View topic - Speed differences

Speed differences

General questions related with the YafaRay Project, 3D computer graphics and about this site.

Speed differences

Post Mon Apr 18, 2011 10:13 am

Hi!
I know it is a very painful question, but i want to ask it so-o-o much! So what is happening to rendering speed? Here is two renders (scene is provided by suomi), upper one rendered in Yafaray 0.1.2 from Blender 2.57 and the second one is rendered in stable Yafaray through Blender 2.49b. No resize done.
The old version render time is 13 minutes and the new one is 49!!! It's almost four times longer! o_O
Image
I Love Yafaray, but it is impossible to understand those difference :(
Tell me at least something, that will reassure me, please!
Last edited by maleficmax on Mon Apr 18, 2011 11:51 am, edited 2 times in total.
My Blender portfolio: http://itcomp.org.ua/?page_id=232
User avatar
maleficmax
 
Posts: 26
Joined: Wed Aug 04, 2010 2:53 pm
Location: Ukraine, Kremenchug

Re: Speed differences

Post Mon Apr 18, 2011 10:19 am

Here are render settings:
I have to say that only 5 resampling passes were done, not 20
Image
My Blender portfolio: http://itcomp.org.ua/?page_id=232
User avatar
maleficmax
 
Posts: 26
Joined: Wed Aug 04, 2010 2:53 pm
Location: Ukraine, Kremenchug

Re: Speed differences

Post Mon Apr 18, 2011 10:33 am

I had the same problem, subscribe under every word.
Wayfaerer
 
Posts: 4
Joined: Thu Apr 15, 2010 6:47 am

Re: Speed differences

Post Mon Apr 18, 2011 1:59 pm

Any time you are testing yafaray for speed and sharing results it is very helpful to include information like what OS you are using, what build of yafaray, etc.

Here are some quick results from me using suomi's scene on ubuntu 10.10 64bit and my own builds of yafaray, the render settings are the same as above except I used two AA passes only:

0.1.1: 5m 22s
0.1.2: 16m 11s

(BTW both builds used the 2.49 exporter, I think this is fairest for comparison.)

I think some significant performance regressions have crept into yafaray relatively recently, I don't remember the disparity being this bad the last time I checked.
User avatar
wizofboz
 
Posts: 235
Joined: Tue Sep 16, 2008 7:00 pm

Re: Speed differences

Post Mon Apr 18, 2011 2:17 pm

I've tested this scene on Windows XP SP3 32 bit (Core2Duo 2.53*2, RAM 2 Gb)
My Blender portfolio: http://itcomp.org.ua/?page_id=232
User avatar
maleficmax
 
Posts: 26
Joined: Wed Aug 04, 2010 2:53 pm
Location: Ukraine, Kremenchug

Re: Speed differences

Post Mon Apr 18, 2011 4:56 pm

I get 35 min. with YafaRay 012 versus 9 min. with YafaRay 011, running on win 7 64 bits. I agree this is not aceptable.

I will do more tests with this scene, to track the source of the slowdown, but I suspect it is the montecarlo integration in path tracing and area light sampling.
User avatar
Samo
 
Posts: 3105
Joined: Tue Dec 20, 2005 10:39 am
Location: Spain

Re: Speed differences

Post Mon Apr 18, 2011 8:19 pm

Edit: rendering 3 passes- 0.1.2: 5min 30 sec vs 0.1.1: 1min 58 sec

Linux 64 - core i7

"Power tends to corrupt, and absolute power corrupts absolutely. Great men are almost always bad men."

User avatar
condar
 
Posts: 771
Joined: Sun Apr 12, 2009 2:41 am

Re: Speed differences

Post Mon Apr 18, 2011 10:02 pm

after doing some tests

Seems that the slowndown happens in both windows and unix systems, it is likely that it happens in Mac OS too.

There is almost no slowdown when Direct Lighting is used,

There is some slowdown when photon mapping is used (between 1-2 times) and much slower when path tracing is used (4 times) than rendered with yafaray 001.

Probably some algorithm used in the indirect lighting calculus is producing this issue.
User avatar
Samo
 
Posts: 3105
Joined: Tue Dec 20, 2005 10:39 am
Location: Spain

Re: Speed differences

Post Tue Apr 19, 2011 3:08 am

About photon mapping: like I posted before, there's the same slowdown in interior renderings, not 1-2 times, but 2-5, depending of method used for illumination (area in window vs portal with sunsky etc.). All materials in the chair scene are shinydiffuse:

http://www.yafaray.org/community/forum/ ... &start=270

Straight comparsion of 0.1.1 and 0.1.2 using the same settings and illuminated by Area light in window shows 2 times difference.

Actually, the Direct Lighting is a bit faster in 0.1.2, as I can see.
suomi
 
Posts: 178
Joined: Mon Sep 13, 2010 4:59 am

Re: Speed differences

Post Tue Apr 19, 2011 3:21 am

suomi wrote:About photon mapping: like I posted before, there's the same slowdown in interior renderings, not 1-2 times, but 2-5, depending of method used for illumination (area in window vs portal with sunsky etc.). All materials in the chair scene are shinydiffuse:

http://www.yafaray.org/community/forum/ ... &start=270

Straight comparsion of 0.1.1 and 0.1.2 using the same settings and illuminated by Area light in window shows 2 times difference.

