xref: /petsc/src/mat/impls/baij/mpi/mpb_baij.c (revision a5b23f4acc7afc99d3844ebd5fb65a81c16e8b8c)
1e8271787SHong Zhang #include <../src/mat/impls/baij/mpi/mpibaij.h>
2e8271787SHong Zhang 
3e8271787SHong Zhang PetscErrorCode  MatGetMultiProcBlock_MPIBAIJ(Mat mat, MPI_Comm subComm, MatReuse scall,Mat *subMat)
4e8271787SHong Zhang {
5e8271787SHong Zhang   PetscErrorCode ierr;
6e8271787SHong Zhang   Mat_MPIBAIJ    *aij  = (Mat_MPIBAIJ*)mat->data;
7e8271787SHong Zhang   Mat_SeqBAIJ    *aijB = (Mat_SeqBAIJ*)aij->B->data;
8e8271787SHong Zhang   PetscMPIInt    commRank,subCommSize,subCommRank;
9e8271787SHong Zhang   PetscMPIInt    *commRankMap,subRank,rank,commsize;
1002e9429cSHong Zhang   PetscInt       *garrayCMap,col,i,j,*nnz,newRow,newCol,*newbRow,*newbCol,k,k1;
11e8271787SHong Zhang   PetscInt       bs=mat->rmap->bs;
1202e9429cSHong Zhang   PetscScalar    *vals,*aijBvals;
13e8271787SHong Zhang 
14e8271787SHong Zhang   PetscFunctionBegin;
15ffc4695bSBarry Smith   ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&commsize);CHKERRMPI(ierr);
16ffc4695bSBarry Smith   ierr = MPI_Comm_size(subComm,&subCommSize);CHKERRMPI(ierr);
17e8271787SHong Zhang 
18*a5b23f4aSJose E. Roman   /* create subMat object with the relevant layout */
19e8271787SHong Zhang   if (scall == MAT_INITIAL_MATRIX) {
20e8271787SHong Zhang     ierr = MatCreate(subComm,subMat);CHKERRQ(ierr);
21e8271787SHong Zhang     ierr = MatSetType(*subMat,MATMPIBAIJ);CHKERRQ(ierr);
22e8271787SHong Zhang     ierr = MatSetSizes(*subMat,mat->rmap->n,mat->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
23e8271787SHong Zhang     ierr = MatSetBlockSizes(*subMat,mat->rmap->bs,mat->cmap->bs);CHKERRQ(ierr);
24e8271787SHong Zhang 
25e8271787SHong Zhang     /* need to setup rmap and cmap before Preallocation */
26e8271787SHong Zhang     ierr = PetscLayoutSetBlockSize((*subMat)->rmap,mat->rmap->bs);CHKERRQ(ierr);
27e8271787SHong Zhang     ierr = PetscLayoutSetBlockSize((*subMat)->cmap,mat->cmap->bs);CHKERRQ(ierr);
28e8271787SHong Zhang     ierr = PetscLayoutSetUp((*subMat)->rmap);CHKERRQ(ierr);
29e8271787SHong Zhang     ierr = PetscLayoutSetUp((*subMat)->cmap);CHKERRQ(ierr);
30e8271787SHong Zhang   }
31e8271787SHong Zhang 
32e8271787SHong Zhang   /* create a map of comm_rank from subComm to comm - should commRankMap and garrayCMap be kept for reused? */
33ffc4695bSBarry Smith   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&commRank);CHKERRMPI(ierr);
34ffc4695bSBarry Smith   ierr = MPI_Comm_rank(subComm,&subCommRank);CHKERRMPI(ierr);
35785e854fSJed Brown   ierr = PetscMalloc1(subCommSize,&commRankMap);CHKERRQ(ierr);
36ffc4695bSBarry Smith   ierr = MPI_Allgather(&commRank,1,MPI_INT,commRankMap,1,MPI_INT,subComm);CHKERRMPI(ierr);
37e8271787SHong Zhang 
38e8271787SHong Zhang   /* Traverse garray and identify blocked column indices [of offdiag mat] that
39e8271787SHong Zhang    should be discarded. For the ones not discarded, store the newCol+1
40e8271787SHong Zhang    value in garrayCMap */
41580bdb30SBarry Smith   ierr = PetscCalloc1(aij->B->cmap->n/bs,&garrayCMap);CHKERRQ(ierr);
42e8271787SHong Zhang   for (i=0; i<aij->B->cmap->n/bs; i++) {
43e8271787SHong Zhang     col = aij->garray[i]; /* blocked column index */
44e8271787SHong Zhang     for (subRank=0; subRank<subCommSize; subRank++) {
45e8271787SHong Zhang       rank = commRankMap[subRank];
46e8271787SHong Zhang       if ((col >= mat->cmap->range[rank]/bs) && (col < mat->cmap->range[rank+1]/bs)) {
4702e9429cSHong Zhang         garrayCMap[i] = (((*subMat)->cmap->range[subRank]- mat->cmap->range[rank])/bs + col + 1);
48e8271787SHong Zhang         break;
49e8271787SHong Zhang       }
50e8271787SHong Zhang     }
51e8271787SHong Zhang   }
