《Temporal AA , SMAA and MSAA》

《Temporal AA , SMAA and MSAA》

Temporal AA

Temporal AA主要是为了修复场景帧率小于运动物体运动速度的锯齿问题，当帧率太低时候，运动的物体就会一卡一卡，为了避免这种造成的锯齿，原理上帧率刷新速度应该大于两倍运动速度才行。TXAA就是为了修复这种情况。比如高速旋转的轮子用这种AA就会有很好的效果。

伪代码：

For each image frame:
        For each object in the frame:
                Calculate the temporal transformation function for each dynamic attribute
                Determine the areas the object covers during the filtered interval
        For each pixel:
                Determine which objects are covering this pixel at some time in the sampled interval
                Determine the subintervals of time during which each object projects onto this pixel Perform hidden surface removal by removing subintervals of occluded objects Determine pixel intensity function based on the remaining subintervals and the object's temporal transformation function Filter resulting pixel intensity function

上面伪代码中最后一步提到的“temporal transformation function”简单指代一个和物体改变相关的函数，比如，在一帧内物体移动到位置。

一些object的属性比如形状，颜色，位置等作为分析并不是很有效率，如果要使用，它们的值之间需要进行插值，利用B-Splines或者简单的线性插值。

TXAA对于非常简单的物体（比如圆盘）可以在image space，对于特别复杂的几何体，就需要在object space之内进行以上算法的估算。

这种AA在实现的时候，可能会依赖motion vector buffer或者前一帧的变换矩阵来找到上一帧像素对应位置，然后做一些blur。

 

wiki的解释：

https://en.wikipedia.org/wiki/Temporal_anti-aliasing

 

关于UE4的Temporal AA的解释：

https://de45xmedrsdbp.cloudfront.net/Resources/files/TemporalAA_small-59732822.pdf

---

 

SMAA

SMAA(Subpixel Morphological Antialiasing) 是目前效果和效率相对最好的AA方法，基于MLAA的寻边算法，相比于传统的MSAA的耗费大，TXAA(tmporal AA)，FXAA造成比较糊的效果，SMAA是AA算法首选。 

https://github.com/iryoku/smaa

/**
 * Copyright (C) 2013 Jorge Jimenez (jorge@iryoku.com)
 * Copyright (C) 2013 Jose I. Echevarria (joseignacioechevarria@gmail.com)
 * Copyright (C) 2013 Belen Masia (bmasia@unizar.es)
 * Copyright (C) 2013 Fernando Navarro (fernandn@microsoft.com)
 * Copyright (C) 2013 Diego Gutierrez (diegog@unizar.es)
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * this software and associated documentation files (the "Software"), to deal in
 * the Software without restriction, including without limitation the rights to
 * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
 * of the Software, and to permit persons to whom the Software is furnished to
 * do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software. As clarification, there
 * is no requirement that the copyright notice and permission be included in
 * binary distributions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */


