VMD-L Mailing List

From: John Stone (johns_at_ks.uiuc.edu)
Date: Tue Mar 04 2014 - 15:39:26 CST

Joshua,
  I suspect that the intent of the SDK directions there is make it easy
to run the CUDA samples, but the unintended consequence is that it
will break _all_ apps like VMD that do ship with their own .dylib,
if that version is not precisely the same rev that was installed
with the toolkit.

Using the static libcudart will cure this, but that requires using
CUDA 5.5, which obsoletes older revs of MacOS, and the versions
of OptiX that I've had the most mileage on were based on CUDA 5.0
which doesn't have the static lib, so it's always tricky with
these library versioning issues.

I'm going to try a VMD MacOS build for CUDA 5.5 using the static CUDA libs
and see how far I get with that as one of the options in the early
beta revs of VMD 1.9.2.

Cheers,
  John Stone
  vmd_at_ks.uiuc.edu

On Tue, Mar 04, 2014 at 12:01:15PM -0700, Joshua Anderson wrote:
> Yes, the NVIDIA docs do say that - but they also tell application developers to bundle libcudart and link to the bundled lib so that users do not need to install the CUDA toolkit.
>
> I haven't checked recently, but you may need DYLD_LIBRARY_PATH set to run simple applications built by "nvcc -o output input.cu". It was needed back in the CUDA 1.0 days. Executables built by cmake link insert an rpath to the lib and do not need a DYLD_LIBRARY_PATH set.
> ------
> Joshua A. Anderson, Ph.D.
> Research Area Specialist, Chemical Engineering, University of Michigan
> Phone: 734-647-8244
> http://www-personal.umich.edu/~joaander/
>
> On Mar 4, 2014, at 10:09 AM, Phillips, Carolyn L. wrote:
>
> > I removed it from my .bashrc, and indeed now VMD starts up without a problem!
> >
> > This is the opposite of what NVIDA advises you to do, btw? curious
> >
> > http://docs.nvidia.com/cuda/cuda-getting-started-guide-for-mac-os-x/index.html
> >
> > ? Define the environment variables.
> > ? The PATH variable needs to include /Developer/NVIDIA/CUDA-5.5/bin
> > ? DYLD_LIBRARY_PATH needs to contain /Developer/NVIDIA/CUDA-5.5/lib
> > To change the environment variables for 32-bit operating systems:
> >
> > export PATH=/Developer/NVIDIA/CUDA-5.5/bin:$PATH
> > export DYLD_LIBRARY_PATH=/Developer/NVIDIA/CUDA-5.5/lib:$DYLD_LIBRARY_PATH
> >
> > To change the environment variables for 64-bit operating systems:
> >
> > export PATH=/Developer/NVIDIA/CUDA-5.5/bin:$PATH
> > export DYLD_LIBRARY_PATH=/Developer/NVIDIA/CUDA-5.5/lib:$DYLD_LIBRARY_PATH
> >
> >
> >
> >
> > On Mar 4, 2014, at 9:55 AM, Maxim Belkin <mbelkin_at_ks.uiuc.edu> wrote:
> >
> >> Hi Carolyn,
> >>
> >> As John suggested, I was able to reproduce your problem by setting DYLD_LIBRARY_PATH. Run ?unset DYLD_LIBRARY_PATH? and try launching VMD then.
> >>
> >> Maxim
> >>
> >> On Mar 4, 2014, at 9:51 AM, John Stone <johns_at_ks.uiuc.edu> wrote:
> >>
> >>> Carolyn,
> >>> Have you checked your environment variables to see if they are
> >>> interfering with VMD's usage of the .dylib that's shipped with it?
> >>> Anything in DYLD_FRAMEWORK_PATH or LD_LIBRARY_PATH that adds CUDA?
> >>>
> >>> Cheers,
> >>> John
> >>>
> >>>
> >>> On Tue, Mar 04, 2014 at 03:45:23PM +0000, Phillips, Carolyn L. wrote:
> >>>> So I have now updated my cuda driver to 5.5.47 and then grabbed the VMD
> >>>> 1.9.1 binary again
> >>>> Same problem
> >>>> /Applications/VMD\ 1.9.1\ 2.app/Contents/MacOS/startup.command ; exit;
> >>>> dyld: Library not loaded: @executable_path/libcudart.dylib
> >>>> Referenced from: /Applications/VMD 1.9.1
> >>>> 2.app/Contents/MacOS/../Resources/VMD.app/Contents/MacOS/VMD
> >>>> Reason: Incompatible library version: VMD requires version 1.1.0 or
> >>>> later, but libcudart.dylib provides version 0.0.0
