1. EENG449b/Savvides Lec 15.1 3/24/04 March 24, 2004 Prof. Andreas Savvides Spring 2004 http://www.eng.yale.edu/courses/eeng449bG EENG 449bG/CPSC 439bG Computer Systems Lecture 15 Software ILP – Chapter 4 Text Sections 4.1 – 4.5

2. EENG449b/Savvides Lec 15.2 3/24/04 Compiler Techniques for Exposing ILP In Chapter 3 we discussed hardware based techniques for ILP –Dynamic scheduling and other hardware based optimizations – mostly apply to superscalar processors In this chapter –Static scheduling techniques at the compiler level –Mostly apply to VLIW processors –Start by examining how to optimize loops

3. EENG449b/Savvides Lec 15.3 3/24/04 Running Example This code, adds a scalar to a vector: for (i=1000; i>0; i=i–1) x[i] = x[i] + s; Assume following latency all examples InstructionInstructionExecutionLatency producing resultusing result in cyclesin cycles FP ALU opAnother FP ALU op 4 3 FP ALU opStore double 3 2 Load doubleFP ALU op 1 1 Load doubleStore double 1 0 Integer opInteger op 1 0

4. EENG449b/Savvides Lec 15.4 3/24/04 FP Loop: Where are the Hazards? Loop:L.DF0,0(R1);F0=vector element ADD.DF4,F0,F2;add scalar from F2 S.D0(R1),F4;store result DADDUIR1,R1,#-8;decrement pointer 8B (DW) BNEZR1,Loop;branch R1!=zero NOP;delayed branch slot Where are the stalls? First translate into MIPS code: -To simplify, assume 8 is lowest address

5. EENG449b/Savvides Lec 15.5 3/24/04 FP Loop Showing Stalls 10 clocks: Rewrite code to minimize stalls? InstructionInstructionLatency in producing resultusing result clock cycles FP ALU opAnother FP ALU op3 FP ALU opStore double2 Load doubleFP ALU op1 1 Loop:L.DF0,0(R1);F0=vector element 2stall 3ADD.DF4,F0,F2;add scalar in F2 4stall 5stall 6 S.DF4, 0(R1);store result 7 DADDUIR1,R1,#-8;decrement pointer 8B (DW) 8stall 9 BNER1,Loop;branch R1!=zero 10stall;delayed branch slot

6. EENG449b/Savvides Lec 15.6 3/24/04 Revised FP Loop Minimizing Stalls 6 clocks, but just 3 for execution, 3 for loop overhead; How make faster? InstructionInstructionLatency in producing resultusing result clock cycles FP ALU opAnother FP ALU op3 FP ALU opStore double2 Load doubleFP ALU op1 1 Loop:L.DF0,0(R1) 2DADDUIR1,R1,#-8 3ADD.DF4,F0,F2 4stall 5BNER1,R2, Loop;delayed branch 6 S.DF4, 8(R1) Swap BNE and S.D by changing address of S.D

7. EENG449b/Savvides Lec 15.7 3/24/04 Unroll Loop Four Times (straightforward way) Rewrite loop to minimize stalls? 1 Loop:L.DF0,0(R1) 2ADD.DF4,F0,F2 3S.D0(R1),F4 ;drop DADDUI & BNE 4L.DF6,-8(R1) 5ADD.DF8,F6,F2 6S.DF8,-8(R1) ;drop DADDUI & BNE 7L.DF10,-16(R1) 8ADD.DF12,F10,F2 9S.DF12,-16(R1) ;drop DADDUI & BNE 10L.DF14,-24(R1) 11ADD.DF16,F14,F2 12S.DF16,-24(R1) 13DADDUIR1,R1,#-32;alter to 4*8 14BNER1,LOOP 14 + (4 x (1+2))+ 2= 28 clock cycles, or 7 per iteration 1 cycle stall 2 cycles stall 1 cycle stall 1 cycle stall (delayed branch)

8. EENG449b/Savvides Lec 15.8 3/24/04 Unrolled Loop Detail Do not usually know upper bound of loop Suppose it is n, and we would like to unroll the loop to make k copies of the body Instead of a single unrolled loop, we generate a pair of consecutive loops: –1st executes (n mod k) times and has a body that is the original loop –2nd is the unrolled body surrounded by an outer loop that iterates (n/k) times –For large values of n, most of the execution time will be spent in the unrolled loop Problem: Although it improves execution performance, it increases the code size substantially!

