SoftwareOnly Bus Encoding Techniques for an Embedded System

1
Software-Only Bus Encoding Techniques for an
Embedded System

• Wei-Chung Cheng, Jian-Lin Liang, Massoud Pedram
• ASC-DAC 2002

2
Introduction

• Goal
• The application of system-level or software-level
  techniques to reduce power consumption on the
  data bus and IO bus.
• NO! Hardware modification using off-the shelf
  product

3
Target System(SA1110 evaluation board)
4
Data Bus (Memory LCD controller)

• LCD background
• 320 240 pixel, 24 bit color depth, 60Hz
• 3202402460 111 M bit / sec (bps)
• Using 256 color palette a simple
  one-dimensional index table, which in turn
  represents a 24 bit color
• 320240860 37 M bit / sec
• Decoding performed by the LCD controller

5
Problem formulation

• Power model on D_bus switching activity
  (Hamming distance) ? minimizing the switching
  activity by changing the palette table

6
Palette Encoding Algorithm (SA)

• State transition graph G

• By a simulated annealing algorithm, we can find
  the solution.
• overhead SA algorithm is run only once per
  image to create the color-mapped image ? ignore

7
Palette Encoding Algorithm (PS)

• In black and white image, the color brightness is
  likely to change continuously
• ? Sort the palette by their brightness and use
  gray code. (palette sorting algorithm)

8
Experimental results (Palladia)

• GUI and image

sparse
continuous
GUI
Image (PS-based Heuristic)
9
IO bus (Compact Flash PCMCIA)

• PCMCIA/CF interface
• Voltage 3.3 V
• Capacitance high (PCMCIA adapter card is
  physically long)
• Frequency high (the speed of solid-state
  memory)
• ? The power dissipation on the bus can be quite
  high

10
PCMCIA bus encoding

• Bus Invert algorithm (stan 95)
• Processing step
• File is divided into blocks
• Each block is analyzed to derive the best pattern
• The 16bit pattern xor 16 bit words in block
• The results are written back to the file
• Difference
• No needing extra hardware
• The calculation of the inversion pattern at the
  compile time
• Redundant information is carried in the file
  header
• The inversion pattern is changed for each block
  of data to adapt to the specific characteristics
  of the block

11
IPS problem formulation

• Goal reduce the power consumption of a mobile
  system in which the applications are predefined.
• The optimization problem can be solved off-line
  without about power dissipation overhead.
• (ex branch-and-bound algorithm)

12
Experimental results
Better results
Marginal improvement is not impressive
?
Variable block size ? Select 4K
10-26 reduction , by very simple
modification of OS.
Block size 4K
13
Conclusions

• Data bus ( LCD memory )
• Reassign the palette and rewrote entries in the
  pixel buffer according to their correlation to
  the image
• 29 reduction in text mode (GUI)
• 17 reduction in graphic mode (image)
• IO bus ( CF - PCMCIA )
• 26 reduction by a very simple modification of
  the OS kernel (fixed version)