Failed to complile MEX-Functions Containing CUDA Code

I am trying to compile the MATLAB provided example file mexGPUExample.cu following:

http://www.mathworks.com/help/distcomp/run-mex-functions-containing-cuda-code.html

The "Set Up for MEX-File Compilation part" was fine. I copied the mex_CUDA_glnxa64.xml file to the current folder and run setenv command.

Then I start to compile mexGPUExample.cu by executing

    mex -largeArrayDims mexGPUExample.cu

After it run for a few seconds it got the error message:

Building with 'nvcc'. 
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

My Machine is Ubuntu 14.04 with gcc 4.8.2 and g++ 4.8.2 The output of gpuDevice is attach in the file.

In order to compatible with mex compile requirement(matlab only support gcc-4.7.x), I set the system default compiler as gcc 4.7 and g++ 4.7 use the method described below:

How to change the default GCC compiler in Ubuntu?

and my output of gpuDevice is: CUDADevice with properties:

                  Name: 'GeForce GTX 650'
                 Index: 1
     ComputeCapability: '3.0'
        SupportsDouble: 1
         DriverVersion: 6.5000
        ToolkitVersion: 6
    MaxThreadsPerBlock: 1024
      MaxShmemPerBlock: 49152
    MaxThreadBlockSize: [1024 1024 64]
           MaxGridSize: [2.1475e+09 65535 65535]
             SIMDWidth: 32
           TotalMemory: 1.0734e+09
       AvailableMemory: 490971136
   MultiprocessorCount: 2
          ClockRateKHz: 1110500
           ComputeMode: 'Default'
  GPUOverlapsTransfers: 1
KernelExecutionTimeout: 1
      CanMapHostMemory: 1
       DeviceSupported: 1
        DeviceSelected: 1
When I searching for the solutions I only found 2 posts waiting for answer desperately. Lots of thanks to any one who can give a hand. >"<

[Joss Knight]

I presume you are running in MATLAB R2014b?

Please can you compile with the -v verbose flag and provide the full output.

mex -v -largeArrayDims mexGPUExample.cu

It looks like whatever you did to set your compiler didn't work because your compiler has linked against libstdc++. One workaround you might try is to add that library (-lstdc++).

[Shaojun]

Yes my matlab version is 2014b the output is:

