I was trying to compile Freesurfer (in a Centos 6.5 VM) with a later gcc to see if the performance improved and noticed that the output gcc 4.4 vs. 4.8 was slightly different. I later compiled with gcc 4.4 *without* openmp support and this version produced the same output as the gcc 4.8 version, as expected. I have attached the recon-all log files, they first differ at line 1664, at iteration 451 of mri_ca_register; after that the differences accumulate. I feel that this shouldn't happen and the outputs should be equivalent.
As a side note, I am unable to compile GEMS and dmri_poistats or any of the GUI programs due to vtk undefined symbols errors. Similar to this http://www.vtk.org/pipermail/vtkusers/2010-September/062426.html. This is apparently a known issue with vtk 5.6. How is this resolved?
PS. The gcc 4.8 binaries were slower than the 4.4 binaries.
Thanks in advance,
Chris.
Dr Chris Adamson Research Officer, Developmental Imaging, Murdoch Childrens Research Institute Murdoch Childrens Research Institute Royal Children's Hospital Flemington Road, Parkville, Victoria 3052, Australia www.mcri.edu.auhttp://www.mcri.edu.au/ E chris.adamson@mcri.edu.au T 03 9936 6780
______________________________________________________________________ This email has been scanned by the Symantec Email Security.cloud service. For more information please visit http://www.symanteccloud.com ______________________________________________________________________
Hello Chris,
Extremely coincidentally I just recently started researching subtle variations between recon-all when run on OSX vs Linux, and just yesterday found the EXACT same thing. The OSX build does not support openmp whereas the Linux build does and I get very minor but detectable difference starting at iteration 128 of mri_ca_register. So I guess all I can say is that I am looking into it as I type this email.
As for the vtk issue, I can not speak to what we did with our vtk build to get the freesurfer code to compile against it (whatever was done it was before my time here) but if you wish I can provide you with a tarball of our prebuilt vtk which you can use to compile against.
-Zeke
On 03/07/2014 01:13 AM, Chris Adamson wrote:
I was trying to compile Freesurfer (in a Centos 6.5 VM) with a later gcc to see if the performance improved and noticed that the output gcc 4.4 vs. 4.8 was slightly different. I later compiled with gcc 4.4 **without** openmp support and this version produced the same output as the gcc 4.8 version, as expected. I have attached the recon-all log files, they first differ at line 1664, at iteration 451 of mri_ca_register; after that the differences accumulate. I feel that this shouldn’t happen and the outputs should be equivalent.
As a side note, I am unable to compile GEMS and dmri_poistats or any of the GUI programs due to vtk undefined symbols errors. Similar to this http://www.vtk.org/pipermail/vtkusers/2010-September/062426.html. This is apparently a known issue with vtk 5.6. How is this resolved?
PS. The gcc 4.8 binaries were slower than the 4.4 binaries.
Thanks in advance,
Chris.
*Dr **Chris Adamson *Research Officer, Developmental Imaging, *Murdoch Childrens Research Institute *Murdoch Childrens Research Institute Royal Children’s Hospital Flemington Road, Parkville, Victoria 3052, Australia www.mcri.edu.au http://www.mcri.edu.au/
E chris.adamson@mcri.edu.au T 03 9936 6780
This email has been scanned by the Symantec Email Security.cloud service. For more information please visit http://www.symanteccloud.com ______________________________________________________________________
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Do you get the same difference if you run an open MP version but force it to use only one thread? doug
On 03/07/2014 03:52 PM, Z K wrote:
Hello Chris,
Extremely coincidentally I just recently started researching subtle variations between recon-all when run on OSX vs Linux, and just yesterday found the EXACT same thing. The OSX build does not support openmp whereas the Linux build does and I get very minor but detectable difference starting at iteration 128 of mri_ca_register. So I guess all I can say is that I am looking into it as I type this email.
As for the vtk issue, I can not speak to what we did with our vtk build to get the freesurfer code to compile against it (whatever was done it was before my time here) but if you wish I can provide you with a tarball of our prebuilt vtk which you can use to compile against.
