Various Rendering Techniques: Sparse Rendering, Temporal Reprojection, Artifacts Filtering

There is a full res depth pass, so there is a cost.

There is also memory footprint to consider whilst developers can also enchant effects that eat up shader/tex power (context of 1800c vs 2160c). Slightly larger streaming buffers might not be a bad thing either.

 

There is a huge distance between it being free and it costing about as much as simply rendering at native. CB is not free, but it is still considerably cheaper than bruteforce.

 

Lets talk about actual numbers instead of doing semantic, from the Dice slides about BF1 and Mass effect (and that's with some optimization, for instance they use RPM for their CBR resolve):

BF1 : ~16ms at 1800cb instead of ~21ms at 1800p (native costs ~5ms more): CBR is 76% cost of native.
Mass effect : ~23.4ms at 1800cb instead of ~36.8ms at 1800p (native costs ~13.4ms more): CBR is 64% cost of native.

Back to semantic: 'CBR is not free, cause free should be 50% cost of native'

 

Who are you arguing with?

 

Certainly would be more interesting to wonder about the techniques themselves and what may be the good and bad in each.

Been quite interested on MSAA trick and Geometry rendering and similar.
You can get decent to good quality edges and original resolution shading. (There could be problems with shading propagation if ID is not expected.)
I'm quite sure that the quad pixel pattern for shading could be fixed with TAA or small shader.
In general overall quality could be better than we have seen thus far with naive methods and no AA. (Even simple filtering for quads and FXAA could be a lot better.)

 

CBR gives you textures that have perceptibly the same clarity than native textures. MSSA trick and Geo rendering do not.

 

I think msaa trick does. They interpolate UV positions.

 

Yes. MSAA trick adds zero blurring to textures. Textures are sampled per pixel. Temporal checkerboarding technique blurs textures slightly in motion (history sampling point almost never hits exactly previous pixel center so it must be interpolated = adds blur). Spatial checkerboard (no temporal reuse) with depth or id map reconstruction keeps edge pixels precise (perfect high contrast silhouettes), but textures are blurred, as texture sampling rate is 1/2 per frame and there's only one frame of texture data used. Odd pixels are simply interpolated.

 

"MSAA trick" is, more or less, just taking advantage of how rendering to g-buffers with certain formats* doesn't require per-pixel texturing and shading ops. Hence, rendering a high-res g-buffer can be implemented as rendering a low-res g-buffer with MSAA.

In this context, it doesn't refer to stuff like simply using a low-res backbuffer with MSAA as a full-res backbuffer, or aliasing a high-res backbuffer as a low-res buffer with MSAA to do things like compromise alpha blend quality for speed.

*In particular, certain formats which inform you of how to sample the textures at any given "pixel," rather than including the sampled texture data directly.

 

Which games do use the MSAA trick ?

 

At least Bound on ps4pro. (They render 1920x1080 8xMSAA to get 4k +2xMSAA buffer.)
For transparency Lost Planet 2.

Pretty sure id-buffer based geometry rendering is used in Gravity Rush 2 and Blue Reflection. (Could be MSAA trick as well, kind of hard to tell them apart in images.)

 

There are multiple different techniques that exploit MSAA hardware. I was referring specifically to our MSAA-trick technique presented at Siggraph 2015 (http://advances.realtimerendering.c...siggraph2015_combined_final_footer_220dpi.pdf).

 

There is more to it in sebbbi's implementation which I'm refering. He does use interpolation to reconstruct some data that was not directly rasterized. But it does that with the g-buffer, before lighting and shading, and as mentioned, by interpolating texture UV instead of actual material values, the reconstruction is near perfect.