SystemLevel Data Format Exploration for Dynamically Allocated Data Structures

1
System-Level Data Format Exploration for
Dynamically Allocated Data Structures

• Peeter Ellervee )
• Ahmed Hemani

Miguel Miranda Francky Catthoor
IMEC vzw Kapeldreef 75 3001 Leuven Belgium
ESD, KTH Electrum 229 S16440 Kista Sweden
) Currently at Tallinn Techn. Univ., Estonia
2
Outline

• Motivation and related work

• Data format exploration

• Results for real-life drivers

• Conclusions

3
Implementing Data-transfer Intensive Dynamic
Applications

• Dynamically allocated data structures
• flexible descriptions of complex applications
• specialized optimization strategies needed

• Past -- applications requiring large storage
  implemented in software on standard CPU
• telecom network applications

• Present -- cost efficient embedded software or
  hardware with custom memory organizations
• real-time, power and area are crucial

4
Related Work

• Foreground memory (register) allocation
• F. J. Kurdahi, et al, 1987
• D. Gajski, et al, 1992
• T. Denk, et al, 1996

• Static array-oriented versus scalar-oriented
• IMEC ATOMIUM   (F. Catthoor, et al, 1998)
• Philips PHIDEO   (J. Van Meerbergen, et al, 1995)

• Recent non-scalar oriented memory allocation
• MeSA -- static memory allocation(U.C.Irvine L.
  Ramachandran, et al, 1994)
• Horizontal and vertical array mapping(CMU H.
  Schmit, et al, 1997)

5
Exploration Environment
6
Data Format Exploration
7
Analysis Modeling Assumptions
8
Collecting Dependency Cases
// structure has three // fields - f1, f2
f3 x1 p1 -gt f1 x2 p1 -gt f2 x3 p1 -gt
f3 // p2 -gt f1 f1(x1,x2,x3) p2 -gt f2
f2(x1,x2,x3) // ...
// structure has three // fields - f1, f2
f3 x1 p1 -gt f1 x2 p1 -gt f2 x3 p1 -gt
f3 // p2 -gt f1 f1(x1,x2,x3) p2 -gt f2
f2(x1,x2,x3) // ...
// structure has three // fields - f1, f2
f3 x1 p1 -gt f1 x2 p1 -gt f2 x3 p1 -gt
f3 // p2 -gt f1 f1(x1,x2,x3) p2 -gt f2
f2(x1,x2,x3) // ...
Source code
Pruned CDFG
Dependency cases
Source code
Pruned CDFG
Dependency cases
9
Compatibility Graph Construction
Dependency cases
Dependency cases
Compatibility graph
Compatibility graph
10
Compatibility Graph Clustering
3 reads 2 writes
3 reads 2 writes, an extra read needed to load
field f3
11
Exploration Results

• OAM - Operation And Maintenance of ATM switch

12
Conclusions

• Data format exploration at system-level enables
  reduction in number of stored/accessed bits
• significant reduction on size, bandwidth and
  power!
• Formal modeling and systematic exploration of
  search space before detailed memory mapping
• Results for four subsets from real-lifeATM cell
  processing
• significant reduction in memory area - up to 20
• significant reduction in memory power consumption
  - up to 60

13
Analysis Assumptions

• Accesses to same field are not considered

• Single execution thread all accesses can be
  statically ordered on a relative time axis

• Only pairs of accesses to the same data structure
  with same pointer (base address) are considered

• Only static dependency analysis

• Only data-flow chains with one access are
  analyzed long dependency chains are covered by
  shorter ones

14
Compatibility GraphConstruction and Clustering
15
Collecting Dependency Cases
16
Compatibility Graph Construction
17
Compatibility Graph Clustering
18
Exploration Results

• Subsets from real-life ATM cell processing
  applications

without formatting