-Zeke
On 03/07/2014 01:13 AM, Chris Adamson wrote:
I was trying to compile Freesurfer (in a Centos 6.5 VM) with a later gcc to see if the performance improved and noticed that the output gcc 4.4 vs. 4.8 was slightly different. I later compiled with gcc 4.4 **without** openmp support and this version produced the same output as the gcc 4.8 version, as expected. I have attached the recon-all log files, they first differ at line 1664, at iteration 451 of mri_ca_register; after that the differences accumulate. I feel that this shouldn’t happen and the outputs should be equivalent.
As a side note, I am unable to compile GEMS and dmri_poistats or any of the GUI programs due to vtk undefined symbols errors. Similar to this http://www.vtk.org/pipermail/vtkusers/2010-September/062426.html. This is apparently a known issue with vtk 5.6. How is this resolved?
PS. The gcc 4.8 binaries were slower than the 4.4 binaries.
Thanks in advance,
Chris.
*Dr **Chris Adamson *Research Officer, Developmental Imaging, *Murdoch Childrens Research Institute *Murdoch Childrens Research Institute Royal Children’s Hospital Flemington Road, Parkville, Victoria 3052, Australia www.mcri.edu.au http://www.mcri.edu.au/
E chris.adamson@mcri.edu.au T 03 9936 6780
This email has been scanned by the Symantec Email Security.cloud service. For more information please visit http://www.symanteccloud.com ______________________________________________________________________
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
The openmp version did use only one thread.
I used the vtk build in the tarball of precompiled support libraries on the Freesurfer ftp site. I tried building vtk 5.8 and still had the same issues. It is not a big deal because this exercise was more to do with the computational components rather than the graphical tools.
-----Original Message----- From: freesurfer-bounces@nmr.mgh.harvard.edu [mailto:freesurfer-bounces@nmr.mgh.harvard.edu] On Behalf Of Douglas N Greve Sent: Saturday, 8 March 2014 7:54 AM To: freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] OpenMP and non-OpenMP output is different at mri_ca_register
Do you get the same difference if you run an open MP version but force it to use only one thread? doug
On 03/07/2014 03:52 PM, Z K wrote:
Hello Chris,
Extremely coincidentally I just recently started researching subtle variations between recon-all when run on OSX vs Linux, and just yesterday found the EXACT same thing. The OSX build does not support openmp whereas the Linux build does and I get very minor but detectable difference starting at iteration 128 of mri_ca_register. So I guess all I can say is that I am looking into it as I type this email.
As for the vtk issue, I can not speak to what we did with our vtk build to get the freesurfer code to compile against it (whatever was done it was before my time here) but if you wish I can provide you with a tarball of our prebuilt vtk which you can use to compile against.
-Zeke
On 03/07/2014 01:13 AM, Chris Adamson wrote:
I was trying to compile Freesurfer (in a Centos 6.5 VM) with a later gcc to see if the performance improved and noticed that the output gcc 4.4 vs. 4.8 was slightly different. I later compiled with gcc 4.4 **without** openmp support and this version produced the same output as the gcc 4.8 version, as expected. I have attached the recon-all log files, they first differ at line 1664, at iteration 451 of mri_ca_register; after that the differences accumulate. I feel that this shouldn't happen and the outputs should be equivalent.
As a side note, I am unable to compile GEMS and dmri_poistats or any of the GUI programs due to vtk undefined symbols errors. Similar to this http://www.vtk.org/pipermail/vtkusers/2010-September/062426.html. This is apparently a known issue with vtk 5.6. How is this resolved?
PS. The gcc 4.8 binaries were slower than the 4.4 binaries.
Thanks in advance,
Chris.
*Dr **Chris Adamson *Research Officer, Developmental Imaging, *Murdoch Childrens Research Institute *Murdoch Childrens Research Institute Royal Children's Hospital Flemington Road, Parkville, Victoria 3052, Australia www.mcri.edu.au http://www.mcri.edu.au/
E chris.adamson@mcri.edu.au T 03 9936 6780
This email has been scanned by the Symantec Email Security.cloud service. For more information please visit http://www.symanteccloud.com ______________________________________________________________________
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Chris Adamson -
Douglas N Greve -
Z K