Segfault when cuda_mem.cu frees memory

[luca]

Hi,

I wished to ask the VASP user community if anybody using VASP on GPUs (Tesla K20X on the Cray XC30 in my case) had a similar issue.

When running the simple CeO2 test of Peter Larsson's test suite on a single core, I got a segmentation fault at line 71 of cuda_mem.cu, when the code tries to free the memory on the device (#else free(*ptr);)
Process 0:
Thread 1 stopped in free from /lib64/libc.so.6 with signal SIGSEGV ​(Segmentation fault)​.
Reason/Origin: address not mapped to object ​(attempt to access invalid address)​

I needed to modify cuda_mem.cu, since the original one with the define statements for NVREGISTERSELF and NVPINNED was giving other error "Failed to free pinned memory!” coded in cuda_mem.cu in the lines immediately before the statement on line 71.
I use the INCAR provided by the VASP test suite, with NCORE=1 and without defining NPAR.

I would appreciate any advice from the VASP community.

Thanks,
Luca

[guiyang_huang1]

Try to remove pinned memory option from the makefile.include.

Or use pgi compiler to compile the gpu vasp.

For gnu or xl compiler, it seems pined memory option can not be used for IBM power9. It is unclear whether it is also true for your system.