1. Handshake protocols for de-synchronization I. Blunno, J. Cortadella, A. Kondratyev, L. Lavagno, K. Lwin and C. Sotiriou Politecnico di Torino, Italy Universitat Politecnica de Catalunya, Barcelona, Spain Cadence Berkeley Lab, Berkeley, USA ICS-FORTH, Crete, Greece

2. Asynchronous for dummies I. Blunno, J. Cortadella, A. Kondratyev, L. Lavagno, K. Lwin and C. Sotiriou Politecnico di Torino, Italy Universitat Politecnica de Catalunya, Barcelona, Spain Cadence Berkeley Lab, Berkeley, USA ICS-FORTH, Crete, Greece

3. Outline What is de-synchronization ? What is de-synchronization ? Behavioral equivalence Behavioral equivalence 4-phase protocols for de-synchronization 4-phase protocols for de-synchronization Concurrency Concurrency Correctness Correctness An example An example

4. Synchronous CLK Asynchronous De-synchronize CLK

5. MS flip-flop Synchronous circuit CLK LLLL L L 00 0 0 11

6. De-synchronization LLLL L L 00 0 0 11 C C C C C C

7. C C C C C C Distributed controllers substitute the clock network The data path remains intact !

8. Design flow Think synchronous Think synchronous Design synchronous: one clock and edge-triggered flip-flops Design synchronous: one clock and edge-triggered flip-flops De-synchronize (automatically) De-synchronize (automatically) Run it asynchronously Run it asynchronously

9. Prior work Micropipelines (Sutherland, 1989) Micropipelines (Sutherland, 1989) Local generation of clocks Local generation of clocks  Varshavsky et al., 1995  Kol and Ginosar, 1996 Theseus Logic (Ligthart et al., 2000) Theseus Logic (Ligthart et al., 2000)  Commercial HDL synthesis tools  Direct translation and special registers Phased logic (Linder and Harden, 1996) (Reese, Thornton, Traver, 2003) Phased logic (Linder and Harden, 1996) (Reese, Thornton, Traver, 2003)  Conceptually similar  Different handshake protocol (2 phase vs. 4 phase)

10. Automatic de-synchronization Devise an automatic method for de-synchronization Devise an automatic method for de-synchronization Identify a subclass of synchronous circuits suitable for de-synchronization Identify a subclass of synchronous circuits suitable for de-synchronization Formally prove correctness Formally prove correctness

11. Outline What is de-synchronization ? What is de-synchronization ? Behavioral equivalence Behavioral equivalence 4-phase protocols for de-synchronization 4-phase protocols for de-synchronization Concurrency Concurrency Correctness Correctness An example An example

14. Synchronous flow

25. De-synchronized flow

27. +

38. Flow equivalence [Guernic, Talpin, Lann, 2003]

39. A B

40. Flow equivalence CLK A 1 3 0 2 1 5 3 1 6 0 B 5 1 2 3 1 4 2 4 3 1 A 1 3 0 2 1 5 3 1 6 0 B 5 1 2 3 1 4 2 4 3 1 Synchronous behavior De-synchronized behavior

41. Flow equivalence CLK A 1 3 0 2 1 5 3 1 6 0 B 5 1 2 3 1 4 2 4 3 1 A 1 3 0 2 1 5 3 1 6 0 B 5 1 2 3 1 4 2 4 3 1 Synchronous behavior De-synchronized behavior

42. Outline What is de-synchronization ? What is de-synchronization ? Behavioral equivalence Behavioral equivalence 4-phase protocols for de-synchronization 4-phase protocols for de-synchronization Concurrency Concurrency Correctness Correctness An example An example

43. LLLL L L 00 0 0 11 C C C C C C

44. C C C C C C

45. C L

46. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0000 A latch cannot read another data item until the successor has captured the current one

47. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0100 A latch cannot read another data item until the successor has captured the current one

48. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0000 A latch cannot read another data item until the successor has captured the current one

49. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 1000 A latch cannot read another data item until the successor has captured the current one

50. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0000

51. ABCD A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0001

52. ABCD A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0000

53. ABCD A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0010 A latch cannot become opaque before having captured the data item from its predecessor

54. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0110 A latch cannot become opaque before having captured the data item from its predecessor

55. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0010 A latch cannot become opaque before having captured the data item from its predecessor

56. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0000 A latch cannot become opaque before having captured the data item from its predecessor

57. ABCD A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- A- B+ C- D+ A+ B- C+ D- 0000

58. ABCD A+ B+ C+ D+ A- B- C- D- A B

59. Outline What is de-synchronization ? What is de-synchronization ? Behavioral equivalence Behavioral equivalence 4-phase protocols for de-synchronization 4-phase protocols for de-synchronization Concurrency Concurrency Correctness Correctness An example An example

60. A+ B+ A- B- Can we increase concurrency ? A+ B+ A- B- A+ B+ A- B- not flow-equivalent

61. A B A B data overrun A B data lost

62. A+ B+ A- B- Can we reduce concurrency ? How much ?

63. A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- (8 states) (6 states) A+ B+ A- B- (5 states) A+ B+ A- B- A+ B+ A- B- (4 states)

64. A B A B A B A B A B A B fully decoupled (Furber & Day) simple 4-phase semi-decoupled (Furber & Day) non-overlapping GasP, IPCMOS de-synchronizationmodel