/**
 *                  _______  ___  ___       ___           ___
 *                 /       ||   \/   |     /   \         /   \
 *                |   (---- |  \  /  |    /  ^  \       /  ^  \
 *                 \   \    |  |\/|  |   /  /_\  \     /  /_\  \
 *              ----)   |   |  |  |  |  /  _____  \   /  _____  \
 *             |_______/    |__|  |__| /__/     \__\ /__/     \__\
 * 
 *                               E N H A N C E D
 *       S U B P I X E L   M O R P H O L O G I C A L   A N T I A L I A S I N G
 *
 *                         http://www.iryoku.com/smaa/
 *
 * Hi, welcome aboard!
 * 
 * Here you'll find instructions to get the shader up and running as fast as
 * possible.
 *
 * IMPORTANTE NOTICE: when updating, remember to update both this file and the
 * precomputed textures! They may change from version to version.
 *
 * The shader has three passes, chained together as follows:
 *
 *                           |input|------------------·
 *                              v                     |
 *                    [ SMAA*EdgeDetection ]          |
 *                              v                     |
 *                          |edgesTex|                |
 *                              v                     |
 *              [ SMAABlendingWeightCalculation ]     |
 *                              v                     |
 *                          |blendTex|                |
 *                              v                     |
 *                [ SMAANeighborhoodBlending ] <------·
 *                              v
 *                           |output|
 *
 * Note that each [pass] has its own vertex and pixel shader. Remember to use
 * oversized triangles instead of quads to avoid overshading along the
 * diagonal.
 *
 * You've three edge detection methods to choose from: luma, color or depth.
 * They represent different quality/performance and anti-aliasing/sharpness
 * tradeoffs, so our recommendation is for you to choose the one that best
 * suits your particular scenario:
 *
 * - Depth edge detection is usually the fastest but it may miss some edges.
 *
 * - Luma edge detection is usually more expensive than depth edge detection,
 *   but catches visible edges that depth edge detection can miss.
 *
 * - Color edge detection is usually the most expensive one but catches
 *   chroma-only edges.
 *
 * For quickstarters: just use luma edge detection.
 *
 * The general advice is to not rush the integration process and ensure each
 * step is done correctly (don't try to integrate SMAA T2x with predicated edge
 * detection from the start!). Ok then, let's go!
 *
 *  1. The first step is to create two RGBA temporal render targets for holding
 *     |edgesTex| and |blendTex|.
 *
 *     In DX10 or DX11, you can use a RG render target for the edges texture.
 *     In the case of NVIDIA GPUs, using RG render targets seems to actually be
 *     slower.
 *
 *     On the Xbox 360, you can use the same render target for resolving both
 *     |edgesTex| and |blendTex|, as they aren't needed simultaneously.
 *
 *  2. Both temporal render targets |edgesTex| and |blendTex| must be cleared
 *     each frame. Do not forget to clear the alpha channel!
 *
 *  3. The next step is loading the two supporting precalculated textures,
 *     'areaTex' and 'searchTex'. You'll find them in the 'Textures' folder as
 *     C++ headers, and also as regular DDS files. They'll be needed for the
 *     'SMAABlendingWeightCalculation' pass.
 *
 *     If you use the C++ headers, be sure to load them in the format specified
 *     inside of them.
 *
 *     You can also compress 'areaTex' and 'searchTex' using BC5 and BC4
 *     respectively, if you have that option in your content processor pipeline.
 *     When compressing then, you get a non-perceptible quality decrease, and a
 *     marginal performance increase.
 *
 *  4. All samplers must be set to linear filtering and clamp.
 *
 *     After you get the technique working, remember that 64-bit inputs have
