xref: /petsc/src/mat/tests/ex239.c (revision 609caa7c8c030312b00807b4f015fd827bb80932)
1a78de62eSJunchao Zhang static char help[] = "Test device/host memory allocation in MatDenseSeqCUDA()\n\n";
2a78de62eSJunchao Zhang 
3a78de62eSJunchao Zhang /* Contributed by: Victor Eijkhout <eijkhout@tacc.utexas.edu> */
4a78de62eSJunchao Zhang 
5a78de62eSJunchao Zhang #include <petscmat.h>
main(int argc,char ** argv)6d71ae5a4SJacob Faibussowitsch int main(int argc, char **argv)
7d71ae5a4SJacob Faibussowitsch {
8a78de62eSJunchao Zhang   PetscInt  global_size = 100;
9a78de62eSJunchao Zhang   Mat       cuda_matrix;
10a78de62eSJunchao Zhang   Vec       input, output;
11a78de62eSJunchao Zhang   MPI_Comm  comm = PETSC_COMM_SELF;
12a78de62eSJunchao Zhang   PetscReal nrm  = 1;
13a78de62eSJunchao Zhang 
14327415f7SBarry Smith   PetscFunctionBeginUser;
159566063dSJacob Faibussowitsch   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
169566063dSJacob Faibussowitsch   PetscCall(MatCreateDenseCUDA(comm, global_size, global_size, global_size, global_size, NULL, &cuda_matrix));
17a78de62eSJunchao Zhang 
189566063dSJacob Faibussowitsch   PetscCall(VecCreateSeqCUDA(comm, global_size, &input));
199566063dSJacob Faibussowitsch   PetscCall(VecDuplicate(input, &output));
209566063dSJacob Faibussowitsch   PetscCall(VecSet(input, 1.));
219566063dSJacob Faibussowitsch   PetscCall(VecSet(output, 2.));
229566063dSJacob Faibussowitsch   PetscCall(MatMult(cuda_matrix, input, output));
239566063dSJacob Faibussowitsch   PetscCall(VecNorm(output, NORM_2, &nrm));
2408401ef6SPierre Jolivet   PetscCheck(nrm <= PETSC_SMALL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PETSc generated wrong result. Should be 0, but is %g", (double)nrm);
259566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&input));
269566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&output));
279566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&cuda_matrix));
289566063dSJacob Faibussowitsch   PetscCall(PetscFinalize());
29b122ec5aSJacob Faibussowitsch   return 0;
30a78de62eSJunchao Zhang }
31a78de62eSJunchao Zhang 
32a78de62eSJunchao Zhang /*TEST
33a78de62eSJunchao Zhang    build:
34a78de62eSJunchao Zhang      requires: cuda
35a78de62eSJunchao Zhang 
36a78de62eSJunchao Zhang    test:
37a78de62eSJunchao Zhang      nsize: 1
38*3886731fSPierre Jolivet      output_file: output/empty.out
39a78de62eSJunchao Zhang 
40a78de62eSJunchao Zhang TEST*/
41