1c4762a1bSJed Brown static char help[] = "Tests for bugs in A->offloadmask consistency for GPU matrices\n\n";
2c4762a1bSJed Brown
3c4762a1bSJed Brown #include <petscmat.h>
4c4762a1bSJed Brown
main(int argc,char ** args)5d71ae5a4SJacob Faibussowitsch int main(int argc, char **args)
6d71ae5a4SJacob Faibussowitsch {
7c4762a1bSJed Brown Mat A;
8c4762a1bSJed Brown PetscInt i, j, rstart, rend, m = 3;
9c4762a1bSJed Brown PetscScalar one = 1.0, zero = 0.0, negativeone = -1.0;
10c4762a1bSJed Brown PetscReal norm;
11c4762a1bSJed Brown Vec x, y;
12c4762a1bSJed Brown
13327415f7SBarry Smith PetscFunctionBeginUser;
14*c8025a54SPierre Jolivet PetscCall(PetscInitialize(&argc, &args, NULL, help));
159566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetInt(NULL, NULL, "-m", &m, NULL));
16c4762a1bSJed Brown
17c4762a1bSJed Brown for (i = 0; i < 2; i++) {
18c4762a1bSJed Brown /* Create the matrix and set it to contain explicit zero entries on the diagonal. */
199566063dSJacob Faibussowitsch PetscCall(MatCreate(PETSC_COMM_WORLD, &A));
209566063dSJacob Faibussowitsch PetscCall(MatSetSizes(A, PETSC_DECIDE, PETSC_DECIDE, m * m, m * m));
219566063dSJacob Faibussowitsch PetscCall(MatSetFromOptions(A));
229566063dSJacob Faibussowitsch PetscCall(MatSetUp(A));
239566063dSJacob Faibussowitsch PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
249566063dSJacob Faibussowitsch PetscCall(MatCreateVecs(A, &x, &y));
259566063dSJacob Faibussowitsch PetscCall(VecSet(x, one));
269566063dSJacob Faibussowitsch PetscCall(VecSet(y, zero));
279566063dSJacob Faibussowitsch PetscCall(MatDiagonalSet(A, y, INSERT_VALUES));
28c4762a1bSJed Brown
29c4762a1bSJed Brown /* Now set A to be the identity using various approaches.
30c4762a1bSJed Brown * Note that there may be other approaches that should be added here. */
31c4762a1bSJed Brown switch (i) {
32d71ae5a4SJacob Faibussowitsch case 0:
33d71ae5a4SJacob Faibussowitsch PetscCall(MatDiagonalSet(A, x, INSERT_VALUES));
34d71ae5a4SJacob Faibussowitsch break;
35c4762a1bSJed Brown case 1:
3648a46eb9SPierre Jolivet for (j = rstart; j < rend; j++) PetscCall(MatSetValue(A, j, j, one, INSERT_VALUES));
379566063dSJacob Faibussowitsch PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
389566063dSJacob Faibussowitsch PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
39c4762a1bSJed Brown break;
40c4762a1bSJed Brown case 2:
4148a46eb9SPierre Jolivet for (j = rstart; j < rend; j++) PetscCall(MatSetValuesRow(A, j, &one));
429566063dSJacob Faibussowitsch PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
439566063dSJacob Faibussowitsch PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
44d71ae5a4SJacob Faibussowitsch default:
45d71ae5a4SJacob Faibussowitsch break;
46c4762a1bSJed Brown }
47c4762a1bSJed Brown
48c4762a1bSJed Brown /* Compute y <- A*x and verify that the difference between y and x is negligible, as it should be since A is the identity. */
499566063dSJacob Faibussowitsch PetscCall(MatMult(A, x, y));
509566063dSJacob Faibussowitsch PetscCall(VecAXPY(y, negativeone, x));
519566063dSJacob Faibussowitsch PetscCall(VecNorm(y, NORM_2, &norm));
5248a46eb9SPierre Jolivet if (norm > PETSC_SQRT_MACHINE_EPSILON) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "Test %" PetscInt_FMT ": Norm of error is %g, but should be near 0.\n", i, (double)norm));
53c4762a1bSJed Brown
549566063dSJacob Faibussowitsch PetscCall(MatDestroy(&A));
559566063dSJacob Faibussowitsch PetscCall(VecDestroy(&x));
569566063dSJacob Faibussowitsch PetscCall(VecDestroy(&y));
57c4762a1bSJed Brown }
58c4762a1bSJed Brown
599566063dSJacob Faibussowitsch PetscCall(PetscFinalize());
60b122ec5aSJacob Faibussowitsch return 0;
61c4762a1bSJed Brown }
62c4762a1bSJed Brown
63c4762a1bSJed Brown /*TEST
64f4747e26SJunchao Zhang testset:
65f4747e26SJunchao Zhang output_file: output/empty.out
66c4762a1bSJed Brown
67c4762a1bSJed Brown test:
68c4762a1bSJed Brown suffix: aijviennacl_1
69f4747e26SJunchao Zhang nsize: {{1 2}}
70c4762a1bSJed Brown args: -mat_type aijviennacl
71c4762a1bSJed Brown requires: viennacl
72c4762a1bSJed Brown
73c4762a1bSJed Brown test:
74c4762a1bSJed Brown suffix: aijcusparse_1
75f4747e26SJunchao Zhang nsize: {{1 2}}
76c4762a1bSJed Brown args: -mat_type aijcusparse
77c4762a1bSJed Brown requires: cuda
78c4762a1bSJed Brown
79c4762a1bSJed Brown test:
80f4747e26SJunchao Zhang suffix: kokkos
81f4747e26SJunchao Zhang nsize: {{1 2}}
82f4747e26SJunchao Zhang args: -mat_type aijkokkos
83f4747e26SJunchao Zhang requires: kokkos_kernels
84f4747e26SJunchao Zhang
85c4762a1bSJed Brown TEST*/
86