Scalability

1
Scalability

• Advanced D3D Programming
• Richard Huddy
• RichardH_at_nvidia.com

2
Basic Objectives

• To produce the best experience on every users
  machine
• To exploit all of the resources available
• To cope with a broad spread of hardware
• To avoid bottoming out during the shelf-life of
  the game / engine

3
What is a high-end PC?

• A 125 mega-texel device
• A 125 mega-pixel device
• A fast CPU ( gt 350MHz)
• AGP 2X/4X Bus
• Lots of system RAM ( gt 64MB)
• Huge frame buffers (16 to 32 MB)
• Multi-Texture at low cost

4
Power Trends
CPU Speed
Fill Rate
?
Appreciate the absolute values and the ratios.
5
So whats the problem?
BeginScene()
time

• Second generation hardware

A
B
C
EndScene()
CPU
a
b
c
Graphics
time
Third generation hardware Wow, 10 faster!
A
B
C
EndScene()
CPU
b
c
a
Graphics
6
What can you do to help?

• Scalability is the key
• Run at higher screen resolutions
• Run at higher color depths
• Use more complex rendering techniques on good
  hardware
• Ship multiple geometry models
• Protect your CPU
• Unlock the frame rate

7
Higher Screen Resolutions

• 1) Include direct support for higher resolution
  modes (uses lots of disk space).
• 2) Store high resolution art and filter down to
  produce lower resolution art.
• 3) Store low resolution art and pixel double
• If you have art at 512x384 use it for 1024x768
• If you have art at 640x480 use it on 1280x1024
• (but only use a 1280x960 viewport)

8
Higher Color Depths

• Runs at much the same speed but gives the user a
  much richer experience
• Uses frame buffer memory constructively
• You can re-use the previous 16 bit assets
• The main performance loss in true color is often
  due to texture management

But beware the Frame Buffer Z Buffer depth
constraint on Riva TNT
9
Complex Rendering Techniques - I

• Environment Mapping
• Beware of spending too much CPU on this.
• Dual Texture Lighting
• Bump Mapping
• Use more alpha transparency
• But see also Alpha sort issues later on
• Please try to use the extra fill rate!

10
Complex Rendering Techniques - II

• Trilinear mipmapping for almost everything
• Use Detail textures
• Large textures for extra realism
• 32 bit textures - where its a quality win
• Compressed textures as long as quality is not
  compromised

11
Protect your CPU

• The big ones
• __ftol and other type conversion nightmares
• sqrt()
• thatll be seventy cycles please...
• Reciprocal square root
• One hundred and nine cycles through the FPU
• Transform and lighting (more on that later)

12
Removing __ftol

• Remember that the compiler doesnt have a choice
  but you can check the output
• Write you own inline assembler conversion routine
  if
• You can accept differing rounding rules
• This doesnt break the optimiser!

13
Replacement for sqrt()

• Sqrt seems natural if you are normalising
  vectors, calculating environment map coordinates
  or calculating distances - but its sloooow
• Sample code is available from the developer web
  site or from me directly and will be in future
  versions of the SDK.

14
Saturation Arithmetic (C)

• Limiting a floating point number to lie in the
  range 0.0 to 1.0 inclusive (traditional method)
• if (f lt 0.0)
• f 0.0
• else if (f gt 1.0)
• f 1.0

15
Saturation Arithmetic (Pentium)

• if ((long )f lt 0)
• (long )f 0
• else if ((long )f gt 0x3f800000)
• (long )f 0x3f800000
• This is faster on a Pentium class processor since
  the FPU is non-optimal (i.e. slow) and the
  integer unit is much faster.

16
Saturation Arithmetic (Pentium II)

• Use the cmov instructions
• cmp f,0
• cmovb f,0
• cmp f,3f800000
• cmova f,3f800000

Faster since unpredictable branches are the
bottleneck here. Unavailable on a Pentium.
17
Unlock the Frame Rate

• Its essential that your physics model can run at
  high refresh rates.
• At least 100fps
• 30 or 60 fps limits are not acceptable and lead
  to flat performance on high end hardware

18
The Value of Batching

• Case Specifics
• The average of Polys Per Call (PPC) to
  DrawPrimitive was 2.6, producing 40fps
• Removing state changes to raise the average PPC
  to 50 produced 58fps
• Most of the removed state changes were
  reasonable, i.e. not logically redundant
• The changes did not reduce visual quality at all
• PPC of 200 is optimal

19
Alpha Sort Issues

• The standard solution is
• 1) Draw all non-alpha polys (sort by texture)
• 2) Draw all alpha polys in back to front order
  with Z compare enabled and Z update disabled.
  This copes with overlapping alpha polys but you
  cant sort by texture. (Intersection requires
  decimation).

20
Alpha Sort with Bounding Boxes

• When you are ready to draw your alpha polys then
  draw non-overlapping sets using the
  sort-by-texture technique as before

A
Here, you can safely draw all of A before any of
B or C BC need sorting
B
Viewport
C
21
Geometry - Part 1

• Use the DX6 Transform and Clip engine - itll be
  nearly as fast as your best efforts
• It takes advantage of CPU specific optimisations
  done by Intel, AMD etc.
• It uses the guard band clipping region to enhance
  performance
• Use the DX7 interface ASAP

22
Geometry - Part 2

• This gets you ready for hardware which can do the
  job much faster than the CPU
• Tell the chip designers if you need anything
  non-standard
• If you think DX is too slow then use a run-time
  benchmark to select between DX and your own code

23
Geometry - Part 3

• Use the DX pipeline for geometry which may be
  rendered
• Use your own transform for bounding boxes,
  collisions, portals etc
• Treat hardware TL as
• Write only
• Not necessarily pixel identical to CPU TL

DIPVB()
24
Geometry - Part 4

• Consider choosing between models at game start-up
  time
• More complex Geometry should be several times
  more complex
• Introduce some LOD management
• Your artists are probably generating more complex
  models and then throwing them away

25
Lighting - Part 1

• If the DX Lighting model is good enough then
  there are people who want to help you
• Multi-texture shadow maps and light maps can be
  very fast now
• remember that (multi-pass ! multi-texture)
• Tell the chip companies what you need

26
Lighting - Part 2

• Support more lights
• User a richer set of light types
• Scale with available power
• If you have more complex geometry you get better
  lighting quality

27
Summary

• Use the D3D pipeline as much as possible
• Use the CPU carefully- Abuse the fill rate
• Get on board with DX7
• Offer the richest experience possible
• You may have to treat the PC as two distinct
  platforms, High-end and Low-end

28
Questions
?
?
?

• ?

?
?
?
Richard Huddy RichardH_at_nvidia.com www.nvidia.com