xref: /petsc/src/mat/tests/ex236.c (revision 40badf4fbc550ac1f60bd080eaff6de6d55b946d)
1 static char help[] = "Test CPU/GPU memory leaks, MatMult and MatMultTransposeAdd during successive matrix assemblies\n\n";
2 
3 #include <petscmat.h>
4 
5 int main(int argc,char **argv)
6 {
7   PetscErrorCode ierr;
8   PetscMPIInt    rank,size;
9   Mat            A;
10   PetscInt       i,j,k,n=3,vstart,rstart,rend,margin;
11   Vec            x,y;
12 
13   ierr = PetscInitialize(&argc,&argv,(char*)0,help);if (ierr) return ierr;
14   CHKERRMPI(MPI_Comm_rank(PETSC_COMM_WORLD,&rank));
15   CHKERRMPI(MPI_Comm_size(PETSC_COMM_WORLD,&size));
16 
17   CHKERRQ(MatCreate(PETSC_COMM_WORLD,&A));
18   CHKERRQ(MatSetSizes(A,n,n,PETSC_DECIDE,PETSC_DECIDE));
19   CHKERRQ(MatSetFromOptions(A));
20 
21   CHKERRQ(MatMPIAIJSetPreallocation(A,n,NULL,0,NULL));
22   CHKERRQ(MatSetOption(A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE));
23   CHKERRQ(MatGetOwnershipRange(A,&rstart,&rend));
24   CHKERRQ(MatCreateVecs(A,&x,&y));
25   CHKERRQ(VecSet(x,1.0));
26 
27   /*
28     Matrix A only has nonzeros in the diagonal block, which is of size 3x3.
29     We do three successive assemblies on A. The first two have the same non-zero
30     pattern but different values, and the third breaks the non-zero pattern. The
31     first two assemblies have enough zero-rows that triggers compressed-row storage
32     in MATAIJ and MATAIJCUSPARSE.
33 
34     These settings are used to test memory management and correctness in MatMult
35     and MatMultTransposeAdd.
36   */
37 
38   for (k=0; k<3; k++) { /* Three assemblies */
39     vstart = (size*k + rank)*n*n+1;
40     margin = (k == 2)? 0 : 2; /* Create two zero-rows in the first two assemblies */
41     for (i=rstart; i<rend-margin; i++) {
42       for (j=rstart; j<rend; j++) {
43         CHKERRQ(MatSetValue(A,i,j,(PetscScalar)vstart,INSERT_VALUES));
44         vstart++;
45       }
46     }
47     CHKERRQ(MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY));
48     CHKERRQ(MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY));
49     CHKERRQ(MatMult(A,x,y));
50     CHKERRQ(MatMultTransposeAdd(A,x,y,y)); /* y[i] = sum of row i and column i of A */
51     CHKERRQ(VecView(y,PETSC_VIEWER_STDOUT_WORLD));
52   }
53 
54   CHKERRQ(MatDestroy(&A));
55   CHKERRQ(VecDestroy(&x));
56   CHKERRQ(VecDestroy(&y));
57   ierr = PetscFinalize();
58 
59   /* Uncomment this line if you want to use "cuda-memcheck --leaf-check full" to check this program */
60   /*cudaDeviceReset();*/
61   return ierr;
62 }
63 
64 /*TEST
65 
66    testset:
67      nsize: 2
68      output_file: output/ex236_1.out
69      filter: grep -v type
70 
71      test:
72        args: -mat_type aij
73 
74      test:
75        requires: cuda
76        suffix: cuda
77        args: -mat_type aijcusparse
78 TEST*/
79