02/25/2021 | News release | Archived content
You can improve the runtime efficiency of your shader without sacrificing the quality of your graphics by packing physically based rendering (PBR) material information into a single texture map and layering it into a compact shader. Check out this experiment.
This experiment works in both the Universal Render Pipeline(URP) and High Definition Render Pipeline(HDRP). To get the most out of this article, you should have some familiarity with Shader Graph. If you are new to Shader Graph, please explore our resources for an introductionand more detailabout this tool for authoring shaders visually.
When working with art assets in a terrain-like environment, multiple layers of tileable material are typically preferred as they produce better blending results. However, the GPU performance cost of multiple texture samples and growth of memory use with each layer added to the shader can be prohibitive for some devices and inefficient in general.
With this experiment, I aimed to:
While the experiment achieved its goals, it comes with some caveats. You'll have to set your priorities according to the demands of your own project in determining which trade-offs are acceptable to you.
1. Single texture-packed PBR material
Before layering, the first thing you need to do is figure out the PBR material packing. PBR material typically comes with the parameters for Albedo (BaseColor), Smoothness mask, Ambient Occlusion, Metalness, and Normal defined. Usually, all five maps are represented in three texture maps. To minimize texture usage, I decided to sacrifice Metalnessand Ambient Occlusionfor this experiment.
The remaining maps - Albedo, Smoothness and Normal Definition - would traditionally be represented by at least two texture maps. Reducing it to a single map requires some preprocessing of each individual channel.
[Link] The final result of the PBR Material packed into a single texture.
Red = dHdu (Derivatives Height Relative to the U direction) for Normal Definition#. Green = dHdv (Derivatives Height Relative to the V direction) for Normal Definition#. Blue = Linear Grayscale shade representing Albedo (color reconstructed in shader). Alpha = Linear Smoothness map (standard Smoothness map).
Note: The texture is imported into Unity with sRGB unchecked and compressed with BC7 format. When porting to other platforms, switch to the platform-supported equivalent 4-channel texture format.
Processing the maps
Albedo is normally defined as an RGB texture; however, many terrain-like materials (rock, sand, mud, grass, etc.) consist of a limited color palette. You can exploit this property by storing Albedo as a grayscale gradient and then color remapping it in the shader.
There is no set method for converting the RGB albedo to a grayscale gradient. For this experiment, The grayscale Albedo was created through selective masking of the original Albedo map channels and Ambient occlusion; to match the prominent color in the shader color reconstruction, just eyeball any manual adjustments. Highlight and Dark coloring with ramp contrast adjustment for Albedo reconstruction.
Smoothness is considered very important for PBR material definition. To define smoothness more precisely, it has its own channel. A simple multiplier was added to the smoothness in the shader for some variation in the material. [Link](You can choose to skip smoothness multiplier to reduce shader instruction.)
The Normal map is important for showing the detailed characteristics of a surface. A typical PBR Material uses a tangent space normal map. In this experiment, I chose a pre-converted derivativesmap using surface gradient framework for the reasons below. (See Morten Mikkelsen's surface gradient frameworkfor more information). To pre-convert tangent space normal maps to derivatives, use this Photoshop action.
Using a pre-converted Derivatives map has several advantages:
In conjunction with the surface gradient framework, the advantages further include:
But pre-converted derivatives from tangent space normal map also have some disadvantages:
Note: Clamping at an angle greater than 45 degrees does not apply to shader-based derivatives conversion.
Depending on your use case, the limitation may have a lesser or greater effect. In this experiment, a normal direction less than 45 degrees does not have a noticeable negative impact on the end result. In fact, in this case, it provides a benefit by reducing unwanted reflection from an extreme normal direction.
[Link]A comparison of a pre-converted derivatives map texture and a tangent space normal map texture.
[Link]Stress test differences between pre-converted derivatives and tangent space normal map. Notice that the Spherical bump exhibits some artifact due to the 45-degree clamping in the pre-converted derivatives map.
The graph to remap the pre-converted derivatives map to surface gradient.
[Link]Note: Make sure the Master stack fragment normal space accepts world space normal instead of the default tangent space normal.
The full unpacking process
[Link]The complete Sub Graph to unpack the Compact PBR texture to output colored Albedo, smoothness and surface gradient. Note:Surface gradient conversion to Normal is done outside the Sub Graphso that the material can be easily blended based on the output of the UnpackedSubGraph.
[Link]The materials sample from the unpacked shader.
[Link]Different angles to better showcase each material's responses to light.
2. Layering the material
For this experiment, I chose a tier-based layering method on a single channel remap. The Sub Graphdoes five linear interpolations (plus the base, forming six layers). There are many ways to blend layer weights. This method has the simplicity of a single vector input, which suits the experiment goal. This allowed lots of layering without burning through multiple channels in splat maps or vertex channels.
The drawback of this method is that you cannot control the weight of an individual layer's contribution. The blending will always be a transition from the previous layer. Depending on the use case, this can be a limiting factor compared to a traditional per-channel blend.
