aij->B,Bnew;
91*2493cbb0SBarry Smith   Mat_AIJ    *Baij = (Mat_AIJ*)B->data;
92*2493cbb0SBarry Smith   int        ierr,i,j,m=Baij->m,n = aij->N,col,ct = 0,*garray = aij->garray;
93*2493cbb0SBarry Smith   Scalar     v;
94*2493cbb0SBarry Smith 
95*2493cbb0SBarry Smith fprintf(stderr,"Warning: you are adding tricky new nonzeros\n");
96*2493cbb0SBarry Smith 
97*2493cbb0SBarry Smith   /* free stuff related to matrix-vec multiply */
98*2493cbb0SBarry Smith   ierr = VecDestroy(aij->lvec); CHKERRQ(ierr); aij->lvec = 0;
99*2493cbb0SBarry Smith   ierr = VecScatterCtxDestroy(aij->Mvctx); CHKERRQ(ierr); aij->Mvctx = 0;
100*2493cbb0SBarry Smith   PETSCFREE(aij->colmap); aij->colmap = 0;
101*2493cbb0SBarry Smith 
102*2493cbb0SBarry Smith   /* make sure that B is assembled so we can access its values */
103*2493cbb0SBarry Smith   ierr = MatAssemblyBegin(B,FINAL_ASSEMBLY); CHKERRQ(ierr);
104*2493cbb0SBarry Smith   MatAssemblyEnd(B,FINAL_ASSEMBLY); CHKERRQ(ierr);
105*2493cbb0SBarry Smith 
106*2493cbb0SBarry Smith   /* invent new B and copy stuff over */
107*2493cbb0SBarry Smith   ierr = MatCreateSequentialAIJ(MPI_COMM_SELF,m,n,0,0,&Bnew); CHKERRQ(ierr);
108*2493cbb0SBarry Smith   for ( i=0; i<m; i++ ) {
109*2493cbb0SBarry Smith     for ( j=Baij->i[i]-1; j<Baij->i[i+1]-1; j++ ) {
110*2493cbb0SBarry Smith       col = garray[Baij->j[ct]-1];
111*2493cbb0SBarry Smith       v = Baij->a[ct++];
112*2493cbb0SBarry Smith       ierr = MatSetValues(Bnew,1,&i,1,&col,&v,INSERTVALUES); CHKERRQ(ierr);
113*2493cbb0SBarry Smith     }
114*2493cbb0SBarry Smith   }
115*2493cbb0SBarry Smith   PETSCFREE(aij->garray); aij->garray = 0;
116*2493cbb0SBarry Smith   ierr = MatDestroy(B); CHKERRQ(ierr);
117*2493cbb0SBarry Smith   aij->B = Bnew;
118*2493cbb0SBarry Smith   aij->assembled = 0;
119*2493cbb0SBarry Smith   return 0;
120*2493cbb0SBarry Smith }
121*2493cbb0SBarry Smith 
122