Errors in ambient calculation in 3R7P1?

I am testing a new benchmark scene to replace the old "mirror pyramid" scene, and I came across some strange behavior in the 3R7P1 release.

The benchmark scene and spase table is at:
http://mark.technolope.org/pages/rad_bench.html

The 3R7P1 image in question is:
http://mark.technolope.org/radmisc/bench4_02.png

The same commands, under 3R6P1, created this:
http://mark.technolope.org/radmisc/bench4_01.png

So, neither of the images are perfect, but the large black areas in the 3R7P1 image worry me. I have tried very similar scenes with nearly similar parameters, and every other time I run it, those dark spots appear on the fractal object in the middle. I am still running more tests on every Radiance install that I can find.

Some notes about the benchmark scene: I tried to include a greater variety of commonly-used Radiance objects and modifiers. The majority of the time spent, though, is still in the ambient calculation---the forte of our favorite software. But I included a large colorpict/mixpict, an obj2mesh object (with normals), plastic, interface, glass and metal, daylight and electric light, and large and small features. The render takes long enough that parallel machines of 2010 still won't be able to finish it in less than a minute. The bench2 scene saw speed improvements of nearly 100-fold over it's 5-year history. Any comments about the benchmark scene are also welcome.

Thanks!

Mark

Hi Mark,

Thanks for putting together a new benchmark. I agree that the dark areas are concerning -- there may be a bug in the irradiance gradient calculation, since this is the only thing that could have such an effect. I made a number of changes to the ambient calculation to accommodate the new rtcontrib program, and it may be that I broke something along the way. I plan to investigate this weekend to see if I can find an error. I performed a test of the ambient calculation including the irradiance gradient calc. when I made the changes, but perhaps I missed something.

I shall get back to you regarding what I find by Monday.

-Greg

Greg,

I am sorry if I am responsible for ruining your weekend. Does the benchmark scene look OK to you (exercise the proper calculations, include enough interesting stuff to be general, etc.)?

Would it help your work if I ran more cases with older Radiance versions?

It's a bummer I missed you all at the conference. Hopefully when I get my degree (and a job), I'll be more able to attend.

Thank you.

Mark

···

On Sat, 20 Aug 2005, Greg Ward wrote:

Hi Mark,

Thanks for putting together a new benchmark. I agree that the dark areas are concerning -- there may be a bug in the irradiance gradient calculation, since this is the only thing that could have such an effect. I made a number of changes to the ambient calculation to accommodate the new rtcontrib program, and it may be that I broke something along the way. I plan to investigate this weekend to see if I can find an error. I performed a test of the ambient calculation including the irradiance gradient calc. when I made the changes, but perhaps I missed something.

I shall get back to you regarding what I find by Monday.

-Greg

_______________________________________________
Radiance-dev mailing list
[email protected]
http://www.radiance-online.org/mailman/listinfo/radiance-dev

On 20/08/05, 06:15:10, Mark Stock <[email protected]> wrote regarding
[Radiance-dev] Errors in ambient calculation in 3R7P1?:

I am testing a new benchmark scene to replace the old "mirror
pyramid" scene, and I came across some strange behavior in the
3R7P1 release.

I too see some anomalies. I've run the test on 4 machines and here
are my pictures:

http://www.blastwave.org/~james/bench4.cthuga.pic
http://www.blastwave.org/~james/bench4.hastur.pic
http://www.blastwave.org/~james/bench4.ra.pic
http://www.blastwave.org/~james/bench4.zeus.pic

My times are here:
http://www.blastwave.org/~james/bench4.results.html

The best is 3454 seconds. With 4 CPUs in the Sun Fire V440 and 2
in the V20z I look forward to an rpiece benchmark test!

I can't seen what is causing the black shadows. Next I will try
some different settings - and rpiece for multi-CPU speed.

James.

James,

Since I am not very familiar with rpice and how to get multiple processes started on a list of machines, maybe you could help me write that portion of the benchmark test? I, too, am eager to try it on the 8-proc Opteron system at work.

Also, the rpict times are for serial runs, right?

I'll put your times up, but recognize that we will have to re-do everything when the bug is fixed.

Mark

···

On Sun, 21 Aug 2005, James Lee wrote:

On 20/08/05, 06:15:10, Mark Stock <[email protected]> wrote regarding
[Radiance-dev] Errors in ambient calculation in 3R7P1?:

I am testing a new benchmark scene to replace the old "mirror
pyramid" scene, and I came across some strange behavior in the
3R7P1 release.

I too see some anomalies. I've run the test on 4 machines and here
are my pictures:

http://www.blastwave.org/~james/bench4.cthuga.pic
http://www.blastwave.org/~james/bench4.hastur.pic
http://www.blastwave.org/~james/bench4.ra.pic
http://www.blastwave.org/~james/bench4.zeus.pic

My times are here:
http://www.blastwave.org/~james/bench4.results.html

The best is 3454 seconds. With 4 CPUs in the Sun Fire V440 and 2
in the V20z I look forward to an rpiece benchmark test!

I can't seen what is causing the black shadows. Next I will try
some different settings - and rpiece for multi-CPU speed.

James.

On 21/08/05, 16:13:30, Mark Stock <[email protected]> wrote regarding Re:
[Radiance-dev] Errors in ambient calculation in 3R7P1?:

Since I am not very familiar with rpice and how to get multiple
processes started on a list of machines, maybe you could help me
write that portion of the benchmark test? I, too, am eager to try
it on the 8-proc Opteron system at work.

OK, I'll see what I can do.

Also, the rpict times are for serial runs, right?

I just ran "make", so just one CPU for rpict.

I'll put your times up, but recognize that we will have to re-do
everything when the bug is fixed.

I understand, that's fine.

BTW, the link to:
http://mark.technolope.org/radmisc/bench4_03.png
is not working.

James.

Hi Mark,

I am sorry if I am responsible for ruining your weekend.

Well, it was a difficult weekend on many counts, and I'd rather hear sooner when a new release is broken than later.

Does the benchmark scene look OK to you (exercise the proper calculations, include enough interesting stuff to be general, etc.)?

I really like the fact that you incorporated many of Radiance's features in a single test. It is really impossible to be comprehensive, but I appreciate your effort and I think you did an admirable job. This is certainly a lot more challenging than any of the test scenes I usually run, which I think is why this problem showed up here rather than in my own prerelease checks.

I may have some tweaks to suggest once I get this bug fixed. I think I've found the problem, and it wasn't the gradient calculation after all but a ramification of some changes I made to accommodate Russian roulette ray termination. There were really quite a few changes to the core renderer in this release, and that's the danger of altering something as complicated as Radiance, even for the original author....

Would it help your work if I ran more cases with older Radiance versions?

No, that's not necessary, thanks. I resurrected 3.6 and verified that it works properly with this scene, unlike 3.7.1.

It's a bummer I missed you all at the conference. Hopefully when I get my degree (and a job), I'll be more able to attend.

Yes, we definitely missed having you there. Christoph showed off some of your artistic Radiance renderings to an appreciative audience, and I related how they were even more impressive in a full-sized print, like the one I saw at SIGGRAPH last year.

Thanks again for your help in discovering this bug. I am running a test now with my hopeful fix, and if that's working, I'll put together a 3.7.2 patch release tomorrow and make an announcement to radiance-general.

-Greg

Greg,

I am looking forward to hearing about these as much as I am the patch release!

Mark

···

On Sun, 21 Aug 2005, Greg Ward wrote:

Does the benchmark scene look OK to you (exercise the proper calculations, include enough interesting stuff to be general, etc.)?

[cut]
I may have some tweaks to suggest once I get this bug fixed.