Scheduling Primitives for Bluespec Arvind Computer Science & Artificial Intelligence Lab

1. Scheduling Primitives for Bluespec Arvind Computer Science & Artificial Intelligence Lab Massachusetts Institute of Technology Work in progress http://csg.csail.mit.edu/arvind

2. outQ Done? mem fifo Motivation rule “recirculate” x = mem.peek(); y = fifo.first(); if done?(x) then fifo.deq() ; mem.deq() ; outQ.enq(x) else mem.deq(); mem.req(addr(x)); fifo.deq(); fifo.enq(y) Need a way of - saying deq happens before enq within the same rule - building such a FIFO http://csg.csail.mit.edu/arvind

3. Conditional action Parallel Composition Sequential Composition Method call Guarded action BS: A Language of Atomic Actions A program is a collection of instantiated modules m1 ; m2 ; ... Module ::= Module name [State variable r] [Rule R a] [Action method g (x) = a] [Read method f (x) = e] a ::= r := e |if e then a | a | a | a ; a | (t = e in a) | m.g(e) | a when e e ::= r | c | t | Op(e , e) | e ? e : e | (t = e in e) | m.f(e) | e when e http://csg.csail.mit.edu/arvind

4. outQ Done? mem fifo Notice LPM in Bluespecusing the sequential connective module lpm rule “recirculate” action method enter(x) = mem.req(addr(x)) | fifo.enq(x) Different made up syntax (x = mem.peek() in (y = fifo.first() in (if done?(x) then fifo.deq() | mem.deq() | outQ.enq(x) else (mem.deq(); mem.req(addr(x))) | (fifo.deq(); fifo.enq(y))) http://csg.csail.mit.edu/arvind

5. Guards vs If’s • Guards affect the surroundings (a1 when p1) | a2 ==> (a1 | a2) when p1 • Effect of an “if” is local (if p1 then a1) | a2 ==> if p1 then (a1 | a2) else a2 p1 has no effect on a2 http://csg.csail.mit.edu/arvind

6. Conditionals & Cases if p then a1 else a2  if p then a1 | if !p then a2 Similarly for cases http://csg.csail.mit.edu/arvind

7. Semantics of a rule execution • Specify which state elements the rule modifies • Let  represent the value of all the registers before the rule executes • Let U be the set of updates implied by the rule execution http://csg.csail.mit.edu/arvind

8. reg-update ├ e  v, (v != ┴) ├ (r := e)  {(r , v)} if-true ├ e  true, ├ a  U ├ (if e then a)  U Like a set union but blows up if there are any duplicates if-false ├ e false ├ (if e then a)  par ├ a1  U1, ├ a2  U2 ├ (a1 | a2) pmerge(U1,U2) seq ├ a1  U1, update(,U1)├ a2  U2 ├ (a1 ; a2) smerge(U1,U2) Like pmerge but U2 dominates U1 BS Action Rules http://csg.csail.mit.edu/arvind

9. a-let-sub ├ e  v, smerge(,{(t,v)})├ a  U ├ ((t = e) in a)  U a-meth-call ├ e  v , (v != ┴) ,x.a = lookup(m.g), smerge(,{(x,v)}) ├ a  U ├ m.g(e)  U BS Action Rules cont Expression rules are similar http://csg.csail.mit.edu/arvind

10. a-when-ready ├ e  true, ├ a  U ├ (a when e)  U e-when-ready-true ├ e1  v, ├ e2  true ├ (e1 when e2)  v e-when-ready-false ├ e1  v, ├ e2  false ├ (e1 when e2) ┴ Guard Semantics • If no rule applies, e.g., if e in (a when e) returns false, then the system is stuck and the effect of the whole atomic action of which (a when e) is a part is “no action”. http://csg.csail.mit.edu/arvind

11. Highly non-deterministic Execution model Repeatedly: • Select a rule to execute • Compute the state updates • Make the state updates Implementation concern: Schedule multiple rules concurrently without violating one-rule-at-a-time semantics http://csg.csail.mit.edu/arvind

12. outQ Done? mem fifo LPM: Scheduling issues module lpm rule “recirculate” action method enter(x) = mem.req(addr(x)) | fifo.enq(x) Can a rule calling method “enter” execute concurrently with the “recirculate” rule ? If not who should have priority? (x = mem.res() in (y = fifo.first() in (if done?(x) then fifo.deq() | mem.deq() | outQ.enq(x) else (mem.deq(); mem.req(addr(x))) | (fifo.deq(); fifo.enq(y))) http://csg.csail.mit.edu/arvind

13. Concurrent Scheduling & Rule composition Rule R1 a1 when p1 Rule R2 a2 when p2 Scheduling Primitives: 1. S = par(R1,R2) where the meaning of rule S is Rule S ((if p1 then a1)|(if p2 then a2)) when (p1 xor p2) 2. S = compose(R1,R2) where the meaning of rule S is Rule S ((a1 when p1);(a2 when p2)) 3. S = restrict(R1,R2) where the meaning of rule S is Rule S (a2 when (!p1 & p2)) http://csg.csail.mit.edu/arvind

14. Schedules Grammar S ::= R | compose(S, S) | par(S, S) | restrict(S, S) 1. The user must specify a schedule 2. A rule can occur multiple times in a schedule Such schedules combined with rule-splitting lets the user specify concurrency safely in execution http://csg.csail.mit.edu/arvind

15. LPM: Split the Internal rule module lpm rule “recirculate” (x = mem.res() in (y = fifo.first() in (if !done?(x) then (mem.deq(); mem.req(addr(x))) | (fifo.deq(); fifo.enq(y))) rule “exit” (x = mem.res() in (y = fifo.first() in (if done?(x) then fifo.deq() | mem.deq() | outQ.enq(x) action method enter(x) = mem.req(addr(x)) | fifo.enq(x) http://csg.csail.mit.edu/arvind

16. LPM Schedules enter has priority over recirculate; no dead cycles • Schedule-1 rc = restrict(enter,recirculate) ee = compose(exit,enter) ex = restrict(ee,exit) en = restrict(exit,enter) S1 = par(rc,ee,en,ex) • Schedule-2 er = restrict(recirculate, enter) ee = compose(exit,er) ex = restrict(ee,exit) en = restrict(er,exit) S2 = par(ee,en,ex) recirculate has priority over enter; no dead cycles Difficult to pick between S1 and S2! http://csg.csail.mit.edu/arvind

17. Final words • Market forces are demanding a much greater variety of SoCs • The design cost for SoCs has to be brought down dramatically by facilitating IP reuse • High-level synthesis tools are essential for architectural exploration and IP development Bluespec and associated tools are a promising solution to this problem http://csg.csail.mit.edu/arvind

18. Resources • The Blusepec Inc Web site http://bluespec.comhttp://bluespec.com/products/Papers.htm • Contact Bluespec Inc to get a free copy of the BSV language manual. For other publications: http://csg.lcs.mit.edu/pubs/bluespecpubs.html • The 6.375 subject at MIT: Complex Digitial Systems http://csg.csail.mit.edu/6.375/ Thanks for listening http://csg.csail.mit.edu/arvind