Source compile error related to half precision?

1

pytorch worked fine when I installed with either anaconda or pip.
However, when I tried to build pytorch from the source (python setup.py install), I got the following error messages. (I am using python 2.7.13, gcc 4.8.5, cuda 8.0)

gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/homes/3/kimjook/pytorch -I/homes/3/kimjook/pytorch/torch/csrc -I/homes/3/kimjook/pytorch/torch/lib/tmp_install/include -I/homes/3/kimjook/pytorch/torch/lib/tmp_install/include/TH -I/homes/3/kimjook/pytorch/torch/lib/tmp_install/include/THPP -I/homes/3/kimjook/pytorch/torch/lib/tmp_install/include/THNN -I/u/drspeech/opt/anaconda2/lib/python2.7/site-packages/numpy/core/include -I/usr/local/cuda/include -I/homes/3/kimjook/pytorch/torch/lib/tmp_install/include/THCUNN -I/u/drspeech/opt/anaconda2/include/python2.7 -c torch/csrc/cuda/Storage.cpp -o build/temp.linux-x86_64-2.7/torch/csrc/cuda/Storage.o -D_THP_CORE -std=c++11 -Wno-write-strings -DWITH_NUMPY -DWITH_CUDA -DCUDA_LIB_PATH=/usr/local/cuda/lib64
cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++ [enabled by default]
torch/csrc/nn/THNN_generic.cpp: In function ‘void torch::nn::Abs_updateOutput(thpp::Tensor*, thpp::Tensor*)’:
torch/csrc/nn/THNN_generic.cpp:45:14: error: ‘THCudaHalfTensor’ was not declared in this scope
             (THCudaHalfTensor*)input->cdata(),
              ^


In file included from /homes/3/kimjook/pytorch/torch/lib/tmp_install/include/THPP/tensors/../storages/THCStorage.hpp:11:0,
                 from /homes/3/kimjook/pytorch/torch/lib/tmp_install/include/THPP/tensors/THCTensor.hpp:22,
                 from torch/csrc/DynamicTypes.cpp:11:
/homes/3/kimjook/pytorch/torch/lib/tmp_install/include/THPP/tensors/../storages/../TraitsCuda.hpp:9:20: error: ‘half’ was not declared in this scope
 struct type_traits<half> {
                    ^

It seems that the compiler cannot correctly recognize “half” precision related data structures or classes. How would I reslove this problem? (I don’t need half precision operations so if possible it’s fine for me to disable it.)

Thanks!

2

That’s weird. Could you please clean the install (watch out, if you have any uncommited changes it will delete them!!) git clean -xfd and then run python setup.py install and upload the full log to pastebin?

3

The full log is as follows.

http://pastebin.com/XqETHPNs

Thanks.

4

i’ve looked at the log and I cant really see what the problem is :frowning:

are you using CUDA 8.0.27 by any chance, instead of the latest 8.0.44?

5

I am using CUDA 8.0.61. Thanks.

6

I have no idea what’s wrong. You have a fresh version of CUDA, so half should be automatically enabled everywhere.

Can you check if adding #include <cuda.h> after this line helps?

7

Tried it, but no differences.

8

Ok, are you sure that /usr/local/cuda-8.0 and /usr/local/cuda contain the same CUDA version? I’m pretty sure there’s something wrong with that. All librs are compiled with /usr/local/cuda-8.0 and claim to detect FP16, but not when the extension is compiled.

Another hack would be to add -DCUDA_HAS_FP16=1 here.

9

Oh. That’s it. My system’s /usr/local/cuda is linked to cuda-6.5 since I am sharing the system with others.

I am using CUDA 8 by setting
export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda-8.0/bin:$LD_PATH

Would there be any other way I can explictly specify the directory path?

Thanks!

10

Great! You could use CUDA_HOME env variable.