(WIP) Issues emerging after gpu-develop is merged
List of issues to be considered after merging gpu-develop.
-
If one wants to use the old CI of QEF/q-e-gpu testing various architectures, runners should be added and default CI file changed. -
Additional message in the output (fft_base.f90 and inside fft_helper_functions in FFTXlib)
IF ( .not. desc%use_pencil_decomposition ) WRITE( stdout,*) ' Using Slab Decomposition'
IF ( desc%use_pencil_decomposition ) WRITE( stdout,*) ' Using Pencil Decomposition'
-
pencil_decomposition_ option is undocumented (line 130 in command_line_options) -
two control flags are undocumented use_gpu
andmany_fft
(line 253 control_flags.f90) -
cuda_subroutines.f90
should be removed. -
workaround in matrix_inversion
in file invmat.f90 -
Undocumented #if defined (__CUDA_OPTIMIZED)
in mp_glibal.f90. Never received confirmation that band parallelism can be safely disabled in "standard SCF" part when doing EXX. -
Accelerated random number generation is disabled and there is no option to enable it (See line 218 wfc_init_gpu) -
device_helper.f90 should be removed from Utilxlib. -
output messages in verbose mode may included unwanted (debug?) information (line 586 electrons.f90) -
CPV source code, PW/src/Coul_cut_2D.f90
and PW/src/exx.f90 do not follow the rule of having separated source files for accelerated subroutines and standard ones. -
there is not point in keeping separate makefiles for GPU. Merge those make.gpu
. -
for some reason OMP is commented in dmxc_lda
qe_drivers_d_lda_lsda.f90 -
in exx_bands.f90 MPI_RECV has been (correctly) moved before MPI_isend -
undocumented use of non standard input option "USEGPU=[yes|no]" read through environment variable.