Actually, the Direct Lighting is a bit faster in 0.1.2, as I can see.


for the benchmark scene I got

0.1.2: 139 secs
0.1.1: 119 secs

"Power tends to corrupt, and absolute power corrupts absolutely. Great men are almost always bad men."

User avatar
condar
 
Posts: 771
Joined: Sun Apr 12, 2009 2:41 am

Re: Speed differences

Post Tue Apr 19, 2011 3:55 am

for the benchmark scene I got

0.1.2: 139 secs
0.1.1: 119 secs


Because this is an open scene with 3-point lighting. The closed scene with one light source shows bigger difference.

Here's the closed environment test. Ubuntu x64, Core2Quad Q6600. 0.1.1 - 7:11, 0.1.2 - 13:46. The scene can be downloaded here: http://www.mf3.ru/samplecontent/mats_close.blend.zip
Attachments
test.png
suomi
 
Posts: 178
Joined: Mon Sep 13, 2010 4:59 am

Re: Speed differences

Post Wed Apr 20, 2011 7:27 am

Ok guys, I think we have found the explanation for the differences in render times.

First of all, in yafaray 012 a shortcoming of the 011 version has been fixed, which is the fact that in Path tracing mode yafaray 011 was unable to reflect the background itself, only the incoming light produced by it. Theres is a simple test to see that:

yafaray011.png
yafaray011.png (372.86 KiB) Viewed 16090 times

yafaray 011, pathtracing with background lighting

yafaray012.png
yafaray012.png (345.81 KiB) Viewed 16090 times

yafaray 012, pathtracing with background lighting

Also the Multiple Importance Sampling code has been changed in yafaray 012, so path tracing should be able to deliver less noise with the same amount of samples than yafaray 011 (click on the images to see them at full resolution):
yafaray011pt16.png

yafaray 011, pathtracing 16 samples, bump maps disabled
yafaray012pt16.png

yafaray 012, pathtracing 16 samples, bump maps disabled

So YafaRay 012 is producing a more correct result and sampling the scene in a different way. In theory in yafaray 012 you would need less samples for a given scene.

This is the chat with Darktide about this issue:
https://gist.github.com/930661
User avatar
Samo
 
Posts: 3105
Joined: Tue Dec 20, 2005 10:39 am
Location: Spain

Re: Speed differences

Post Wed Apr 20, 2011 9:06 am

So will "downgrading" samples count compensate that render times? It's hard to believe that we are loosing so much speed. :(
My Blender portfolio: http://itcomp.org.ua/?page_id=232
User avatar
maleficmax
 
Posts: 26
Joined: Wed Aug 04, 2010 2:53 pm
Location: Ukraine, Kremenchug

Re: Speed differences

Post Wed Apr 20, 2011 4:49 pm

maleficmax wrote:So will "downgrading" samples count compensate that render times? It's hard to believe that we are loosing so much speed. :(


yes. 0.1.2 needs less samples or less render passes to get the same result as with 0.1.1 and so you can get similar render times for both versions.... And Glossy reflections in 0.1.2 with PathTracing are consistent with Direct Lighting and Photon Mapping.
Attachments
difference.png

"Power tends to corrupt, and absolute power corrupts absolutely. Great men are almost always bad men."

User avatar
condar
 
Posts: 771
Joined: Sun Apr 12, 2009 2:41 am

Re: Speed differences

Post Wed Apr 20, 2011 10:17 pm

I did some testing using number of resampled pixels as a measure of quality while path tracing and my findings are this:

1.2 is more efficient, with the same settings in both, there is a significant increase in quality of the render after each pass and at the end of the render.

However, this is not enough to make up the speed difference.

with the same render settings...
render time: 1.1:78 seconds 1.2 225 seconds 180% increase
resampled pixels for last pass: 1.1:49000 1.2 45600

setting the number of passes on 1.2 so that the number of resampled pixels would be the same for both:
render time: 1.1 78 seconds 1.2 111 seconds 40% increase

part of the slow down could be attributed to over sampling some of the pixels in the first pass. If they were smooth enough with the 1.1 sampling method then there is no need to spend more time on them. I first suspected this when I was using more samples per pass. However the first pass takes only about 22 seconds for the render from which I got this data and only about 10% of the pixels didn't need to be resampled after the first pass. So that rules it out as the entire slow down.

I used the glossy test.blend from the scene pack. AA threshold =.01, path tracing samples = 8 (small number to save time with lots of passes to see the resample levels) AA samples per pass 1

the 1.1 render was done using standard blender 2.49 the 1.2 render was done using the latest 2.57 build from graphicall. It is very possible that this is not a fair comparison, but it was the most objective method I could think of.
loh
 
Posts: 250
Joined: Sat Sep 20, 2008 3:22 am

Next

Return to News & Discussion



Who is online

Users browsing this forum: No registered users and 4 guests