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> 6a78de62eSJunchao Zhang int main(int argc, char** argv) 7a78de62eSJunchao Zhang { 8a78de62eSJunchao Zhang PetscErrorCode ierr; 9a78de62eSJunchao Zhang PetscInt global_size=100; 10a78de62eSJunchao Zhang Mat cuda_matrix; 11a78de62eSJunchao Zhang Vec input,output; 12a78de62eSJunchao Zhang MPI_Comm comm = PETSC_COMM_SELF; 13a78de62eSJunchao Zhang PetscReal nrm = 1; 14a78de62eSJunchao Zhang 15a78de62eSJunchao Zhang ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr; 16a78de62eSJunchao Zhang ierr = MatCreateDenseCUDA(comm,global_size,global_size,global_size,global_size,NULL,&cuda_matrix);CHKERRQ(ierr); 17a78de62eSJunchao Zhang ierr = MatAssemblyBegin(cuda_matrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 18a78de62eSJunchao Zhang ierr = MatAssemblyEnd(cuda_matrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 19a78de62eSJunchao Zhang 20a78de62eSJunchao Zhang ierr = VecCreateSeqCUDA(comm,global_size,&input);CHKERRQ(ierr); 21a78de62eSJunchao Zhang ierr = VecDuplicate(input,&output);CHKERRQ(ierr); 22a78de62eSJunchao Zhang ierr = VecSet(input,1.);CHKERRQ(ierr); 23a78de62eSJunchao Zhang ierr = VecSet(output,2.);CHKERRQ(ierr); 24a78de62eSJunchao Zhang ierr = MatMult(cuda_matrix,input,output);CHKERRQ(ierr); 25a78de62eSJunchao Zhang ierr = VecNorm(output,NORM_2,&nrm);CHKERRQ(ierr); 26*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(nrm > PETSC_SMALL,PETSC_COMM_SELF,PETSC_ERR_PLIB,"PETSc generated wrong result. Should be 0, but is %g",(double)nrm); 27a78de62eSJunchao Zhang ierr = VecDestroy(&input);CHKERRQ(ierr); 28a78de62eSJunchao Zhang ierr = VecDestroy(&output);CHKERRQ(ierr); 29a78de62eSJunchao Zhang ierr = MatDestroy(&cuda_matrix);CHKERRQ(ierr); 30a78de62eSJunchao Zhang ierr = PetscFinalize(); 31a78de62eSJunchao Zhang return ierr; 32a78de62eSJunchao Zhang } 33a78de62eSJunchao Zhang 34a78de62eSJunchao Zhang /*TEST 35a78de62eSJunchao Zhang build: 36a78de62eSJunchao Zhang requires: cuda 37a78de62eSJunchao Zhang 38a78de62eSJunchao Zhang test: 39a78de62eSJunchao Zhang nsize: 1 40a78de62eSJunchao Zhang 41a78de62eSJunchao Zhang TEST*/ 42