Game Development Reference
Figure 3.1. In this example our deferred shading method (a) achieves equivalent an-
tialiasing quality to 8 × MSAA, but (c) significantly reduces the number of shader eval-
uations. (b) To the same antialiasing quality, classic deferred shading needs a super-
instead we deduplicate the data itself, ensuring that a visible surface is shaded
only once, regardless of the number of subsamples it covers.
This article is based on our recent research paper, presented at the 2012 ACM
Symposium on Interactive 3D Graphics and Games [Liktor and Dachsbacher 12].
We cover the basic theory of decoupled sampling, and then focus on the imple-
mentation details of our new G-buffer in the OpenGL pipeline.
Decoupled Sampling in a Rasterization Pipeline
The Nature of Aliasing
To understand the motivation of decoupled sampling, let us consider the ren-
dering of a 2D image as a signal-processing problem. Rasterization uses point
sampling to capture visible surfaces that causes problems if the sampled signal is
not band-limited: frequencies higher than the sampling frequency lead to aliasing
in the rendered image. Antialiasing methods can prefilter the signal to eliminate
frequencies above the sampling limit, increase the frequency of sampling, or al-
ternatively apply reconstruction filters to supress aliasing artifacts.
Any rendering method using point sampling must first solve the visibility
problem to find the surface points that determine the colors at each sample.