The Sub Graphshown above is predefined for six layers of tier-based blending. To create more layers, divide 1 by the desired number of layers blended, subtract 1, and then remap each layer based on that value range. For example, for a nine-layer blend material, each layer remap range is 1/(9-1) = 0.125.
|Layer||In Min||In Max|
Be aware that as you divide the single channel into smaller portions, you have less shading range.
[Link]The layering method tested using color to showcase the blending characteristic.
[Link]The material to composite the 6 layers of the unpacked material with a single vertex color (in this case, the red channel).
Note: It's simpler to blend the layers of a single sample-packed PBR material in the Shader Graph than to blend three PBR textures. [Link]Since layer blending is tier-based, it's a good idea to plan the order of the layers before you set up the blend material.
3. Extending the functionality of material
Layer blending requires only a single channel (the red vertex channel). The remaining three vertex channels offer extra functionalities.
[Link]The final Shader Graph produces results using the remaining vertex channels.
In this experiment, vertex painting was done inside Unity Editor using Polybrush (available from the Package Manager).
[Link] Suggested Vertex Paint color palette for this shader.
Red:Used to weight the layer contribution.
Red vertex channel painting demo
Green:Sets the surface gradient property, to flip, reduce or add normal bump contribution (remapped to -1 and 1).
Green vertex channel painting demo
Blue:Controls smoothness and surface gradient bump scale to create a wet water look
Blue vertex channel painting demo
Alpha:Controls the weight of the Albedo layer, setting the base color to white, with the contribution based on the y axis of the surface normal. It does not alter the smoothness and takes advantage of the original surface layer smoothness and bump property.
Alpha vertex channel painting combined with previous channels to showcase how the whole layers interact with the snow
The combined results of the different vertex painting channels:[Link]
You can adjust the shader blending method and the settings for the various vertex channel/splat map functionalities according to your project's requirements.
4. Comparing performance
The purpose of this experiment was to extend the functionality of the Shader Graph while minimizing resources. The texture was preprocessed and unpacked, but is there a payoff in runtime efficiency? Performance profiling shows the efficiencies these efforts produced.
A standard six-layer blend shader was created for comparison with the compact six-layer blend shader. Both shaders were created using an identical blending method with the same functionalities. The main difference is that the standard shader uses three different textures to represent a single layer. For profiling, a single mesh was rendered on screen with blend material using the Universal Render Pipeline in the targeted platform.
Mobile memory and performance profile
Texture compression for mobile (Android):
Standard PBR with Albedo, Mask and Normal map at 1024×1024 for mobile:
Total Texture memory usage 5.431 MB
Compact PBR at 1024×1024 for mobile:
Total Texture memory usage 2.048 MB
With the compact six-layer material, there is approximately 62% Less texture memory consumption on Mobile (Android), savings of more than half.Mobile Android/Vulcan with Adreno 630 (Snapdragon 845); Snapdragon profile results:
Compact material renders on screen approximately 30% faster.
[Link] Profiling result from Snapdragon Profiler.
PC memory and performance profile
Standard PBR with Albedo, Mask and Normal map at 1024×1024:
Compact PBR at 1024×1024:
Total Texture memory usage 7.8 MB
The compact six-layer material uses 60% less texture memory consumption on PC (savings of more than half).
PC laptop with Radeon 460 Pro rendering at 2880×1800; RenderDoc profile results:
PC desktop with nVidia GTX 1080 rendering at 2560×1440; nSight profile results:
Compact material renders on screen approximately 45% faster.
[Link]Profiling results from nSight.
Console performance profile
On PlayStation 4, using compact material yields 60% memory savings, identical to that for PC as the PS4 uses the same compression.
PS4 base rendering at 1920x 1080; Razer profile results:
[Link]Profiling result from PS4 Razor profiler.
In summary, using a compact six-layer PBR shader offers performance gain and significant memory savings. The variation of GPU performance is interesting but expected, as unpacking the material consumes more ALUs than sampling more textures.
Download the sample project
This sample project with Shader Graphs and Sub Graphs can be downloaded here:
[DOWNLOAD HERE], Unity 2020.2.5f1 with HDRP 10.3.1 [DOWNLOAD HERE], Unity 2020.2.5f1 with URP 10.3.
[DOWNLOAD HERE], Photoshop action to pre-convert tangent space normal map to derivatives.
[Link]Screenshot from Universal Render Pipeline version of the project.
The main components of this experiment are:
This experiment showcases how you can use Shader Graph to produce beautiful graphics that are also efficient. Hopefully, this example can inspire artists and developers to push aesthetic boundaries with their Unity projects.
About the author
Rinaldo Tjan (Technical Art Director, R&D, Spotlight Team) is a real-time 3D artist with an extreme passion for real-time lighting and rendering systems.
[Link]Having started his career in the PlayStation 2 days, he has more than a decade of end-to-end artist workflow knowledge, from texturing to final rendered scene creation. Prior to joining Unity Technologies, he helped deliver AAA games such as BioShock 2, The Bureau: XCOM Declassified, and Mafia III.
He currently works with Unity clients to help them augment their projects and realize their true potential using Unity while helping drive the internal development and standards of Unity rendering features.