I would like to officially declare omnidirectional shadow maps as being practical

Cypher19 · 2013-09-13T02:03:53

Okay, omnidirectional (aka point) lights are hard to do right? You go with shadow maps and you get crap quality (even with fairly high resolution maps, imo), you go with shadow volumes and you get crap performance. Well, in the last couple months I've been developing a VERY powerful algorithm with the intent of increasing performance and image quality with shadow maps. It all started with my huge frustration with how God-awful 1024x1024x6 (cube maps for omni lights of course) shadow maps looked (see below), and only got okay performance on them (~50fps on my X800Pro, before optimizations) Anyways, I was banging nails into the wall with my forehead for awhile, trying to think of an alternate method. And an alternate method I did come up with. I call it "skewbe" shadow mapping (because I skew the cubemap, get it?). I won't write up all the details here, I want to save that for a paper later, but I figured you guys would like to see it. So, here goes nothing: The shots on the left are done with 512x512x6 cubemaps, and the shots on the right are done with 512x512x6 skewbemaps, and (STILL completely unoptimized) performance on my X800Pro is 90fps. If you don't think that's awesome, then leave. Executable: http://www.eng.uwaterloo.ca/~dcrooks/ShadowMapping.rar And if you don't have the DX9SDK April update installed, get this too and put it in your system32 directory: http://www.eng.uwaterloo.ca/~dcrooks/d3dx9_25.rar

Graphics and GPU Programming Programming

Started by Cypher19 June 18, 2005 01:38 PM

35 comments, last by BornToCode 10 years, 7 months ago

Promit

13,404

June 18, 2005 11:49 PM

Quote:Original post by Toji
Here's a question: Could your algorithim be use for directional lights as well? If so this could be a great all-purpouse technique!

I haven't implemented it myself, but isn't PSMs supposedly a perfectly good algorithm for directional shadow maps?

SlimDX | Ventspace Blog | Twitter | Diverse teams make better games. I am currently hiring capable C++ engine developers in Baltimore, MD.

Cypher19

768

Author

June 19, 2005 12:11 AM

Quote:Original post by Toji
Here's a question: Could your algorithim be use for directional lights as well? If so this could be a great all-purpouse technique!

No, this algorithm is purely for omnidirectional lights, (and I guess you could use it for spotlights...) as it involves manipulating the cubemap in a very specific way. When I publish the paper, you'll see why it would be impossible, or at the very least EXTREMELY unpractical (so much so that a single CSM would produce okay results)

_DarkWIng_

602

June 19, 2005 03:23 AM

Cypher19, nice to see you finish this. The difference is amazing.

Any chance you could make a small demo without Blinn/Fresnel effects and with option to change shadowmap size and number of samples (for softening). I would really like to see how it would run on older machines (GF-FX). I know I could get almost all this by hacking .fx file but I really don't want to.

Any date or release of the paper?

You should never let your fears become the boundaries of your dreams.

Monder

993

June 19, 2005 04:02 AM

Nice demo, I'm looking forward to reading the paper on it.

Oh how do you get the demo to tell you your FPS, I wanna know how well it performs on my machine?

iMalc

2,466

June 19, 2005 04:52 AM

Keep up the good work!

"In order to understand recursion, you must first understand recursion."
My website dedicated to sorting algorithms

Luke Philpot

200

June 19, 2005 07:04 AM

Wow! Very nice! Can't wait for the paper :)

My Website | Everything you need to know about 3D Graphics | Google | Search the Forums

Cypher19

768

Author

June 19, 2005 09:35 AM

Quote:Original post by Monder
Oh how do you get the demo to tell you your FPS, I wanna know how well it performs on my machine?

http://www.fraps.com/

Quote:Original post by _DarkWIng_
Cypher19, nice to see you finish this. The difference is amazing.

Any chance you could make a small demo without Blinn/Fresnel effects and with option to change shadowmap size and number of samples (for softening). I would really like to see how it would run on older machines (GF-FX). I know I could get almost all this by hacking .fx file but I really don't want to.

Any date or release of the paper?

If you're desperate to change the lighting model, you could modify the .fx file (which is just text). All of the lighting is done at around line 400 and is organized fairly neatly. You can just comment out individual lines.

For the number of samples, 4 is the default, and is the limit for PS2.0. If you REALLY want to up the sample count, around line 50 you'll find what you need to change (it's pretty obvious) but you'll have to scroll down to line 617 and change the PS compile target to 2.b or 3.0.

Lastly, I might include the option to change the SM size at a later date.

Oh, and someone was wondering about the release date of the paper. I hope to finish it in one (possible) or two (probable) weeks.

mikeman

3,017

June 19, 2005 09:59 AM

It looks good, but it seems to work only in some cases. When I look straight to the vertical walls from close, the difference in quality is huge. But when place the camera a little above the floor level and looking at the center of the scene, the shadows seem pretty much the same with or without skewbe mapping.
To give you an idea what I'm talking about, I took a screenshot(it's with SkewMaps on, PCF and lighting off):

Also, there are cases where the quality is lower when skewmapping is on. That happens when I look shadows casted at a far vertical wall, and the shadow is on the corner of my FOV. With skewmapping off it looks ok, when I turn it on it looks blocky. If I rotate the camera so the shadow is at the center of my FOV, the quality of the shadow suddenly increases.

It certainly looks very promising though, keep working on it.

[Edited by - mikeman on June 19, 2005 10:59:49 AM]

Hyperdrive - Futuristic Combat Racing

RipTorn

722

June 20, 2005 12:30 AM

Since I assume there is a lot of projection going on, does this method suffer any depth fighting issues with highly complex geometry and large shadow radius? dual-paraboloid shadow mapping had issues with that if I recal correctly... Are the walls casting shadows?
It would also seem there are some minor issues with object silhouettes when anisotropic filtering is enabled..