Post Your Irrlicht Screenshots / Render Here.

[RdR]

Started on IK for my animation system.

Implemented CCD (Cyclic Coordinate Descent):
http://youtu.be/V4qO_CcbIYU
http://youtu.be/CjPxA2vuhVg

Edit:
http://youtu.be/xinNK498OxA

[rubenwardy]

ARSA Framework - Sponza Rendering
RenderTexture: 1920x1080, ShadowMap: 512 (2 split), SSAO: 512, FXAA
FPS:31, Triangles 837,537
CPU: i7-930, Redon4670
final:
...

Very nice! That is amazing!

[wing64]

Thanks

[hendu]

Untuned rectilinear shadow map vs traditional shadow map. Both 512x512. Guess which is which.

[devsh]

I tell you what you should do..

While PSSM promises the best pixel usage/coverage
You should use CSM because you can make the ShadowMaps clamp to world coordinates and jump across instead of moving smoothly (which is better for temporal aliasing, less pixel swimming as you move the camera)

Then, you should also use VSM or something like that for smooth shadows, maybe it can be enough to hide PSSM swimming artifacts (unless you can clamp separate splits' pixels to world coordinates)

Finally your frustum splits should be split at logarithmic boundaries

Code: Select all

 
splitNearDistance = pow(fullFrustumNearDist,1.0-float(splitID)/numOfSplits)*pow(fullFrustumFarDist,float(splitID)/numOfSplits);
splitFarDistance = pow(fullFrustumNearDist,1.0-float(splitID+1)/numOfSplits)*pow(fullFrustumFarDist,float(splitID+1)/numOfSplits);
 

where spltID starts counting at 0 like an index
THIS IS REALLY THE BEST SPLIT SCHEME GIVING BEST SAMPLING RAITOS

If using deferred rendering, or at least occlusion queries with a depth pass... you can know the expected maximum and minimum expected depths of all pixels within the image
You need to pingpong the depth attachment of the render target minimizing each time by performing the max() and min() functions on groups of pixels (maybe 4x4, shrink by factor of 4 each time)
Obviously ignore the "background" pixel depths, i.e. whatever the depth gets cleared to

Then you can get a LOT BETTER COVERAGE

Code: Select all

 
splitNearDistance = pow(minDetectedDist,1.0-float(splitID)/numOfSplits)*pow(maxDetectedDist,float(splitID)/numOfSplits);
splitFarDistance = pow(minDetectedDist,1.0-float(splitID+1)/numOfSplits)*pow(maxDetectedDist,float(splitID+1)/numOfSplits);
 

because usually if you split logarithmically, none of the pixels on the screen will use the first split Shadow Map

THEN if you want even more quality... Trapezoid Shadow Maps (improvement upon Perspective Shadow Maps which provide more detail near the camera), then you can stretch the shadowmap in a way that will give you more pixels near the viewer.

After performing all of the 4 above, you're at Crysis Quality shadowmapping



P.S. if you want performance, use geometry shaders and MRTs to draw to all splits of the ShadowMaps at once (4 splits should almost always be enough, and most of the time 4 is excessive)

[hendu]

This is funny You should look up the term before comparing other techniques.

Rectilinear shadow maps are far superior to all the ones you mentioned. It's a newer technique (2012) that creates a single shadow map. No splits. No geometry shaders. No perspective aliasing. No multiple rendering.

It is faster than CSM while having better quality. It does not need splits or multiple maps. It's much simpler, which is simply great when it's also faster and better quality (Note: my picture is untuned as mentioned. See the paper for a proper comparison.)

[devsh]

I would dispute the claim of better quality than CSM

sure, its clever but the distortion introduces the need for tesellation and basically gives you the same issues as using non-linear transform shadowmaps for point lights in deferred rendering (Dual Paraboloid or Spheremap)

plus if you really claim rectilinear is superior... why not swap Light Perspective Shadow Mapping for Rectilinear Shadowmapping as your method of perspective adjustment

[hendu]

sure, its clever but the distortion introduces the need for tesellation and basically gives you the same issues as using non-linear transform shadowmaps for point lights in deferred rendering (Dual Paraboloid or Spheremap)

While it does need somewhat tesselated data (no huge triangles), it works fine without dynamic tesselation - the minor errors aren't that noticable once the shadow is softened. For the particular errors of paraboloid (the infinity/singularity), those don't appear.

plus if you really claim rectilinear is superior... why not swap Light Perspective Shadow Mapping for Rectilinear Shadowmapping as your method of perspective adjustment

It seems like a step down to me. While also a type of warping, it leaves it at a perspective setting, while rectilinear allows any number of importance parameters, starting from only covering visible pixels.

[devsh]

I knew something was wrong with rectilinear shadowmapping

it took a while to get.... all soft-shadow techniques will fail or be terribly hard to implement due to random size changes and you will get really bad non-uniform size penumbra as well as the penumbra will be blurred in really weird skewed elipses

[hendu]

Huh? What are you talking about? Soft shadows are completely separate, RT only does hard shadows. Penumbra width depends on your soft shadow tech.

You are correct that percentage closer filtering and its derivatives wouldn't work well with the changing texel sizes, but there are many other soft shadow techniques.

[devsh]

MMMM>..... guess what I'm doing and why

[ent1ty]

Spamming the forum with lots of hi-res pictures, and no, no idea why

[chronologicaldot]

Oo! Oo! I can add to that! I see white-edged boxes! XP

[ent1ty]

white-edged are my favorite kind of boxes

[hendu]

I quite like the bloom on top of the blue lines. It gives a tron-like feel. Now if they only were AA'd.