QUOTE(cosmonaut @ Jan 16 2007, 01:34 PM)

Just out curiosity Yves, what kind of speed up might we be able to expect?
Well, I only took about one week to optimize. That is not much and that was mainly just enough time to write down my ideas. So far I got about 15% improvement on an equal setting comparison. I tried a technique that gave my 35% improvement but in order to really use that, I would need to write quite a bit of new code for hair, particles, blobies, booleans, etc., and I don't have this time. SSS is waiting. So for now, I would not promise more than that 15% because I don't think I will have more time to dedicate to optimizing in the near future.
QUOTE
For instance, how long did that last render with hair take?
About 2 hours for a 9 passes. Hair is taking a lot of that time. The same render, 9 passes, without hair is taking a little less than 8 minutes. I think we will need to think about tweaking the hair density if we want lower render times.
QUOTE
I did a little experimenting this weekend with different shadow types comparing render times/results. I guess I still feel unless we are using soft raytraced shadows (are you planning to?) and you are able to get some significant speedup, that shadow maps still have the advantage.
We haven't decided anything yet. We are still in the experimenting phase to see what works and what doesn't.
QUOTE
Oh, and by harsh I meant that standard raytraced shadows have no blurring, which in most cases doesn't look very appealing. Don't get me wrong, I love using soft raytraced shadows for stills/when I have time to wait. They certainly are a lot easier to use, no messing with shadow bias, adjusting map res, etc, and can look way better, but they still add so much time to the render that it doesn't seem to be worth it, especially with so many frames to render.
There is a new property in raytraced shadows in v14 called "Distribute among passes" or something like that. This can cut down render time quite a bit when rendering multipasses because with that ON' you can set the number of ray cast to just one and the light sample will still be jittered on each pass producing soft shadows. This could cut render time by at least half.
Actual number of ray cast to produce a nice soft shadow depends greatly on the width of light and the distance from the subject or more to the point it depends on the width of the projected shadow penumbra zone. I usually set the number of rays to one in V14 and render at 9 passes. For most situations, this gives good results.
Also, this is little known but we now have a de-noiser post processing plugin in V14, So it is now possible to set the number of rays lower than for perfect soft shadows and then run the de-noiser plugin on the final render.
QUOTE
I was missing a lot of shadows with my lighting, which I think I've been able to fix in a lot of cases now.
Yes. It looks like you've been to fix a lot of them. z-buffer shadows precision depends on the klieg cone angle and the shaodw map resolution so when we can use a tight spot, like you did on Woot, we can get good shadow precision. But if the klieg spot angle was to be larger to cover the whole of Woot, then we would start loosing the shadows that separates the packsac shoulder strap from his collar and it would look flat. This is actually, the result we see on Tin Woodman shoulder plate where with ray-traced shadows, the relationship between the shoulder plate and the arm is obvious but that same relationship is missing from the z-buffer shadows because the cone angle is larger and covers more scene area.
QUOTE
Of course, all this is moot if you've got some good speedups going for raytracing. And if you guys decide we really should use raytraced shadows I'll go ahead and make the switch.
We haven't decided anything. We are examining the results and still debating the pros and cons. Your tests are part of that and us discussing that here is also part of that.
QUOTE
btw, are we planning to use multipass rendering for everything, and if so how many passes?
There are no hard cut rules on that either right now and I don't think there will ever be. The decision between multipass vs a-buffer and how many passes will be decided on a scene by scene basis. Scenes with little tiny details will probably require multipass with more passes but in all event, we will need to examine all the other alternatives for optimizing the renders. I don't mean programming optimization alone but I mean selecting alternative approaches like using decals instead of materials which not only render faster but are also easier to antialias, reducing the density of hair when the characters are not in a close-up, etc.