[buckmaster]

Slick as! Coming along very nicely

[andrewmac]

I've been waiting for this presentation to come out ever since SIGGRAPH 2015 courses were announced:

http://advances.realtimerendering.com/s ... Remedy.pdf

It essentially validates my approach here. While they don't use the exact same method, they use the same overall approach. The approach is to use a rough approximation for large scale GI diffuse contributions (in their case irradiance volumes, in my case conservative voxelization) and then use a screenspace technique to fill in the micro details and short range light bounce and reflections (SSR + SSDO in my case). It saves you from having to use large resolution voxel grids to get fine details since they'll be filled in by screenspace techniques.

In other news I started to improve shadows last night and while I certainly made some strides I'm really not happy with the results. I have a few ideas to get to the quality I'd like though.

[HeadClot]

So question - Will this Global Illumination solution require an Nvidia Card to use or will it be hardware agnostic?

[andrewmac]

It's hardware agnostic, however I should mention none of the GI techniques are specific to any hardware, they just run better on certain cards. NVidia showed off VXGI because their new cards are powerful enough to voxelize the whole scene and still get 30fps. My method is more conservative (at a sacrifice of quality and accuracy) so it's less demanding on the GPU. It does however require compute shaders so it won't work on older DX9 hardware.

[andrewmac]

I tried to implement SMAA yesterday but I failed. I don't want to spend too much time on a single antialiasing technique when there are so many options available so I abandoned ship on SMAA for now and implemented DLAA, the solution developed for The Force Unleashed 2:


Full Size: http://i.imgur.com/Kal2WKh.png

As you can see it does a pretty damn good job of getting rid of the aliasing. It's a bit slower than FXAA but not by a lot. FXAA just wasn't quite cutting it in my opinion. Looks like the first version of Torque 6 will come with FXAA and DLAA for antialiasing options.

[Gav]

I don't know if it's worth it but somewhere i read that DX10+ has support for multisampled textures which can be used to implement MSAA along with deferred shading.

[Timmy]

I implemented SMAA in DM a few months back, worked great. Although i haven't got around to implementing the better S2x etc version of it. None the less worked far better than FXAA as it was. I'm guessing it's far easier in T3D though considering you can use HLSL directly

[Chelaru]

Looks good, but why is there light bleed ?

[andrewmac]

@

Timmy : The implementation was straight forward but it just didn't work. The edge detection pass looks correct, the weighting pass doesn't quite look right (but not super far off either) and then the neighborhood pass does nothing. The way it's setup my next step would be to debug his actual code and I don't want to do that. When I looked around on the internet I found 4 or 5 forum posts who got to the same point. Weight pass just doesn't look right and nothing interesting happens in the final pass. I didn't scrap the code though so I'll still end up pushing it, just disabled. Maybe myself or someone else can find the problem later but it was just turning into too much of a time sink.

@

Chelaru : Right now we can easily fix shadow bleed by adjusting the depth bias but that's not really how I'd like to see it handled. I think in a modern engine tweaking shadowmap settings is not something a person should have to do. Other than perhaps an overall quality setting that chooses shadowmap resolution I don't think an artist should have to concern themselves with things like number of cascades, depth bias, split distribution, etc. I'm hoping to settle on a solution that won't require people to think about shadow settings.

Simply put I keep the light bleed visible so it's a stain on the screenshots so I'll force myself to implement a proper solution and not just tweak to fix it in one scene. It's next on the chopping block though. EVSM + SDSM seems the ideal solution, we'll see how that goes.

[Timmy]

The exact same thing happened on my end when I tested it with T3D's new OpenGL renderer, D3D worked perfectly. Like you I didn't bother digging any deeper as there is a good chance we won't be using GL anyway ( I have revived the D3D11 stuff that anis started).