[BUG] Need to wrap cub:: and thrust:: namespace in nvcomp namespace
lilohuang opened this issue · 7 comments
Describe the bug
Download nvcomp_2.4.1_x86_64_11.x.tgz from NVIDIA website, and dump symbols from NVCOMP shared objects with nm. I still can see a lot cub:: and thrust:: functions are not wrapped in nvcomp namespace. Please refer to NVIDIA/thrust#1401 to recompile NVCOMP shared objects to prevent crash in CUDA 11.
e.g.
bash-4.2$ nm -CD libnvcomp.so | grep cub 000000000002df70 W void cub::EmptyKernel() 0000000000035950 W void nvcomp::cub::EmptyKernel()
Expected behavior
All cub:: and thrust:: functions must be wrapped in nvcomp namespace.
Thanks for pointing this out and for linking to a page containing the solution! From the post, it looks like it should be fairly straightforward to address, but I'll see what I can do.
This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.
This is of course needed and cannot be closed.
Thanks for pinging this again! It ended up being not quite as simple as it looked at first, and somehow, some other people using both nvCOMP and cub/thrust weren't running into the same issue, so it fell by the wayside for a bit. I'll try asking the person who gave the originally suggested fix. There's probably something simple that I missed.
Yep. There were a couple simple things that I'd missed. I have a change that seems to work, and it should be in a release fairly soon.
@ndickson-nvidia thanks for your feedback. :)
nvCOMP 2.5.1 is now out and should hopefully fix this, but please let us know if it doesn't, and I'll try to investigate further. Thanks again for pointing this out!