Decomposing Specifications with Concurrent Outputs to Resolve State Coding Conflicts in Asynchronous

1
Decomposing Specifications with Concurrent
Outputs to Resolve State Coding Conflicts in
Asynchronous Logic Synthesis

• Hemangee K. Kapoor Mark B. Josephs
• Centre for Concurrent Systems and VLSI,
• London South Bank University, UK

2
Synthesis Path
Separate out Forks
Signal Transition Graph (STG)
petrify
3
The problem

• Concurrent outputs in a specification can
  bedifficult for Petrify to handle
• Complete state coding (CSC) conflicts may arise
  when the environment is allowed to react
  beforereceiving every output in a burst
• Petrifys own heuristics for introducing internal
  signals (state variables)
• do not always solve CSC, and
• do not necessarily solve CSC in a way that yields
  the best circuits

4
Solution

• Decompose process P into process P Forks
• P exhibits less concurrency in its
  output-burststhan P, so there may be fewer CSC
  conflicts
• Fork-based decomposition can be readily applied
  when P and P are expressed in our DISP language

5
Decomposition Heuristics

• Consider list of signals yy1yn that always
  occur together in output bursts
• (H1)
• n gt 1
• Replace y by fresh signal x
• Create a fork component
• FORK forever do x/y end
• (H2)
• n gt 0
• Replace each output burst y,t by x followed by
  z/t
• (x and z are fresh signals)
• Create a fork component
• FORK forever do x/y,z end

6
Example of H1

• P forever do a,b/c,d end

Apply H1 to fork out c and d P1 forever do
a,b/x end Fork forever do x/c,d end
7
Example of H2

• Q forever do
• select a0,b/d0,c
• alt a1,b/d1,c
• end
• end

8
Benchmarks

• A suite of SCSI bus controller benchmarks were
  re-expressed in DISP
• Heuristic H2 (with n1) was successfully applied
  in all cases
• In the case of the pipelined SCSI target receive
  protocol, Petrify could not solve CSC before
  decomposition
• PTRCV
• forever do
• select StartDMARcv/EndDMAInt then
• StartDMARcv/ReqOutN AckInN/ReqOutN,DRQ
• alt DRAckNormN,AckInN/ReqOutN,DRQ then
• DRAckNormN,AckInN/ReqOutN,DRQ
• alt DRAckLastN,AckInN/DRQ then
• DRAckLastN/EndDMAInt
• end
• end

9
Decomposition of PTRCV

• Fork out signal DRQ
• PTRCV1
• forever do
• select StartDMARcV/EndDMAInt then
• StartDMARcv/ReqOutN AckInN/x z/ReqOutN
• alt DRAckNormN,AckInN/x then
• z/ReqOutN DRAckNormN,AckInN/x z/ReqOutN
• alt DRAckLastN,AckInN/x then
• z/- DRAckLastN/EndDMAInt
• end
• end
• Fork forever do x/DRQ,z end

10
Experimental Results
11
Conclusion

• Petrify applies its own heuristics to resolve
  statecoding conflicts
• Specifications exhibiting concurrent outputs
  aredifficult for Petrify to handle
• New Fork-based decomposition heuristics
• Results from applying H2 to benchmarks
• 25 saving in area
• 52 fewer state variables
• 39 fewer set-reset pins
• 2.7 times faster synthesis

12
Conclusion

• Forks maintain delay-insensitive behaviour
• Offers practical approach to DI-decomposition
  where each component can be implemented as SI
  circuit
• Future work
• Wire-based decomposition
• Each wire introduces a new state variable
• Applicable to self-contained blocks found in
  counters
• Details of benchmark circuits and synthesis
  results
• http//www.bcim.lsbu.ac.uk/ccsv/dac04/
• It remains to ack support from EC FP5 contract
  IST-1999-29119 benchmarks from Nowick Petrify
  from Cortadella di2pn from Furey