Chapter 4: Retiming Keshab K. Parhi Ret iming : Moving around - PowerPoint PPT Presentation

Chapter 4: Retiming Keshab K. Parhi

Ret iming : Moving around exist ing delays • Does not alt er t he lat ency of t he syst em • Reduces t he crit ical pat h of t he syst em • Node Ret iming D 3D 5D 3D 2D •Cut set Ret iming D D 2D B D D F A D C E D Chap. 4 2

Retiming • Generalization of Pipelining • Pipelining is Equivalent to Introducing Many delays at the Input followed by Retiming Chap. 4 3

• Ret iming Formulat ion Ret iming r(U) r (V) ω ω ’ U V U V Source node Dest inat ion node ω ’ = ω + r(V) - r(U) •Propert ies of ret iming –The weight of t he ret imed pat h p = V 0 --> V 1 --> … .. V k is given by ω r (p)= ω (p) + r (V k ) - r(V 0 ) –Ret iming does not change t he number of delays in a cycle. –Ret iming does not alt er t he it erat ion bound in a DFG as t he number of delays in a cycle does not change –Adding t he const ant value j t o t he ret iming value of each node does not alt er t he number of delays in t he edges of t he ret imed graph. •Ret iming is done t o meet t he f ollowing – Clock period minimizat ion – Regist er minimizat ion Chap. 4 4

• Ret iming f or clock period minimizat ion – Feasibilit y const raint ω ’(U,V) ≥ 0 ⇒ causalit y of t he syst em ⇒ ω (U,V) ≥ r(U) - r(V) (one inequalit y per edge) – Crit ical Pat h const raint r(U) - r(V) ≤ W(U,V) - 1 f or all vert ices U and V in t he graph such t hat D(U,V) > c where c = t arget clock period. The t wo quant it ies W(U,V) and D(U,V) are given as: W(U,V) = min{ w (p) : U → V} D(U,V) = max{ t (p) : U → V and w (p) = W(U,V) (1) G D (1) (1) 2D A B C D E (1) (1) (1) D W(A,E) = 1 & D(A,E) = 5 F (2) Chap. 4 5

• Algorit hm t o comput e W(U,V) and D(U,V): • Let M = t max n, where t max is t he maximum comput at ion t ime of t he nodes in G and n is t he # of nodes in G. • Form a new graph G’ which is t he same as G except t he edge weight s are replaced by w’(e) = Mw(e) – t (u) f or all edges U � V. • Solve f or all pair short est pat h problem on G’ by using Floyd Warshall algorit hm. Let S’ UV be t he short est pat h f orm U � V. I f U ≠ V, t hen W(U,V) =  S’ UV / M  and D(U,V) = MW(U,V) - • S’ UV + t (V). I f U = V, t hen W(U,V) = 0 and D(U,V) = t (U). • Using W(U,V) and D(U,V) t he f easibilit y and crit ical pat h const raint s are f ormulat ed t o give cert ain inequalit ies. The inequalit ies are solved using const raint graphs and if a f easible solut ion is obt ained t hen t he circuit can be clocked wit h a period ‘c’. Chap. 4 6

• Solving a syst em of inequalit ies : Given M inequalit ies in N variables where each inequalit y is of t he f orm r i – r j ≤ k f or int eger values of k. � Draw a const raint graph � Draw t he node i f or each of t he N variables r i , I = 1, 2, … , N. � Draw t he node N+1. � For each inequalit y r i – r j ≤ k , draw t he edge j � i of lengt h k. � For each node i, i = 1, 2, … , n, draw t he edge N+1 � i f rom t he node N+1 t o node I wit h lengt h 0. � Solve using a short est pat h algorit hm. � The syst em of inequalit ies have a solut ion if f t he const raint graph cont ains no negat ive cycles. � I f a solut ion exist s, one solut ion is where r i is t he minimum lengt h pat h f rom t he node N+1 t o node i. Chap. 4 7

• K-slow t ransf ormat ion – Replace each D by kD Clock (1) (1) A0 → B0 0 A B T iter = 2ut A1 → B1 1 A2 → B2 2 D Af t er 2-slow t ransf ormat ion Clock (1) (1) A0 → B0 0 A B T clk = 2ut 1 T iter = 2 × 2ut=4ut A1 → B1 2D 2 3 A2 → B2 4 *I nput new samples every alt ernat e cycles. *null operat ions account f or odd clock cycles. *Hardware ut ilized only 50% t ime Chap. 4 8

• Ret iming 2-slow graph D A B D T clk = 1ut T it er = 2 × 1=2ut *Hardware Ut ilizat ion = 50 % *Hardware can be f ully ut ilized if t wo independent operat ions are available. Chap. 4 9

2-Slow Lattice Filter (Fig. 4.7) A 100 stage Lattice Filter with critical path 2 multiplications and 101 additions The 2-slow version Chap. 4 10

A ret imed version of t he 2 slow circuit wit h crit ical pat h of 2 mult iplicat ions and 2 addit ions I f T m = 2 u.t . and T a = 1 u.t ., t hen T clk = 6 u.t ., T it er = 2X6 = 12 u.t . I n Original Lat t ice Filt er, T it er = 105 u.t . I t erat ion Period Bound = 7 u.t . Chap. 4 11

Other Applications of Retiming • Retiming for Register Minimization (Section 4.4.3) • Retiming for Folding (Chapter 6) • Retiming for Power Reduction (Chap. 17) • Retiming for Logic Synthesis (Beyond Scope of This Class) • Multi-Rate/Multi-Dimensional Retiming (Denk/Parhi, Trans. VLSI, Dec. 98, Jun.99) Chap. 4 12