Last change
on this file since 2654 was
2654,
checked in by jderouillat, 9 days ago
|
Implement Kokkos kernels compatible with GPU : removing all C++ object from kernels (access to class members).
The behavior of NumTraits?<double>::isNan is not correct on GPU, a solution must be found (temporary solution consists in checking that values are not too large).
Revert some CArray_WF introduction in the initialisation (mask management in CGrid).
For a given process, the CArray_WF class can't address host and device. Clients should be compile with a GPU backend, while servers with a CPU backend (serial, openMP).
Some transfer methods of connectors have been duplicated to manage CArray (initialisation) and CArray_WF (workflow).
A gpu_computing parameter (gpu_computing, 0 by default) has been introduced in the generic_testcase to upload field3D management on GPU.
Other xios_send_field have been temporarily disabled.
|
File size:
542 bytes
|
Line | |
---|
1 | module purge |
---|
2 | |
---|
3 | module load gcc/9.1.0 nvidia-compilers/23.1 cuda/12.1.0 |
---|
4 | |
---|
5 | export KOKKOS_HOME=${LSCEWORK}/KOKKOS/INSTALL/kokkos.traccs.oMP |
---|
6 | export PATH=${KOKKOS_HOME}/bin:${PATH} |
---|
7 | export LD_LIBRARY_PATH=${KOKKOS_HOME}/lib64:${LD_LIBRARY_PATH} |
---|
8 | export CPATH=${KOKKOS_HOME}/include:${CPATH} |
---|
9 | |
---|
10 | module load openmpi/4.1.1-cuda |
---|
11 | export OMPI_CXX=nvcc_wrapper |
---|
12 | module load hdf5/1.12.0-mpi-cuda |
---|
13 | module load netcdf-c/4.7.4-mpi-cuda |
---|
14 | |
---|
15 | export TMP=/dev/shm # to compile on node with enough memory |
---|
16 | export TMPDIR=/dev/shm # to compile on node with enough memory |
---|
Note: See
TracBrowser
for help on using the repository browser.