EENG449b/Savvides Lec /22/05 March 22, 2005 Prof. Andreas Savvides Spring EENG 449bG/CPSC 439bG Computer Systems Lecture 15 Instruction Level Parallelism I

EENG449b/Savvides Lec /22/05 Instruction Level Parallelism Reading for this lecture: Chapter 3, pages 172 – 192 Chapter 3: ILP in hardware Recall ILP tries to minimize these terms through the overlapped execution of instructions

EENG449b/Savvides Lec /22/05 Recap So far we discussed pipelining as a basic form of ILP Now we will see some more advanced topics Trend: –Dynamic approaches mostly dominate desktops (hardware implementation) –Static scheduling approaches (Chap. 4) more popular in embedded systems.

EENG449b/Savvides Lec /22/05 Where is the maximal gain in ILP? Basic block – a straight line code sequence with no branches in it except the entry and the exit point Limited amount of parallelism within a basic block –Instructions depend on each other so they cannot be reordered –In typical MIPS programs dynamic branch frequency between 15 – 25% ( 4 – 7 ) instructions between a pair of branches Need to exploit parallelism across multiple basic blocks

EENG449b/Savvides Lec /22/05 Loops : an example for parallelism for (i=1; i <= 1000; i=i+1) x[i] = x[i] + y[i]; Loop iterations can overlap – loop level parallelism Main technique – loop unrolling –Can be done either in hardware or software So what kind of dependencies do we need to worry about?

EENG449b/Savvides Lec /22/05 Data Dependences & Hazards An instruction i is data depended on instruction j if: –Instruction i produces a result used by instruction j –Instruction j is data dependent on instruction k, and instruction k is data depended on instruction i

EENG449b/Savvides Lec /22/05 Data Dependences Data dependencies are properties of programs Detection of hazards and stalls are properties of the pipeline organization A dependence can be overcomed by: Maintaining the dependence and avoiding the hazard Transforming the code to eliminate the dependence

EENG449b/Savvides Lec /22/05 Detecting Data Dependences Data values can flow through registers or memory Data dependences that flow through registers are easy to detect –Register names are the same so it is easy to check –More complicated when branches intervene Data dependences are harder to detect in memory 100(R4) and 20(R6) may point to the same memory location!! Crucial aspect to consider in compiler techniques

EENG449b/Savvides Lec /22/05 Name Dependences Name dependence: two instructions use the same register or memory, without any flow of data that is actually associated with that register or memory location Types of name dependences Antidependence (WAR) – instruction j writes a register that instruction i reads Output dependence (WAW) – instruction i and instruction j write the same memory location or register Name dependences are not real dependences Just change the names – register renaming – can be done by the hardware or the compiler

EENG449b/Savvides Lec /22/05 Data Hazards & Dependences Changes the access to the operand ordering Read After Write (RAW) – j tries to read a source before i writes it – program order must be reserved Write After Write (WAW) – j tried to write an operand before it is written by i – output dependence. Can only happen in pipelines that write in more than one stage or let an instruction to proceed when another instruction is stalled Write After Read (WAR) – j tries to write an instruction before it is read by i – antidependence – mostly occurs when instructions write results early in the pipeline, or when instructions are reordered

EENG449b/Savvides Lec /22/05 Control Dependences Control dependences control the ordering of instructions with respect to branch instructions –Instructions should execute in correct program order –Ex. Should not execute instructions from the then clause of an if statement if not needed Control dependence constraints –Instructions control dependent on a branch cannot be moved before a branch »E.g an instruction from the then component of a statement cannot be move before the if component –An instruction that is not control dependent on a branch cannot be moved after the branch so that is execution is depended on the branch

EENG449b/Savvides Lec /22/05 Control Dependence Control dependence is not the critical property to preserve –May be willing to execute extra instructions if that does not compromise program correctness Need to preserve –Exception behavior – the way exceptions raise in a program should not be altered –Data flow – flow of data among instructions that produce results and those that consume them

EENG449b/Savvides Lec /22/05 Potential Control Hazard Issue DADDUR1, R2, R3 BEQZR4, L DSUBUR1, R5, R6 L:…. ORR7, R1, R8 R1 in the OR instruction depends on both the ADD and the SUB instruction. Data dependence is not enough. Data flow must also be preserved. Speculation will be used to avoid this: Idea builds on dynamic scheduling concepts. Bet the outcome of a branch and start executing under the provision that the result could be invalidated.

