2D3D Graphics Acceleration on Low Power Devices

1
2D/3D Graphics Acceleration on Low Power Devices

• Iosif Antochi
• Computer Engineering Laboratory
• Electrical Engineering Department
• TU Delft, NL

2
Summary

• Introduction (Why do we need graphics
  accelerators?)
• The Software Interface
• Software implementation of a 3D graphics library
  (Mesa)
• What Needs to be Accelerated?
• Logical Organization of a Graphics Accelerator
• Advanced Features of a Graphics Accelerator
• Our Work
• Conclusions

3
Why Do We Need Graphics Accelerators?

• The human-machine interaction is based more and
  more on graphics interfaces (towards virtual 3D
  worlds representations), which are more natural
  than alphanumeric displays.
• Current (consumer level) general processors are
  not capable of rendering in real-time the
  current graphics applications
• To free the central processing unit(s) from
  intensive and specialized computations

4
Typical Applications for Computer Graphics

• Simulators
• Flight simulators
• Military training
• Hazardous conditions simulations
• Games
• Interactive Applications
• User friendly interfaces
• Video conferencing

5
Generic Graphics Application

• Behaviour
• InitSystem
• While (running)
• Generate_a_scene
• Display_the_generated_scene
• Update_scene (automatic or on user input)

6
Graphics Application Data Structures

• All properties of a scene have
• a discrete representation!

World
Object 1
Object 2
Object n

Position
Shape

7
Display Types

• Vectorial Displays
• Raster Displays

RGB component
8
Sample Real-Time Rendered Image
9
The Way You See a Scene
10
The Way It Was Painted
11
The Software Interface

• Software requirements for applications in order
  to use an accelerator

Portable Library Calls (API)
High level
Device Driver
Low level
Graphic Accelerator
12
A 3D Graphics Pipeline
Model View Transformation
Lighting
Normalization
Perspective Projection
Rasterization
13
Mesa Structure

Mesa Core
Device Driver Entry Points
Triangle Acceleration
Line Acceleration
Point Acceleration
Software Rasterizer (OS Mesa)
Accelerated Functions
A Graphics Accelerator
14
What Needs to Be Accelerated?

• 2D graphics
• - BitBlt (Blitter operations)
• - Lines and Points
• 3D graphics
• - Triangle setup processing
• - Texture mapping

15
2D Operations

• Fixed BitBlt used for
• - copying
  rectangular areas from system memory or
• video memory to
  video memory or from video memory to
  system memory.
• - double buffered animation.
• Variable BitBlt used for
• - stretching an image
• - expanding an image

16
Performance Bottlenecks in 3D Graphics Rendering
Systems

• Geometry processing and lighting models.
• Triangle setup calculating the triangle edge
  slopes and increments necessary for
  scan-conversion.
• The rate at which the fixed points iterators
  generate pixel values.
• The bandwidth to the frame buffer and texture
  memory.

17
Our Graphics Accelerator
AMBA BUS
AMBA BUS INTERFACE
Instruction Decoder
Triangle Setup Pipeline
Span Generator ( Iterator )
2D Engine
Texture Unit
Texture Cache
BitBlt
Lines
Points
3D Engine
Pixel Engine
Pixel and Z buffer Cache
Local Memory
18
Triangle Setup Pipeline

P1
Scanline
y
P4
P5
P3
P2
x
19
Triangle Setup Pipeline (cont)

• Vertices
• - (x0 , y0 , z0), (x1 , y1 , z1), (x2 , y2
  , z2) .
• Colors
• - (r0, g0, b0, a1), (r1, g1, b1 , a1), (r2, g2
  , b2 , a2) . ( specular color components ).
• Texture Coordinates
• - (u0, v0), (u1, v1), (u2, v2).
• Interpolation increments
• - dr /dx, dr /dy, dg /dx, dg /dy, db /dx,
  db /dy, da /dx, da /dy,
• - du /dx, du /dy, dv /dx, dv /dy.
  (perspective correction increments)

20
Span Generator

• Linear interpolation
• For x, y coordinates and eventually color
  components (r, g, b, a)
• Hyperbolic Interpolation (perspective corrected)
• For texture coordinates sets (u 0, v 0 ), (u 1 ,
  v 1) and other values that need a perspective
  correction.
• Usually this type of interpolation uses the w
  coordinate from the clipping unit.
• For each parameter k we want to
  perform hyperbolic interpolation
• We linearly interpolate kk/w and w1/w along
  polygon edges and across scanlines,
• At each pixel we have to divide k by w to
  obtain the proper k value

21
Texture Unit

• Compute physical texture coordinates and extract
  required texels according to a specified sampling
  method.
• Texture sampling methods
• - Point sampling. (one texel needed)
• - Linear interpolation. (4 texels )
• - Bilinear interpolation. (4 texels)
• - Trilinear interpolation. (8 texels)
• Combine the filtered color obtained from the
  texture map with the primary color using a
  specified method such as
• Replace
• Modulate
• Decal
• Blend

Using Mip-mapping
22
Bilinear Texture Filtering
23
Trilinear filtering using Mip-maps
24
Tiny Texture Cache

• Conventional texture cache size for a PC graphics
  accelerator usually is at least 4KB.
• We have evaluated different cache sizes and block
  sizes.
• Even for a 256B cache we obtained reasonable
  performance at a smaller power consumption.

25
Tiny Texture Cache Statistics
Original
Original
Realistic
Realistic
Point sampling
Trilinear interpolation
26
Pixel Engine

• Fog Unit
• Color Sum
• Scissor Test
• Alpha Test
• Stencil Test
• Depth Buffer Test (Z Test)
• Blending Operations
• Logical Operations

27
Advanced Features

• Tile Rendering
• Texture Processor
• Antialiasing

28
Tile Rendering

• In software.
• At driver level.
• In hardware
• Pros
• Smaller onchip memory
• Extensibility
• Potential parallelism
• Cons
• Data overhead

Graphic Accelerator
Tile Memory
Graphics Memory
29
Texture Processor

• Traditional texture combining methods
• Replace
• Modulate
• Decal
• Blend
• Texture Processor
• Pros
• An increased number of combinations (user
  programmable)
• Cons
• More computational power
• More memory bandwidth required ( bigger caches)

Fixed number of combinations
30
Antialiasing

• At primitive level Points, Lines, Triangles
• Texture level (eg. Using Mip-maps)
• Full scene anti-aliasing using more frames
• Pros
• Might increase rendering quality
• Cons
• More processing power and memory are required

31
Conclusions

• Hardware graphics accelerators are needed for
  real-time graphics on consumer level devices .
• Performance is not the only metric for evaluating
  a graphics accelerator.
• Low power graphics accelerators need different
  design strategies.
• Work in progress.