1c4762a1bSJed Brown static char help[] = "Test CPU/GPU memory leaks, MatMult and MatMultTransposeAdd during successive matrix assemblies\n\n";
2c4762a1bSJed Brown
3c4762a1bSJed Brown #include <petscmat.h>
4c4762a1bSJed Brown
main(int argc,char ** argv)5d71ae5a4SJacob Faibussowitsch int main(int argc, char **argv)
6d71ae5a4SJacob Faibussowitsch {
7c4762a1bSJed Brown PetscMPIInt rank, size;
8c4762a1bSJed Brown Mat A;
9c4762a1bSJed Brown PetscInt i, j, k, n = 3, vstart, rstart, rend, margin;
10c4762a1bSJed Brown Vec x, y;
11c4762a1bSJed Brown
12327415f7SBarry Smith PetscFunctionBeginUser;
13c8025a54SPierre Jolivet PetscCall(PetscInitialize(&argc, &argv, NULL, help));
149566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
159566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
16c4762a1bSJed Brown
179566063dSJacob Faibussowitsch PetscCall(MatCreate(PETSC_COMM_WORLD, &A));
189566063dSJacob Faibussowitsch PetscCall(MatSetSizes(A, n, n, PETSC_DECIDE, PETSC_DECIDE));
199566063dSJacob Faibussowitsch PetscCall(MatSetFromOptions(A));
20c4762a1bSJed Brown
219566063dSJacob Faibussowitsch PetscCall(MatMPIAIJSetPreallocation(A, n, NULL, 0, NULL));
229566063dSJacob Faibussowitsch PetscCall(MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
239566063dSJacob Faibussowitsch PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
249566063dSJacob Faibussowitsch PetscCall(MatCreateVecs(A, &x, &y));
259566063dSJacob Faibussowitsch PetscCall(VecSet(x, 1.0));
26c4762a1bSJed Brown
27c4762a1bSJed Brown /*
28a3234186SStefano Zampini Matrix A only has nonzeros in the diagonal block, which is of size 3x3.
29c4762a1bSJed Brown We do three successive assemblies on A. The first two have the same non-zero
30c4762a1bSJed Brown pattern but different values, and the third breaks the non-zero pattern. The
31c4762a1bSJed Brown first two assemblies have enough zero-rows that triggers compressed-row storage
32c4762a1bSJed Brown in MATAIJ and MATAIJCUSPARSE.
33c4762a1bSJed Brown
34c4762a1bSJed Brown These settings are used to test memory management and correctness in MatMult
35c4762a1bSJed Brown and MatMultTransposeAdd.
36c4762a1bSJed Brown */
37c4762a1bSJed Brown
38c4762a1bSJed Brown for (k = 0; k < 3; k++) { /* Three assemblies */
39c4762a1bSJed Brown vstart = (size * k + rank) * n * n + 1;
40c4762a1bSJed Brown margin = (k == 2) ? 0 : 2; /* Create two zero-rows in the first two assemblies */
41c4762a1bSJed Brown for (i = rstart; i < rend - margin; i++) {
42c4762a1bSJed Brown for (j = rstart; j < rend; j++) {
439566063dSJacob Faibussowitsch PetscCall(MatSetValue(A, i, j, (PetscScalar)vstart, INSERT_VALUES));
44c4762a1bSJed Brown vstart++;
45c4762a1bSJed Brown }
46c4762a1bSJed Brown }
479566063dSJacob Faibussowitsch PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
489566063dSJacob Faibussowitsch PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
499566063dSJacob Faibussowitsch PetscCall(MatMult(A, x, y));
509566063dSJacob Faibussowitsch PetscCall(MatMultTransposeAdd(A, x, y, y)); /* y[i] = sum of row i and column i of A */
519566063dSJacob Faibussowitsch PetscCall(VecView(y, PETSC_VIEWER_STDOUT_WORLD));
52c4762a1bSJed Brown }
53c4762a1bSJed Brown
549566063dSJacob Faibussowitsch PetscCall(MatDestroy(&A));
559566063dSJacob Faibussowitsch PetscCall(VecDestroy(&x));
569566063dSJacob Faibussowitsch PetscCall(VecDestroy(&y));
579566063dSJacob Faibussowitsch PetscCall(PetscFinalize());
58c4762a1bSJed Brown
59*74df5e01SJunchao Zhang /* Uncomment this line if you want to use "compute-sanitizer --tool memcheck [sanitizer_options] app_name [app_options]" to check this program */
60c4762a1bSJed Brown /*cudaDeviceReset();*/
61b122ec5aSJacob Faibussowitsch return 0;
62c4762a1bSJed Brown }
63c4762a1bSJed Brown
64c4762a1bSJed Brown /*TEST
65c4762a1bSJed Brown
66c4762a1bSJed Brown testset:
67c4762a1bSJed Brown nsize: 2
68c4762a1bSJed Brown output_file: output/ex236_1.out
69c4762a1bSJed Brown filter: grep -v type
70c4762a1bSJed Brown
71c4762a1bSJed Brown test:
72c4762a1bSJed Brown args: -mat_type aij
73c4762a1bSJed Brown
74c4762a1bSJed Brown test:
75c4762a1bSJed Brown requires: cuda
76c4762a1bSJed Brown suffix: cuda
77c4762a1bSJed Brown args: -mat_type aijcusparse
78c4762a1bSJed Brown TEST*/
79