Code Optimization

1
Code Optimization

• Witawas Srisa-an
• CSCE 496 Embedded Systems Design and
  Implementation

2
Agenda

• Talk about possible exam ideas
• Code optimization techniques
• Not everyone has reconfigurable processors!
• Credits
• Most of slides in this lecture are based on
  slides created by Profs. Raj Rajkumar and
  Professor Priya Narasimhan from ECE Dept at
  Carnegie Mellon

3
Exam Ideas
4
Code Optimization

• Programmers can improve program performance by
  writing better code
• Improve data structure and/or algorithms
• Merge vs. bubble sorts
• Reorganize code or provide flags to help
  compilers
• Last option is to write in assembly

5
Better Algorithms

• Merge vs. bubble sorts
• Which one runs faster?
• Which one causes more cache misses?

6
Common Optimization Techniques

• Sub-expression elimination
• Dead code elimination
• Induction variables
• Strength reduction
• Loop unrolling
• In-lining

7
Common Techniques (cont.)

• Sub-expression elimination

myfunction index1 8 i x a index1
temp 8 i index2 4 j t aindex2
atemp t temp2 4 j atemp2 x
goto myfunction
8
Common Techniques (cont.)

• Dead code elimination

int i 0 i i 1 if (i 0) j j
8 else j j 10
use ASSERT and ifdef to advice the compiler
about deadcode
9
Common Techniques (cont.)

• Induction variables and strength reduction

i 0 j 0 label j j 1 i 4 j ai
2 b i if (i lt 1000) goto label
10
Optimization Techniques (cont.)

• In-lining

main addi s0, t1, 0 addi s1, t2, 0 jal
mult add t3, v0, 0 mult addi sp, sp -12 sw
s1, 4(sp) sw s0, 8(sp) sw ra, 12(sp) sll
v0, s0, s1 lw s1, 4(sp) lw s0, 8(sp) lw
ra, 12(sp) addi sp, sp, 12 jr ra
Whats wrong with this picture?
11
Optimization Techniques (cont.)

• Loop unrolling
• Eliminate branches (why?)

12
Architecture Dependent Optimizations
X Y 64 Convert 8-bit RGB to 8-bit YCC Y
0.299R 0.587G 0.114B Cb -0.169R - 0.331G
0.500B 128 Cr 0.500R - 0.419G 0.082B 128
13
Architecture Dependent Optimizations (cont.)
Address Register
Addr Incrementer
Incrementer Bus
ALU Bus
Register Bank
Write Buffer (holds address and data)
A Bus
Barrel Shifter
B Bus
32-bit ALU
Mem Addr Register
Write Data Register
Read Data/Instr Reg
Dout310
Data310
RAM
14
Summary

• No magic bullet
• optimizations sometimes dont work
• programmers need to help
• various techniques that may require prior
  knowledge of the hardware