Comments on: Strided blur and other tips for SSAO

By: Jason Hoerner

Jason Hoerner — Tue, 06 Apr 2010 20:00:45 +0000

Thank you so much for sharing this! I was banging my head against the wall for the last week trying to figure out how to do a decent blur. All the websites on SSAO spend a lot of time discussing the shader algorithm, but gloss over the blur. I think the blur is the hard part. Once I found this website, and tried a 2-stride blur, it magically solved all my problems!

I figured out a way to completely eliminate the interleave pattern with an additional blur pass. You can use a 3×3 mean filter, using the same depth / normal aware sampling algorithm. You need to adjust the weights of the samples depending on whether you use a 3-stride or 2-stride.

For a 3-stride pattern, you can just use constant weights (1 for all). For a 2-stride, you can use weights in a pattern like this:

[ 1 2 1 ]
[ 2 4 2 ]
[ 1 2 1 ]

This is equivalent to averaging the four 2×2 blocks of pixels surrounding the center point (you can’t just do 2×2 because it will shift the image diagonally). Enjoy!

By: Timothy Farrar

Timothy Farrar — Fri, 18 Sep 2009 14:36:22 +0000

Would be interesting to try adjusting filter tap location by a function of VPOS to attempt to break up the interleave pattern (assuming TEX bound with free ALU capacity). Might play havoc on texture cache however.

Might also be interesting to play with temporal feedback in the filtering (like Gears II) but in combination with Ignacio’s suggestion of importance sampling and changing filter position temporally.

By: Aras Pranckevičius

Aras Pranckevičius — Fri, 18 Sep 2009 04:28:04 +0000

@imbusy: ok ok. But it works! :)

@icastano: yeah, makes sense.

By: icastano

icastano — Fri, 18 Sep 2009 00:15:58 +0000

Something that I’ve suggested a few times, but never tried myself, is to use some sort of importance sampling. Pixels closer to the center should be sampled at a higher frequency, since they have more important contributions, than pixels with lower contributions. Note that you would also have to readjust the weights. This is basically like the 2D poison disk distribution that ATI used in some of the old demos, but on one dimension only.

By: imbusy

imbusy — Thu, 17 Sep 2009 19:32:58 +0000

Using bilateral or just discarding – it’s still not separable.

By: Aras Pranckevičius

Aras Pranckevičius — Thu, 17 Sep 2009 17:46:18 +0000

@imbusy: well yeah, it’s not bilateral in fact. I’m just doing kinda-Gaussian, and discarding samples that don’t meet some criteria.

@steve: word!

@ReJ: I don’t remember us using it, actually. Or I have already forgot about it :)

By: Real-Time Rendering » Blog Archive » Clearing the Queue

Real-Time Rendering » Blog Archive » Clearing the Queue — Thu, 17 Sep 2009 17:35:39 +0000

[...] This is one more reason the Internet is great: an in-depth article on normal compression techniques, weighing the pros and cons of each. This sort of article would probably not see the light of day in traditional publications, even Game Developer – too long for them, but all the info presented here is worthwhile for a developer making this decision. Aras’ blog has other nice bits such as packing a float into RGBA and SSAO blurring. [...]

By: ReJ

ReJ — Thu, 17 Sep 2009 16:24:46 +0000

@Aras: I can’t believe you didn’t knew this “trick”… I’m pretty sure we’ve been doing it in late Interamotion days even.

By: steve

steve — Thu, 17 Sep 2009 09:24:45 +0000

I discovered something similar when doing PCSS on a project last year – that although I should have needed to increase the number of PCF samples as the kernel widened, it was surprising what you could actually get away with in practice. And interestingly, the more detailed / non-uniform your diffuse textures are, the more you can get away with, because the diffuse channel makes it even harder to spot the ‘dithering’ patterns.

Moral of the story, theory sucks, random experimentation FTW :)

By: imbusy

imbusy — Thu, 17 Sep 2009 09:00:53 +0000

It shouldn’t be separable, but hey, you said it yourself, we’re not building nuclear reactors :)

By: Aras Pranckevičius

Aras Pranckevičius — Thu, 17 Sep 2009 08:55:56 +0000

@imbusy: yeah, something like that. I’m not sure if it’s exactly “bilateral”, but basically it discards samples based on depth/normal difference.

By: imbusy

imbusy — Thu, 17 Sep 2009 08:50:20 +0000

So that’s bilateral blur you’re using, not gaussian, right? Otherwise how do you get it to respect the boundaries?