Here we provide examples of cuSOLVERMp library API usage
Distributed decompositions and linear system solutions
Dense matrix LU factorization and linear system solve
Dense matrix Cholesky factorization and linear system solve
Dense matrix Symmetric Eigensolver
Dense matrix Symmetric Generalized Eigensolver
Dense matrix QR factorization and linear system solve
Examples are bootstrapped by MPI and use it to set up distributed data. Those examples are intended just to show how API is used and not for performance benchmarking. For same reasons process grid is hardcoded to 2x1 in the examples, however you can change it to other values in following lines:
/* Define grid of processors */
const int numRowDevices = 2;
const int numColDevices = 1;
Based on your distributed setup you can choose how your GPU devices are mapped to processes - change following line in the example to suit your needs:
const int localDeviceId = getLocalRank();
In these samples each process will use CUDA device ID equal to the local MPI rank ID of the process.
Linux
x86_64
cuSOLVERMp is distributed through NVIDIA Developer Zone and also as a part of HPC SDK. cuSOLVERMp requires CUDA Toolkit, HPC-X, NCCL and GDRCOPY to be installed on the system. The samples require c++11 compatible compiler.
Build examples using make command:
make
Run examples with mpi command and number of processes according to process grid values, i.e.
mpirun -n 2 ./mp_getrf_getrs
mpirun -n 2 ./mp_potrf_potrs
mpirun -n 2 ./mp_syevd
mpirun -n 2 ./mp_sygvd
mpirun -n 2 ./mp_geqrf
mpirun -n 2 ./mp_gels