Benchmarking Gimp/GEGL
This discussion is connected to the gimp-user-list.gnome.org mailing list which is provided by the GIMP developers and not related to gimpusers.com.
This is a read-only list on gimpusers.com so this discussion thread is read-only, too.
| Benchmarking Gimp/GEGL | Jernej Simončič | 29 Apr 08:44 |
| Benchmarking Gimp/GEGL | Øyvind Kolås | 29 Apr 10:54 |
| Benchmarking Gimp/GEGL | Partha Bagchi | 29 Apr 11:16 |
| Benchmarking Gimp/GEGL | Daniel Smith | 29 Apr 12:52 |
| Benchmarking Gimp/GEGL | Partha Bagchi | 12 May 17:25 |
| Benchmarking Gimp/GEGL | Patrick Horgan | 14 May 13:41 |
| Benchmarking Gimp/GEGL | Partha Bagchi | 14 May 22:14 |
| Benchmarking Gimp/GEGL | Jernej Simončič | 29 Apr 13:20 |
| Benchmarking Gimp/GEGL | Partha Bagchi | 29 Apr 13:36 |
Benchmarking Gimp/GEGL
On Sun, 29 Apr 2012 03:31:26 +0200, Øyvind Kolås wrote:
It also turns out that babl and GEGL on win32 seem to be compiled practically without optimization and without taking modern instruction sets into account, making any testing of them on windows unrepresentative of their actual performance.
What are the recommended optimization flags?
Benchmarking Gimp/GEGL
On Sun, Apr 29, 2012 at 10:44 AM, Jernej Simončič wrote:
On Sun, 29 Apr 2012 03:31:26 +0200, Øyvind Kolås wrote:
It also turns out that babl and GEGL on win32 seem to be compiled practically without optimization and without taking modern instruction sets into account, making any testing of them on windows unrepresentative of their actual performance.
What are the recommended optimization flags?
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
/Ø
Benchmarking Gimp/GEGL
On Sun, Apr 29, 2012 at 6:54 AM, Øyvind Kolås wrote:
On Sun, Apr 29, 2012 at 10:44 AM, Jernej Simončič wrote:
On Sun, 29 Apr 2012 03:31:26 +0200, Øyvind Kolås wrote:
It also turns out that babl and GEGL on win32 seem to be compiled practically without optimization and without taking modern instruction sets into account, making any testing of them on windows unrepresentative of their actual performance.
What are the recommended optimization flags?
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
/Ø --
«The future is already here. It's just not very evenly distributed» -- William Gibson http://pippin.gimp.org/ http://ffii.org/ _______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list
I am using -Ofast for my builds. I will add the flags you mention when I build RC1. Right now, I have other headaches with RC1. :(
One thing I can report is that with RC1 on a Mac, Snow Leopard (64-bit build only), I get a timing of 5 minutes with the same scenario as indicated at the top of the thread, far faster than the timing on Windows (64-bit as mentioned before).
Benchmarking Gimp/GEGL
What model of Mac is it?
And what is the model number exactly of the HP laptop?
I don't suppose anyone has ever built a rendering farm or cluster for Gimp, have they? Could such a thing be done?
Just wondering. Dan
On 4/29/12, Partha Bagchi wrote:
On Sun, Apr 29, 2012 at 6:54 AM, Øyvind Kolås wrote:
On Sun, Apr 29, 2012 at 10:44 AM, Jernej Simončič wrote:
On Sun, 29 Apr 2012 03:31:26 +0200, Øyvind Kolås wrote:
It also turns out that babl and GEGL on win32 seem to be compiled practically without optimization and without taking modern instruction sets into account, making any testing of them on windows unrepresentative of their actual performance.
What are the recommended optimization flags?
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
/Ø --
«The future is already here. It's just not very evenly distributed» -- William Gibson http://pippin.gimp.org/ http://ffii.org/ _______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-listI am using -Ofast for my builds. I will add the flags you mention when I build RC1. Right now, I have other headaches with RC1. :(
One thing I can report is that with RC1 on a Mac, Snow Leopard (64-bit build only), I get a timing of 5 minutes with the same scenario as indicated at the top of the thread, far faster than the timing on Windows (64-bit as mentioned before). _______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list
Benchmarking Gimp/GEGL
On Sun, 29 Apr 2012 12:54:37 +0200, Øyvind Kolås wrote:
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
The compiler is still gcc, so the same optimization flags that apply to Linux apply to Windows, too. Right now I'm using "-O2 -mtune=barcelona -mtune=core2 -mfpmath=sse -msse2" for 32-bit build and "-O2 -mtune=barcelona -mtune=core2" for 64-bit build.
