cuda fails to build when c++11

[soumith]

Hey,

I am trying to build nnForge on my machine: Ubuntu 14.04 64-bit with CUDA 6.0

I am running into two different issues: (1) When CPP11COMPILER=yes the CUDA modules fail to build with the error:

nvcc -c absolute_layer_hessian_cuda.cu -use_fast_math -DBOOST_NOINLINE='__attribute__ ((noinline))' -O3 -DENABLE_CUDA_PROFILING -Xcompiler="-I/usr/local/include -I/usr/local/cuda/include -I/usr/local/include -ffast-math -march=native -mfpmath=sse -msse2  -O3 -DENABLE_CUDA_PROFILING" -gencode=arch=compute_20,code=sm_20 -gencode=arch=compute_30,code=sm_30 -gencode=arch=compute_35,code=\"sm_35,compute_35\" -o absolute_layer_hessian_cuda.o
In file included from /usr/include/c++/4.8/random:35:0,
                 from ../nn_types.h:20,
                 from ../layer_data.h:20,
                 from ../layer.h:21,
                 from layer_hessian_cuda.h:19,
                 from absolute_layer_hessian_cuda.h:19,
                 from absolute_layer_hessian_cuda.cu:17:
/usr/include/c++/4.8/bits/c++0x_warning.h:32:2: error: #error This file requires compiler and library support for the ISO C++ 2011 standard. This support is currently experimental, and must be enabled with the -std=c++11 or -std=gnu++11 compiler options.
 #error This file requires compiler and library support for the \

(2) With CPP11COMPILER=no the core library itself fails to build with the error:

g++ -I/usr/local/include/boost/tr1/tr1 -I/usr/local/include -I/usr/local/include -ffast-math -march=native -mfpmath=sse -msse2  -O3 -DENABLE_CUDA_PROFILING   -c -o absolute_layer.o absolute_layer.cpp
In file included from /usr/include/c++/4.8/random:35:0,
                 from nn_types.h:20,
                 from layer_data.h:20,
                 from layer.h:21,
                 from absolute_layer.h:19,
                 from absolute_layer.cpp:17:
/usr/include/c++/4.8/bits/c++0x_warning.h:32:2: error: #error This file requires compiler and library support for the ISO C++ 2011 standard. This support is currently experimental, and must be enabled with the -std=c++11 or -std=gnu++11 compiler options.
 #error This file requires compiler and library support for the \
  ^
In file included from layer_data.h:20:0,
                 from layer.h:21,
                 from absolute_layer.h:19,
                 from absolute_layer.cpp:17:
nn_types.h:41:25: error: ‘tr1’ in namespace ‘std’ does not name a type
 #define nnforge_mt19937 std::tr1::mt19937
                         ^
rnd.h:23:10: note: in expansion of macro ‘nnforge_mt19937’
  typedef nnforge_mt19937 random_generator;
          ^

[soumith]

This is my Settings.mk

--- a/Settings.mk
+++ b/Settings.mk
@@ -1,18 +1,18 @@
 BUILD_MODE=release
 ENABLE_CUDA_BACKEND=yes
-ENABLE_CUDA_PROFILING=no
-CPP11COMPILER=no
+ENABLE_CUDA_PROFILING=yes
+CPP11COMPILER=yes
 BOOST_PATH=/usr/local
 OPENCV_PATH=/usr/local
-NETCDF_INSTALLED=yes
+NETCDF_INSTALLED=no
 NETCDF_PATH=
-MATIO_INSTALLED=yes
+MATIO_INSTALLED=no
 MATIO_PATH=
 CUDA_PATH=/usr/local/cuda
 NVCC=nvcc
 NNFORGE_PATH=../..
-NNFORGE_INPUT_DATA_PATH=/home/max/nnforge/input_data
-NNFORGE_WORKING_DATA_PATH=/home/max/nnforge/working_data
+NNFORGE_INPUT_DATA_PATH=./nnforge/input_data
+NNFORGE_WORKING_DATA_PATH=./nnforge/working_data

[soumith]

If you've seen these issues before, or if there's a standard script to install nnForge with CUDA support for Ubuntu, please let me know.

Thanks!

[soumith]

My main goal is to benchmark all the ConvNet libraries with reproducible instructions/scripts. Located here: https://github.com/soumith/convnet-benchmarks

[milakov]

Hi Soumith,

CPP11COMPILER=no # CUDA doesn't support C++11 yet BOOST_PATH=/usr # If you install Boost through deb packages they are here

This is not the first time developers encounter this problem. It looks like I should set BOOST_PATH=/usr in Settings.mk

[soumith]

Thank you!

[soumith]

Also, the boost libraries are no longer post-fixed by "-mt" in the latest versions, they are built multi-threaded by default, had to change that as well, fyi.

[milakov]

Yeah, thanks, good point.

[milakov]

This issue is fixed with CUDA 6.5: you can now build CUDA backend with -std=c++11

If you see AVX related errors when CPP11COMPILER=yes then set CPP_HW_ARCHITECTURE=-march=corei7 This workaround should work.