CUDA and Matlab on Linux : mexcuda not working

Hi,

I'm triing to use the GPU sparse librairie from http://ch.mathworks.com/matlabcentral/fileexchange/44423-gpu-sparse--accumarray--non-uniform-grid but it is impossible for me to correctly compile any file. I tried with both Matlab2015a and Matlab2015b but i've got some dark errors comming from the cusp library like :

Warning: The selected C++ compiler is not supported for CUDA compilation. Searching for a supported compiler. 
> In mexcuda (line 89) 
Building with 'nvcc'.
Error using mex
In file included from /home/louissarazin/Desktop/lts5usgroup/matlabDev/src/CUsparse/gspmv_coo.cu:44:0:
/usr/local/cuda-7.0/bin/..//include/cusp/blas.h:31:6: warning: #warning | WARNING: cusp/blas.h is deprecated as of 0.4.0; use
cusp/blas/blas.h instead | [-Wcpp]
/usr/local/cuda-7.0/bin/..//include/cusp/system/cuda/detail/multiply/coo_flat_spmv.h(257): error: no operator "=" matches
these operands
            operand types are: volatile thrust::complex<float> = thrust::complex<float>
          detected during:
            instantiation of "void cusp::system::cuda::spmv_coo_flat_kernel<IndexType,ValueType,BLOCK_SIZE>(IndexType,
            IndexType, const IndexType *, const IndexType *, const ValueType *, const ValueType *, ValueType *, IndexType *,
            ValueType *) [with IndexType=int, ValueType=thrust::complex<float>, BLOCK_SIZE=256U]"
(409): here
            instantiation of "void
            cusp::system::cuda::__spmv_coo_flat<InitializeY,DerivedPolicy,MatrixType,VectorType1,VectorType2,UnaryFunction,BinaryFunction1,BinaryFunction2>(thrust::system::cuda::detail::execution_policy<DerivedPolicy>
            &, MatrixType &, VectorType1 &, VectorType2 &, UnaryFunction, BinaryFunction1, BinaryFunction2) [with
            InitializeY=false, DerivedPolicy=thrust::system::cuda::detail::tag, MatrixType=const
            cusp::coo_matrix_view<cusp::array1d_view<thrust::device_ptr<const int>>,
            cusp::array1d_view<thrust::device_ptr<const int>>, cusp::array1d_view<thrust::device_ptr<const
            thrust::complex<float>>>, int, thrust::complex<float>, thrust::device_system_tag>, VectorType1=const
            cusp::array1d_view<thrust::device_ptr<const thrust::complex<float>>>,
            VectorType2=cusp::array1d_view<thrust::device_ptr<thrust::complex<float>>>,
            UnaryFunction=cusp::detail::zero_function<thrust::complex<float>>,
            BinaryFunction1=thrust::multiplies<thrust::complex<float>>,
            BinaryFunction2=thrust::plus<thrust::complex<float>>]"
(462): here

Ok so first it warns me than my compiler is not supported for CUDA compilation. I'm currently using gcc-4.8.2 (or 4.8.4, is it really matter ?). I also tried to compile the simple exemple and this is what I get :