EENG449b/Savvides Lec /22/05 Dynamic Scheduling Statically scheduled pipelines –When a data dependence cannot be hidden with bypassing or forwarding, the processor stalls until the data is cleared Dynamic scheduling –Hardware reorders instructions to reduce the stalls while maintaining data flow and instruction behavior Advantages –Handles dependences not known at compile time »Simplifies compiler design –Allows code compiled for one pipeline to run efficiently on another Disadvantage – hardware complexity

EENG449b/Savvides Lec /22/05 Dynamic Scheduled Pipelines Simple pipelines result in hazards that require stalling. Static scheduling – compilers rearrange instructions to avoid stalls. Dynamic scheduling – processor executes instructions out-of-order to minimize stalls Dynamic scheduling requires splitting the ID stage into stages: –Issue – Decode instructions, check for structural hazards –Read operands – Wait until there are no data hazards, then read operands –Also need to know when each instruction begins and ends execution Requires a lot more bookkeeping!

EENG449b/Savvides Lec /22/05 Examples DIV.DF0, F2, F4 ADD.DF10, F0, F8 SUB.DF12, F8, F14 SUB.D cannot execute because of antidependence of ADD.D on DIV.D The only reason that SUB.D cannot execute before is the in-order execution requirement Split ID stage in 2 parts –Check for stuctural hazards –Check for the absence of data hazards

EENG449b/Savvides Lec /22/05 Lookout for WAR and WAW hazards DIV.D F0, F2, F4 ADD.D F6, F0, F8 SUB.D F8, F10, F14 MUL.D F6, F10, F8 Both hazards can be avoided with register renaming Antidependence Potential WAW Hazard

EENG449b/Savvides Lec /22/05 Preserving Exception Behavior Out-of-order completion must also preserve exception behavior. No instruction can raise an exception until the processor knows that the instruction will be executed –Pipeline may have completed instructions that are later in the program order –Pipeline may have not yet completed instructions earlier in the program that may case exceptions when executed

EENG449b/Savvides Lec /22/05 ID-Stage Changes Split the ID stage in two parts 1.Issue – Decode instructions and check for structural hazards 2.Read operands – wait until no data hazards, read operands Need to distinguish between when an instruction begins execution and when it completes execution. Pipeline allows instructions to be in execution at the same time – this allows for dynamic optimizations All instructions enter the issue stage in order, but may bypass each other in the second stage

EENG449b/Savvides Lec /22/05 Scoreboarding Scoreboarding – a technique that allows out- of-order execution when resources are available and there are no data dependencies – originated in CDC6600 in the mid 60s. Scoreboard fully responsible for instruction execution and hazard detection –Requires changes in # of functional units and latency of operations –Needs to keep track of status of all instructions in execution

EENG449b/Savvides Lec /22/05 Scoreboarding II

EENG449b/Savvides Lec /22/05 Tomasulo’s Algorithm Hardware based technique for ILP –Tracks when operands are available to avoid RAW hazards –Introduces register renaming to avoid WAW and WAR hazards »What does this mean? More sophisticated approach than the scoreboard from Appendix A Initially designed for the IBM 360/91 –Designed in the late 60s –Scoreboarding + register renaming –4 FP registers, long memory access delays, long FP times – compiler level optimizations were limited

EENG449b/Savvides Lec /22/05 Register Renaming DIV.DF0, F2, F4 ADD.DF6, F0, F8 S.DF6, 0(R1) SUB.DF8, F10, F14 MUL.DF6, F10, F8 Where is the antidependence (WAR)?

EENG449b/Savvides Lec /22/05 Register Renaming DIV.DF0, F2, F4 ADD.DF6, F0, F8 S.DF6, 0(R1) SUB.DF8, F10, F14 MUL.DF6, F10, F8 Where is the output dependence (WAW)?

EENG449b/Savvides Lec /22/05 Register Renaming DIV.DF0, F2, F4 ADD.DF6, F0, F8 S.DF6, 0(R1) SUB.DF8, F10, F14 MUL.DF6, F10, F8 Where are the true data dependences (RAW)?