9. EENG449b/Savvides Lec 15.9 3/24/04 Unrolled Loop That Minimizes Stalls (scheduled based on the latencies from slide 4) What assumptions made when moved code? –OK to move store past DSUBUI even though changes register –OK to move loads before stores: get right data? –When is it safe for compiler to do such changes? 1 Loop:L.DF0,0(R1) 2L.DF6,-8(R1) 3L.DF10,-16(R1) 4L.DF14,-24(R1) 5ADD.DF4,F0,F2 6ADD.DF8,F6,F2 7ADD.DF12,F10,F2 8ADD.DF16,F14,F2 9S.DF4, 0(R1) 10S.DF8, -8(R1) 11S.DF12, -16(R1) 12DADDUIR1,R1,#-32 13BNER1,LOOP 14S.DF16, 8(R1) ; 8-32 = -24 14 clock cycles, or 3.5 per iteration Better than 7 before scheduling and 6 when scheduled and not unrolled

10. EENG449b/Savvides Lec 15.10 3/24/04 Compiler Perspectives on Code Movement Compiler concerned about dependencies in program Whether or not a HW hazard depends on pipeline Try to schedule to avoid hazards that cause performance losses (True) Data dependencies (RAW if a hazard for HW) –Instruction i produces a result used by instruction j, or –Instruction j is data dependent on instruction k, and instruction k is data dependent on instruction i. If dependent, can’t execute in parallel Easy to determine for registers (fixed names) Hard for memory (“memory disambiguation” problem): –Does 100(R4) = 20(R6)? –From different loop iterations, does 20(R6) = 20(R6)?

11. EENG449b/Savvides Lec 15.11 3/24/04 Compiler Perspectives on Code Movement Name Dependencies are Hard to discover for Memory Accesses –Does 100(R4) = 20(R6)? –From different loop iterations, does 20(R6) = 20(R6)? Our example required compiler to know that if R1 doesn’t change then: 0(R1)  -8(R1)  -16(R1)  -24(R1) There were no dependencies between some loads and stores so they could be moved by each other

12. EENG449b/Savvides Lec 15.12 3/24/04 Steps Compiler Performed to Unroll Check OK to move the S.D after DADDUI and BNEZ, and find amount to adjust S.D offset Determine unrolling the loop would be useful by finding that the loop iterations were independent Rename registers to avoid name dependencies Eliminate extra test and branch instructions and adjust the loop termination and iteration code Determine loads and stores in unrolled loop can be interchanged by observing that the loads and stores from different iterations are independent –requires analyzing memory addresses and finding that they do not refer to the same address. Schedule the code, preserving any dependences needed to yield same result as the original code

13. EENG449b/Savvides Lec 15.13 3/24/04 Where are the name dependencies? 1 Loop:L.DF0,0(R1) 2ADD.DF4,F0,F2 3S.DF4,0(R1) ;drop DADDUI & BNE 4L.DF0,-8(R1) 5ADD.DF4,F0,F2 6S.DF4, -8(R1) ;drop DADDUI & BNE 7L.DF0,-16(R1) 8ADD.DF4,F0,F2 9S.DF4, -16(R1) ;drop DADDUI & BNE 10L.DF0,-24(R1) 11ADD.DF4,F0,F2 12S.DF4, -24(R1) 13DADDUIR1,R1,#-32;alter to 4*8 14BNER1,LOOP 15NOP How can remove them? (See pg. 310 of text)

14. EENG449b/Savvides Lec 15.14 3/24/04 Where are the name dependencies? 1 Loop:L.DF0,0(R1) 2ADD.DF4,F0,F2 3S.D0(R1),F4 ;drop DSUBUI & BNEZ 4L.DF6,-8(R1) 5ADD.DF8,F6,F2 6S.D-8(R1),F8 ;drop DSUBUI & BNEZ 7L.DF10,-16(R1) 8ADD.DF12,F10,F2 9S.D-16(R1),F12 ;drop DSUBUI & BNEZ 10L.DF14,-24(R1) 11ADD.DF16,F14,F2 12S.D-24(R1),F16 13DSUBUIR1,R1,#32;alter to 4*8 14BNEZR1,LOOP 15NOP The Orginal “register renaming” – instruction execution can be overlapped or in parallel