52e8271787SHong Zhang 
53e8271787SHong Zhang   if (scall == MAT_INITIAL_MATRIX) {
54e8271787SHong Zhang     /* Now compute preallocation for the offdiag mat */
55580bdb30SBarry Smith     ierr = PetscCalloc1(aij->B->rmap->n/bs,&nnz);CHKERRQ(ierr);
56e8271787SHong Zhang     for (i=0; i<aij->B->rmap->n/bs; i++) {
57e8271787SHong Zhang       for (j=aijB->i[i]; j<aijB->i[i+1]; j++) {
58e8271787SHong Zhang         if (garrayCMap[aijB->j[j]]) nnz[i]++;
59e8271787SHong Zhang       }
60e8271787SHong Zhang     }
61e8271787SHong Zhang     ierr = MatMPIBAIJSetPreallocation(*(subMat),bs,0,NULL,0,nnz);CHKERRQ(ierr);
62e8271787SHong Zhang 
63e8271787SHong Zhang     /* reuse diag block with the new submat */
64e8271787SHong Zhang     ierr = MatDestroy(&((Mat_MPIBAIJ*)((*subMat)->data))->A);CHKERRQ(ierr);
65e8271787SHong Zhang 
66e8271787SHong Zhang     ((Mat_MPIBAIJ*)((*subMat)->data))->A = aij->A;
67e8271787SHong Zhang 
68e8271787SHong Zhang     ierr = PetscObjectReference((PetscObject)aij->A);CHKERRQ(ierr);
69e8271787SHong Zhang   } else if (((Mat_MPIBAIJ*)(*subMat)->data)->A != aij->A) {
70e8271787SHong Zhang     PetscObject obj = (PetscObject)((Mat_MPIBAIJ*)((*subMat)->data))->A;
71e8271787SHong Zhang 
72e8271787SHong Zhang     ierr = PetscObjectReference((PetscObject)obj);CHKERRQ(ierr);
73e8271787SHong Zhang 
74e8271787SHong Zhang     ((Mat_MPIBAIJ*)((*subMat)->data))->A = aij->A;
75e8271787SHong Zhang 
76e8271787SHong Zhang     ierr = PetscObjectReference((PetscObject)aij->A);CHKERRQ(ierr);
77e8271787SHong Zhang   }
78e8271787SHong Zhang 
79e8271787SHong Zhang   /* Now traverse aij->B and insert values into subMat */
80dcca6d9dSJed Brown   ierr = PetscMalloc3(bs,&newbRow,bs,&newbCol,bs*bs,&vals);CHKERRQ(ierr);
81e8271787SHong Zhang   for (i=0; i<aij->B->rmap->n/bs; i++) {
82e8271787SHong Zhang     newRow = (*subMat)->rmap->range[subCommRank] + i*bs;
83e8271787SHong Zhang     for (j=aijB->i[i]; j<aijB->i[i+1]; j++) {
84e8271787SHong Zhang       newCol = garrayCMap[aijB->j[j]];
85e8271787SHong Zhang       if (newCol) {
86e8271787SHong Zhang         newCol--; /* remove the increment */
87e8271787SHong Zhang         newCol *= bs;
88e8271787SHong Zhang         for (k=0; k<bs; k++) {
89e8271787SHong Zhang           newbRow[k] = newRow + k;
90e8271787SHong Zhang           newbCol[k] = newCol + k;
91e8271787SHong Zhang         }
92e8271787SHong Zhang         /* copy column-oriented aijB->a into row-oriented vals */
9302e9429cSHong Zhang         aijBvals = aijB->a + j*bs*bs;
94e8271787SHong Zhang         for (k1=0; k1<bs; k1++) {
9502e9429cSHong Zhang           for (k=0; k<bs; k++) {
9602e9429cSHong Zhang             vals[k1+k*bs] = *aijBvals++;
97e8271787SHong Zhang           }
98e8271787SHong Zhang         }
99e8271787SHong Zhang         ierr = MatSetValues(*subMat,bs,newbRow,bs,newbCol,vals,INSERT_VALUES);CHKERRQ(ierr);
100e8271787SHong Zhang       }
101e8271787SHong Zhang     }
102e8271787SHong Zhang   }
103e8271787SHong Zhang   ierr = MatAssemblyBegin(*subMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
104e8271787SHong Zhang   ierr = MatAssemblyEnd(*subMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
105e8271787SHong Zhang 
106e8271787SHong Zhang   /* deallocate temporary data */
10702e9429cSHong Zhang   ierr = PetscFree3(newbRow,newbCol,vals);CHKERRQ(ierr);
108e8271787SHong Zhang   ierr = PetscFree(commRankMap);CHKERRQ(ierr);
109e8271787SHong Zhang   ierr = PetscFree(garrayCMap);CHKERRQ(ierr);
110e8271787SHong Zhang   if (scall == MAT_INITIAL_MATRIX) {
111e8271787SHong Zhang     ierr = PetscFree(nnz);CHKERRQ(ierr);
112e8271787SHong Zhang   }
113e8271787SHong Zhang   PetscFunctionReturn(0);
114e8271787SHong Zhang }
115