A Crash Course on Programmable Graphics Hardware

1
A Crash Course on Programmable Graphics Hardware

• Li-Yi Wei
• 2005
• at Tsinghua University, Beijing

2
Why do we need graphics hardware?
3
The evolution of graphics hardware
SGI Origin 3400
NVIDIA Geforce 7800
4

• 7 years of graphics
• accelenation.com/?doc123page1

5
Ray tracing

• General flexible
• Intuitive
• Global illumination
• Hard to accelerate

6
Polygonal graphics pipeline

• Local computation
• Easy to accelerate
• Not general
• Unintuitive

7
Graphics hierarchy

• Layered approach
• Like network layers
• Encapsulation
• Easy programming
• Driver optimization
• Driver workaround
• Driver simulation
• Protection
• Hardware error check

8
Overview

• Graphics pipeline
• Only high level overview (so you can program),
  not necessarily real hardware
• GPU programming

9
Graphics pipeline
10
Application

• Mostly on CPU
• High level work
• User interface
• Control
• Simulation
• Physics
• Artificial intelligence

11
Host

• Gatekeeper of GPU
• Command processing
• Error checking
• State management
• Context switch

12
Geometry

• Vertex processor
• Primitive assembly
• Clip cull
• Viewport transform

13
Vertex Processor

• Process one vertex at one time
• No information on other vertices
• Programmable
• Transformation
• Lighting

14
Transformation

• Global to eye coordinate system

15
Lighting

• Diffuse
• Specular

16
Transform Light on Vertex Processor

• A sequence of assembly instructions
• (more on this later)

17
Primitive Assembly

• Assemble individual vertices into triangle (or
  line or point)
• Performance implication
• A triangle is ready only when all 3 vertices are
• Vertex coherence caching

18
Clipping Culling

• Backface culling
• Remove triangles facing away from view
• Eliminate ½ of the triangles in theory
• Clipping against view frustum
• Triangles may become quadrilaterals

19
Viewport transform

• From floating point range -1, 1 x -1, 1 to
  integer range 0, height-1 x 0, width-1

20
Rasterization

• Convert primitives (triangles, lines) into pixels
• Barycentric coordinate
• Attribute interpolation

21
Triangles into pixels
22
Attribute interpolation

• Interpolation
• Barycentric

23
Perspective correct interpolation
correct
incorrect
24
Fragment processor

• Fragment corresponds to a single pixel and
  includes color, depth, and sometimes
  texture-coordinate values.
• Compute color and depth for each pixel
• Most interesting part of GPU

25
Texture

• Optional
• (though hard to avoid)
• Cache data
• Hide latency from FB
• Sampling/filtering
• I told you this last time

26
ROP (Raster Operation)

• Write to framebuffer
• Comparison
• Z, stencil, alpha, window

27
Framebuffer

• Storing buffers and textures
• Connect to display
• Characteristics
• Size
• Bandwidth
• Latency

28
Conceptual programming model

• Inputs (read-only)
• Attributes
• Constants
• Textures
• Registers (read-write)
• Used by shader
• Outputs (write-only)
• Attributes

29
Simple example

• HPOS position
• COL0 diffuse color
• MOV oHPOS, vHPOS
• MOV oCOL0, vCOL0

30
More complex example

• oCOL0 vCOL0 constantvHPOS
• MOV oHPOS, vHPOS
• MOV R0, vCOL0
• MAD R0, vHPOS, c0, R0
• MOV oCOL0, R0

31
Sample instruction set
32
A real example
33
High-level shading language

• Writing assembly is
• Painful
• Not portable
• Not optimize-able
• High level shading language solves these
• Cg, HLSL

34
Cg example
35
Applications

• Too many of them for me to describe here
• The only way to learn is try to program
• Useless for you even if I try to describe
• Look at developer website
• NVIDIA, ATI, GPGPU

36
Homework

• Try to program GPU!
• Even without NVIDIA GPU, you can download the
  emulator
• Stanford course on graphics hardware
• http//www.graphics.stanford.edu/courses/cs448a-01
  -fall/
• History of graphics hardware
• 7 years of graphics
• accelenation.com/?doc123page1

37
(No Transcript)
38
S2005 Arial, Bold, 37 points

• This subtitle is 27 points
• Bullets are orange
• They have 110 line spacing, 6 points
  before/after
• Longer bullets in the form of a paragraph are
  harder to read if there is insufficient line
  spacing. This is the maximum recommended number
  of lines per slide (seven).
• Sub bullets look like this