65. A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- simple 4-phase non-overlapping semi-decoupled (Furber & Day) fully decoupled (Furber & Day) GasP, IPCMOS de-synchronizationmodel

66. AB cntrl Ri Ai Rx Ax Ro Ao Ri+ A- Rx+ B- Ro+ Ai+ Ax+ Ao+ Ri- A+ Rx- B+ Ro- Ai- Ax- Ao- (semi-decoupled 4-phase protocol)

67. AB cntrl Ri Ai Rx Ax Ro Ao A- B- A+ B+ (semi-decoupled 4-phase protocol)

68. AB cntrl Ri Ai Rx Ax Ro Ao A- B- A+ B+ (semi-decoupled 4-phase protocol)

69. AB cntrl Ri Ai Rx Ax Ro Ao A- B- A+ B+ (semi-decoupled 4-phase protocol)

70. AB cntrl Ri Ai Rx Ax Ro Ao A- B- A+ B+ (semi-decoupled 4-phase protocol)

71. AB cntrl Ri Ai Rx Ax Ro Ao A- B- A+ B+ (semi-decoupled 4-phase protocol)

72. AB cntrl Ri Ai Rx Ax Ro Ao A- B- A+ B+ (semi-decoupled 4-phase protocol)

73. A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B-

74. Outline What is de-synchronization ? What is de-synchronization ? Behavioral equivalence Behavioral equivalence 4-phase protocols for de-synchronization 4-phase protocols for de-synchronization Concurrency Concurrency Correctness Correctness An example An example

76. Which protocols are valid for de-synchronization ?

77. A+ B+ A- B- Theorem: the de-synchronization protocol preserves flow-equivalence Proof: by induction on the length of the traces Induction hypothesis: same latch values at reset Induction step: same values at cycle i  same values at cycle i+1

78. A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B-

79. A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- Theorem: any reduction in concurrency preserves flow-equivalence

80. Any hybrid approach preserves flow-equivalence ! Semi-decoupledSemi-decoupled Semi-decoupled non-overlappingnon-overlapping FullydecoupledFullydecoupled

81. ABCD A+ B+ C+ D+ A- B- C- D-

82. ABCD A+ B+ C+ D+ A- B- C- D- semi- decoupled non- overlapping fully decoupled Flow-equivalence is preserved, … but …

83. Liveness Preservation of flow-equivalence: all the generated traces are equivalent Preservation of flow-equivalence: all the generated traces are equivalent Are all traces generated ? (Is the marked graph live ?) Not always ! Are all traces generated ? (Is the marked graph live ?) Not always !

84. A+ B+ C+ D+ A- B- C- D- Liveness: all cycles have at least one token [Commoner 1971] Semi-decoupled 4-phase handshake protocol

85. A+ B+ C+ D+ A- B- C- D- Simple 4-phase handshake protocol

86. Results about liveness At least three latches in a ring are required with only one data token circulating [Muller 1962] At least three latches in a ring are required with only one data token circulating [Muller 1962] (this paper): any hybrid combination of protocols is live if the simple 4-phase protocol is not used any cycle has at least one token Theorem (this paper): any hybrid combination of protocols is live if the simple 4-phase protocol is not used Proof: any cycle has at least one token

87. A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- simple 4-phase non-overlapping semi-decoupled (Furber & Day) fully decoupled (Furber & Day) GasP, IPCMOS de-synchronizationmodel

88. Outline What is de-synchronization ? What is de-synchronization ? Behavioral equivalence Behavioral equivalence 4-phase protocols for de-synchronization 4-phase protocols for de-synchronization Concurrency Concurrency Correctness Correctness An example An example

90. Async DLX block diagram

91. Synchronous RTL = Synchronous Desynchronized Cycle: 4.4ns Power: 70.9mW Area: 372,656  m Cycle: 4.45ns Power: 71.2mW Area: 378,058  m All numbers are after Placement & Routing All numbers are after Placement & Routing Total of 1500 flip-flops, 3000 latches Total of 1500 flip-flops, 3000 latches DE-SYNC design includes 5 controllers, each driving 2 clock trees DE-SYNC design includes 5 controllers, each driving 2 clock trees Power numbers include the clock tree Power numbers include the clock tree Technology: UCM/Virtual Silicon 0.18 µm Technology: UCM/Virtual Silicon 0.18 µm

92. De-synchronized DLX on FPGA (demo outside the conference room)

93. Discussion The de-synchronization model provides an abstraction of the timing behavior The de-synchronization model provides an abstraction of the timing behavior

94. [2,3] [1,2][8,9] [5,7] [3,5] [2,4] ABE F G C D [0,0][3,5] [5,7] [2,3] [2,4] [1,2] [8,9] [1,2] Timing analysis Exploration of the design space

95. A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- A+ B+ A- B- simple 4-phase non-overlapping semi-decoupled (Furber & Day) fully decoupled (Furber & Day) GasP, IPCMOS de-synchronizationmodel

96. Conclusions EDA tools require a formal support (they must work for all circuits) EDA tools require a formal support (they must work for all circuits) A complete characterization of 4-phase protocols has been presented (partial order based on concurrency) A complete characterization of 4-phase protocols has been presented (partial order based on concurrency) Design flow developed at Cadence Berkeley Labs Design flow developed at Cadence Berkeley Labs  Automated from gate netlist  Static timing analysis to derive matched delays  Constrained P&R to meet timing constraints