RepeatNormal+Transformation+Mask not accalerated on GF8?

Linuxhippy · 12-08-08, 03:53 PM

Hi,

1.) I am currently writing a benchmark to test RENDER acceleration. The 180.11 driver performs incredible well, and on GF6/7 it supports all operations we depend on.

However on GF8+ it seems to fallback for the RepeatNormal+Transformation+Mask case, at least for RenderMark's test.
Is this a known limitation, and/or are there plans to support that combination?

The benchmark can be found at:
http://78.31.67.79:8080/jxrender/RenderMark.html

On my GF6600 I get:

Code:

./JXRenderMark-0.6 11 256
7505,59 Ops/s; transformed texture paint (!); 256x256

On GF8+ it seems to fallback, a GF9600 gets:

Code:

10.563380 Ops/s; transformed texture paint (!); 250x250

2.) Gradient performance is terrible.
I attached a screenshot of Java's new cross-platform look-and-feel.
Although gradients are cached as much as possible, I get 30% better benchmark results when I replace gradient with a 1x1 solid picture.
And this is a general benchmark, testing a lot of text-stuff - the button-performance test shown is twice as fast without gradients.
It would be really great if you could implement gradients soon, and also for GF6/7.

Thanks a lot for the great new drivers, Clemens

Linuxhippy · 12-10-08, 05:55 AM

*push*

Plagman · 12-10-08, 08:05 AM

How large are the source and mask in the case you're talking about?

Linuxhippy · 12-10-08, 08:32 AM

Source: 12x12 - 32bit
Mask: 256x256 - 8bit
Destination: 768x768 - 24-bit

Billinear Filter + Transformation set on Source.

Precompiled version (as well as source) of the benchmark is available, if you would like to give it a try:
./JXRenderMark-0.6 11 256

Thanks, Clemens

Plagman · 12-10-08, 10:15 AM

What if you make the source bigger (50x50) ? I can't run your test right now.

Linuxhippy · 12-10-08, 06:52 PM

Sorry, I can't test too - I don't own a GeForce8+.

I am just verifying that my software will work great with the nvidia binary driver, and this is so far the only problem I've found. (and the extremly slow gradients).

Thanks, Clemens

Linuxhippy · 12-13-08, 05:17 AM

Ok, I've reported it, thats so far all I can do.
If my users complain, I'll simply point to you

Linuxhippy · 12-22-08, 01:33 PM

*Push*
Would you test it if I would create an isolated testcase instead of referring to the whole test-case?

Thanks, Clemens

Lysius · 12-22-08, 01:40 PM

I think there should be enough users around that would run the testcase, if you made an isolated test or would at least post the parameters for your test.

I would have run "make the source bigger (50x50)" on my 8800GTS 512 if I had figured out how to do that.

Linuxhippy · 01-02-09, 01:59 PM

Hi again,

I extended the benchmark a bit to allow small and larger source-images, would be really great if GF8+ owners could give it a try:

Code:

./JXRenderMark 11 256 12 256

A x86 binary is attached, the source can be compiled with:

Code:

gcc -lXrender -lX11 JXRenderMark.c -o JXRenderMark

Thank you in advance and a happy new year, Clemens

**AaronP** · 01-02-09, 02:23 PM

Happy New Year, Clemens! I should note that there's a limitation that I was in the process of fixing before everyone went on vacation that prevents small (<32x32) pixmaps from being placed in vidmem on G80 and up.

pcjason · 01-02-09, 02:27 PM

Here is the results of the tests on my system (Athlon x2 5050e + nVidia 8200 IGP + 4GB RAM):

13.140162 Ops/s; small transformed TexturePaint (10x10) (!); 256x256
14.502762 Ops/s; large transformed TexturePaint (256x256) (!); 256x256

This was using driver 173.14.15. Please let me know if there are any other tests I can help with.

