VirtualSync: Timing Optimization by Synchronizing Logic Waves with Sequential and Combinational Components as Delay Units

Transcription

1 VirtualSync: Timing Optimization by Synchronizing Logic Waves with Sequential and Combinational Components as Delay Units Grace Li Zhang 1, Bing Li 1, Masanori Hashimoto 2 and Ulf Schlichtmann 1 1 Chair of Electronic Design Automation, Technical University of Munich (TUM) 2 Department of Information Systems Engineering, Osaka University

2 Overview Motivation VirtualSync Timing Model Timing Optimization Framework of VirtualSync Experimental Results Summary 2

3 The Traditional Timing Paradigm Clock-to-q delay t cq : 3 Setup time t su : 1 Hold time t h : 1 T min = t cq +d max +t su =3+17+1=21 Sequential components such as flip-flops synchronize signal propagations. Combinational gates perform logic computations. Reduce design effort Disadvantages Flip-flops have clock-to-q delays and impose setup time. Delay imbalances between flipflop stages degrade performance. 3

4 Timing Optimization Methods Gate Sizing T min = =16 Retiming T min = 3+7+1=11 The limit in the traditional timing paradigm VirtualSync T min =(3+13+1)/2= % reduction compared with retiming&sizing 4

5 VirtualSync Concept fast path must be delayed loops must be blocked boundary F3 boundary F5 boundary F6 Circuit under optimization VirtualSync: Step 1: Remove all flip-flops except those at the boundary of the module Step 2: Block fast signals for timing synchronization, including signals arriving at boundary flip-flops too early through fast paths signals traveling across combinational loops 5

6 VirtualSync Concept delay fast path by buffers relative reference points for timing checking F/L loop blocked by flip-flop/latch boundary boundary boundary Circuit under optimization Delay units (logic gates, flip-flops and latches) are used to slow down signals on fast paths and loops. Relative reference points provide relative timing information. 6

7 Delay Units in VirtualSync s u s v s u s v s u s v t d s v output gap s v T+t cq t su s v output gap T/2+t cq input gap t su D: duty cycle t d input gap s u t h T s u t h D*T T s u Linear delaying effect of a combinational delay unit Constant delaying effect of a flip-flop Piecewise delaying effect of a latch Input gap: the difference of arrival times of two signals at a delay unit Output gap: the difference between their arrival times after they pass through the unit 7

8 Overall Flow of VirtualSync sequential circuits remove all flip-flops mark reference points create selection variables for delay units at each circuit node maximize performance and minimize area using ILP decrease lower bound of inserted delay no set lower bound of inserted delay All required delays are padded? yes Optimized circuit 8

9 Results of VirtualSync Speed increase (%) Area change (%) Speed increase and area results compared with ideally balanced design 9

10 Summary A new timing model, VirtualSync, with sequential components and combinational logic gates as delay units is proposed. By viewing flip-flops and latches as delay units, circuit performance can be pushed even beyond the limit of the traditional timing paradigm. VirtualSync demonstrates a good potential for high-performance designs. 10

11 Thank you for your attention!

12 Heuristic method in VirtualSync Emulation of sequential delay units with different padding delays for long and short paths Model approximation with clock/data-to-q delays Yes Different padding delays are needed? No Model legalization using accurate delay models Different padding delays are needed? Yes Buffer replacement using sequential units and delay discretization No Optimized circuit 12

13 Results of VirtualSync Circuit Critical part Optimized circuit Comparison #gates #flipflop #flipflop #latch #buffer clock period reduction area increase s % 2.84% s % -5.17% s % -1.09% s % 6.01% s % -0.5% systemcdes % 2.43% mem_ctrl % 0.97% usb_funct % 0.21% ac97_ctrl % -9.76% pci_bridge % 0.05% The comparison was made with extreme retiming and sizing, with which the timing performance has reached the limit in the traditional timing paradigm. 13

14 Relative Timing References in VirtualSync s o =3 s u =14 s v =4 s w =7 s t =3 s z =5 o u v w t z F1 11 F2 3 F3 2 F4 T=10 t cq =3 t su =1 t h =1 boundary boundary removed after optimization kept after optimization s z t h s z +t su T The location of the removed flip-flops such as F2 and F3 are called anchor points. The anchor points allow to relate timing information to boundary flip-flops. Every time when a signal passes an anchor point, its arrival time is converted by subtracting T. If F3 is removed, the arrival time s z becomes -3+2=-1, violating the hold time constraint. The timing constraints at the boundary flip-flops force the usage of the internal sequential delay units! 14

15 Synchronizing Logic Waves by Delay Units comb. delay? ξ uv seq. delay? anchor? λ tz u v d vw sizing? w t z 1. Combinational delay unit and gate sizing s w s u +ξ uv *r u + d vw *r u s w s u +ξ uv *r l + d vw *r l (1) (2) s u, s u,s w, s w are the latest and earliest arrival time of node u and w. is the delay of an inserted buffer. r u and r l are two constants to reserve a guard band for process variations. ξ uv 15

16 Synchronizing Logic Waves by Delay Units comb. delay? ξ uv seq. delay? anchor? λ tz u v d vw sizing? w t z 2. Insertion of sequential delay units Case 1: No sequential delay unit is inserted between w and t s t s w s t s w (3) (4) s t, s t,s w, s w are the latest and earliest arrival time of node t and w. 16

17 Synchronizing Logic Waves by Delay Units comb. delay? ξ uv seq. delay? anchor? λ tz u v d vw sizing? w t z 2. Insertion of sequential delay units Case 1: A flip-flop is inserted between w and t s w, s w N wt *T +φ wt +t h *r u (5) s t (N +1)*T +φ +t *r wt wt cq s w, s w (N wt +1)*T +φ wt t su *r u (6) s t (N wt +1)*T +φ wt +t cq *r l u (7) (8) φ wt is the phase shift of the clock signal A flip-flop only works in a region t h after the rising clock edge and t su before the next rising clock edge. The signal always starts to propagate from the next active clock edge. 17

18 Synchronizing Logic Waves by Delay Units comb. delay? ξ uv seq. delay? anchor? λ tz u v d vw sizing? w t z 2. Insertion of sequential delay units Case 1: A level-sensitive latch is inserted between w and t s t N wt *T +φ wt + D*T +t cq *r u s t s w +t dq *r u (9) s t max(n wt *T +φ wt + D*T +t cq *r l, (10) s w +t dq *r l ) (11) D is the duty cycle of the clock signal The upper is the case that the latch is non-transparent; the lower is the case that the latch is transparent. The signal starts to propagate from the maximum of the earliest time. 18

19 Synchronizing Logic Waves by Delay Units comb. delay? ξ uv seq. delay? anchor? λ tz u v d vw sizing? 3. Reference shift with respect to anchor points s z = s t λ tz *T (12) 4.Wave non-interference condition s u +t stable s u +T Overall formulation w t z (13) Objective: find a solution to make the circuit work at a given clock period Subject to: (1)-(13) and setup and hold time constraints at the boundary flip-flops NP-hard! 19

20 Results of seq. delay units after buffer replacement Number of seq. units Before rep. After rep. 0 20

21 Runtime Circuit T r (s) s s s s s systemcdes mem_ctrl usb_funct ac97_ctrl pci_bridge