EA Namco Bandai Flagship Studios Hellgate: London

1
EA / Namco BandaiFlagship StudiosHellgate
London

• Thread for Features

2
Thread for features

• Why not reward users with high-end CPUs?
• Well look at how Hellgate London tackled this
  issue

3
What do you do when FPS enhancement is not enough?

• Threading for performance has been done.
• Framerate is maxed out on 2-core.
• What about 4-core? What about n-core?
• Add features that only appear or become more
  intense on multi-core.
• Dont change gameplay!

4
Amdahl was a buzzkill

• Speedup is limited by serial portion of code.
• Features are limited only by creativity and
  effort.

Faster
Awesomer
5
Short on ideas? Use these!
6
Hellgate London was a natural choice to thread
for features

• Hellgate London was already targeting a broad
  range of hardware.
• Graphics would scale depending upon GPU.
• How should it scale depending upon CPU?
• Threading for performance deemed not enough
  return on investment.
• Hellgate London didnt have a profile which
  seemed to benefit from threading the graphics
  engine.
• Lets thread for features!

7
Hellgate London is architected for the future

• Hellgate London has two scaled thread pools
• First thread pool is for loading assets.
• Speeds up level loading.
• Avoids framerate hitches when new sounds or
  models are loaded.
• Second thread pool is for arbitrary jobs.
• Texture color swizzling to make armor sets appear
  cohesive.
• What else could be done here?

8
Let it snow?
9
Weather particles were a small part of the
overall frame time

• Weather had a small time slice of the frame
  calculation.
• Calculation of the weather was moderately
  time-consuming, rendering not so much.
• Just ramping up particles impacted CPU severely,
  and started to impact GPU somewhat.
• 2x, 4x weather wouldnt be a noticeable
  improvement must go big!
• Comparing weather particle time to overall frame
  time showed a ratio of 150.
• 50x weather particles possible on dual-core, in
  theory.
• Must remain aware of GPU impact.

10
An initial fork-and-join approach didnt work out

• First try was to do a fork-and-join approach.
• Primary thread would do half the work on
  dual-core.
• Dedicated thread would do the other half.
• No data copying! Memory is consistent.

Render
Render
Original
Fork-and-Join

• Half of 50x is still 25x!

11
Second try asynchronous update, synchronous
render

• Asynchronous update wont stall the primary
  thread.
• Memory is no longer consistent.
• Weather particles are redrawn when new positions
  are available (probably every frame).

Render
Render
Original
Asynchronous

• Trickier, but looks like a winner.

12
Heres the code changes that were made

• A callback was defined to update weather
  particles.
• This used the existing job task pool.
• Vertex buffers were created to hold the weather
  particles.
• Particle systems were marked as regular or
  asynchronous.
• Asynchronous systems were split into the good
  twin and the evil twin
• During draw, good twin draws the evil twins
  particles from last frame, then captures draw
  state for the evil twin.
• During update, evil twin is processed in the
  callback and fills the vertex buffer.

13
Impact to memory and GPU prevents theoretical 50x
maximum

• One frame delay means that more memory is
  required.
• Also, there are more weather particles than any
  other kind of particle.
• 50x particles impacts framerate, but 30x...

14
Wheres my hot chocolate?
15
We could keep going...

• This could be taken further by
• Using additional weather particle systems for
  other ambient effects.
• Using the same technique to scale the weapon
  effects.
• Adding forces to the weather particles to
  simulate localized wind.

16
Threading for features has a lot of potential

• Every additional core supplies an additional
  thread in the pool.
• Every additional thread an additional feature?
• Well see an example in the next section, after
  the break.

17
We learned some dos and donts

• Do
• Target features in addition to performance.
• Take advantage of existing threading
  infrastructure.
• Dont
• Give up on threading if your first approach fails.

18
What do you think?

• Have you tried something like this?
• Have you rejected trying something like this?