An implementation of Checkerboard Rendering using Magnum and OpenGL.
Checkerboard rendering is a temporal upsampling technique that reduces shading work with minimal impact on visual quality. Expensive fragment shader operations are run at half-resolution and then combined with the previous frame's output in a resolve shader.
By always shading 2 MSAA samples and jittering the viewport by one pixel in every odd frame, we can combine the current frame's samples A and B with last frame's samples C and D to fill the full-res framebuffer:
even frame: odd frame: combined frame:
(current) (previous) (full-res)
+---+---+ +---+---+ +---+---+
| | A | | | C | | C | A |
+---+---+ + +---+---+ = +---+---+---+
| B | | | D | | | D | B |
+---+---+ +---+---+ +---+---+
While this is trivial for static scenes, complications arise with moving objects/cameras and (dis)occlusion. The resolve step needs to consider per-pixel velocity between frames and detect and handle fragments being occluded.
For a more thorough introduction, check out Intel's white paper.
This implementation is loosely based on the Intel D3D11 forward shading example, while also introducing improvements from several other papers (see References). Here's a rough overview of all the steps:
- Velocity pass
- Render full-res per-pixel screenspace velocity buffer
- Only dynamic objects; for static objects camera reprojection is used in the reconstruction pass
- Jitter camera viewport
- Translate a full-res pixel to the right
- Happens every second (= odd) frame
- Checkerboard pass: Render scene at quarter-res
- MSAA 2X framebuffer with fixed sample positions
- Fragment shader is run on both samples
- Reuses downsampled velocity pass depth to reduce overdraw
- Reconstruction pass: Combine previous and current quarter-res
- Fullscreen pass, one thread for each full-res fragment
- See Reconstruction shader for more details
The reconstruction shader is run for each full-res fragment. It performs the following steps:
- Determine if the current fragment was shaded by this frame
- This is true if this fragment falls on the current sample positions. Sample positions depend on the current camera jitter
- If so, return the color and exit
- Resolve the current color from last frame's output
- Sample velocity buffer to calculate the current fragment's position in the previous frame
- For static objects the velocity is instead calculated by reprojecting the current world space position through the previous camera matrices
- Check if the previous fragment position doesn't fall on the previous frame's sample positions
- This is the case when any movement cancelled the camera jitter
- If so, return the neighborhood average and exit
- Check for occlusion
- Done by comparing the reprojected depth against its neighbors with a configurable tolerance value
- If occluded, return the neighborhood average and exit
- Fetch the reprojected fragment color
- Sample velocity buffer to calculate the current fragment's position in the previous frame
- Clamp result
- Clamp the reprojected color to the min/max values of the neighborhood [El Mansouri]
- To allow for small-scale details, blend back to the actual color using a confidence value
- Confidence is based on the neighborhood depth difference
- To prevent bright samples from dominating the result, a simple tonemapping operator is used before averaging
- Using a differential blend operator [Wihlidal, Jimenez] here greatly reduces checkerboard artifacts at color discontinuities
There are a few more edge cases and some extra debug output not mentioned here. The full GLSL code can be found in ReconstructionShader.frag.
OpenGL doesn't have a fixed set of MSAA sample positions, but several extensions exist that allow you to manually specify them:
The ARB extension is based on the NV extension and is in practice only supported by Nvidia hardware.
To enable per-sample fragment shader invocations, ARB_sample_shading (or OpenGL 4.0) is required. Setting a MIN_SAMPLE_SHADING_VALUE of 1.0 causes per-sample fragment evaluation, as well as exact interpolation of input values to the sample position.
Since screen-space derivatives in the fragment shader are calculated at half-res, they have twice the magnitude compared to full-res rendering. This is especially detrimental for texturing since larger UV derivatives cause higher MIP levels and therefore blurriness. To fix this, use textureGrad with corrected gradients or add a LOD bias of -0.5 to all texture samplers.
- Transparent objects cause artifacts since the velocity used for reprojection accounts for the transparent object, not anything behind it. Look into ways to improve this.
- Extensive benchmarks. Provide some hard numbers on frame times using a more complex scene and realistic shading work.
- Implement a deferred rendering variant. Intel's implementation uses separate render targets for storing opaque lighting and the transparent forward pass which doesn't seem entirely necessary.
- Use a different color space than RGB. YCoCg might be a better candidate for color clamping/clipping.
- Variable Rate Shading (through e.g. NV_shading_rate_image) could be used in place of per-sample shading. A 2x1 shading rate would allow interlacing columns similar to SMAA TU2x.
- Checkerboard Rendering for Real-Time Upscaling on Intel Integrated Graphics (Trapper Mcferron, Adam Lake, 2018)
- Rendering Rainbow Six Siege (Jalal El Mansouri, 2016)
- 4K Checkerboard in Battlefield 1 and Mass Effect (Graham Wihlidal, 2017)
- Dynamic Temporal Antialiasing and Upsampling in Call of Duty (Jorge Jimenez, 2020)
CMake (>= 3.4) is required for building.
-
Clone repository:
git clone --recursive https://github.com/pezcode/mosaiikki.git cd mosaiikki -
Generate project files:
mkdir build cd build # e.g. VS 2019, compile for x64 platform cmake -G "Visual Studio 16 2019" -A x64 .. cd ..
-
Build. Open the project files with your IDE/build tool, or use CMake:
cmake --build build/ --parallel --config Release
- Magnum for rendering and asset import
- Dear ImGui for the UI