Design of ASIP Architectures for Image
Processing in CMOS VLSI Technologies
*S. Saponara, *L. Fanucci, *P. Terreni, S. Marsi, G. Ramponi
*Dip. Ingegneria della Informazione - Università di Pisa
Dip. Elettrotecnica, Elettronica, Informatica – Università di Trieste
NEWCOM: NoE Wireless Communications
Riunione Annuale GE 2006
Ischia, 21-23 giugno 2006
The Retinex Non Linear Filtering Class
Y
R
I
Problem: correcting images acquired in bad lighting conditions and
emphasizing the details
Applications: Consumer electronics, Medical imaging, Surveillance systems,
Assisted drive systems
State of art: Real time processing based on DSP (Nasa Langley Lab)
Riunione Annuale GE 2006
Ischia, 21-23 giugno 2006
Design of Retinex-like Filters
Algorithmic design: Non linear filters (inspired by the Human Visual System
behaviour) with optimal performance – complexity trade-off vs. state-of-art
Challenge: Reducing implementation cost (area, power), real-time processing,
keeping the flexibility of the SW layer within the specific filtering class
(r ) 
I_in
1
1  e blog r
β
R
/

1
2
Rmod
X
Contrast
enhnance
RRF

Luminance
Estimation
Y
Luminance
Correction
Ymod
I_out
Linear
Stretch
y 

 1

255
 

 y
( y)  255  

255


Riunione Annuale GE 2006
Ischia, 21-23 giugno 2006
Solution: the ASIP Design Approach
ASIPs
SW
Design
ASICs
HW Design
Log
Physically
Optimized
ICs
105 . . . 106
Digital
Signal
Processors
Log P ower consumption
Log Flexibility
General
Purpose
Processors
P erformance
103 . . . 104
Riunione Annuale GE 2006
Ischia, 21-23 giugno 2006
The ASIP Design Flow
Automatic generation:
LISA processor design
- HDL code
- C Compiler
- Simulator
- Debugger
RTL automatically generated
Gate-level synthesis
Riunione Annuale GE 2006
Ischia, 21-23 giugno 2006
Designed VLSI architecture
7-stage pipeline: Fetch, Instruction Decode plus 5 execution stages
42 instructions (32-bit)
14-bit fixed point arithmetic
LUT (ROM)-based realization of non linear operators:
By-pass mechanism to avoid pipeline stall
Riunione Annuale GE 2006
Ischia, 21-23 giugno 2006
ASIP vs. DSP and HW Dedicated Designs
Target technology: 180 nm CMOS standard-cells 1.8 V
Two man-month for both ASIP (LISA) and dedicated HDL designs
ASIP vs. Dedicated (semi-custom flow)
Same algorithmic quality of the dedicated HW design
Increased complexity (area, power) x 1.6
Higher flexibility due to the SW layer
ASIP vs. DSP (Nasa Langley lab)
Reduced power cost by one order of magnitude
Same flexibility for the Retinex filtering class
Riunione Annuale GE 2006
Ischia, 21-23 giugno 2006
Scarica

Application-specific Instruction-set Processor