Installation error: ‘memcpy’ was not declared in this scope

1

When I install the Pytorch package from source, it reports the following problem.

My platform is: Ubuntu 16.06 + CUDA 7.5

[ 1%] Linking CXX shared library libTHCUNN.so
[100%] Built target THCUNN
Install the project…
– Install configuration: “Release”
– Installing: /home/yuhang/pytorch/pytorch-master/torch/lib/tmp_install/lib/libTHCUNN.so.1
– Installing: /home/yuhang/pytorch/pytorch-master/torch/lib/tmp_install/lib/libTHCUNN.so
– Set runtime path of “/home/yuhang/pytorch/pytorch-master/torch/lib/tmp_install/lib/libTHCUNN.so.1” to “”
– Up-to-date: /home/yuhang/pytorch/pytorch-master/torch/lib/tmp_install/include/THCUNN/THCUNN.h
– Up-to-date: /home/yuhang/pytorch/pytorch-master/torch/lib/tmp_install/include/THCUNN/generic/THCUNN.h
– Configuring done
– Generating done
– Build files have been written to: /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl
[100%] Generating lib/libnccl.so
Compiling src/libwrap.cu > /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl/obj/libwrap.o
Compiling src/core.cu > /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl/obj/core.o
Compiling src/all_gather.cu > /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl/obj/all_gather.o
Compiling src/all_reduce.cu > /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl/obj/all_reduce.o
Compiling src/broadcast.cu > /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl/obj/broadcast.o
Compiling src/reduce_scatter.cu > /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl/obj/reduce_scatter.o
Compiling src/reduce.cu > /home/yuhang/pytorch/pytorch-master/torch/lib/build/nccl/obj/reduce.o
ptxas warning : Too big maxrregcount value specified 96, will be ignored
ptxas warning : Too big maxrregcount value specified 96, will be ignored
ptxas warning : Too big maxrregcount value specified 96, will be ignored
ptxas warning : Too big maxrregcount value specified 96, will be ignored
/usr/include/string.h: In function ‘void* __mempcpy_inline(void*, const void*, size_t)’:
/usr/include/string.h:652:42: error: ‘memcpy’ was not declared in this scope
** return (char ) memcpy (__dest, __src, __n) + __n;*
** ^**
/usr/include/string.h: In function ‘void __mempcpy_inline(void, const void*, size_t)’:**
/usr/include/string.h:652:42: error: ‘memcpy’ was not declared in this scope
** return (char ) memcpy (__dest, __src, __n) + __n;*

Note that in Caffe installation, similar problem can be solved in here:

https://groups.google.com/forum/#!msg/caffe-users/Tm3OsZBwN9Q/XKGRKNdmBAAJ

By changing the CMakeLists.txt, I wonder whether we have some similar solutions in Pytorch.

Thanks,
Yuhang

2

Further into this problem I found that it is caused by the version of gcc is too new:

Caffe:
https://github.com/BVLC/caffe/issues/4046
Torch:
https://github.com/szagoruyko/imagine-nn/issues/42

A usual way to solve it is to add a flag as: flags=-D_FORCE_INLINES before compiling. Is there any place I could insert this command installing Pytorch ?

3

CCFLAGS="-D_FORCE_INLINES" CXXFLAGS="-D_FORCE_INLINES" python setup.py install should do it.

4

Thanks for reply but the error is still there. I’m trying ‘CFLAGS’ instead of ‘CCFLAGS’.

5

I tried this:
CCFLAGS="-D_FORCE_INLINES" CFLAGS="-D_FORCE_INLINES" CXXFLAGS="-D_FORCE_INLINES" python setup.py install

But still not working. Might because some other problems…

6

Any update?

Even I’m getting the same error.

7

update your CUDA to 8.0

8

Maybe you can follow this disscussion:

In my case, I am using Ubuntu 16.04 and CUDA 7.5 and I add -D_FORCE_INLINES in CXXFLAGS in file torch/lib/nccl/Makefile.

5912c61fbc84e.png1859×588 115 KB

@shiningsurya @ywu36