文章

Snapdragon™ Game Super Resolution

Snapdragon™ Game Super Resolution

Snapdragon™ Game Super Resolution

Introduction

Snapdragon™ Game Studios developed Snapdragon™ Game Super Resolution (Snapdragon™ GSR or SGSR), which integrates upscaling and sharpening in one single GPU shader pass. The algorithm uses a 12-tap Lanczos-like scaling filter and adaptive sharpening filter, which presents smooth images and sharp edges.

Our solution provides an efficient solution for games to draw 2D UI at device resolution for better visual quality, while rendering the 3D scene at a lower resolution for performance and power savings.

The technique has visual quality on par with other spatial upscaling techniques while being highly optimized for Adreno™ GPU hardware

Integration

Our shader provide a set of configurations that can be adjusted to fine tune your needs, these are briefly explained below and can be found directly on the shader file.

The shader requires an input texture (your low-res 3D scene) and viewport rect, it doesn’t require an anti-aliased input but works best with one.

Operation mode

SGSR comes with 3 operation modes: RGBA, RGBY and LERP, by default the shader is set to use RGBA (mode 1).

Edge threshold

We suggest keeping its value as 8.0/255 for mobile applications but changing it to 4.0/255 if targeting VR.

Edge sharpness

By default this value is set to 2, but any number in the range [1.0, 2.0] can be used.

Vulkan

If targeting vulkan, uncomment the UseUniformBlock define.

Performance

Runtime cost on Snapdragon

SGSR is well optimized to run on Adreno™ and generally can provide maximum wave occupancy. Considering input content with a resolution of 1240x576, the following frame times were measured using our most recent Snapdragon™ chips:

SGSR

Edge ThresholdSD888SD8Gen1SD8Gen2
40.4756 ms0.3563 ms0.2952 ms
80.4754 ms0.3543 ms0.2939 ms

BILINEAR/LERP

SD888SD8Gen1SD8Gen2
0.2476 ms0.1963 ms0.1558 ms

Device configurations at maximum GPU frequency:

  • Snapdragon 888 (2340x1080)
  • Snapdragon 8 Gen1 (2340x1080)
  • Snapdragon 8 Gen2 (2400x1600)

SGSR performance scales lineally with the configured scaling ouput region (configurable on the shader). RGBA shares the exact same speed with RGBY.

Besides the provided configurations, the frame time also depends on the input texture resolution and final output resolution.

Details visual comparison

With Snapdragon GSR, 1080p games can become sharper 4K games. Games that were only 30 FPS can be played at 60+ FPS so graphics look even smoother. And since performance is correlated with power, you can get these features while extending battery life and gameplay time.

Visual comparison with Bilinear

1706492387597snapdragon_gsr_video.gif

Artifact (Ray-Tracing enabled)

Performance and visual comparison of rendering at Native (1080p) resolution vs. using Snapdragon GSR upscaling. This demonstration was created in Unreal Engine 5 with Ray Query shadows at 2 rays per pixel.

21.7FPS alt text

50% screen res upscaled - 60FPS alt text

Discovery

Visual comparison of rendering at Native (4K) vs. Bilinear vs. Snapdragon GSR.

Original alt text

Bilinear (lerp) alt text

SGSR alt text

Closeup alt text

Other Applications

The high performance, low latency, single-pass nature of SGSR makes it well suited for XR applications. Unlike other popular super resolution algorithms, SGSR uses only a single GPU pass, so it can be combined with other operations minimizing latency and additional bandwidth.

Although not provided here, our shader can be mixed with AA techniques, potentially saving a lot of application bandwidth.

Version history

VersionDate
1.0.02023-06-29

References

Introducing Snapdragon Game Super Resolution, https://www.qualcomm.com/news/onq/2023/04/introducing-snapdragon-game-super-resolution

Using Super Resolution to Boost Resolution in Virtual Reality, https://developer.qualcomm.com/blog/using-super-resolution-boost-resolution-virtual-reality

License

Snapdragon™ Game Super Resolution is licensed under the BSD 3-clause “New” or “Revised” License. Check out the LICENSE for more details.

本文由作者按照 CC BY 4.0 进行授权