[Trilinos-Users] Tpetra and Kokkos with CUDA backend

[Radu Popescu]

Hello,

I am trying to compile Trilinos 10.4 with Tpetra and Kokkos enabled. I
am interested in testing Tpetra with the Kokkos backend, on GPU and
hybrid nodes.
The process fails when nvcc is compiling CUDA files in Kokkos (error:
calling a host function from a __device__/__global__ function is not
allowed).
Is there any sort of extra configuration required for Trilinos when using nvcc?

The following is the setup I am using:

Linux 64bit

Trilinos 10.4 - Enabled packages: Epetra, EpetraExt, Tpetra, Teuchos,
Kokkos, ThreadPool, AztecOO, Amesos, Ifpack, ML, Zoltan, Triutils,
Didasko

Third-party libraries for Trilinos: UMFPACK, Pthread, GotoBLAS,
LAPACK, MPI, ParMETIS, HDF5, CUDA, Thrust

gcc 4.4.4

NVIDIA developer driver 256.40
CUDA 3.1 - compiled examples in SDK for both C for CUDA and OpenCL
with gcc 4.4.4 - no problems reported, and no extra configuration
required (no -fpermissive compiler option required)

Thrust 1.2.1 - tested and works, without any extra configuration

Thank you.

Regards,
Radu Popescu