EENG449b/Savvides Lec /22/05 Getting Rid of Name Dependencies Assume we have 2 temporary registers S and T the code sequence can be re-written as: DIV.DF0, F2, F4DIV.D F0, F2, F4 ADD.DF6, F0, F8ADD.D S, F0, F8 S.DF6, 0(R1)S.D S, 0(R1) SUB.DF8, F10, F14SUB.D T, F10, F14 MUL.DF6, F10, F8MUL.D F6, F10, T Any subsequent uses of F8 should be replaced with register T –Requires sophisticated compiler analysis since intervening branches may change the meaning of F8 –Tomasulo’s algorithm can handle renaming across branches

EENG449b/Savvides Lec /22/05 Tomasulo’s Scheme for Avoiding Name Dependences Use Reservation Stations & Issue Logic –Buffer the operands of instructions waiting to issue –Buffers the operand as soon as it is available, eliminating the need to get an operand from a register –Operands are renamed to the names of the reservation station, avoiding register name conflicts –There are more reservation stations than registers »Eliminates more hazards than the compiler Pending instructions designate the reservation station that will provide their input

EENG449b/Savvides Lec /22/05 Advantages of Tomasulo’s Algorithm Hazard Detection and Execution Control are distributed –The information held in each functional unit will determine when an instruction is ready to execute Reservation stations buffer pass the results directly to functional units –Do not have to go through registers

EENG449b/Savvides Lec /22/05 MIPS FPU with Tomasulo Issue: In order instructions to Preserve correct data flow If there is an empty reservation station issue the instruction with operands Else stall –stuctural hazard

EENG449b/Savvides Lec /22/05 MIPS FPU with Tomasulo If operands not available, keep track of the FUs that produce them – Register renaming

EENG449b/Savvides Lec /22/05 MIPS FPU with Tomasulo All the results are made available on CDC that goes everywhere. All reservation stations have tag fields controlled by the pipeline control

EENG449b/Savvides Lec /22/05 An Instruction goes through 3 basic steps 1.Issue – described in the previous slide Instructions are issued in order from the instruction FIFO If there is an empty reservation station for that operation, issue the instruction and operands to the reservation station. If there is no reservation station available, stall and wait for one If the operands are not in the registers, track the operational units that will produce the result.

EENG449b/Savvides Lec /22/05 Step 2 2. Execute – Operands placed in the reservation tables as they become available When all operands available the instruction is executed - this execution delay eliminates RAW hazards Loads and stores have 2 execution steps 1. Compute the effective address and place in load or store buffer 2. Execute as soon as memory unit is available No instruction is executed until all preceding branches have been determined to preserve exception behavior This prevents the occurrence of exceptions Exceptions get generated only if commands will execute. Next lecture we will discuss speculation for more efficient handling of branches

EENG449b/Savvides Lec /22/05 Step 3 3. Write result –Results written on common data bus (CDB) »End up in corresponding registers and reservation tables –Write data to memory also happens at this step

EENG449b/Savvides Lec /22/05 Things to note about Tomasulo’s Scheme Data structures to detect and eliminate hazards are attached to: –Reservation stations –Register file –Load and store buffers Reservation stations act as a set of virtual registers –More than FP registers so register renaming is possible

EENG449b/Savvides Lec /22/05 Reservation Table Fields To track the state of the algorithm: Op – operation to perform on source operands Q j, Q k – the reservation stations that will produce the operand V j, V k – The value of the source operands A – holds information on the memory address calculation (immediate and address calculation are stored here) Busy – Reservation station and its accompanying functional unit is busy The register file also contains a field Q i – The number of the reservation station that contains the value that should be stored in this register

EENG449b/Savvides Lec /22/05 Tomasulo Algorithm Example L.D F6, 34(R2) L.D F2, 45(R3) MUL.D F0, F2, F4 SUB.D F8, F2, F6 DIV.D F10, F0, F6 ADD.D F6, F8, F2

EENG449b/Savvides Lec /22/05

EENG449b/Savvides Lec /22/05 Tomasulo’s Advantages No checking needed for WAR or WAW as registers are renamed Hazard detection logic is distributed Loads and stores are treated as basic functional units Has larger register sets – reservation tables Exploits ILP well but requires more complex hardware than scoreboarding

EENG449b/Savvides Lec /22/05 Next Lecture Tomasulo Algorithm Details Speculation