Synchronization Detection for Multi-Process Hierarchical Synthesis

1
Synchronization Detection for Multi-Process
Hierarchical Synthesis

• Oliver Bringmann
• Wolfgang Rosenstiel
• Dirk Reichardt

2
Outline

• Resource sharing in hierarchical synthesis
• Communication model and synchronization points
• Synchronization point detection
• Synchronicity condition
• Initial synchronization points
• Loop-based synchronization points
• Conclusion

3
Hierarchical Synthesis
System Specification
Specification
RT-Library
Schedule
RT-Structure
Bottom-up Synthesis
High-Level Synthesis

• Synthesis of hierarchical composed system
  specifications
• Providing of optimizations across different
  hierarchical levels with respect to already
  synthesized module structures
• Automatic integration of optimized modules
• Resource sharing across process boundaries
• Multiple processes have to be considered during
  synthesis

4
Resource Sharing in Multi-Process Specifications
Prerequisite Common Clock and Reset
Clock
Reset
FSM
FSM
FSM
Goal Only one implementation of module
Main tasks 1. Which states synchronize/de-synchro
nize different processes? 2. In which clock steps
are the demanded submodules in use? 3. Can a
resource conflict occur when sharing
submodules? (2. 3. Þ Resource sharing in
hierarchical synthesis, ICCAD 97)
5
Synchronization and De-Synchronization Points
6
Communication Model
Communication model bases on message passing
7
Synchronicity Condition (SC)
Synchronicity Check by Simulation
Process P2
Process P1
Analytically Synchronicity Check

• Communication Ci is a synchronization point Û
  is reached before / simultaneously with
  , for all paths, starting from previous
  synchronization points

8
Communication Dependency Graph
Communication Dependency Graph
Hierarchical Schedule
I
I
I
I
1
2
S1
1
S1
C1
C1
R1
1
R1
1
R2
2
R2
4
1


C2
C2
2
3
3
2
S2
3
S2
1
C3
C3
R3
S3
R3
S3
Process P2
Process P1
Process P2
Process P1

• Reduction to communication states , and
  unbounded loops
• Sequential timing are represented by edge labels

R
S
U
3
S2
S1
9
Node Collapsing during Graph Construction
10
Applicability of the Synchronicity Condition
I
I
1
2
S1
C1
R1
1
R2
3
4
1
12
C2
U1
2
1
u1
S2
1
C3
S3
R3

• Problem Synchronicity condition (SC) are only
  applicable between SPs
• Þ SC gives just a criteria for verifying SPs
• Þ SC can not directly be used to determine SPs

11
Initial Synchronization Points (ISPs)
I
I
1
2
S1
C1
R1
1
R2
3
4
1
12
C2
U1
2
1
u1
S2
1
C3
S3
R3

• Applicability problem not existing for initial
  synchronization points
• Representation of the SCs bases on linear
  inequalities
• Transformation to linear equalities by use of
  slack variables
• Slack variables represent the wait states of the
  receive nodes
• Negative slack values indicate a violation of the
  SCs

12
Loop-based SP Calculation by Cyclic Implication
Synchronization Cycles
I
I
SC(C1 C2)
1
2
S1
C1
R1
SC(C2 C3)
SC(C3 C1)
1
R2
3
4
1
SC(C2 C3)
12
C2
U1
2
SC(C3 C2)
1
u1
S2
1
C3
S3
R3
SC(C1 C1)

• Assumption All communications are SPs
• SCs are applied to cyclic dependent
  communications
• Each cycle are considered independently
• If all SCs of a synchronization cycle are valid Þ
  related Ci are LSPs

13
Synchronization Point Detection (Outline)
I
I
1
2
S1
C1
x1, x2, x3 ³ 0
Þ x2 2
R1
1
Þ x3 12
R2
3
4
1
Þ x1 -1 Þ x1 0
12
C2
U1
2
1
u1
S2
1
C3
S3
R3
1. Choose greatest unconsidered communication
cycle
2. Evaluate synchronicity conditions for the
involved communications
14
Conclusion

• Synchronization point detection for multi-process
  specifications
• Inter-process communication bases on the
  message-passing model
• Blocking and non-blocking communications are
  allowed
• No restrictions of structured control structures
  are required
• Especially unbounded data-dependent loops can be
  handled
• SPs can be used for multi-process resource
  sharing
• Further area of application is for instance, the
  minimization of the communication overhead

15
Related Work

• Filo, Ku, Coelho, De Micheli (IEEE TCAD, 1993)
• Minimization of the communication overhead
• Communication model restricted to basic blocks
• Amon, Hulgaard, Burns, Borriello (ICCD, 1993)
• Determination of timing bounds between different
  consecutive communications
• Conditional branches are not allowed
• Dey, Bommmu (ICCAD, 1997)
• Calculation of the worst-case execution time
  inmulti-process specifications
• Limited to specifications without unbounded loops