> >>>> /Applications/VMD 1.9.1 2.app/Contents/MacOS/startup.command: line 7:
> >>>> 47884 Trace/BPT trap: 5 "$p/../Resources/VMD.app/Contents/MacOS/VMD"
> >>>> $*
> >>>> logout
> >>>> From the error message above I wonder if your install_name_tool and otool
> >>>> is finding the libcudart.dylib at all. (version 0.0.0 ??)
> >>>> Also, I noticed that the the /usr/local/cuda directory all points to the
> >>>> Developer folder as well
> >>>> lrwxr-xr-x 1 root wheel 35 Nov 21 09:06 EULA.txt ->
> >>>> /Developer/NVIDIA/CUDA-5.5/EULA.txt
> >>>> lrwxr-xr-x 1 root wheel 30 Nov 21 09:06 bin ->
> >>>> /Developer/NVIDIA/CUDA-5.5/bin
> >>>> lrwxr-xr-x 1 root wheel 30 Nov 21 09:06 doc ->
> >>>> /Developer/NVIDIA/CUDA-5.5/doc
> >>>> lrwxr-xr-x 1 root wheel 33 Nov 21 09:06 extras ->
> >>>> /Developer/NVIDIA/CUDA-5.5/extras
> >>>> lrwxr-xr-x 1 root wheel 34 Nov 21 09:06 include ->
> >>>> /Developer/NVIDIA/CUDA-5.5/include
> >>>> drwxr-xr-x 28 root wheel 952 Mar 4 08:34 lib
> >>>> lrwxr-xr-x 1 root wheel 36 Nov 21 09:06 libnsight ->
> >>>> /Developer/NVIDIA/CUDA-5.5/libnsight
> >>>> lrwxr-xr-x 1 root wheel 34 Nov 21 09:06 libnvvp ->
> >>>> /Developer/NVIDIA/CUDA-5.5/libnvvp
> >>>> lrwxr-xr-x 1 root wheel 31 Nov 21 09:06 nvvm ->
> >>>> /Developer/NVIDIA/CUDA-5.5/nvvm
> >>>> lrwxr-xr-x 1 root wheel 33 Nov 21 09:06 open64 ->
> >>>> /Developer/NVIDIA/CUDA-5.5/open64
> >>>> lrwxr-xr-x 1 root wheel 34 Nov 21 09:06 samples ->
> >>>> /Developer/NVIDIA/CUDA-5.5/samples
> >>>> lrwxr-xr-x 1 root wheel 30 Nov 21 09:06 src ->
> >>>> /Developer/NVIDIA/CUDA-5.5/src
> >>>> lrwxr-xr-x 1 root wheel 32 Nov 21 09:06 tools ->
> >>>> /Developer/NVIDIA/CUDA-5.5/tools
> >>>> and in lib
> >>>> lrwxr-xr-x 1 root wheel 46 Nov 21 09:06 libcudart.dylib ->
> >>>> /Developer/NVIDIA/CUDA-5.5/lib/libcudart.dylib
> >>>> On Mar 3, 2014, at 9:08 AM, John Stone <johns_at_ks.uiuc.edu> wrote:
> >>>>
> >>>> Carolyn,
> >>>> Let us know if Maxim's suggestion makes any difference.
> >>>> Maxim, Thanks for the tip.
> >>>>
> >>>> Another thing to check is to see which CUDA runtime .dylib your locally
> >>>> compiled VMD is actually picking up when it runs.
> >>>> Since you have several on your system, one question I would ask is
> >>>> whether
> >>>> anything would have changed your DYLD_FRAMEWORK_PATH or LD_LIBRARY_PATH
> >>>> or similar environment variables in such a way that VMD is now picking
> >>>> up the wrong .dylib where it had not previously.
> >>>>
> >>>> In the binaries we produce, we use the MacOS X 'install_name_tool' and
> >>>> 'otool' commands to set some of our library dependency paths in such a
> >>>> way that they are unlikely to be overridden by environment variable
> >>>> changes like the ones above, but if you compile VMD from source and
> >>>> don't do this on your own, then anything could potentially happen if
> >>>> system-wide paths are changed such that the wrong .dylib shows up
> >>>> earlier
> >>>> in the search path.
> >>>>
> >>>> It should be easy to attach a debugger to your VMD during startup to
> >>>> help track down what's going on there. When the debugger attaches, it
> >>>> will
> >>>> usually print out what shared libraries the app is linked against and
> >>>> what
> >>>> they were resolved to, and this would help track down if there's
> >>>> something
> >>>> bad going on with mismatched .dylib files...
> >>>>
> >>>> Cheers,
> >>>> John Stone
> >>>>
> >>>> On Sun, Mar 02, 2014 at 05:54:44PM -0600, Maxim Belkin wrote:
> >>>>
> >>>> Hi Carolyn,
> >>>>
> >>>> The issue you are facing is caused by CUDA drivers. Try installing the
> >>>> latest CUDA 5.5.43 drivers. Until this version, CUDA didn?t work on a
> >>>> great deal of GPUs under Mavericks.
> >>>>
> >>>> Maxim
> >>>>
> >>>> On Mar 2, 2014, at 4:54 PM, Phillips, Carolyn L. <cphillips_at_anl.gov>
