VFX Optimization: Core Concepts

Identifying GPU, Render Thread, and Game Thread issues

There are several features of UE3 in place which allow effects artists to trouble shoot performance issues. First off identifying where your effects are creating bottlenecks will go a long way towards optimizing the correct issue, thereby retaining visual integrity/quality.

One way to monitor for issues while playing through levels is to pull up the STAT UNIT debug overlay.

1. launch the game on your console
2. hit tab and enter 'stat unit'
   

The stat unit command displays the time spent for the current frame on the Game Thread, the time spent for the Render thread, and the time spent for the GPU. It is possible to monitor the different times updated by stat unit to see areas of the level with performance issues.

Keep an eye on the different thread times, if your target framerate is 30fps, all threads should remain below 33.3ms.

If any thread goes over the 33ms and your particles are a suspected culprit, you can use the STAT PARTICLES command to gauge particle times on the different threads.

STAT PARTICLES lists out multiple statistics relative to Game Thread and Render Thread. Pay attention to Draw Calls (Render Thread) and Tick Time (Game Thread).

Targeting Game Thread related issues

Once you have determined you have a Game Thread issue, use the STAT PARTICLES command to verify your findings.

This stat gives quite a bit of information so read over the descriptions on the linked page to help you determine where your costs are highest. Keep an eye on Particle Tick Time and your Particle Counts.

Note: It is important to remember that a particle system is the sum of the components (emitters/modules) which make up the effect. The more unique emitters and modules in your system, and the more of those systems in your scene, the higher the Game Thread cost.

If your tick time is high use the ParticleTickStats command to identify which systems are your main offenders.

The ParticleTickStats command has 3 arguments:

• Single - This argument will capture a single frame and write out a file
• Start - This argument will begin capturing stats
• Stop - This argument will end stats so you can see tick time for a duration

The ParticleTickStats report is then written out to this directory:

[UE3 Directory]\[YourGame]\Profiling\PSTick-[sys time].csv

The spreadsheet is broken into columns and rows outlining the cost of your effects

Total Tick ms

Total time spent ticking all instances that Particle System

Avg Tick ms

Avg. time for ticking that Particle System (Total Tick/Ticks)

Max Tick ms

The HIGHEST tick time for an instance of that Particle System

Ticks

Number of INSTANCES of that Particle System

Avg Active/Tick

Avg # of particles active during the Tick of that Particle System (Total Particles/Ticks)

Max Active/Tick

Max # of particles active during the Tick of that particle (in a single instance)

Depending on your platform you will need to browse to the directory through the appropriate means. This file is comma delineated and can be opened in Excel etc. etc.

Once you have identified Game Thread systems which are offenders there are several paths to take.

1. Reduce particle emission counts on the systems being used.
2. Reduce the number of particle systems in the scene.
3. Reduce lifetime (the amount of time particle evaluation is calculated) for some of your emitters.
4. Check expensive calculations such as collision etc. to ensure the settings you are using are optimal, reduce collisions or the amount of particles/meshes colliding.
5. Remove expensive calculations such as collision, dynamic parameters, etc. if necessary.
6. Replace Game Thread calculated particle effects with static mesh effects if possible.
7. Set Fixed Bounds on your particle system so you do not calculate them each frame.
8. Increase LodDistanceCheckTime so you are checking LODs less often (for looping effects with LODMethod set to automatic.)

Using Unreal Front End to launch the build on console

Unreal FrontEnd can be found in your projects binaries folder

[UE3 Directory]\Binaries\UnrealFrontend.exe

Unreal Front End is a GUI which allows any discipline to quickly cook and deploy a build or a level to see the results of local data on a target platform. Unreal Front End also can be used as a launching point for your project, with links to the editor and many other tools. See UnrealFrontend for more information on using this tool.

In relation to effects performance we are going to be focusing on using two parameters when launching the game to get a relatively accurate gauge of how well the game is running. In the Launch Options box the commands -novsync, -noverifygc, and -noailogging.

-novsync disables vertical sync (locking fps @ 30) so we can see how much overhead there is in the scene, -noverifygc disables periodic garbage collection which produces a visible hitch every so often, and -noailogging disables ai logging which slows the game considerably, aiLogging and garbabge collection run in the background and slow performance when enabled. Disabling these will give a closer reproduction of actual frame rate in a final release build, minus disc load time.

When doing any of this testing in a debug build, it is important to remember there is a cost to running the game in debug which will not be visible in a final release build. Using tools like the stat overlays have a cost and will cause your results to be skewed slightly.

Troubleshooting batches of FX

Several Commandlets exist for the purpose of checking batches of effects for common problems, and outputting spreadsheets which list the offenders.

For an overview of commandlets and explanations of several of the built-in commandlets, see Commandlet Home.

These Commandlets include:

ParticleSystemAudit

This commandlet will perform an audit on all particle systems that are contained in the database resulting from MineCookedPackages.

NOTE: This will be changed to use the 'On DVD' tag, as well as simply checking all existing particle systems.

It will generate csv files containing the following to assist in optimizing/fixing content:

All particle systems w/ NO LOD levels

No LOD levels means the particle system is 'empty'.

All particle systems w/ a single LOD level

These systems should be reviewed to ensure they are not environmental (level placed) effects. In those cases, you typically want to utilize LODs to turn off emitters in the distance.

All particle systems w/out fixed bounds set

All particle systems that can should leverage the fixed relative bounds.

All particle systems w/ LOD Method of Automatic & a check time of 0.0

This indicates a scenario where the particle system will perform its LOD check every frame.

All particle systems w/ bOrientZAxisTowardCamera enabled

These systems will cause problems if your game will support split-screen.

All particle systems w/ missing materials

These systems will render with the default material.

All particle systems w/ no emitters

These systems are 'empty' - should they be deleted?

All particle systems w/ collision on in at least one emitter

Current collision methods are a VERY high cost operation for particles.

FindOnDvDPsysDynamicParameter

This Commandlet finds any effects using the dynamic parameter module, referencing a material without a functioning dynamic parameter material expression. The process is not entirely fool proof, it is good practice to verify the result through the material editor prior to removal of a dynamic parameter module.

Back to VFX Optimization