Having trouble using on AWS

[dela3499]

I've installed the ArrayFire libraries, pip-installed the python wrapper, and followed the directions on the wiki, but I haven't been able to run any code yet.

When I run af.info(), I get the following:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python2.7/site-packages/arrayfire/device.py", line 32, in info
    safe_call(backend.get().af_info())
  File "/usr/local/lib/python2.7/site-packages/arrayfire/util.py", line 75, in safe_call
    raise RuntimeError(to_str(err_str), af_error)
RuntimeError: ('', 501)

I'm on an AWS g2.2xlarge, with the NVIDIA/Amazon Linux image.

Any debugging advice?

[pavanky]

Did you go through this link as part of Post install instructions ? You need to set LD_LIBRARY_PATH to /usr/local/cuda/lib64:/usr/local/cuda/nvvm/lib64

[dela3499]

I did. Perhaps I've just made some dumb mistake like a typographical error. Has anyone else had any trouble?

On Dec 23, 2015, at 7:31 PM, Pavan Yalamanchili notifications@github.com wrote:

Did you go through this link as part of Post install instructions ? You need to set LD_LIBRARY_PATH to /usr/local/cuda/lib64:/usr/local/cuda/nvvm/lib64

—
Reply to this email directly or view it on GitHub.

[pavanky]

@dela3499 Can you echo $LD_LIBRARY_PATH and verify ?

[dela3499]

Sure thing. I'll do it first thing tomorrow morning.

On Dec 23, 2015, at 7:44 PM, Pavan Yalamanchili notifications@github.com wrote:

@dela3499 Can you echo $LD_LIBRARY_PATH and verify ?

—
Reply to this email directly or view it on GitHub.

[dela3499]

So, here's what I've got in my bashrc file:

# Source global definitions
if [ -f /etc/bashrc ]; then
    . /etc/bashrc
fi

# User specific aliases and functions
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/nvvm/lib64:/home/ec2-user/arrayfire-3/lib:$LD_LIBRARY_PATH

# without symlink, still doesn't work
# export LD_LIBRARY_PATH=/opt/nvidia/cuda/lib64:/opt/nvidia/cuda/nvvm/lib64:/home/ec2-user/arrayfire-3/lib:$LD_LIBRARY_PATH

My CUDA libraries are in /opt/nvidia/cuda, rather than /usr/local/cuda, so I've put a symbolic link in place with

sudo ln -s /opt/nvidia/cuda /usr/local/cuda

[pavanky]

@dela3499 Can you tell me what the output of ldd libafcuda.so from arrayfire-3/lib ?

[dela3499]

It looks like I've got the CUDA 6.5 toolkit, and AF needs 7.5. I suppose I just need to install the latest CUDA toolkit?

[ec2-user@ip-172-31-14-119 lib]$ ldd libafcuda.so
    linux-vdso.so.1 =>  (0x00007ffc63592000)
    libcudart.so.7.5 => not found
    libcublas.so.7.5 => not found
    libcusolver.so.7.5 => not found
    libcufft.so.7.5 => not found
    libnvvm.so.3 => not found
    libcuda.so.1 => /usr/lib64/libcuda.so.1 (0x00007fc1fc36a000)
    libforge.so => /home/ec2-user/arrayfire-3/lib/libforge.so (0x00007fc1fc0bd000)
    libGL.so.1 => /usr/lib64/libGL.so.1 (0x00007fc1fbd70000)
    libGLU.so.1 => not found
    libstdc++.so.6 => /usr/lib64/libstdc++.so.6 (0x00007fc1fba6b000)
    libm.so.6 => /lib64/libm.so.6 (0x00007fc1fb768000)
    libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007fc1fb552000)
    libc.so.6 => /lib64/libc.so.6 (0x00007fc1fb190000)
    libdl.so.2 => /lib64/libdl.so.2 (0x00007fc1faf8b000)
    libpthread.so.0 => /lib64/libpthread.so.0 (0x00007fc1fad6f000)
    librt.so.1 => /lib64/librt.so.1 (0x00007fc1fab67000)
    libfontconfig.so.1 => /usr/lib64/libfontconfig.so.1 (0x00007fc1fa931000)
    libfreetype.so.6 => /usr/lib64/libfreetype.so.6 (0x00007fc1fa695000)
    /lib64/ld-linux-x86-64.so.2 (0x0000559111836000)
    libglfw.so.3 => not found
    libnvidia-tls.so.340.32 => /usr/lib64/tls/libnvidia-tls.so.340.32 (0x00007fc1fa491000)
    libnvidia-glcore.so.340.32 => /usr/lib64/libnvidia-glcore.so.340.32 (0x00007fc1f787e000)
    libX11.so.6 => /usr/lib64/libX11.so.6 (0x00007fc1f7543000)
    libXext.so.6 => /usr/lib64/libXext.so.6 (0x00007fc1f7330000)
    libexpat.so.1 => /lib64/libexpat.so.1 (0x00007fc1f7107000)
    libxcb.so.1 => /usr/lib64/libxcb.so.1 (0x00007fc1f6ee9000)
    libXau.so.6 => /usr/lib64/libXau.so.6 (0x00007fc1f6ce6000)

[ec2-user@ip-172-31-14-119 lib64]$ pwd
/opt/nvidia/cuda/lib64
[ec2-user@ip-172-31-14-119 lib64]$ ls
be                   libcudart.so.6.5.14  libcuinj64.so.6.5      libnppc.so         libnpps_static.a
gfec                 libcudart_static.a   libcuinj64.so.6.5.14   libnppc.so.6.5     libnvblas.so
inline               libcufft.so          libculibos.a           libnppc.so.6.5.14  libnvblas.so.6.5
libcublas_device.a   libcufft.so.6.5      libcurand.so           libnppc_static.a   libnvblas.so.6.5.14
libcublas.so         libcufft.so.6.5.14   libcurand.so.6.5       libnppi.so         libnvToolsExt.so
libcublas.so.6.5     libcufft_static.a    libcurand.so.6.5.14    libnppi.so.6.5     libnvToolsExt.so.1
libcublas.so.6.5.14  libcufftw.so         libcurand_static.a     libnppi.so.6.5.14  libnvToolsExt.so.1.0.0
libcublas_static.a   libcufftw.so.6.5     libcusparse.so         libnppi_static.a   libOpenCL.so
libcudadevrt.a       libcufftw.so.6.5.14  libcusparse.so.6.5     libnpps.so         libOpenCL.so.1
libcudart.so         libcufftw_static.a   libcusparse.so.6.5.14  libnpps.so.6.5     stubs
libcudart.so.6.5     libcuinj64.so        libcusparse_static.a   libnpps.so.6.5.14

[pavanky]

@dela3499 you can build arrayfire with 6.5 if you cant upgrade cuda.

Additionally, you will also need to set AF_PATH to /opt/arrayfire-3 to get it to pick up the libraries.

I fixed the issue in the python wrapper to not require the path, but it looks like the upstream libraries seem to require it.

[dela3499]

Thanks for the help!

[pavanky]

@dela3499 btw I am going to update the package in an hour or so that should fix some of the issues.