 *     half-rate linear filtering on GCN.
 *
 *     If SMAA is applied to 64-bit color buffers, switching to point filtering
 *     when accesing them will increase the performance. Search for
 *     'SMAASamplePoint' to see which textures may benefit from point
 *     filtering, and where (which is basically the color input in the edge
 *     detection and resolve passes).
 *
 *  5. All texture reads and buffer writes must be non-sRGB, with the exception
 *     of the input read and the output write in
 *     'SMAANeighborhoodBlending' (and only in this pass!). If sRGB reads in
 *     this last pass are not possible, the technique will work anyway, but
 *     will perform antialiasing in gamma space.
 *
 *     IMPORTANT: for best results the input read for the color/luma edge 
 *     detection should *NOT* be sRGB.
 *
 *  6. Before including SMAA.h you'll have to setup the render target metrics,
 *     the target and any optional configuration defines. Optionally you can
 *     use a preset.
 *
 *     You have the following targets available: 
 *         SMAA_HLSL_3
 *         SMAA_HLSL_4
 *         SMAA_HLSL_4_1
 *         SMAA_GLSL_3 *
 *         SMAA_GLSL_4 *
 *
 *         * (See SMAA_INCLUDE_VS and SMAA_INCLUDE_PS below).
 *
 *     And four presets:
 *         SMAA_PRESET_LOW          (%60 of the quality)
 *         SMAA_PRESET_MEDIUM       (%80 of the quality)
 *         SMAA_PRESET_HIGH         (%95 of the quality)
 *         SMAA_PRESET_ULTRA        (%99 of the quality)
 *
 *     For example:
 *         #define SMAA_RT_METRICS float4(1.0 / 1280.0, 1.0 / 720.0, 1280.0, 720.0)
 *         #define SMAA_HLSL_4
 *         #define SMAA_PRESET_HIGH
 *         #include "SMAA.h"
 *
 *     Note that SMAA_RT_METRICS doesn't need to be a macro, it can be a
 *     uniform variable. The code is designed to minimize the impact of not
 *     using a constant value, but it is still better to hardcode it.
 *
 *     Depending on how you encoded 'areaTex' and 'searchTex', you may have to
 *     add (and customize) the following defines before including SMAA.h:
 *          #define SMAA_AREATEX_SELECT(sample) sample.rg
 *          #define SMAA_SEARCHTEX_SELECT(sample) sample.r
 *
 *     If your engine is already using porting macros, you can define
 *     SMAA_CUSTOM_SL, and define the porting functions by yourself.
 *
 *  7. Then, you'll have to setup the passes as indicated in the scheme above.
 *     You can take a look into SMAA.fx, to see how we did it for our demo.
 *     Checkout the function wrappers, you may want to copy-paste them!
 *
 *  8. It's recommended to validate the produced |edgesTex| and |blendTex|.
 *     You can use a screenshot from your engine to compare the |edgesTex|
 *     and |blendTex| produced inside of the engine with the results obtained
 *     with the reference demo.
 *
 *  9. After you get the last pass to work, it's time to optimize. You'll have
 *     to initialize a stencil buffer in the first pass (discard is already in
 *     the code), then mask execution by using it the second pass. The last
 *     pass should be executed in all pixels.
 *
 *
 * After this point you can choose to enable predicated thresholding,
 * temporal supersampling and motion blur integration:
 *
 * a) If you want to use predicated thresholding, take a look into
 *    SMAA_PREDICATION; you'll need to pass an extra texture in the edge
 *    detection pass.
 *
 * b) If you want to enable temporal supersampling (SMAA T2x):
 *
 * 1. The first step is to render using subpixel jitters. I won't go into
 *    detail, but it's as simple as moving each vertex position in the
 *    vertex shader, you can check how we do it in our DX10 demo.
 *
 * 2. Then, you must setup the temporal resolve. You may want to take a look
 *    into SMAAResolve for resolving 2x modes. After you get it working, you'll
 *    probably see ghosting everywhere. But fear not, you can enable the
 *    CryENGINE temporal reprojection by setting the SMAA_REPROJECTION macro.
