Parallelization in next G'MIC [1.5.7.0]

YAFU · **Joined:** Nov 09, 2011 **Posts:** 726

No problem David. I like being a beta tester when I can do it

A question. Does the code is very different in each filter, regarding multi core support?

That must be a laborious task to do those modificacuiones in each filter.

Ronounours · **Joined:** Jan 03, 2011 **Posts:** 1656

YAFU : Well, I'm able to introduce parallelization in some filters quite quickly, because I've done a generic routine which simply divide an image in four quarters, and apply the filter on each of these four quarters simultaneously (in different threads). The resulting image is then recombined afterwards from these 4 processed quarters.
And this simple technique can be used in a lot of filters without much troubles.
By the way, it's not exactly perfect quarters as I allow some pixel overlapping between the processed quarters, so that filters needing neighboring pixel information can render something nice, without weird effects at the quarter borders.

PhotoComix · **Joined:** Apr 12, 2010 **Posts:** 5870

I tried with my laptop,Dual core, speed is not double but anyway is much faster

lylejk · **Joined:** Sep 24, 2010 **Posts:** 12850

Ronounours wrote:

Arghh. Just found a bug when trying to parallelize the 'Rodilius' filter. Calls to the FFTW3 library functions are not thread-safe, and doing this simultaneously in multiple threads make the library crash. I've fixed it, but I'll need to recompile the binary packages to fix it definitely.
Sorry for people who installed the beta

(but after all, this was a beta, so there are some risks to use it!).

Will disable the multi-process option when running Rodilius then until I can get a up to date compile. Definitely still appreciate your efforts on this David and look forward to the speed improvements as all presets gets this ability.

Ronounours · **Joined:** Jan 03, 2011 **Posts:** 1656

I've updated the source and binary files for the 1.5.7.0 beta, Rodilius should now be able to run in parallel too.

Caz · **Posted:** Fri Aug 30, 2013 12:05 pm **(#26)**

I just updated the filters list and G'Mic now crashes when I try to run the Rodilius filter, I didn't actually get a chance to run it as the crash happens as soon as I hit on the Rodilius link in the filters list.
I tried a few of the other filters and it appears to run them okay, dream smoothing was one and smooth anisotropic was another.
Though it was best to post the results of my updated G'mic install.
Running Linux Mint 15.

K1TesseraEna · **Joined:** Sep 22, 2012 **Posts:** 2089 **Location:** US

I finally found time to install and test-run it.
Image size 3650X3296 RGBA
Filter Anisotropic Smoothing
Render times: single thread 00:01:19:96
multi-core, multi-thread 00:00:20:56

Stunning!!! 6x times faster!!!
Win 7 64-bit, i7-3770K @3.5GHz here

GnuTux · **Posted:** Fri Aug 30, 2013 12:18 pm **(#28)**

Caz wrote:

I just updated the filters list and G'Mic now crashes when I try to run the Rodilius filter, I didn't actually get a chance to run it as the crash happens as soon as I hit on the Rodilius link in the filters list.

Worse results for me, as GIMP core dumps as soon as Rodilus is selected. Running Fedora 19 (x86_64).

Ronounours · **Joined:** Jan 03, 2011 **Posts:** 1656

gnutux : did you try to install the latest version (from today!).

GnuTux · **Posted:** Fri Aug 30, 2013 1:38 pm **(#30)**

I'm using 1.5.7.0 Beta (x86_64). Perhaps the 64bit version has not been updated.

This is the error I get when I select rodilus..

Quote:

*** Error in `gimp-2.8/plug-ins/gmic_gimp': double free or corruption (fasttop): 0x00007f926c0780f0 ***
/.gimp-2.8/plug-ins/gmic_gimp: fatal error: Segmentation fault

Ronounours · **Joined:** Jan 03, 2011 **Posts:** 1656

yes, the 64bits version of samj has not been updated today.

GnuTux · **Posted:** Fri Aug 30, 2013 1:50 pm **(#32)**

Ah, OK, that explains it. :bigthup

YAFU · **Joined:** Nov 09, 2011 **Posts:** 726

I have compiled the new beta version to 64 bits. Rodilius is working well with parallelization.

thanks!

YAFU · **Joined:** Nov 09, 2011 **Posts:** 726

Anyway the speed boost does not seem as good as it with smooth anisotropic filter.

Rodilius one core: 17sec
Rodilius Parallelization: 13sec

In the system monitor I can see that I have only two cores working at maximum at the same time. Is this filter limited to two cores?

Ronounours · **Joined:** Jan 03, 2011 **Posts:** 1656

The problem is that Rodilius uses some function of the FFTW library, which is *not* thread-safe. So basically, all FFTW operations are done sequentially, and you loose quite a lot of speed in doing that.

YAFU · **Joined:** Nov 09, 2011 **Posts:** 726

Oh, I see

YAFU · **Joined:** Nov 09, 2011 **Posts:** 726

I've noticed that the result in Rodilius is not the same with parallelization:

Source:

samj · **Joined:** May 09, 2012 **Posts:** 914

Bonjour,

Update the G'MIC beta version for Win 64-bit - August 30, 2013
http://www.aljacom.com/~gmic/gmic-beta_ ... ts_win.zip

lylejk · **Joined:** Sep 24, 2010 **Posts:** 12850

Just what I was waiting for Samj. The progress bar doesn't work, but the filters do and that's all that really matters to me. Thanks a heap. Do see a speedup in Rodillius; not much but still a speedup.

Ronounours · **Joined:** Jan 03, 2011 **Posts:** 1656

Updated filters to allow the use of up to 8 cores for parallelized filters.
Anyone able to check the speed gain, if any ?

Attachment:

gmic_8cores.png [ 273.58 KiB | Viewed 1634 times ]