15. EENG449b/Savvides Lec 15.15 3/24/04 Limits to Loop Unrolling Decrease in the amount of loop overhead amortized with each unroll – After a few unrolls the loop overhead amortization is very small Code size limitations – memory is not infinite especially in embedded systems Compiler limitations – shortfall in registers due to excessive unrolling – register pressure – optimized code may loose its advantage due to the lack of registers

16. EENG449b/Savvides Lec 15.16 3/24/04 Static Branch Prediction Simplest: Predict taken –average misprediction rate = untaken branch frequency, which for the SPEC programs is 34%. –Unfortunately, the misprediction rate ranges from not very accurate (59%) to highly accurate (9%) Predict on the basis of branch direction? –choosing backward-going branches to be taken (loop) –forward-going branches to be not taken (if) –SPEC programs, however, most forward-going branches are taken => predict taken is better Predict branches on the basis of profile information collected from earlier runs –Misprediction varies from 5% to 22%

17. EENG449b/Savvides Lec 15.17 3/24/04 Basic VLIW Architectures Does not require the hardware for making dynamic issue decisions – the compiler is responsible for scheduling Has as an advantage in wider issue processors –Small size instructions (2 or 3) superscalar overhead is minimal –For larger instructions hardware complexity grows »Better off with VLIW –Typical instruction width – 5 »1 Integer OP, 2 FP Ops and 2 Memory Refs »12 – 24 bits per unit, instruction width 112 – 168 bits

18. EENG449b/Savvides Lec 15.18 3/24/04 Basic VLIW Architectures II There must be enough parallelism to fill the slots –Unroll loops –Use local optimizations on straight line code –If code has many branches – need global optimizations (e.g trace scheduling) VLIW disadvantage –Harder to update compiler between different versions of the hardware »Object code translation is a possible solution General advantage of multiple issue processors vs. vector processors –Potential to extract parallelism from less structured code –Ability to use a more conventional and typically less expensive, cache based memory system

19. EENG449b/Savvides Lec 15.19 3/24/04 VLIW: Very Large Instruction Word Each “instruction” has explicit coding for multiple operations –In IA-64, grouping called a “packet” –In Transmeta, grouping called a “molecule” (with “atoms” as ops) Tradeoff instruction space for simple decoding –The long instruction word has room for many operations –By definition, all the operations the compiler puts in the long instruction word are independent => execute in parallel –E.g., 2 integer operations, 2 FP ops, 2 Memory refs, 1 branch »16 to 24 bits per field => 7*16 or 112 bits to 7*24 or 168 bits wide –Need compiling technique that schedules across several branches

20. EENG449b/Savvides Lec 15.20 3/24/04 When Safe to Unroll Loop? Example: Where are data dependencies? (A,B,C distinct & nonoverlapping) for (i=0; i<100; i=i+1) { A[i+1] = A[i] + C[i]; /* S1 */ B[i+1] = B[i] + A[i+1]; /* S2 */ } 1. S2 uses the value, A[i+1], computed by S1 in the same iteration. 2. S1 uses a value computed by S1 in an earlier iteration, since iteration i computes A[i+1] which is read in iteration i+1. The same is true of S2 for B[i] and B[i+1]. This is a “loop-carried dependence”: between iterations For our prior example, each iteration was distinct Implies that iterations can’t be executed in parallel, Right????

21. EENG449b/Savvides Lec 15.21 3/24/04 Some Loop Carried Dependences can be Parallelized Example: for (i=0; i<=100; i=i+1) { A[i+1] = A[i] + B[i]; /* S1 */ B[i+1] = C[i] + D[i]; /* S2 */ } S1 uses a value assigned by S2 in the previous iteration – loop carried dependence HOWEVER – dependence is not circular –No statement depends on itself –S1 depends on S2 but S2 does not depend on S1 –Absence of cycle gives partial ordering in statements – loop is parallel

22. EENG449b/Savvides Lec 15.22 3/24/04 Parallel Version of Loop 1. There is no dependence from S1 to S2. This means that S1 and S2 can be interchanged 2. On the first iteration S1 depends on B[1] computed prior to initiating the loop B[1] = A[1] + B[1]; for (i=0; i<=99; i=i+1) { B[i+1] = C[i] + D[i]; /* S2 */ A[i+1] = A[i+1] + B[i]; /* S1 */ } B[101] = C[100] + D[100]; Loop iterations can now be overlapped if statements inside the loop are executed in order.