 *    Check out SMAA_DECODE_VELOCITY if your velocity buffer is encoded.
 *
 * 3. The next step is to apply SMAA to each subpixel jittered frame, just as
 *    done for 1x.
 *
 * 4. At this point you should already have something usable, but for best
 *    results the proper area textures must be set depending on current jitter.
 *    For this, the parameter 'subsampleIndices' of
 *    'SMAABlendingWeightCalculationPS' must be set as follows, for our T2x
 *    mode:
 *
 *    @SUBSAMPLE_INDICES
 *
 *    | S# |  Camera Jitter   |  subsampleIndices    |
 *    +----+------------------+---------------------+
 *    |  0 |  ( 0.25, -0.25)  |  float4(1, 1, 1, 0)  |
 *    |  1 |  (-0.25,  0.25)  |  float4(2, 2, 2, 0)  |
 *
 *    These jitter positions assume a bottom-to-top y axis. S# stands for the
 *    sample number.
 *
 * More information about temporal supersampling here:
 *    http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf
 *
 * c) If you want to enable spatial multisampling (SMAA S2x):
 *
 * 1. The scene must be rendered using MSAA 2x. The MSAA 2x buffer must be
 *    created with:
 *      - DX10:     see below (*)
 *      - DX10.1:   D3D10_STANDARD_MULTISAMPLE_PATTERN or
 *      - DX11:     D3D11_STANDARD_MULTISAMPLE_PATTERN
 *
 *    This allows to ensure that the subsample order matches the table in
 *    @SUBSAMPLE_INDICES.
 *
 *    (*) In the case of DX10, we refer the reader to:
 *      - SMAA::detectMSAAOrder and
 *      - SMAA::msaaReorder
 *
 *    These functions allow to match the standard multisample patterns by
 *    detecting the subsample order for a specific GPU, and reordering
 *    them appropriately.
 *
 * 2. A shader must be run to output each subsample into a separate buffer
 *    (DX10 is required). You can use SMAASeparate for this purpose, or just do
 *    it in an existing pass (for example, in the tone mapping pass, which has
 *    the advantage of feeding tone mapped subsamples to SMAA, which will yield
 *    better results).
 *
 * 3. The full SMAA 1x pipeline must be run for each separated buffer, storing
 *    the results in the final buffer. The second run should alpha blend with
 *    the existing final buffer using a blending factor of 0.5.
 *    'subsampleIndices' must be adjusted as in the SMAA T2x case (see point
 *    b).
 *
 * d) If you want to enable temporal supersampling on top of SMAA S2x
 *    (which actually is SMAA 4x):
 *
 * 1. SMAA 4x consists on temporally jittering SMAA S2x, so the first step is
 *    to calculate SMAA S2x for current frame. In this case, 'subsampleIndices'
 *    must be set as follows:
 *
 *    | F# | S# |   Camera Jitter    |    Net Jitter     |   subsampleIndices   |
 *    +----+----+--------------------+-------------------+----------------------+
 *    |  0 |  0 |  ( 0.125,  0.125)  |  ( 0.375, -0.125) |  float4(5, 3, 1, 3)  |
 *    |  0 |  1 |  ( 0.125,  0.125)  |  (-0.125,  0.375) |  float4(4, 6, 2, 3)  |
 *    +----+----+--------------------+-------------------+----------------------+
 *    |  1 |  2 |  (-0.125, -0.125)  |  ( 0.125, -0.375) |  float4(3, 5, 1, 4)  |
 *    |  1 |  3 |  (-0.125, -0.125)  |  (-0.375,  0.125) |  float4(6, 4, 2, 4)  |
 *
 *    These jitter positions assume a bottom-to-top y axis. F# stands for the
 *    frame number. S# stands for the sample number.
 *
 * 2. After calculating SMAA S2x for current frame (with the new subsample
 *    indices), previous frame must be reprojected as in SMAA T2x mode (see
 *    point b).
 *
 * e) If motion blur is used, you may want to do the edge detection pass
 *    together with motion blur. This has two advantages:
 *
 * 1. Pixels under heavy motion can be omitted from the edge detection process.
 *    For these pixels we can just store "no edge", as motion blur will take
 *    care of them.
 * 2. The center pixel tap is reused.
 *
 * Note that in this case depth testing should be used instead of stenciling,
 * as we have to write all the pixels in the motion blur pass.
 *
 * That's it!
 */

