[Neurodebian-users] fsl-5.0-gpu

[Jonathan Berrebi]

Hi Alex,

Hopefully it will take less time on a TESLA card since it is the aim to use such one on the server. In that case we don't have any Intel HD/ Optimus issue. I assume anyway that people having NVIDIA cards in their laptop use it more to have 3D acceleration or testing cuda program than to actually use GPU power for computing.

We looked at the dyad2.nii.gz file and it looked globally the same. But the histograms are slightly different and substracting the images is (obvisously) not on the whole image. I guess this is more a question to the FSL mailing list, isn't it?

In the patch I sent earlier today I commented out the line at the origin of the zero padding warning. In the bedpostx_gpu file (and hence in the "monitor" file created by the latter), in the "check" routine, the parameters $FSLDIR and $part are not passed. So the program is looking for /bin/zeropad instead of $FSLDIR/bin/zeropad. Anyway the variable partzp does not seem to be used by "monitor" so I commented it out.

And about Optimus, it seems that nvidia does not develop for linux so the recommended settings for cuda by NVIDIA are the nvidia driver (you see an nvidia screen before login at boot). Then when I run samples from the SDK it runs on GPU without issues on gdm. I can use the mouse, play around with windows. But that works only with the following bios settings ("Discrete Graphics" and not "Integrated Graphics" nor "Optimus"). But maybe is also dependent on the laptop?

I'll probably contact you soon for the other gpu routines in fsl.

Thank you very much,

Jonathan

________________________________________
De : Neurodebian-users [neurodebian-users-bounces+jonathan.berrebi=ki.se at lists.alioth.debian.org] de la part de Alex Waite [alexqw85 at gmail.com]
Envoyé : mercredi 26 février 2014 09:50
À : neurodebian-users at lists.alioth.debian.org
Objet : Re: [Neurodebian-users] fsl-5.0-gpu

> I replaced all the #/bin/sh with #/bin/bash in the bedpostx_gpu and bedpostx_postproc_gpu.sh and fsl_sub files. I think it worked now.

Excellent.

> It took 6 hours on a 96 cores GPU instead of one day on Intel Xeon.

That sounds good at least.

> The only warning I get is on the monitor "routine" ${subjdir}.bedpostX/monitor about zeropadding:
>
> <<FP02.bedpostX/monitor: line 21: /bin/zeropad: No such file or directory>>

That's odd... I greped through all of FSL's scripts, and every instance
of zeropad is prepended with $FSLDIR; it should find zeropad. Did you,
by chance, change that line?

> I will check that the results are numerically not different than when they CPU computed.

Always a good idea.

> Once thing though is that I had to stop gdm3 in order to get it to work. I really tried to use the intel HD 3000
 > integrated GPU for the Graphics by changing the settings in the bios.
But then the nvidia-detect routine did not
 > detect the GPU. So I thought it wouldn't work. Then I even installed
bumblebee  and changed the settings in the
 > bios into Optimus but that become more a bumblebee than a fsl/cuda
problem so I gave up. I might retry later.

The laptop could be disabling the nvidia GPU outright if you select the
intel GPU (saving power and all). Optimus may actually be the correct
BIOS setting. You do not want to mess with bumblebee, however, since
that is about switching GPUs on-the-fly, which you have no desire to do
here. You want to use Intel for display and nvidia for computing.

Sounds like things are coming along well for you!

---Alex


_______________________________________________
Neurodebian-users mailing list
Neurodebian-users at lists.alioth.debian.org
http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/neurodebian-users