Update package.py #20
Conversation
cuda access issue fix
| if 'cuda' in self.spec: | ||
| env.prepend_path("LD_LIBRARY_PATH", self.spec["cuda"].prefix.lib64) |
There was a problem hiding this comment.
Setting LD_LIBRARY_PATH like this should not be required.
libcudart.so is linked with an rpath so it is found.
I believe it then locates libcuda.so and others via dynamic loading.
In particular libcuda.so is not provided by Spack (beyond the dummy "stub" instance). Rather this lib is taken from the system. On Debian, the system libcuda.so gets found with no special help.
If libcuda.so is provided at "weird" location then LD_LIBRARY_PATH would be needed. Or, likewise if the user is relocating the Spack install area then a/many LD_LIBRARY_PATH settings would be needed.
But in both cases the .prefix value would not supply the appropriate location.
There was a problem hiding this comment.
Ok, so this may have only been an issue on our system on Polaris. It may be the case that it is stored in a weird location. If this is only an issue on our systems, than I agree it should not be necessary. I will poke around a bit and follow up. Thanks!
| @@ -218,5 +218,8 @@ def install(self, spec, prefix): | |||
| python(*ins) | |||
|
|
|||
| def setup_run_environment(self, env): | |||
| env.prepend_path("WIRECELL_PATH", self.prefix.share.wirecell); | |||
| env.prepend_path("WIRECELL_PATH", self.prefix.share.wirecell) | |||
There was a problem hiding this comment.
Setting WIRECELL_PATH by default is a bit controversial. Here's the issue:
Most users/experiments maintain their configs separate from their WCT release. They include the subset of top level files that get installed into share/wirecell/. Setting WIRECELL_PATH here is incomplete and then the user must make sure to prepend their real cfg area to assure any updates that they provide take precedence over those in share/wirecell/. If they instead append, they can get hard to find bugs.
I know FNAL UPS packages do set this env var. I think that is also a net, if a small, badness.
There was a problem hiding this comment.
Thanks, that is good information and I didn't know that was an issue. I wasn't intending to change the behavior of WIRECELL_PATH in this line, I was just removing a semicolon which I thought was unnecessary. Maybe I am underestimating the role of that semi-colon in the problem you described?
There was a problem hiding this comment.
Oh! Duh, sorry. I misread the diff. Yeah, the semicolon is wrong, if only cosmetically.
I suggest we let this PR stew for a while and if we find out there is some benefit from adjusting the LD_LIBRARY_PATH we can revive things.
Cuda access issue fix. Minor issue found a few months ago working with Haiwang on installing wire-cell with gpus available on spack.