//-----------------------------------------------------------------------------
// SMAA Presets

/**
 * Note that if you use one of these presets, the following configuration
 * macros will be ignored if set in the "Configurable Defines" section.
 */

#if defined(SMAA_PRESET_LOW)
#define SMAA_THRESHOLD 0.15
#define SMAA_MAX_SEARCH_STEPS 4
#define SMAA_DISABLE_DIAG_DETECTION
#define SMAA_DISABLE_CORNER_DETECTION
#elif defined(SMAA_PRESET_MEDIUM)
#define SMAA_THRESHOLD 0.1
#define SMAA_MAX_SEARCH_STEPS 8
#define SMAA_DISABLE_DIAG_DETECTION
#define SMAA_DISABLE_CORNER_DETECTION
#elif defined(SMAA_PRESET_HIGH)
#define SMAA_THRESHOLD 0.1
#define SMAA_MAX_SEARCH_STEPS 16
#define SMAA_MAX_SEARCH_STEPS_DIAG 8
#define SMAA_CORNER_ROUNDING 25
#elif defined(SMAA_PRESET_ULTRA)
#define SMAA_THRESHOLD 0.05
#define SMAA_MAX_SEARCH_STEPS 32
#define SMAA_MAX_SEARCH_STEPS_DIAG 16
#define SMAA_CORNER_ROUNDING 25
#endif

//-----------------------------------------------------------------------------
// Configurable Defines

/**
 * SMAA_THRESHOLD specifies the threshold or sensitivity to edges.
 * Lowering this value you will be able to detect more edges at the expense of
 * performance. 
 *
 * Range: [0, 0.5]
 *   0.1 is a reasonable value, and allows to catch most visible edges.
 *   0.05 is a rather overkill value, that allows to catch 'em all.
 *
 *   If temporal supersampling is used, 0.2 could be a reasonable value, as low
 *   contrast edges are properly filtered by just 2x.
 */
#ifndef SMAA_THRESHOLD
#define SMAA_THRESHOLD 0.1
#endif

/**
 * SMAA_DEPTH_THRESHOLD specifies the threshold for depth edge detection.
 * 
 * Range: depends on the depth range of the scene.
 */
#ifndef SMAA_DEPTH_THRESHOLD
#define SMAA_DEPTH_THRESHOLD (0.1 * SMAA_THRESHOLD)
#endif

/**
 * SMAA_MAX_SEARCH_STEPS specifies the maximum steps performed in the
 * horizontal/vertical pattern searches, at each side of the pixel.
 *
 * In number of pixels, it's actually the double. So the maximum line length
 * perfectly handled by, for example 16, is 64 (by perfectly, we meant that
 * longer lines won't look as good, but still antialiased).
 *
 * Range: [0, 112]
 */
#ifndef SMAA_MAX_SEARCH_STEPS
#define SMAA_MAX_SEARCH_STEPS 16
#endif

/**
 * SMAA_MAX_SEARCH_STEPS_DIAG specifies the maximum steps performed in the
 * diagonal pattern searches, at each side of the pixel. In this case we jump
 * one pixel at time, instead of two.
 *
 * Range: [0, 20]
 *
 * On high-end machines it is cheap (between a 0.8x and 0.9x slower for 16 
 * steps), but it can have a significant impact on older machines.
 *
 * Define SMAA_DISABLE_DIAG_DETECTION to disable diagonal processing.
 */
#ifndef SMAA_MAX_SEARCH_STEPS_DIAG
#define SMAA_MAX_SEARCH_STEPS_DIAG 8
#endif

/**
 * SMAA_CORNER_ROUNDING specifies how much sharp corners will be rounded.
 *
 * Range: [0, 100]
 *
 * Define SMAA_DISABLE_CORNER_DETECTION to disable corner processing.
 */
#ifndef SMAA_CORNER_ROUNDING
#define SMAA_CORNER_ROUNDING 25
#endif

/**
 * If there is an neighbor edge that has SMAA_LOCAL_CONTRAST_FACTOR times
 * bigger contrast than current edge, current edge will be discarded.
 *
 * This allows to eliminate spurious crossing edges, and is based on the fact
 * that, if there is too much contrast in a direction, that will hide
 * perceptually contrast in the other neighbors.
 */
#ifndef SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR
#define SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR 2.0
#endif

/**
 * Predicated thresholding allows to better preserve texture details and to
 * improve performance, by decreasing the number of detected edges using an
 * additional buffer like the light accumulation buffer, object ids or even the
 * depth buffer (the depth buffer usage may be limited to indoor or short range
 * scenes).
 *
 * It loca