The FRIDGE Project overview and perspectives Presentation at Berkeley Wireless Research Center, Apri

1
The FRIDGE Projectoverview and
perspectivesPresentation at Berkeley Wireless
Research Center, April 4th, 2000

• Institute for Integrated Signal Processing
  Systems,Prof. Dr. Heinrich Meyr,Aachen
  University of Technology

Holger Keding, Martin Coors, Oliver Wahlen, Olaf
Lüthje
ISS
2
Motivation
Why bother about Fixed-Point Design?

• chip area
• power consumption
• price-per-device
• throughput

Implementation Advantages
Application domains

• high-volume products
• mobile applications

3
Fixed-Point Design Process
OFDM frequency synchronization unit
Design space exploration
5 Weeks
3 Weeks
4
FRIDGE
Fixed-Point Programming andDesign Environment
float
fixed

• Bit-true modeling SystemC (www.SystemC.org)

• interpolative transformation

• Architecture Mapping

5
Algorithmic Transformation
floating-point description
ANSI-C/SystemC
fixed-point description
SystemC
6
Fixed-Point Specification
7
FRIDGE Design Methodology Local Annotation
local annotation
hybrid
SystemC
8
SystemC Local Annotations

• incremental transformation
• efficient modeling
• starting point for an implementation

SystemC
Bit-true ANSI C
float two_tap_fir(float in) state0
in ...
float two_tap_fir(float in) int tmp
(int)((in0.125)4) state0 0.25
(tmpgt31)?max((tmplt-32)?-32tmp) ...
sc_fix(in,6,4,SC_RND,SC_SAT)
9
FRIDGE Design Methodology Hybrid Simulation
local annotation
hybrid
SystemC
10
FRIDGE Design Methodology Wordlength
Interpolation
local annotation
hybrid
SystemC

• Range propagation
• Precision propagation

11
Interpolation example
a sc_fix(in1,4,1) b sc_fix(in2,4,4) .... c
a bd 1.0625 c
global annotations wl_max 8 MSB SAT LSB
RND
c sc_fix(a b,8,5)
d sc_fix(sc_ufix(1.0625,5,1)c,
8,5,SC_RND,SC_NO)
-1,0.875
-9,7.875
-7.875,9
-8,7
Correlation between range and fixed-point
attributes (2s complement) minval ? -2 iwl-1,
maxval ? 2 iwl-1 - 2-fwl ? iwl ? max( ldminval,
ldmaxval 2-fwl) 1
12
The FRIDGE Framework
floating-point algorithm (ANSI-C)
Annotation
hybrid algorithm (SystemC)
simulation system
Algorithmic level
Interpolation
fixed-point algorithm (SystemC)
Implementation
Implementation level
fixed-point algorithm (SystemC)
integer algorithm (ANSI-C)
integer algorithm (C62x C code)
13
Research Topics

• link to implementation
• target specific backends needed
• designers have to write the implementation
  manually
• simulation speed
• SystemC fixed-point simulation classes provide a
  high modeling efficiency
• C based generic fixed-point simulation is far
  too slow

14
Fast Simulation Code Generation
0
30
60
90
15
HW Fixed-Point Operations
Example Addition
gtgt sa
a

gtgt sb
b

machine word length (mwl)
c
Restrictions mwl - iwla gt
lbpa - sa gt fwla (1)
mwl - iwlb gt lbpb - sb gt fwlb
(2)
mwl - iwlc gt lbpc gt fwlc
(3)
lbpc lbpa - sa lbpb - sb
(4)
16
Restrictions for data type mapping
User defined or given set of data types e.g.
short int long
User defined or given data path e.g. adder
multiplier


17
Embedding Shift Optimization

• Goal Minimize costs
• number of shift operations
• number of cast operations

Expression statements t1 a b t2 t1
c e t2 d


e

d
c

b
a
18
Define-Use and Use-Define Chains
Expression statements S1 t1 a b S2
t2 t1 c S3 e t2 d
Data Flow
DFG
S1
S2
S3
19
Example DU Chains for Loops
S1
S2
S3

• Syntax directed data flow analysis
• state of the art in compiler design
• in general syntax oriented
• ambiguous by principle

20
Example DU Chains for Loops
S1
S2

• semantic oriented data flow analysis
• distinguish between statement and execution
• in general semantic oriented
• ambiguous only for control structures

21
Advanced Data and Control Flow Analysis

• Operation level data and control flow modeling
• more precise information
• more aggressive optimization
• analysis of pointers and arrays
• handling of ambiguity supported
• DU/UD chains allow data type selection and
  embedding of bit-true words
• higher computational costs
• different analysis levels required

22
Summary

• FRIDGE provides efficient fixed-point modeling
  capabilities and data analysis support
• Designer can concentrate on the (few) fixed-point
  parameters that really matter
• Automatic transformation from hybrid to bit-true
  code enable a faster exploration of the design
  space.
• Fast Simulation is crucial for system development
• Link to implementation closes the gap in the
  design flow
• Tool supported targeting of different platforms
  shortens the design cycles

floating- point algorithm (ANSI-C)
Annotation
hybrid algorithm (SystemC)
simulation system
Algorithmic level
Interpolation
fixed-point algorithm (SystemC)
Implementation
Implementation level
Integer algorithm (ANSI-C)
integer algorithm (C62x C
code)
fixed-point algorithm (behavioral
SystemC)