Benchmarking Gimp/GEGL
On Sun, Apr 29, 2012 at 9:20 AM, Jernej Simončič wrote:
On Sun, 29 Apr 2012 12:54:37 +0200, Øyvind Kolås wrote:
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
The compiler is still gcc, so the same optimization flags that apply to Linux apply to Windows, too. Right now I'm using "-O2 -mtune=barcelona -mtune=core2 -mfpmath=sse -msse2" for 32-bit build and "-O2 -mtune=barcelona -mtune=core2" for 64-bit build.
-- < Jernej Simončič ><><><><>< http://eternallybored.org/ >
_______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list
I agree.
I just read here (http://pointclouds.org/news/ffast-math.html), it might not be safe to use -ffast-math. I might be worthwhile to turn on -ftree-vectorize for Windows.
Thanks, Partha
Benchmarking Gimp/GEGL
I followed your suggestions and built babl-0.1.11/gegl-0.2.1/Gimp-2.8.0 using -Ofast -ffast-math -ftree-vectorize. Using these flags, I reran the test at the top of the thread. The c2g rendering on my machine is now 36 seconds as compared to 6 minutes. I am impressed. :)
Now to test out gimp 2.9.1. That should be fun.
On Sun, Apr 29, 2012 at 6:54 AM, Øyvind Kolås wrote:
On Sun, Apr 29, 2012 at 10:44 AM, Jernej Simončič wrote:
On Sun, 29 Apr 2012 03:31:26 +0200, Øyvind Kolås wrote:
It also turns out that babl and GEGL on win32 seem to be compiled practically without optimization and without taking modern instruction sets into account, making any testing of them on windows unrepresentative of their actual performance.
What are the recommended optimization flags?
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
/Ø --
«The future is already here. It's just not very evenly distributed» -- William Gibson http://pippin.gimp.org/ http://ffii.org/ _______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list
Benchmarking Gimp/GEGL
On 05/12/2012 10:25 AM, Partha Bagchi wrote:
I followed your suggestions and built babl-0.1.11/gegl-0.2.1/Gimp-2.8.0 using -Ofast -ffast-math -ftree-vectorize. Using these flags, I reran the test at the top of the thread. The c2g rendering on my machine is now 36 seconds as compared to 6 minutes. I am impressed. :)
If it makes such a difference, shouldn't they be defaults?
Patrick
Now to test out gimp 2.9.1. That should be fun.
On Sun, Apr 29, 2012 at 6:54 AM, Øyvind Kolås wrote:
On Sun, Apr 29, 2012 at 10:44 AM, Jernej Simončič wrote:
On Sun, 29 Apr 2012 03:31:26 +0200, Øyvind Kolås wrote:
It also turns out that babl and GEGL on win32 seem to be compiled practically without optimization and without taking modern instruction sets into account, making any testing of them on windows unrepresentative of their actual performance.
What are the recommended optimization flags?
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
/Ø --
«The future is already here. It's just not very evenly distributed» -- William Gibson http://pippin.gimp.org/ http://ffii.org/ _______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list_______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list
Benchmarking Gimp/GEGL
I am using them as default. Now the only issue with the test above is that the original gaussian filter is still very slooooow. I have to figure out the correspondence between gegl-gaussian-blur and the gaussian blur filter.
On Mon, May 14, 2012 at 9:41 AM, Patrick Horgan wrote:
On 05/12/2012 10:25 AM, Partha Bagchi wrote:
I followed your suggestions and built babl-0.1.11/gegl-0.2.1/Gimp-2.8.0 using -Ofast -ffast-math -ftree-vectorize. Using these flags, I reran the test at the top of the thread. The c2g rendering on my machine is now 36 seconds as compared to 6 minutes. I am impressed. :)
If it makes such a difference, shouldn't they be defaults?
Patrick
Now to test out gimp 2.9.1. That should be fun.
On Sun, Apr 29, 2012 at 6:54 AM, Øyvind Kolås wrote:
On Sun, Apr 29, 2012 at 10:44 AM, Jernej Simončič wrote:
On Sun, 29 Apr 2012 03:31:26 +0200, Øyvind Kolås wrote:
It also turns out that babl and GEGL on win32 seem to be compiled practically without optimization and without taking modern instruction sets into account, making any testing of them on windows unrepresentative of their actual performance.
What are the recommended optimization flags?
On win32, no idea, perhaps look at what compiler flags are being used on linux? This is signal processing code and everything from -ffast-math to -ftree-vectorize and probably more are important.
/Ø --
«The future is already here. It's just not very evenly distributed» -- William Gibson http://pippin.gimp.org/ http://ffii.org/ _______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list_______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list_______________________________________________ gimp-user-list mailing list
gimp-user-list@gnome.org
http://mail.gnome.org/mailman/listinfo/gimp-user-list