    Verbose mode is on.
... Looking for compiler 'nvcc' ...
... Looking for environment variable 'MW_NVCC_PATH' ...No.
... Executing command 'which nvcc' ...Yes ('/usr/local/cuda-6.5/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-6.5/bin/nvcc
  Options file: /home/shaojun/projects/voice/mex_CUDA_glnxa64.xml
  CMDLINE1 : /usr/local/cuda-6.5/bin/nvcc -c --compiler-options=-D_GNU_SOURCE,-DMATLAB_MEX_FILE -I"/home/shaojun/tools/MATLAB2014b/extern/include" -I"/home/shaojun/tools/MATLAB2014b/simulink/include" -I"/home/shaojun/tools/MATLAB2014b/toolbox/distcomp/gpu/extern/include/" -gencode=arch=compute_20,code=sm_20 -gencode=arch=compute_30,code=sm_30 -gencode=arch=compute_50,code=\"sm_50,compute_50\" $NVCC_FLAGS --compiler-options=-ansi,-fexceptions,-fPIC,-fno-omit-frame-pointer,-pthread -O -DNDEBUG /home/shaojun/projects/voice/mexGPUExample.cu -o /tmp/mex_6677493866400559_14758/mexGPUExample.o
  CMDLINE2 : /usr/bin/g++ -pthread -Wl,--no-undefined  -shared -O -Wl,--version-script,"/home/shaojun/tools/MATLAB2014b/extern/lib/glnxa64/mexFunction.map" /tmp/mex_6677493866400559_14758/mexGPUExample.o   -Wl,-rpath-link,/home/shaojun/tools/MATLAB2014b/bin/glnxa64 -L"/home/shaojun/tools/MATLAB2014b/bin/glnxa64" -lmx -lmex -lmat -lm -lmwgpu /home/shaojun/tools/MATLAB2014b/bin/glnxa64/libcudart.so.6.0 -o mexGPUExample.mexa64
  CMDLINE3 : rm -f /tmp/mex_6677493866400559_14758/mexGPUExample.o
  CXX : /usr/local/cuda-6.5/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_30,code=sm_30 -gencode=arch=compute_50,code=\"sm_50,compute_50\" $NVCC_FLAGS
  CXXFLAGS : --compiler-options=-ansi,-fexceptions,-fPIC,-fno-omit-frame-pointer,-pthread
  INCLUDE : -I"/home/shaojun/tools/MATLAB2014b/extern/include" -I"/home/shaojun/tools/MATLAB2014b/simulink/include" -I"/home/shaojun/tools/MATLAB2014b/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,"/home/shaojun/tools/MATLAB2014b/extern/lib/glnxa64/mexFunction.map"
  LINKLIBS : -Wl,-rpath-link,/home/shaojun/tools/MATLAB2014b/bin/glnxa64 -L"/home/shaojun/tools/MATLAB2014b/bin/glnxa64" -lmx -lmex -lmat -lm -lmwgpu /home/shaojun/tools/MATLAB2014b/bin/glnxa64/libcudart.so.6.0
  LDOPTIMFLAGS : -O
  LDDEBUGFLAGS : -g
  OBJEXT : .o
  LDEXT : .mexa64
  NVCC : /usr/local/cuda-6.5/bin/nvcc
  GCC : /usr/bin/g++
  CPPLIB_DIR : /usr/lib/gcc/x86_64-linux-gnu/4.7/libstdc++.so
  MATLABROOT : /home/shaojun/tools/MATLAB2014b
  ARCH : glnxa64
  SRC : /home/shaojun/projects/voice/mexGPUExample.cu
  OBJ : /tmp/mex_6677493866400559_14758/mexGPUExample.o
  OBJS : /tmp/mex_6677493866400559_14758/mexGPUExample.o 
  SRCROOT : /home/shaojun/projects/voice/mexGPUExample
  DEF : /tmp/mex_6677493866400559_14758/mexGPUExample.def
  EXP : mexGPUExample.exp
  LIB : mexGPUExample.lib
  EXE : mexGPUExample.mexa64
  ILK : mexGPUExample.ilk
  MANIFEST : mexGPUExample.mexa64.manifest
  TEMPNAME : mexGPUExample
  EXEDIR : 
  EXENAME : mexGPUExample
  OPTIM : -O -DNDEBUG
  LINKOPTIM : -O
-------------------------------------------------------------------
Building with 'nvcc'.
/usr/local/cuda-6.5/bin/nvcc -c --compiler-options=-D_GNU_SOURCE,-DMATLAB_MEX_FILE -I"/home/shaojun/tools/MATLAB2014b/extern/include" -I"/home/shaojun/tools/MATLAB2014b/simulink/include" -I"/home/shaojun/tools/MATLAB2014b/toolbox/distcomp/gpu/extern/include/" -gencode=arch=compute_20,code=sm_20 -gencode=arch=compute_30,code=sm_30 -gencode=arch=compute_50,code=\"sm_50,compute_50\" $NVCC_FLAGS --compiler-options=-ansi,-fexceptions,-fPIC,-fno-omit-frame-pointer,-pthread -O -DNDEBUG /home/shaojun/projects/voice/mexGPUExample.cu -o /tmp/mex_6677493866400559_14758/mexGPUExample.o
/usr/bin/g++ -pthread -Wl,--no-undefined  -shared -O -Wl,--version-script,"/home/shaojun/tools/MATLAB2014b/extern/lib/glnxa64/mexFunction.map" /tmp/mex_6677493866400559_14758/mexGPUExample.o   -Wl,-rpath-link,/home/shaojun/tools/MATLAB2014b/bin/glnxa64 -L"/home/shaojun/tools/MATLAB2014b/bin/glnxa64" -lmx -lmex -lmat -lm -lmwgpu /home/shaojun/tools/MATLAB2014b/bin/glnxa64/libcudart.so.6.0 -o mexGPUExample.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

[Joss Knight]

What is the output of !/usr/bin/g++ --version?

Answer 1

[Joss Knight]

The problem appears to be that you're using a more recent version of the CUDA Toolkit than is supported by MATLAB. You are using version 6.5, whereas R2014b only supports version 6.0. You are therefore using an incompatible version of nvcc.

Make sure you've got CUDA 6.0 installed, then set MW_NVCC_PATH to point to it:

>> setenv MW_NVCC_PATH /usr/local/cuda-6.0/bin

Then try again.

[Samarth]

Thanks for your suggestion. I installed CUDA 6.0 and modified MW_NVCC_PATH, but the error persists. Do you have any more ideas?

Thanks!

Answer 2

[Kai]

I met withe exactly the same problem. Did you finally figure out a solution?

Thanks in advance!

Answer 3

[Jan Barowski]

I got the same error using Matlab 2015 Prerelease on Ubuntu 14.04 with CUDA 6.5 and g++4.8. Thanks to Matlab support we figured out, that manually linking does the trick. Here is my compile script for the matlab mex/gpu example:

myArch = computer('arch'); pathToOpts = fullfile(matlabroot, ... 'toolbox', 'distcomp', 'gpu', ... 'extern', 'src', 'mex', myArch,'gcc',... ['mex_CUDA_' myArch '.xml']); copyfile(pathToOpts,'.','f')

mex -v -largeArrayDims mexGPUExample.cu -lstdc++ -lc

Hope this helps!