CUSolver implementation fails, EIGENSOLVER_GPU works
Description of the problem:
The same input file works with EIGENSOLVER_GPU, fails when using CU_SOLVER interface (real valued case).
Please fill the points below:
-
Version of the code: commit 46d2a63b
-
Compilers and libraries used: PGI compiler, CUDA 10.1
-
has the code been build with configure or handcrafted
make.inc
?
make.inc has been modified to compile CU_SOLVER interface.
- Is the problem reproducible?
Attached input and output files (showing and not showing errors).
I uploaded the Hamiltonian and overlap H and S (700 Mb each) that makes CU_solver fail here:
http://www.fis.unipr.it/home/pietro.bonfa/bug/20208_H.dat
http://www.fis.unipr.it/home/pietro.bonfa/bug/20208_S.dat
These two files are generated by the following code:
call get_environment_variable("SLURM_JOB_ID", jname)
call get_environment_variable("SLURM_ARRAY_TASK_ID",tname)
fname = trim(jname) // trim(tname) //"_H.dat"
OPEN(UNIT=13, FILE=fname, ACTION="write", STATUS="replace", FORM="unformatted")
fname = trim(jname) // trim(tname) // "_S.dat"
OPEN(UNIT=14, FILE=fname, ACTION="write", STATUS="replace", FORM="unformatted")
allocate(h_bkp_h, source=h_bkp_d)
allocate(s_bkp_h, source=s_bkp_d)
write(13), n, m, ldh
write(13) h_bkp_h
write(14), n, m, ldh
write(14) s_bkp_h
close(13)
CLOSE(UNIT=14)
Edited by Pietro