Compilation Techniques for Multimedia Processors

1
Compilation Techniques for Multimedia Processors

• Andreas Krall and Sylvain Lelait
• Technische Universitat Wien

2
Motivation

• High processing power needed by multimedia
  applications
• Special instruction sets for multimedia data
• Implemented by special processors
• Multimedia instruction set extensions
• Visual instruction set(VIS) of UltraSPARC
• Altivec extension of PowerPC
• MMX extension of Pentium
• MAX-2 instruction set of HP PA-RISC

3
Problems and Solutions

• New functionality has not been exploited
  properly
• Have to code in assembly language
• Use provided system libraries
• Call macros in high-level language
• Compile a program coded in high-level language
  into multimedia instructions
• Classic vectorization
• Vectorization by loop unrolling

4
Classic Vectorization

• Loop analysis
• Loop normalization
• Scalar expansion
• Dependence analysis
• Vectorization
• Alignment management
• Strip mining
• Constant expansion
• Lower iteration space
• Lower alignment
• Instruction selection and register allocation

5
Classic Vectorization

• Loop analysis
• Loop normalization
• Scalar expansion
• Dependence analysis
• Vectorization
• Alignment management
• Strip mining
• Constant expansion
• Lower iteration space
• Lower alignment
• Instruction selection and register allocation

6
Classic Vectorization

• Loop analysis
• Loop normalization
• Scalar expansion
• Dependence analysis
• Vectorization
• Alignment management
• Strip mining
• Constant expansion
• Lower iteration space
• Lower alignment
• Instruction selection and register allocation

7
Classic Vectorization

• Loop analysis
• Loop normalization
• Scalar expansion
• Dependence analysis
• Vectorization
• Alignment management
• Strip mining
• Constant expansion
• Lower iteration space
• Lower alignment
• Instruction selection and register allocation

8
Vectorization by Loop Unrolling

• Loop analysis
• Compute unrolling degree
• Loop unrolling
• Dependence analysis
• Dependence verification
• Generation of vector instructions
• Alignment management
• Lower iteration space
• Lower alignment
• Instruction selection and register allocation

9
Vectorization by Loop Unrolling

• Loop analysis
• Compute unrolling degree
• Loop unrolling
• Dependence analysis
• Dependence verification
• Generation of vector instructions
• Alignment management
• Lower iteration space
• Lower alignment
• Instruction selection and register allocation

10
Vectorization by Loop Unrolling

• Loop analysis
• Compute unrolling degree
• Loop unrolling
• Dependence analysis
• Dependence verification
• Generation of vector instructions
• Alignment management
• Lower iteration space
• Lower alignment
• Instruction selection and register allocation

11
Experimental Results