> >>>> wrote:
> >>>>
> >>>> A little search turned up these versions.
> >>>>
> >>>> ./NVIDIA/CUDA-5.5/lib/libcudart.dylib
> >>>> ./VMD 1.9.1.app/Contents/vmd/libcudart.dylib
> >>>> ./MATLAB_R2013b.app/bin/maci64/libcudart.dylib
> >>>> /local/cuda/lib/libcudart.dylib
> >>>>
> >>>> I think the first it the one it should be finding.
> >>>> I don?t know why such a library is included in vmd, but I am going
> >>>> to assume its legitimate
> >>>> Same for the Matlab one
> >>>> I assume it may be the /local version that is creating the problem.
> >>>>
> >>>> FYI,
> >>>>
> >>>> which nvcc
> >>>> /Developer/NVIDIA/CUDA-5.5/bin/nvcc
> >>>>
> >>>> -Carolyn
> >>>> On Mar 2, 2014, at 3:20 PM, Axel Kohlmeyer <akohlmey_at_gmail.com>
> >>>> wrote:
> >>>>
> >>>> On Sun, Mar 2, 2014 at 3:42 PM, Phillips, Carolyn L.
> >>>> <cphillips_at_anl.gov> wrote:
> >>>>
> >>>> Hello,
> >>>>
> >>>> I am trying to pin down why my version of VMD 1.91.1 built from
> >>>> source has
> >>>> started seg faulting. (Note, i cannot run vmd -debug because
> >>>> gdb is no
> >>>> longer included with Xcode on Mavericks as a command line tool,
> >>>> you can get
> >>>> it from macports but it is now called ggdb )
> >>>>
> >>>> I started with verifying if I download and run the VMD1.9.1
> >>>> binary that it
> >>>> still works. What I found is that this version does work,
> >>>> MacOS X OpenGL
> >>>> (32-bit Intel x86) (Apple MacOS-X 10.5.x or later).
> >>>>
> >>>> This version crashes, MacOS X OpenGL, CUDA (32-bit Intel x86)
> >>>> (Apple
> >>>> MacOS-X 10.5.x or later with CUDA)
> >>>>
> >>>> It cannot seem to interface with the cuda library correctly
> >>>>
> >>>> /Volumes/VMD-1.9.1/VMD\ 1.9.1.app/Contents/MacOS/startup.command
> >>>> ; exit;
> >>>> dyld: Library not loaded: @executable_path/libcudart.dylib
> >>>> Referenced from: /Volumes/VMD-1.9.1/VMD
> >>>> 1.9.1.app/Contents/MacOS/../Resources/VMD.app/Contents/MacOS/VMD
> >>>> Reason: Incompatible library version: VMD requires version 1.1.0
> >>>> or later,
> >>>> but libcudart.dylib provides version 0.0.0
> >>>>
> >>>> hi carolyn,
> >>>>
> >>>> this looks like VMD picks up some old and incompatible version of
> >>>> the
> >>>> cuda runtime library from somewhere.
> >>>>
> >>>> axel.
> >>>>
> >>>> /Volumes/VMD-1.9.1/VMD 1.9.1.app/Contents/MacOS/startup.command:
> >>>> line 7:
> >>>> 2533 Trace/BPT trap: 5
> >>>> "$p/../Resources/VMD.app/Contents/MacOS/VMD" $*
> >>>> logout
> >>>>
> >>>> [Process completed]
> >>>>
> >>>> My system is
> >>>>
> >>>> Mac OSX 10.9.1
> >>>> Graphics NVIDIA GeForce GT 750M 2048 MB
> >>>> Processor 2.3 GHz Intel Core i7
> >>>>
> >>>> -Carolyn
> >>>>
> >>>> --
> >>>> Dr. Axel Kohlmeyer akohlmey_at_gmail.com http://goo.gl/1wk0
> >>>> College of Science & Technology, Temple University, Philadelphia
> >>>> PA, USA
> >>>> International Centre for Theoretical Physics, Trieste. Italy.
> >>>>
> >>>> --
> >>>> NIH Center for Macromolecular Modeling and Bioinformatics
> >>>> Beckman Institute for Advanced Science and Technology
> >>>> University of Illinois, 405 N. Mathews Ave, Urbana, IL 61801
> >>>> http://www.ks.uiuc.edu/~johns/ Phone: 217-244-3349
> >>>> http://www.ks.uiuc.edu/Research/vmd/
> >>>
> >>> --
> >>> NIH Center for Macromolecular Modeling and Bioinformatics
> >>> Beckman Institute for Advanced Science and Technology
> >>> University of Illinois, 405 N. Mathews Ave, Urbana, IL 61801
> >>> http://www.ks.uiuc.edu/~johns/ Phone: 217-244-3349
> >>> http://www.ks.uiuc.edu/Research/vmd/
> >>
> >
> 

-- 
NIH Center for Macromolecular Modeling and Bioinformatics
Beckman Institute for Advanced Science and Technology
University of Illinois, 405 N. Mathews Ave, Urbana, IL 61801
http://www.ks.uiuc.edu/~johns/           Phone: 217-244-3349
http://www.ks.uiuc.edu/Research/vmd/