1. Roman LyseckyUniversity of California, Riverside1 Techniques for Reducing Read Latency of Core Bus Wrappers Roman L. Lysecky, Frank Vahid, & Tony D. Givargis Department of Computer Science University of California Riverside, CA 92521 {rlysecky, vahid, givargis}@cs.ucr.edu This work was supported in part by the NSF and a DAC scholarship.

2. Roman LyseckyUniversity of California, Riverside2 Introduction Core Library MIPS MEM Cache DSP DMA Core XCore Y Core-based designs are becoming common –available as both soft and hard Problem - How can interfacing be simplified to ease integration?

3. Roman LyseckyUniversity of California, Riverside3 Introduction One Solution - One standard on-chip bus –All cores have same interface –Appears to be unlikely (VSIA) Another Solution - Divide core into a bus wrapper and internal parts –Rowson and Sangiovanni-Vincentelli ‘97 - Interface-Based Design –VSIA developing standard for interface between wrapper and internals Far simpler than standard on-chip bus –Refer to bus wrapper as an interface module(IM)

4. Roman LyseckyUniversity of California, Riverside4 Previous Work - Pre-fetching Analogous to caching, store local copies of registers inside the interface module Enable quick response time Eliminates extra cycles for register reads Transparent to system bus and core internals Easily integrate with different busses No performance overhead Acceptable increases in size and power Pre-fetching was manually added to each core

5. Roman LyseckyUniversity of California, Riverside5 Previous Work - Architecture of IM pre-fetch registers Pre-fetch Unit - Implements the pre- fetching heuristic Goal: maximize the number of hits Controller - Interfaces to system bus How can we automate the design of the PFU?

6. Roman LyseckyUniversity of California, Riverside6 Outline “Real-time” Pre-fetching –Mapping to real-time scheduling Update Dependency Model –General Register Attributes –Petri Net model construction –Petri Net model refinement –Pre-fetch Scheduling Experiments Conclusions

7. Roman LyseckyUniversity of California, Riverside7 Real-time Pre-fetching A - Age Constraint = 4 B - Age Constraint = 6 Access-time Constraint = 2 Naïve Schedule More Efficient Schedule Age constraint –Number of cycles old data may be when read Access-time constraint –Maximum number of cycles a read access may take

8. Roman LyseckyUniversity of California, Riverside8 Real-time Pre-fetching Mapping to Real-time scheduling –Register -> Process –Internal bus -> Processor –Pre-fetch -> Process execution –Register age constraint -> Process period –Register Access-time constraint -> Process deadline –Pre-fetch time -> Process computation time Assume a pre-fetch requires 2 cycles

9. Roman LyseckyUniversity of California, Riverside9 Real-time Pre-fetching Cyclic Executive –Major cycle = time required to pre-fetch all registers –Minor cycle = rate at which highest priority process will be executed –Problems Sporadic writes All process periods must be multiples of the minor cycle Computationally infeasible for large register sets

10. Roman LyseckyUniversity of California, Riverside10 Real-time Pre-fetching Rate monotonic priority assignment –Register with smallest register age constraint will have the highest priority

11. Roman LyseckyUniversity of California, Riverside11 Real-time Pre-fetching Ci = Computation Time for register i Ai = Pre-fetch Time for register i Utilization-based schedulability test

12. Roman LyseckyUniversity of California, Riverside12 Real-time Pre-fetching Ri = Response Time for register i Ci = Computation Time for register i Ii = Maximum interference in interval [t, t+Ri) Response Time Analysis –Response of register I is defined as follows –Register set is schedulable if for each register the response time is less than or equal to its age constraint

13. Roman LyseckyUniversity of California, Riverside13 Real-time Pre-fetching Sporadic register writes –Writes to registers are sporadic –Take control of internal bus, thus delaying pre-fetching of registers Deadline monotonic priority –Register with smallest register access-time constraint will have the highest priority –Add a write register WR to register set Access-time constraint = Deadline Age constraint = maximum rate at which write will occur

14. Roman LyseckyUniversity of California, Riverside14 Experiments - Area(Gates) Note: To better evaluate the effects of IM’s, our cores were kept simple, thus resulting in a smaller than normal size. Average increase of IM w/ RTPF over IM w/ BW of 1.4K gates

15. Roman LyseckyUniversity of California, Riverside15 Experiments - Performance(ns)

16. Roman LyseckyUniversity of California, Riverside16 Experiments - Energy(nJ)

17. Roman LyseckyUniversity of California, Riverside17 Register Attributes –Update type, access type, notification type, and structure type Update dependencies –Internal dependencies dependencies between registers –External dependencies updates to register via reads and writes from on-chip bus updates from external ports to internal core register Petri Nets –Determined that we could use Petri Nets to model our update dependencies

18. Roman LyseckyUniversity of California, Riverside18 Petri Net Based Dependency Model Bus Place Random Transition Register Places Update Dependencies

19. Roman LyseckyUniversity of California, Riverside19 Refined Petri Net Model Data Dependency Refined Transition

20. Roman LyseckyUniversity of California, Riverside20 Pre-fetch Schedule Create a heap registers to be pre-fetched Create a list for update arcs Repeat –if request detected then add outgoing arcs to heap set write register access-time to 0 and add to heap –if read request detected then add outgoing arcs to update arc list –for register at top of heap do if access-time = 0 then pre-fetch register, remove from heap if current age = 0 then pre-fetch register, reset current age, add register to heap –while update arcs list is not empty do if transition fires then set register’s access-time to 0 and add to heap

21. Roman LyseckyUniversity of California, Riverside21 Experiments - Area(Gates) Note: To better evaluate the effects of IM’s, our cores were kept simple, thus resulting in a smaller than normal size. Average increase of IM w/ PF over IM w/ BW of 1.5K gates Average increase of IM w/ PF over IM w/ RTPF of.1K gates

22. Roman LyseckyUniversity of California, Riverside22 Experiments - Performance(ns)

23. Roman LyseckyUniversity of California, Riverside23 Experiments - Energy(nJ)

24. Roman LyseckyUniversity of California, Riverside24 Conclusions Real-time pre-fetching and update dependency pre- fetching produce good results Update dependency model is more efficient in pre- fetching registers Two approaches are complementary Enable the automatic generation of pre-fetching unit