Anti-aliasing Anti-Alasing game in refining summary

The root cause of sawtooth (aliasing distortion) generated

Graphics is a fundamental process graphics transformed into an image of the process, the process is a continuous signal after sampling converted into discrete signal (pixel display device is discrete) process, had been born aliasing (distortion, aliasing) because discrete signal recognized by human eyes. Anti-aliasing is through some means to reconstruct the lost signal. We usually use anti-aliasing algorithms actually reduce the jagged, not anti-aliased.

Aliasing of two sources: the color information of the sampling point is lost and the visibility of the information of the sampling points (including coverage and occlusion Occlusion cover) is missing ;
Antialiasing (reset signal) of the basic method: multi-sample (multi-sampling) and the pre-filter (pre-filter)

Classification of anti-aliasing algorithms

In the preceding line (forward rendering), by increasing the sample, and precisely in the edge antialiasing method using depth information: SSAA, MSAA, CSAA (MSAA optimized version of the five-point sampling), RGSS (sampling point of rotation, irregular )Wait.
After the line segment (delayed rendering) by workup pure broken, i.e. perform graphics antialiasing algorithm: FXAA, MLAA, SMAA like.
The anti-aliasing filter types: anti-noise similar to the image restoration and image processing techniques (high frequency noise can be regarded as a sawtooth), there are many filtering algorithm, for example: a geometric or arithmetic mean filter, median filter, adaptive mean filter etc., depending on the difference between the filter weights convolution operator redesign and mixed mode. NOTE: This filtering is performed in screen space, of course, also possible to take the edge detection (Laplacian convolution kernel and the like), filtered and then optimizes the routes only at the edges.
Design pre-filter, the aliasing information is passed to the rear section of the pipeline as possible, the pipeline may have to be modified: NSAA, AGAA.
Antialiasing (antialiasing above all other screen space) in the time domain: temprial anti-alasing, TXAA.
Other: Based DLAA deep learning, real-time anti-aliasing.

Ground simple and crude and primitive SSAA MSAA, anti-aliasing technology

SSAA (Super Sample Anti-Aliasing), super-sampling anti-aliasing

Simple multiples of excessive force to more accurately restore the signal samples, 4x, 8x, 16x and the like.
GPU, shader iteration mode points per pixel and per sample, if you use the SSAA GPU uses per sample mode.
Finally coverage mask downsampling.

MSAA (Multi-Sample AntiAliasing), multi-sample anti-aliasing

Rasterization stage occlusion coverage information calculated binary mask and stores the color information from the color of the pixel from the center of the nearest sampling points, the pixel depth test fails sampling points is not colored.
MSAA is generally supported by the hardware anti-aliasing (anti-aliasing, AA) technology, which uses a plurality of sample depth value of each pixel in a pixel region rasterization stage, the depth of these sample values in each pixel sharing a shading calculations, i.e., each pixel is still performed only once fragment shader, the result of the calculation is then copied to each of the sub-sampling point depth, such a depth test fails pixel sub-sampling points will not include color information, can be more that the final color transition smoother.
MSAA SSAA different and each pixel is that only MSAA shading calculations performed once (for visibility, color aliasing sources of weakness), and therefore have greater than SSAA performance advantages, but can not be processed and therefore MSAA coloring due to the function calculation (such as high light) due to insufficient sampling aliasing (e.g. Shader aliasing), but because the human eye sensitivity to a higher edge geometry in color sensitivity , and therefore is a very favored MSAA techniques.

https://www.cnblogs.com/ghl_carmack/p/8245032.html

This super-sampling anti-aliasing methods have occurred in the preceding line, it does not support a delay in rendering.

## after processing type antialiasing

FXAA (Fast Approximate Anti-Aliasing), Fast Approximate Anti-Aliasing

A partial empirical algorithm, algorithm designers do not know why the anti-aliasing algorithm will be effective.
Fuzzy inevitably present their own problems after treatment antialiasing method, for complex objects like fluff effect a good class.
The main luminance luma edge detection, and then anti-aliasing operation on the edge;

http://blog.simonrodriguez.fr/articles/30-07-2016_implementing_fxaa.html

MLA

MLAA basic idea is: an edge detection on each frame image (usually luminance, color, depth, or normal edge detection), then these edges pattern recognition, classification that Z, U, L three shapes, the shape of the edge is re-vectoring ** (re-vectorization) ** reverse restore its original shape, and the upper edge of the pixel is calculated according to the coverage area of the mixing weights, which mixed with the color of the surrounding, so as to achieve a smooth serration .

SMAA(Subpixel Monorplogical Anti-Aliasing)

http://www.iryoku.com/smaa/

https://blog.csdn.net/qezcwx11/article/details/78426052

SMAA core principles derived from the MLAA, edge detection added contrast consideration. SMAA is binding and MLAA SSAA / MSAA respective advantages, reduce post-processing blur caused FXAA antialiasing.
Edge Detection added contrast considerations;
Vertical and horizontal directions into zigzag.

Compare

Compared FXAA SMAA will be more blurred, the contrast effect as follows: https://images.nvidia.com/geforce-com/international/comparisons/just-cause-3/just-cause-3-anti-aliasing-interactive-comparison- 001-smaa-vs-fxaa.html

The new anti-aliasing technology

TAA / TXAA (Temporal Anti-Aliasing), time anti-aliasing

NVIDIA design of higher quality anti-aliasing technology.

https://zhuanlan.zhihu.com/p/24063209

https://docs.nvidia.com/gameworks/content/gameworkslibrary/postworks/util.html

https://zhuanlan.zhihu.com/p/20786650

https://developer.download.nvidia.cn/gameworks/events/GDC2016/msalvi_temporal_supersampling.pdf

SSAA and achieve the effect, but is based on multiple sampling frame buffer history, the history samples from the frame, the pixel weighting jitter range.
The disadvantage is that with the accumulation of historical colors, can cause blurred (motion blur) can not be absolutely eliminated, especially in fast moving object or lens cases, can lead to ghosting, especially VR device This method does not apply.
AGAA problem for the birth of the TAA, but is still not mature, high hardware requirements.

AGAA (Aggregate G-Buffer Anti-Aliasing), G buffer enhanced anti-aliasing

AGAA using MSAA idea, but the polymerization prefilter through pre-filter will result in reduced sampling MASS multiple G-Buffer memory is too big problem, the multiple subsampling filter points to less aggregation.

NSAA (None-Sampling Anti_Aliasing), the non-sample anti-aliasing

Special for visibility analysis type prefilter, edge antialiasing.
Instead of applying the traditional Chinese raster rendering, the new pipeline pipeline, currently not applicable.

DLAA

。。。

The overall idea of the algorithm improvements:

Edge detection, only the edge antialiasing (Contrast Adaptation)
Jagged edges form of pattern recognition, information restored again vectorization
The sensitivity of the human eye color geometric edges higher
pre-filter prefilter, a direct multiple samples to reduce large amounts of memory problems
Gathering information on the historical frame of time domain oversampling to reduce the oversampling memory had caused a big problem