dynamic load libnvrtc.so

[leezu]

Description

Compiling MXNet with nvrtc support dynamically links libcuda.so. However, libcuda.so is part of the cuda driver and not typically available on cpu machines. This means, on cpu machines libmxnet.so will have missing dependencies and can't be loaded.

Unfortunately, parts of MXNet compilation require to dlopen(libmxnet.so). In particular OpWrapperGenerator.py for the cpp package. This breaks compiling mxnet with nvrtc + cpp-package on cpu machines (as used by our CI).

As a workaround, the stub driver can be added to LD_LIBRARY_PATH. A better solution may be to dlopen libnvrtc and thereby drop the dependency on libcuda.so.

CC: @ptrendx

References

• dlopen libnvrtc instead of dynamically linking against it pytorch/pytorch#3395
• Dynamic load of nvidia libs NervanaSystems/ngraph#765

[ptrendx]

I agree it makes sense to dlopen it.

[ChaiBapchya]

@mxnet-label-bot add [bug]
This work-around [adding to LD_LIBRARY_PATH] is great. But without the workaround, the mxnet breaks. So I guess, it needs to be fixed and is a bug rather than a feature-request.

[ptrendx]

Hi @mseth10, do you think the utilities that you introduced in dynamic library loading support PR could be used for this?

[mseth10]

Hi @ptrendx if you need a generic dlopen for this, I'll suggest using ctypes CDLL.

[ptrendx]

But that would be just for Python then, right? I would need it to call functions from C++ and the load_lib function from your PR seems like it already handles all the differences in OSs and stuff.

[mseth10]

Sorry, I thought you were referring to the python api introduced by the PR. That has a lot of additional functionality.
We can use the c++ utility function - lib_load https://github.com/apache/incubator-mxnet/blob/9d440868603ad26b702e12ddd2587e5c4b56e42b/src/initialize.cc#L113

[ptrendx]

I thought about this and the real problem seems to actually be libcuda.so, right? libnvrtc.so does not depend on libcuda.so directly, it is just that to actually use the kernels compiled with libnvrtc.so one needs to use the driver API. So it should be fine to link normally to libnvrtc.so, and just dlopen libcuda.so, right @leezu?