>> mexcuda -v cusp_test.cu
Warning: The selected C++ compiler is not supported for CUDA compilation. Searching for a supported compiler. 
> In mexcuda (line 89) 
Trying MEX options '/usr/local/MATLAB/R2015b/toolbox/distcomp/gpu/extern/src/mex/glnxa64/nvcc_g++.xml'...SUCCESS
mex -largeArrayDims -f /usr/local/MATLAB/R2015b/toolbox/distcomp/gpu/extern/src/mex/glnxa64/nvcc_g++.xml NVCC_FLAGS="" -v cusp_test.cu 
Verbose mode is on.
... Looking for compiler 'nvcc' ...
... Looking for environment variable 'MW_NVCC_PATH' ...No.
... Looking for environment variable 'CUDA_BIN_PATH' ...No.
... Looking for environment variable 'CUDA_PATH' ...No.
... Executing command 'which nvcc' ...Yes ('/usr/local/cuda-7.0/bin/nvcc').
... Executing command '/usr/local/cuda-7.0/bin/nvcc --version | egrep -E 'release 7.0' > /dev/null && echo /usr/local/cuda-7.0/bin/nvcc' ...Yes ('/usr/local/cuda-7.0/bin/nvcc').
... Executing command '/usr/local/cuda-7.0/bin/nvcc --version | egrep -E 'release 7.0' > /dev/null && echo /usr/local/cuda-7.0/bin/nvcc' ...Yes ('/usr/local/cuda-7.0/bin/nvcc').
... Executing command 'which g++' ...Yes ('/usr/bin/g++').
... Executing command 'g++ -print-file-name=libstdc++.so' ...Yes ('/usr/lib/gcc/x86_64-linux-gnu/4.7/libstdc++.so').
Found installed compiler 'nvcc'.
Options file details
-------------------------------------------------------------------
  Compiler location: /usr/local/cuda-7.0/bin/nvcc
  Options file: /usr/local/MATLAB/R2015b/toolbox/distcomp/gpu/extern/src/mex/glnxa64/nvcc_g++.xml
  CMDLINE1 : /usr/local/cuda-7.0/bin/nvcc -c --compiler-options=-D_GNU_SOURCE,-DMATLAB_MEX_FILE -I"/usr/local/MATLAB/R2015b/extern/include" -I"/usr/local/MATLAB/R2015b/simulink/include" -I"/usr/local/MATLAB/R2015b/toolbox/distcomp/gpu/extern/include/" -gencode=arch=compute_20,code=sm_20 -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=\"sm_50,compute_50\" -std=c++11  --compiler-options=-ansi,-fexceptions,-fPIC,-fno-omit-frame-pointer,-pthread -O -DNDEBUG /home/louissarazin/Desktop/lts5usgroup/matlabDev/src/CUsparse/cusp_test.cu -o /tmp/mex_108099938166401_13970/cusp_test.o
  CMDLINE2 : /usr/bin/g++ -pthread -Wl,--no-undefined  -shared -O -Wl,--version-script,"/usr/local/MATLAB/R2015b/extern/lib/glnxa64/mexFunction.map" /tmp/mex_108099938166401_13970/cusp_test.o   -Wl,-rpath-link,/usr/local/MATLAB/R2015b/bin/glnxa64 -L"/usr/local/MATLAB/R2015b/bin/glnxa64" -lmx -lmex -lmat -lm -lmwgpu /usr/local/MATLAB/R2015b/bin/glnxa64/libcudart.so.7.0 -o cusp_test.mexa64
  CMDLINE3 : rm -f /tmp/mex_108099938166401_13970/cusp_test.o
  CXX : /usr/local/cuda-7.0/bin/nvcc
  DEFINES : --compiler-options=-D_GNU_SOURCE,-DMATLAB_MEX_FILE
  MATLABMEX : -DMATLAB_MEX_FILE
  NVCCFLAGS : -gencode=arch=compute_20,code=sm_20 -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=\"sm_50,compute_50\" -std=c++11 
  CXXFLAGS : --compiler-options=-ansi,-fexceptions,-fPIC,-fno-omit-frame-pointer,-pthread
  INCLUDE : -I"/usr/local/MATLAB/R2015b/extern/include" -I"/usr/local/MATLAB/R2015b/simulink/include" -I"/usr/local/MATLAB/R2015b/toolbox/distcomp/gpu/extern/include/"
  CXXOPTIMFLAGS : -O -DNDEBUG
  CXXDEBUGFLAGS : -g
  LDXX : /usr/bin/g++
  LDFLAGS : -pthread -Wl,--no-undefined 
  LDTYPE : -shared
