Instruction Level Parallelism

1. Instruction Level Parallelism

2. Outline • What’s ILP • ILP vs Parallel Processing • Sequential execution vs ILP execution • Limitations of ILP • ILP Architectures • Sequential Architecture • Dependence Architecture • Independence Architecture • ILP Scheduling • Open Problems • References

3. What’s ILP • Architectural technique that allows the overlap of individual machine operations ( add, mul, load, store …) • Multiple operations will execute in parallel (simultaneously) • Goal: Speed Up the execution • Example: load R1  R2 add R3  R3, “1” add R3  R3, “1” add R4  R3, R2 add R4  R4, R2 store [R4]  R0

4. Example: Sequential vs ILP • Sequential execution (Without ILP) Add r1, r2  r8 4 cycles Add r3, r4  r7 4 cycles 8 cycles • ILP execution (overlap execution) Add r1, r2  r8 Add r3, r4  r7 Total of 5 cycles

5. ILP Overlap individual machine operations (add, mul, load…) so that they execute in parallel Transparent to the user Goal: speed up execution Parallel Processing Having separate processors getting separate chunks of the program ( processors programmed to do so) Nontransparent to the user Goal: speed up and quality up ILP vs Parallel Processing

6. ILP Challenges • In order to achieve parallelism we should not have dependences among instructions which are executing in parallel: • H/W terminology Data Hazards ( RAW, WAR, WAW) • S/W terminology Data Dependencies

7. Dependences and Hazards • Dependences are a property of programs • If two instructions are data dependent they can not execute simultaneously • A dependence results in a hazard and the hazard causes a stall • Data dependences may occur through registers or memory

8. Types of Dependencies • Name dependencies • Output dependence • Anti-dependence • Data True dependence • Control Dependence • Resource Dependence

9. Name dependences • Output dependence When instruction I and J write the same register or memory location. The ordering must be preserved to leave the correct value in the register add r7,r4,r3 div r7,r2,r8 • Anti-dependence When instruction j writes a register or memory location that instruction I reads i: add r6,r5,r4 j: sub r5,r8,r11

10. An instruction j is data dependent on instruction i if either of the following hold: instruction i produces a result that may be used by instruction j , or instruction j is data dependent on instruction k, and instruction k is data dependent on instruction i LOOP LD F0, 0(R1) ADD F4, F0, F2 SD F4, 0(R1) SUB R1, R1, -8 BNE R1, R2, LOOP Data Dependences

11. Control Dependences • A control dependence determines the ordering of an instruction i, with respect to a branch instruction so that the instruction i is executed in correct program order. • Example: If p1 { S1; }; If p2 { S2; }; • Two constraints imposed by control dependences: • An instruction that is control dependent on a branch cannot be moved before the branch • An instruction that is not control dependent on a branch cannot be moved after the branch

12. Resource dependences • An instruction is resource-dependent on a previously issued instruction if it requires a hardware resource which is still being used by a previously issued instruction. • e.g. • div r1, r2, r3 • div r4, r2, r5

13. ILP Architectures • Computer Architecture: is a contract (instruction format and the interpretation of the bits that constitute an instruction) between the class of programs that are written for the architecture and the set of processor implementations of that architecture. • In ILP Architectures: + information embedded in the program pertaining to available parallelism between instructions and operations in the program

14. ILP Architectures Classifications • Sequential Architectures: the program is not expected to convey any explicit information regarding parallelism. (Superscalar processors) • Dependence Architectures: the program explicitly indicates the dependences that exist between operations (Dataflow processors) • Independence Architectures: the program provides information as to which operations are independent of one another. (VLIW processors)

15. Sequential architecture and superscalar processors • Program contains no explicit information regarding dependencies that exist between instructions • Dependencies between instructions must be determined by the hardware • It is only necessary to determine dependencies with sequentially preceding instructions that have been issued but not yet completed • Compiler may re-order instructions to facilitate the hardware’s task of extracting parallelism

16. Superscalar Processors • Superscalar processors attempt to issue multiple instructions per cycle • However, essential dependencies are specified by sequential ordering so operations must be processed in sequential order • This proves to be a performance bottleneck that is very expensive to overcome

17. Dependence architecture and data flow processors • The compiler (programmer) identifies the parallelism in the program and communicates it to the hardware (specify the dependences between operations) • The hardware determines at run-time when each operation is independent from others and perform scheduling • Here, no scanning of the sequential program to determine dependences • Objective: execute the instruction at the earliest possible time (available input operands and functional units).

