Error in routine cdiaghg_gpu (1): cusolverDnCreate

  1. QE version, or git commit hash if the develop branchis used; Tested on v6.8 and v7.1

  2. input data and pseudopotential files, or (better) links to them; si.scf.in

input command:

mpirun -np 1 pw.x < ./si.scf.in > ./si.scf.out

change -np option to 2~4 also stopped

  1. output file(s), especially if execution time is long. si.scf.out

Also useful:

Hi!

I'm trying to use qe with my gpu machine. My machine setting is,

  • CPU: i9-10980XE
  • RAM: 128gb
  • GPU: rtx3090 24gb x 4ea
  • OS: Ubuntu 20.04.5 LTS

And, I tried with qe v6.8 and v7.1, both are built well (I mean, no errors) but faced same error. My build command is,

module purge
module load /opt/nvidia/hpc_sdk/modulefiles/nvhpc/23.1

cmake -DCMAKE_C_COMPILER=mpicc \
      -DCMAKE_Fortran_COMPILER=mpif90 \
      -DCMAKE_INSTALL_PREFIX=~/.local \
      -DQE_FFTW_VENDOR=Internal \
      -DCMAKE_CUDA_ARCHITECTURES=86 \
      -DQE_ENABLE_CUDA=ON \
      -DQE_ENABLE_MPI_GPU_AWARE=ON \
      ..

cmake --build . --config Release --target install --parallel

What I did for solve but nothing worked are,

  1. tried to turn off mpi support, e.g., QE_ENABLE_MPI=OFF or QE_ENABLE_MPI_GPU_AWARE=OFF etc.
  2. reinstall nvidia driver, hpc libs

Please help ;)

p.s. oh, and if I tried to run the command above, mpi binary always warn me about below.

libibverbs: Warning: couldn't open config directory '/usr/etc/libibverbs.d'.
--------------------------------------------------------------------------
[[60028,1],0]: A high-performance Open MPI point-to-point messaging module
was unable to find any relevant network interfaces:

Module: OpenFabrics (openib)
  Host: user1

Another transport will be used instead, although this may result in
lower performance.

NOTE: You can disable this warning by setting the MCA parameter
btl_base_warn_component_unused to 0.
--------------------------------------------------------------------------
Edited by Yohan LEE