libtorch with xeus-cling results in "symbol '__emutls_v._ZSt11__once_call' unresolved while linking"

[justinblaber]

Some functionality seems to work, like torch::zeros. However, if I try to use torch::from_blob I get the following errors:

IncrementalExecutor::executeFunction: symbol '__emutls_v._ZSt11__once_call' unresolved while linking [cling interface function]!
IncrementalExecutor::executeFunction: symbol '__emutls_v._ZSt15__once_callable' unresolved while linking [cling interface function]!

Any ideas how I can go about solving this? I'd really like to be able to use libtorch with xeus-cling... but its proving to be very difficult. I manually built this pytorch, but I've also tried downloading the pre-compiled nightly builts. I was even able to do some very hacky stuff to get libtorch.so to compile from the cling environment, but the above two errors still appear...

[JohanMabille]

You probably need to load the library (and tell xeus-cling where to find it). See here.

[justinblaber]

@JohanMabille I've added all the relevant libtorch librarys (library/include paths and loaded the libtorch library) to my knowledge.

I checked and emutls is something inside libgcc or something? Which makes me think the problem is more deeply rooted.

There are some other similar issues:

https://root-forum.cern.ch/t/unresolved-symbol-emutls-v-zst11-once-call/30595/2 https://github.com/jupyter-xeus/xeus-cling/issues/262

I tried using pre-built binaries, recompiling from source, lots of stuff really. Nothing seems to work.

If libtorch works well with xeus-cling that would be pretty amazing for pytorch users that want to experiment with deploying in C++...

[cauachagas]

It's a cling issue .

Possible solution https://github.com/llvm/llvm-project/commit/8eec91e96d4a03ee1a3287405dda53e60d0c9933

I don't use libtorch with xeus-cling, but it works fine. Only from_blob does not work (I think it works on OSX).

@justinblaber, correct me if I'm wrong, with from_blob you point the address (an OpenCV matrix, for example) instead of making a copy to initialize a Tensor? I don't know what your case is, but you can initialize a Tensor with std::vector, for example, and then reshape. In cell 7 and 8 of this notebook there is an example of how to do it.

[justinblaber]

@cauachagas Do you know how I can use that solution you linked to fix the issue? I'm not familiar with llvm/cling so I don't know where to begin...

But, I'm basically using it because I have a memory buffer (i.e. just a pointer) that stores segmentation results from tensorrt. I wanna do an argmax operation on it, so I was just gonna from_blob() it then use libtorch to do the argmax.

I think this is pretty important functionality because I could see a lot of use cases where you have a memory buffer and want to use a libtorch operation on it without having to reallocate it or something.

[cauachagas]

Cling uses LLVM5 with custom Clang. The possible solution uses version 6 of the LLVM. If this is the only solution, then, in theory, you can solve it using a more updated version of LLVM.

I may be talking nonsense, but I believe that this problem does not happen on OSX, because it natively uses Clang. For Linux, the build of the custom Clang is done with GCC.

I use Arch Linux and tried to compile Cling using LLVM5 and Clang5 from the arch4edu repository. I told Cmake the path of the variable LLVM_CONFIG among other things. He compiled everything, but at the end, the cling gives error 😢 . Apparently, Clang has to be provided here

git clone --depth=1 -b "cling-patches" http://root.cern.ch/git/clang.git

Update:

See this script from AUR

https://aur.archlinux.org/cgit/aur.git/tree/PKGBUILD?h=cling

Do this

cd ~/
srcdir=$HOME
git clone --depth=1 -b "cling-patches" http://root.cern.ch/git/clang.git
git clone --depth=1 -b "cling-patches" http://root.cern.ch/git/llvm.git
git clone --depth=1 -b "v0.7" http://root.cern.ch/git/clang.git

and then follow as in the script and modify the necessary files https://github.com/llvm/llvm-project/commit/8eec91e96d4a03ee1a3287405dda53e60d0c9933