Surface based anti-aliasing

Abstract

Figure 1: From left to right: 8x super-sample anti-aliasing (SSAA), 8x multi-sample anti-aliasing (MSAA) and surface-based anti-aliasing (SBAA) with 8 visibility and 2 surface samples per pixel. The circles represent visibility samples, while the blue and red discs represent shading samples from two different surfaces. The four red primitives sharing the same vertex are part of the same foreground surface. Our MERGE2 algorithm exploits this configuration and shades only one sample for all four red primitives while reserving a second surface sample for the blue background surface. Unlike multi-sampling, SBAA based algorithms impose an upper bound on the number of captured, stored and shaded surfaces rather than primitives in each pixel, therefore significantly reducing storage and shading costs. AbstractWe present surface based anti-aliasing (SBAA), a new approach to real-time anti-aliasing for deferred renderers that improves the performance and lowers the memory requirements for anti-aliasing methods that sample sub-pixel visibility. We introduce a novel way of decoupling visibility determination from shading that, compared to previous multi-sampling based approaches, significantly reduces the number of samples stored and shaded per pixel. Unlike postprocess anti-aliasing techniques used in conjunction with deferred renderers, SBAA correctly resolves visibility of sub-pixel features, minimizing spatial and temporal artifacts.

Decoupled coverage anti-aliasing

“…As with some previous approaches [Salvi and Vidimče 2012;Kerzner and Salvi 2014], our algorithm exploits the fact that we cannot distinguish many distinct surfaces within a pixel, by merging similar fragments together to reduce the number of per-sample : (a,b,c) projecting the triangle onto the pixel plane in either perspective space or normalized device coordinates, (d,e,f) using each edge to index into a lookup table to determine coverage for a half plane, and bitwise AND the three edge coverages to get the fragment's sub-pixel coverage map.…”

“…We compare, variously, with SBAA using "merge" strategy and 4 surfaces per pixel [Salvi and Vidimče 2012], 8× MSAA, and a 512× samples per pixel ground truth. …”

“…To reduce the size of the G-Buffer, Surfaced-Based Anti-Aliasing [Salvi and Vidimče 2012] introduced the notion of consolidating fragments into a small set of surfaces that could be independently shaded. To do this, they analyze the result of a MSAA prepass, then find the most important n fragments to store in the G-Buffer.…”

“…For example, switching a 64×MSAA system at 1920 × 1080 resolution from regular forward rendering to deferred shading would increase the memory footprint from 1GB to 2GB for a typical G-Buffer configuration. Although researchers have begun to explore approaches to compress the G-Buffer by storing a small set of aggregate surfaces [Salvi and Vidimče 2012;Kerzner and Salvi 2014;Crassin et al 2015], these approaches do not address the first problem with the large size of the visibility map for high-quality anti-aliasing.…”

Aggregate G-buffer anti-aliasing

“…Our work improves on the approach of Salvi et al [2012], which analyses the results of dense geometry sampling during rasterization to identify and retain exactly one fragment from the n "most important" surfaces per pixel. However their method discards information from all other surfaces, which leads to aliasing in situations where many surfaces contribute to a pixel's appearance.…”