gaohebing / potatobenchmark

The Potato Benchmark Project

Potato Benchmark is an benchmark project to make it easy to compare performance between LWRP and Built-in pipelines. It contains some baseline performance test scenes that specifically stress parts of the graphics pipeline such as test fill rate of built-in shaders, batching/driver cost and bandwidth.

The benchmark is still WIP. It builds for Android (GLES2, GLES3, Vulkan) and iOS (Metal) a set of test scenes in LWRP and Built-in. A main scene allows to select which pipeline to test. Each tests scene warms up for 2s. and measure frame times (average, median, min, max) in a period of 8s.

How to use it

Building to mobile

In the Unity toolbar select Tools then select the target API to built. This will either generate an .APK file or an XCode project.

Building to other platforms

There's no custom build script for other platforms other than mobile. You have to switch manually to the platform and build scenes. That's it.

With Test Runner

Go to Window -> General -> Test Runner. Then tests can run on either editor or deploy to target device. Then to see results go to Window -> Analysis -> Performance Test Report

Running Tests

An main menu will display two buttons to select with Pipeline benchmark tests to run. Results will be outputted to console. You can filter console output with [Benchmark]. F.ex on in Android use the following command line: adb logcat -s Unity | grep [Benchmark].

Note: If you are running in mobile, please remember to give 2min rest between each pipeline test.

Detail about current Tests

Fill rate tests

Stress GPU fragment (ALU/Texture) by rendering a native device resolution set of fullscreen quads. Test is designed to achieve 2.5x overdraw which is acceptable rate for a mobile game. The purpose of the fill rate tests is to compare performance of single vs multiple passes strategies for each pipeline built-in shaders in different devices.

• Lit shader (Directional Light Only)
• Lit shader (1 Directional + 4 point lights per-pixel)

Draw Call (GPU)

Stress GPU by issueing a considerable number of drawcalls that can be batched.

Tests have about 1k block building and one directional shadow casting light + PCF Filtering. The purpose of the realtime shadow test scene is to compare how each pipeline handles shadow culling, rendering and depth pre-pass.

Test with and without cascades.

• Realtime Shadows
• Realtime Shadows Cascades

Draw call (CPU)

Test batching/cost of drawcall setup by stressing CPU with 2.7k simple drawcalls each with a different materials. The purpose of this tests scene is to compare the SetPass cost for each pipeline.

Bandwidth

Stress bandwidth cost.

The purpose of this test is to figure out points of improvements in terms of bandwidth cost to have a very lean pipeline in term of LOAD/STORE. Currently we have a lot of room for improvement in terms of bandwidth when using MSAA.

Credits

• Spawn Object script adapted from Arnaud's Batcher.