Презентация Computer Architecture and Implementation онлайн

Содержание слайда: COMP 206: Computer Architecture and Implementation Montek Singh Mon, Oct 3, 2005 Topic: Instruction-Level Parallelism (Dynamic Scheduling: Introduction)

Содержание слайда: Instruction-Level Parallelism Relevant Book Reading (HP3): Dynamic Scheduling (in hardware): Appendix A & Chapter 3 Compiler Scheduling (in software): Chapter 4

Содержание слайда: Hardware Schemes for ILP Why do it in hardware at run time? Works when can’t know dependences at compile time Simpler compiler Code for one machine runs well on another machine Key idea: Allow instructions behind stall to proceed DIV.D F0, F2, F4 ADD.D F10, F0, F8 SUB.D F8, F8, F14 Enables out-of-order execution Implies out-of-order completion ID stage check for both structural and data dependences

Содержание слайда: Dynamic Scheduling Instructions are issued in order (leftmost I) Execution can begin out of order (leftmost E) Execution can terminate out of order (W) What is I?

Содержание слайда: Explanation of I To be able to execute the SUB.D instruction A function unit must be available Adder is free in example There should be no data hazards preventing early execution None in this example We must be able to recognize the two previous conditions Must examine several instructions before deciding on what to execute I represents the instruction window (or issue window) in which this examination happens If every instruction starts execution in order, then I is superfluous Otherwise: Instruction enter the issue window in order Several instructions may be in issue window at any instant Execution can begin out of order

Содержание слайда: Out-of-order Execution and Renaming WAW hazard on register F10: prevents out-of-order execution on machine like CDC 6600 If processor was capable of register renaming: the WAW hazard would be eliminated SUB.D could execute early as before example: IBM 360/91

Содержание слайда: Memory Consistency Memory consistency refers to the order of main memory accesses as compared to the order seen in sequential (unpipelined) execution Strong memory consistency: All memory accesses are made in strict program order Weak memory consistency: Memory accesses may be made out of order, provided that no dependences are violated Weak memory consistency is more desirable leads to increased performance In what follows, ignore register hazards Q: When can two memory accesses be re-ordered?

Содержание слайда: Four Possibilities for Load/Store Motion Load-Load can always be interchanged (if no volatiles) Load-Store and Store-Store are never interchanged Store-Load is the only promising program transformation Load is done earlier than planned, which can only help Store is done later than planned, which should cause no harm Two variants of transformation If load is independent of store, we have load bypassing If load is dependent on store through memory (e.g., (R1) == (R4)), we have load forwarding

Содержание слайда: More on Load Bypassing and Forwarding Either transformation can be performed at compile time if the memory addresses are known, or at run-time if the necessary hardware capabilities are available Compiler performs load bypassing in loop unrolling example (next lecture) In general, if compiler is not sure, it should not do the transformation Hardware is never “not sure”

Содержание слайда: Load Bypassing in Hardware Requires two separate queues for LOADs and STOREs Every LOAD has to be checked for every STORE waiting in the store queue to determine whether there is a hazard on a memory location assume that processor knows original program order of all these memory instructions In general, LOAD has priority over STORE For the selected LOAD instruction, if there exists a STORE instruction in the store queue such that … LOAD is behind STORE (in program order), and their memory addresses are the same … then the LOAD cannot be sent to memory, and must wait to be executed only after the store is executed

Содержание слайда: Example of Load Bypassing Memory access takes four cycles Actions at various points in time End of cycle 1: LQ = [(0)]; SQ = [ ]; execute first load End of cycle 5: LQ = [(3), (4)]; SQ = [(1)]; execute first load End of cycle 9: LQ = [(4), (5)]; SQ = [(1), (7)]; execute first load End of cycle 13: LQ = [(5)]; SQ = [(1), (7)]; load yields to store We are assuming that no LOADs or STOREs issue between instructions 7 and 22

Содержание слайда: History of Dynamic Scheduling First implemented on CDC 6600 and IBM 360/91 in the early 1960s Fetched and decoded single instr. at a time in program order Between decoding and execution, instructions stayed in issue window where hazards were detected and resolved Some hazards resolved before issuing (e.g., availability of FU) Most data hazards resolved only after issuing Hazard resolution done with sophisticated hardware scheme For each instruction in issue window, separately track, monitor, enforce, and eventually resolve all hazards affecting the instruction before it could begin execution Result: Instructions started execution when they were ready to execute, rather than starting in program order