  LINKEXPORT : -Wl,--version-script,"/usr/local/MATLAB/R2015b/extern/lib/glnxa64/mexFunction.map"
  LINKLIBS : -Wl,-rpath-link,/usr/local/MATLAB/R2015b/bin/glnxa64 -L"/usr/local/MATLAB/R2015b/bin/glnxa64" -lmx -lmex -lmat -lm -lmwgpu /usr/local/MATLAB/R2015b/bin/glnxa64/libcudart.so.7.0
  LDOPTIMFLAGS : -O
  LDDEBUGFLAGS : -g
  OBJEXT : .o
  LDEXT : .mexa64
  NVCC : /usr/local/cuda-7.0/bin/nvcc
  GCC : /usr/bin/g++
  CPPLIB_DIR : /usr/lib/gcc/x86_64-linux-gnu/4.7/libstdc++.so
  MATLABROOT : /usr/local/MATLAB/R2015b
  ARCH : glnxa64
  SRC : /home/louissarazin/Desktop/lts5usgroup/matlabDev/src/CUsparse/cusp_test.cu
  OBJ : /tmp/mex_108099938166401_13970/cusp_test.o
  OBJS : /tmp/mex_108099938166401_13970/cusp_test.o 
  SRCROOT : /home/louissarazin/Desktop/lts5usgroup/matlabDev/src/CUsparse/cusp_test
  DEF : /tmp/mex_108099938166401_13970/cusp_test.def
  EXP : cusp_test.exp
  LIB : cusp_test.lib
  EXE : cusp_test.mexa64
  ILK : cusp_test.ilk
  MANIFEST : cusp_test.mexa64.manifest
  TEMPNAME : cusp_test
  EXEDIR : 
  EXENAME : cusp_test
  NVCC_FLAGS : 
  OPTIM : -O -DNDEBUG
  LINKOPTIM : -O
-------------------------------------------------------------------
Building with 'nvcc'.
/usr/local/cuda-7.0/bin/nvcc -c --compiler-options=-D_GNU_SOURCE,-DMATLAB_MEX_FILE -I"/usr/local/MATLAB/R2015b/extern/include" -I"/usr/local/MATLAB/R2015b/simulink/include" -I"/usr/local/MATLAB/R2015b/toolbox/distcomp/gpu/extern/include/" -gencode=arch=compute_20,code=sm_20 -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=\"sm_50,compute_50\" -std=c++11  --compiler-options=-ansi,-fexceptions,-fPIC,-fno-omit-frame-pointer,-pthread -O -DNDEBUG /home/louissarazin/Desktop/lts5usgroup/matlabDev/src/CUsparse/cusp_test.cu -o /tmp/mex_108099938166401_13970/cusp_test.o
/usr/bin/g++ -pthread -Wl,--no-undefined  -shared -O -Wl,--version-script,"/usr/local/MATLAB/R2015b/extern/lib/glnxa64/mexFunction.map" /tmp/mex_108099938166401_13970/cusp_test.o   -Wl,-rpath-link,/usr/local/MATLAB/R2015b/bin/glnxa64 -L"/usr/local/MATLAB/R2015b/bin/glnxa64" -lmx -lmex -lmat -lm -lmwgpu /usr/local/MATLAB/R2015b/bin/glnxa64/libcudart.so.7.0 -o cusp_test.mexa64
Error using mex
/usr/lib/x86_64-linux-gnu/libc_nonshared.a(atexit.oS): In function `atexit':
(.text+0x12): undefined reference to `__cxa_atexit'
collect2: error: ld returned 1 exit status

I tried to change the nvcc_g++ but really successfully. Any clue ?

Thanks,

louis.

Answer 1

[Joss Knight]

I believe this is due to missing C++ libraries that aren't linked by default by mexcuda because they can cause issues with nvcc in CUDA 7.0 and before. Try adding -lstdc++ to your compilation and report back.

[Louis Louis]

Effectively one library was not linked correctly :

>> mexcuda -lc mexGPUExample.cu

does the trick. Thank you !

So now I go back to my first problem which seems to be more relative to CUDA, Thrust and CUSp than to MATLAB :

/usr/local/cuda-7.0/bin/..//include/cusp/system/cuda/detail/multiply/coo_flat_spmv.h(257): error: no operator "=" matches
these operands
            operand types are: volatile thrust::complex<float> = thrust::complex<float>

It should be an incompatibility error between Thrust and Cusp version. Maybe I should also linked these libraries for the compilation ? I will investigate it.