12-08-08, 03:53 PM	#1
Linuxhippy Registered User Join Date: Feb 2004 Posts: 585	RepeatNormal+Transformation+Mask not accalerated on GF8? Hi, 1.) I am currently writing a benchmark to test RENDER acceleration. The 180.11 driver performs incredible well, and on GF6/7 it supports all operations we depend on. However on GF8+ it seems to fallback for the RepeatNormal+Transformation+Mask case, at least for RenderMark's test. Is this a known limitation, and/or are there plans to support that combination? The benchmark can be found at: http://78.31.67.79:8080/jxrender/RenderMark.html On my GF6600 I get: Code: ./JXRenderMark-0.6 11 256 7505,59 Ops/s; transformed texture paint (!); 256x256 On GF8+ it seems to fallback, a GF9600 gets: Code: 10.563380 Ops/s; transformed texture paint (!); 250x250 2.) Gradient performance is terrible. I attached a screenshot of Java's new cross-platform look-and-feel. Although gradients are cached as much as possible, I get 30% better benchmark results when I replace gradient with a 1x1 solid picture. And this is a general benchmark, testing a lot of text-stuff - the button-performance test shown is twice as fast without gradients. It would be really great if you could implement gradients soon, and also for GF6/7. Thanks a lot for the great new drivers, Clemens Attached Thumbnails

12-10-08, 05:55 AM	#2
Linuxhippy Registered User Join Date: Feb 2004 Posts: 585	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? push

12-10-08, 08:05 AM	#3
Plagman NVIDIA Corporation Join Date: Sep 2007 Posts: 254	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? How large are the source and mask in the case you're talking about?

12-10-08, 08:32 AM	#4
Linuxhippy Registered User Join Date: Feb 2004 Posts: 585	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? Source: 12x12 - 32bit Mask: 256x256 - 8bit Destination: 768x768 - 24-bit Billinear Filter + Transformation set on Source. Precompiled version (as well as source) of the benchmark is available, if you would like to give it a try: ./JXRenderMark-0.6 11 256 Thanks, Clemens

12-10-08, 10:15 AM	#5
Plagman NVIDIA Corporation Join Date: Sep 2007 Posts: 254	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? What if you make the source bigger (50x50) ? I can't run your test right now.

12-10-08, 06:52 PM	#6
Linuxhippy Registered User Join Date: Feb 2004 Posts: 585	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? Sorry, I can't test too - I don't own a GeForce8+. I am just verifying that my software will work great with the nvidia binary driver, and this is so far the only problem I've found. (and the extremly slow gradients). Thanks, Clemens

12-13-08, 05:17 AM	#7
Linuxhippy Registered User Join Date: Feb 2004 Posts: 585	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? Ok, I've reported it, thats so far all I can do. If my users complain, I'll simply point to you

12-22-08, 01:33 PM	#8
Linuxhippy Registered User Join Date: Feb 2004 Posts: 585	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? Push Would you test it if I would create an isolated testcase instead of referring to the whole test-case? Thanks, Clemens

12-22-08, 01:40 PM	#9
Lysius Registered User Join Date: Oct 2008 Posts: 98	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? I think there should be enough users around that would run the testcase, if you made an isolated test or would at least post the parameters for your test. I would have run "make the source bigger (50x50)" on my 8800GTS 512 if I had figured out how to do that.

01-02-09, 02:23 PM	#11
AaronP NVIDIA Corporation Join Date: Mar 2005 Posts: 2,487	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? Happy New Year, Clemens! I should note that there's a limitation that I was in the process of fixing before everyone went on vacation that prevents small (<32x32) pixmaps from being placed in vidmem on G80 and up.

01-02-09, 02:27 PM	#12
pcjason Registered User Join Date: Dec 2007 Posts: 17	Re: RepeatNormal+Transformation+Mask not accalerated on GF8? Here is the results of the tests on my system (Athlon x2 5050e + nVidia 8200 IGP + 4GB RAM): 13.140162 Ops/s; small transformed TexturePaint (10x10) (!); 256x256 14.502762 Ops/s; large transformed TexturePaint (256x256) (!); 256x256 This was using driver 173.14.15. Please let me know if there are any other tests I can help with.