Login  Register

Re: SPIM multi view deconvolution via GPU

Posted by Douglas Richardson on Jun 13, 2014; 8:42pm
URL: http://imagej.273.s1.nabble.com/SPIM-multi-view-deconvolution-via-GPU-tp5008097p5008214.html

Sorry it took me awhile to reply, I didn't have a lot of time to work on
this this week.  Here is the good news, everything seems to be working well
now.  Thanks to both Stephan and Brian for your suggestions.  Here's what I
discovered in the end:

The current code is not compatible with CUDA6.0 as Stephan had suggested.
CUDA5.5 is needed.  As Brian had recomended, I was able to instal CUDA5.5,
then update the driver for the GTX Titan Black GPU (I'm now using Graphics
Driver 337.88)

Another note for others trying this: For a complete instal of the CUDA5.5
(or 6.0) toolbox, an installed copy of Visual Studio 2012 (not 2013!) is
required.  (I currently only have the 30 day trial, hopefully everything
doesn't crash after 30 days.)

The .dll and .so files only needed to be placed in the /Fiji.app directory,
not their own lib folder

The CUDA library is added to the system path upon installation in Windows 7.

I now have this running on two systems:

1) Windows 7 (64bit), 16GB RAM, GTX 285 GPU, CUDA 5.5, Graphics Driver
320.57, Visual Studio 2012
2) Windows 7 (64bit), 196GB RAM, GTX Titan Black GPU, CUDA 5.5, Graphics
Driver 337.88, Visual Studio 2012

Thanks again for all the help!

-Doug


On Wed, Jun 11, 2014 at 8:04 PM, Stephan Preibisch <[hidden email]>
wrote:

> Hi Brian,
>
> it seemed to be a somewhat inconsistent behavior between platforms and
> versions, that’s why I changed it to the manual selection in the new
> version. I would try lib/linux64, lib, folder directly. One of them should
> work. At least it did back then on Windows and Ubuntu.
>
> I was under time pressure because of the paper when I finished that code,
> so it is not that well done. I am mostly working on the new version now
> which should not have these problems any more.
>
> I hope by the end of next week it can be released.
>
> Cheers,
> Stephan
>
> On Jun 11, 2014, at 17:48 , Brian Northan <[hidden email]> wrote:
>
> > Hi Stephan
> >
> > Thanks, I will try it out.  Also, looking at your code in
> > "Multi_View_Deconvolution.java" I saw a comment (around line 825 where it
> > loads the libary) "under linux automatically checks lib/linux64".  So
> that
> > mean it only checks the Fiji/lib directory under linux??
> >
> >
> > On Wed, Jun 11, 2014 at 5:20 PM, Stephan Preibisch <[hidden email]
> >
> > wrote:
> >
> >> Thanks a lot Brian for the help!
> >>
> >> the behavior is unfortunately very different under Windows and Linux as
> we
> >> figured out.
> >>
> >> The new version of my plugin that is currently developed handles it much
> >> better I think, it actively looks for it and you can change the path. It
> >> will be released very soon, if you want to try it ahead of time let me
> >> know, I can send you jars (or here is the source
> >> https://github.com/fiji/spimreconstruction/tree/revised-spimdata)! It
> is
> >> much more flexible and powerful and iterates with the BigDataViewer.
> >>
> >> Cheers,
> >> Stephan
> >>
> >> On Jun 11, 2014, at 17:11 , Brian Northan <[hidden email]> wrote:
> >>
> >>> Hi Douglas
> >>>
> >>> Any luck fixing your problem?  I had a similar problem working with a
> >> cuda
> >>> plugin.
> >>>
> >>> First I was getting an error something like "can't find library.so"
> >>>
> >>> 1.  I had to put the native libraries in a "lib" directory directly
> under
> >>> the Fiji directory.  So perhaps put the .dll in the "lib" directory.
> >>>
> >>> Then I got an error something like "can't find cudart.so"
> >>>
> >>> 2.  I had to add the location of the Cuda library to my system library
> >>> path.  The ubuntu cuda installer didn't do this automatically.  Not
> sure
> >>> the behaviour is the same in Windows or not.  But you might want to
> check
> >>> and see if the location of the cuda libaries has been added to the
> system
> >>> path.
> >>>
> >>> 3.  You mention you couldn't get the newest graphics card to work with
> >> Cuda
> >>> 5.5.  I think that might be because cuda 5.5 installs an older version
> of
> >>> the nvidia drivers.  I think you can get the newer cards working with
> >> Cuda
> >>> 5.5.  This info is for GTX 750 Ti but perhaps relevant for you as well.
> >>>
> >>> "The GTX 750 Ti is not supported by the driver (319.37) that is
> included
> >>> with the cuda 5.5 toolkit installer package. Download the latest driver
> >>> installer package for GeForce 750 Ti from www.nvidia.com and run that
> to
> >>> replace your 319.37 driver (only). You should have better results. Yes,
> >>> CUDA 5.5 is compatible with the newer driver you are installing."
> >>>
> >>>
> >>>
> >>>
> >>> On Tue, Jun 10, 2014 at 11:17 AM, Johannes Schindelin <
> >>> [hidden email]> wrote:
> >>>
> >>>> Hi Douglas,
> >>>>
> >>>> On Mon, 9 Jun 2014, Douglas Richardson wrote:
> >>>>
> >>>>> Wayne: Sorry, I'll stick to the Fiji bug reporter from now on.......
> >>>>
> >>>> The ImageJ mailing list is the best place to discuss these things: as
> >> you
> >>>> pointed out, Stephan keeps an eye on this list. He does not keep an
> eye
> >> on
> >>>> the Fiji bugzilla. SPIM registration is an ImageJ plugin after all, a
> >>>> plugin that just happens to be available in the Fiji distribution of
> >>>> ImageJ for users' convenience.
> >>>>
> >>>> There are other users of the SPIM deconvolution plugin that will
> benefit
> >>>> from seeing this discussion and its outcome. There is a lot of
> expertise
> >>>> on this mailing list and the use of CUDA is on the rise in biological
> >>>> image processing. I think this is in fact a perfect place to discuss
> >> this.
> >>>>
> >>>> The strength of the ImageJ mailing list has always been that it is a
> >>>> friendly place with a wide range of expertise about everything related
> >> to
> >>>> ImageJ, and sometimes even beyond.
> >>>>
> >>>> Let's continue to keep it that way,
> >>>> Johannes
> >>>>
> >>>> --
> >>>> ImageJ mailing list: http://imagej.nih.gov/ij/list.html
> >>>>
> >>>
> >>> --
> >>> ImageJ mailing list: http://imagej.nih.gov/ij/list.html
> >>
> >>
> >> --
> >> ImageJ mailing list: http://imagej.nih.gov/ij/list.html
> >>
> >
> > --
> > ImageJ mailing list: http://imagej.nih.gov/ij/list.html
>
> --
> ImageJ mailing list: http://imagej.nih.gov/ij/list.html
>

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html