23. EENG449b/Savvides Lec 15.23 3/24/04 Another possibility: Software Pipelining Observation: if iterations from loops are independent, then can get more ILP by taking instructions from different iterations Software pipelining: reorganizes loops so that each iteration is made from instructions chosen from different iterations of the original loop (~ Tomasulo in SW)

24. EENG449b/Savvides Lec 15.24 3/24/04 Software Pipelining Example Loop: L.D F0,0(R1) ADD.D F4,F0,F2 S.D F4, 0(R1) DADDUI R1, R1, #-8 BNE R1, R2, Loop

25. EENG449b/Savvides Lec 15.25 3/24/04 Software Pipelining Example Before: Unrolled 3 times 1 L.DF0,0(R1) 2 ADD.DF4,F0,F2 3 S.D0(R1),F4 4 L.DF6,-8(R1) 5 ADD.DF8,F6,F2 6 S.D-8(R1),F8 7 L.DF10,-16(R1) 8 ADD.DF12,F10,F2 9 S.D-16(R1),F12 10 DADDUIR1,R1,#-24 11 BNEZR1,LOOP After: Software Pipelined 1 S.D0(R1),F4 ;Stores M[i] 2 ADD.DF4,F0,F2 ;Adds to M[i-1] 3 L.DF0,-16(R1);Loads M[i-2] 4 DADDUIR1,R1,#-8 5 BNEZR1,LOOP Symbolic Loop Unrolling – Maximize result-use distance – Less code space than unrolling – Fill & drain pipe only once per loop vs. once per each unrolled iteration in loop unrolling SW Pipeline Loop Unrolled overlapped ops Time 5 cycles per iteration

26. EENG449b/Savvides Lec 15.26 3/24/04 Loop Unrolling vs. Software Pipelining Both provide a better scheduled inner loop Loop Unrolling –Reduces the overhead of the loop, the branch and counter update code Software Pipelining –Reduces the number the loop is not running at peak speed to once per loop at the beginning and end –Easier when the body of a loop is a basic block, much more complex when it contains internal flow control If we unroll a loop that does 100 iterations a constant number of times i.e 4 Then we have to pay the overhead 100/4=25 times

27. EENG449b/Savvides Lec 15.27 3/24/04 Software Pipelining vs. Loop Unrolling

28. EENG449b/Savvides Lec 15.28 3/24/04 Trace Scheduling Parallelism across IF branches vs. LOOP branches? Trace scheduling incurs cost to less frequent paths. Two steps: –Trace Selection »Find likely sequence of basic blocks (trace) of (statically predicted or profile predicted) long sequence of straight-line code –Trace Compaction »Squeeze trace into few VLIW instructions »Need bookkeeping code in case prediction is wrong This is a form of compiler-generated speculation –Compiler must generate “fixup” code to handle cases in which trace is not the taken branch –Needs extra registers: undoes bad guess by discarding Subtle compiler bugs mean wrong answer vs. poorer performance; no hardware interlocks So far it has been successfully applied to scientific code with intensive loops but still unclear if it is suitable for programs with less loops.

29. EENG449b/Savvides Lec 15.29 3/24/04 Advantages of HW (Tomasulo) vs. SW (VLIW) Speculation HW advantages: –HW better at memory disambiguation since knows actual addresses –HW better at branch prediction since lower overhead –HW maintains precise exception model –HW does not execute bookkeeping instructions –Same software works across multiple implementations –Smaller code size (not as many nops filling blank instructions) SW advantages: –Window of instructions that is examined for parallelism much higher –Much less hardware involved in VLIW (unless you are Intel…!) –More involved types of speculation can be done more easily –Speculation can be based on large-scale program behavior, not just local information

30. EENG449b/Savvides Lec 15.30 3/24/04 Superscalar v. VLIW Smaller code size Binary compatibility across generations of hardware Simplified Hardware for decoding, issuing instructions No Interlock Hardware (compiler checks?) More registers, but simplified Hardware for Register Ports (multiple independent register files?)

31. EENG449b/Savvides Lec 15.31 3/24/04 Next Time Hardware support for exposing more parallelism, examples and conclusion of ILP Discussion of project reports HWK3 out next lecture Next Week Memory Hierarchies – Chapter 5 – last chapter for the course!