A RealTime Procedural Shading System for Programmable Graphics Hardware

1
A Real-Time Procedural Shading System for
Programmable Graphics Hardware

• Kekoa ProudfootWilliam R. Mark
• Svetoslav Tzvetkov
• Pat Hanrahan

2
Introduction

• Graphics processors are becoming programmable
• Programmability makes it possible to make
  real-time graphics more lifelike
• But somebody has to do the programming

3
Programmable graphics hardware

• Problem programming graphics hardware is
  difficult
• Low-level interfaces
• Different manufacturers produce hardware with
  different interfaces and functionality
• Sounds familiar microprocessors
• Assembly language interface
• Intel processor very different from Alpha

4
Programmable graphics hardware

• Programming in assembly is difficult
• Have to worry about register allocation,
  instruction scheduling, etc.
• With normal processors, we program in
  higher-level languages, and compile down to
  assembly
• Can we apply this idea to programmable graphics
  hardware?

5
Shading languages

• High level description programmer describes how
  objects should be drawn, instead of how to draw
  them
• Kind of like C you still need to know what
  youre doing, but you dont need to worry about
  low-level details
• Makes programming easier, and makes programs
  portable

6
Compiling the shading language
shading language

• Traditional compiler structure
• Programmable pipeline program is an intermediate
  language

compiler front end
programmable pipeline program
retargetable compiler back end
hardware-specific code
7
Front end analysis - computation frequencies

• Primitives (cubes, etc) get converted to vertices
  (triangles)
• Vertices get converted to fragments (pixels)
• Computation can occur at any of these stages
• Use type inference to perform computation as
  early as possible

constant
primitive group
vertex
fragment
8
Compiler back end

• Modular structure
• To support new hardware, make a new module
• One module per computation frequency
• Modules can be interchanged
• Arbitrarily complex computations can be done
• Render in multiple passes if necessary
• Use the CPU if all else fails

9
Back end optimizations

• Group parallel scalar operations into vector
  operations
• Copy propagation
• Collapse MUL and ADD to MAD (multiply-and-add)
• Use negate-on-read when possible

10
Conclusion

• Shaders provide an abstraction barrier between
  the programmer and the hardware
• But, the programmer no longer has direct control
  over the performance of the program

11
Discussion

• The goal is to produce lifelike real-time
  graphics. Can we afford an extra layer of
  abstraction?