Just to let you guys know, the demo for variance shadow maps that we showed at GDC is now available (executable and source) at http://www.punkuser.net/vsm/.
Enjoy!
Enjoy!
Variance Shadow Maps Demo
- Thread starter Andrew Lauritzen
- Start date
DudeMiester
Regular
Nice, I've been following your thread at GameDev.net. :smile: imho, this is the future of shadowing, well unless the IHVs deliever a hardware filter like AF just for shadow maps
(I'm assuming that FX16 is shorthand for fixed-point 16-bit... I could be wrong.)fellix said:
The fixed-point surface should actually have better precision than floating point, since we effectively waste the exponent bits (our range *is* bounded). The reason why both are included is that fixed-point filters in hardware on ATI, and floating-point filters in hardware on NVIDIA. (Note that for the cases where hardware filtering is unsupported, bilinear is simulated in the shader, although mipmapping and aniso are not for performance reasons.)
Oh yeah? Well you were on the right track as it turns out that moments *are* the representation that filters linearly.Simon F said:
Yeah it shouldn't be too hard to work out the few remaining artifacts - we have some ideas, just haven't had the time. I'm also convinced that this is the way to go...DudeMeister said:
Hardware PCF working with AF would be nice, but it only solves part of the problem. It still doesn't solve mipmapping (and there's no way to do that efficiently using brute-force PCF) and it still does not allow arbitrary filtering (stuff like multisampling and mipmap LOD bias is very useful for shadow maps too).
Thanks for the feedback so far.
Yeah that's a bit misleading - as I mention in the README, I couldn't get derivative biasing to work consistently on both NVIDIA and ATI, so I just disabled it for now. PCF should produce effectively the same shadows as VSM, except that it is far more expensive and thus is clamped to relatively small filter kernels (ex. 5x5).Bludd said:
Some shots of mine, without filtering:
PCF - FX16 - FP16 - FP32 ;
Brief note: Filter kernels over 5x5 is overkill, at least for the given scene with a single or two spot lights at such a close range.
p.s.
Fixed-point format is obviously producing artifacts at low view angles and hi-res shadow maps:
FX16 - FP16 ;
PCF - FX16 - FP16 - FP32 ;
Brief note: Filter kernels over 5x5 is overkill, at least for the given scene with a single or two spot lights at such a close range.
p.s.
Fixed-point format is obviously producing artifacts at low view angles and hi-res shadow maps:
FX16 - FP16 ;
Last edited by a moderator:
It also depends on the shadow map resolution obviously... you can pick anything from 64x64 to 2048x2048 IIRC.fellix said:
That's really odd... I'll have to check that. What card do you have? It's possible that fixed-point just isn't working correctly at all on NVIDIA (I've not tested it much... the idea is to use FP on NVIDIA and FX on ATI). If that's on ATI then I'll have to try and reproduce that.fellix said:
Also the numeric artifacts in the fp16 shot (the first one) look odd... nothing like what I'm used to, and there shouldn't be that many of them... in fact they should be practically invisible without blurring.
Oh ok... in any case it's just a constant in the main program naturally. There are some limits of what are renderable in GL though - I don't recall what they are for each vendor.fellix said:
Should be just fine then, although as I mentioned, fp16 is the ideal VSM mode for NVIDIA.fellix said:
Huh, well I'm not sure then. I know that NVIDIA and ATI have different fp16 formats though, so I may have to retest on NVIDIA after a made a few changes for ATI. In any case, it is possible to eliminate those without much workfellix said:
We actually use a 4xRGBA16 since 2xRGBA16 isn't currently renderable (at least in OpenGL) anyways. I honestly haven't done a lot of testing with fixed-point on NVIDIA though so it may not work as expected... that said, use fp16Zengar said:
Nice job. Enjoyed looking through it.AndyTX said:
epic
AndyTX said:We actually use a 4xRGBA16 since 2xRGBA16 isn't currently renderable (at least in OpenGL) anyways. I honestly haven't done a lot of testing with fixed-point on NVIDIA though so it may not work as expected... that said, use fp16
Sorry, I meant that nvidia doesn't support any fixed-point precision above 8 bits. There is a pdf somewhere at their page displaying all texture formats nvidia GPUs support. All above RGBA8 are not supported. Fp16 is the correct way
Ooh wow, that's awesome, thanks for the link! Evil auto-substitution thoughZengar said:
Well 8-bit precision would explain why the result is so poor on NVIDIA. Perhaps I'll release a demo that forces the use of fp16 on NVIDIA and fx16 on ATI since that's really the way to go.