18. Dependence architectures Dataflow processors • Dataflow processors are representative of Dependence architectures • Execute instruction at earliest possible time subject to availability of input operands and functional units • Dependencies communicated by providing with each instruction a list of all successor instructions • As soon as all input operands of an instruction are available, the hardware fetches the instruction • The instruction is executed as soon as a functional unit is available • Few Dataflow processors currently exist

19. Dataflow strengths and limitations • Dataflow processors use control parallelism alone to fully utilize the FU. • Dataflow processor is more successful than others at looking far down the execution path to find control parallelism • When successful its better than speculative execution: • Every instruction is executed is useful • Processor does not have to deal with error conditions, because of speculative operations

20. Independence architecture and VLIW processors • By knowing which operations are independent, the hardware needs no further checking to determine which instructions can be issued in the same cycle • The set of independent operations >> the set of dependent operations • Only a subset of independent operations are specified • The compiler may additionally specify on which functional unit and in which cycle an operation is executed • The hardware needs to make no run-time decisions

21. VLIW processors • Operation vs instruction • Operation: is an unit of computation (add, load, branch = instruction in sequential ar.) • Instruction: set of operations that are intended to be issued simultaneously • Compiler decides which operation to go to each instruction (scheduling) • All operations that are supposed to begin at the same time are packaged into a single VLIW instruction

22. VLIW strengths • In hardware it is very simple: • consisting of a collection of function units (adders, multipliers, branch units, etc.) connected by a bus, plus some registers and caches • More silicon goes to the actual processing (rather than being spent on branch prediction, for example), • It should run fast, as the only limit is the latency of the function units themselves. • Programming a VLIW chip is very much like writing microcode

23. VLIW limitations • The need for a powerful compiler, • Increased code size arising from aggressive scheduling policies, • Larger memory bandwidth and register-file bandwidth, • Limitations due to the lock-step operation, binary compatibility across implementations with varying number of functional units and latencies

24. Summary: ILP Architectures

25. Static Scheduling boosted by parallel code optimization ILP Scheduling Dynamic Scheduling without static parallel code optimization Dynamic Scheduling boosted by static parallel code optimization • done by the compiler • The processor receives dependency-free and optimized code for parallel execution • Typical for VLIWs and a few pipelined processors (e.g. MIPS) • done by the processor • The code is not optimized for parallel execution. The processor detects and resolves dependencies on its own • Early ILP processors (e.g. CDC 6600, IBM 360/91 etc.) • done by processor in conjunction with parallel optimizing compiler • The processor receives optimized code for parallel execution, but it detects and resolves dependencies on its own • Usual practice for pipelined and superscalar processors (e.g. RS6000)

26. ILP Scheduling: Trace scheduling • An optimization technique that has been widely used for VLIW, superscalar, and pipelined processors. • It selects a sequence of basic blocks as a trace and schedules the operations from the trace together. • Example: Instr1 Instr2 Branch x Instr3

27. Trace Scheduling • Extract more ILP • Increase machine fetch bandwidth by storing logically consecutive blocks in physically contiguous cache location (possible to fetch multiple basic blocks in one cycle) • Trace scheduling can be implemented by hardware or software

28. Trace Scheduling in HW • Hardware technique makes use of a large amount of information in dynamic execution to format traces dynamically and schedule the instructions in trace more efficiently. • Since the dependency and memory access addresses have been solved in dynamic execution, instructions in trace can be reordered more easily and efficiently. • Example: trace cache approach

29. Trace scheduling in SW • Supplement to machines without hardware trace scheduling support. • Formats traces based on static profiled data, and schedules instructions using traditional compiler scheduling and optimization technique. • It faces some difficulties like code explosion and exception handling.

30. ILP open problems • Pipelined scheduling : Optimized scheduling of pipelined behavioral descriptions. Two simple type of pipelining (structural and functional). • Controller cost : Most scheduling algorithms do not consider the controller costs which is directly dependent on the controller style used during scheduling. • Area constraints : The resource constrained algorithms could have better interaction between scheduling and floorplanning. • Realism : • Scheduling realistic design descriptions that contain several special language constructs. • Using more realistic libraries and cost functions. • Scheduling algorithms must also be expanded to incorporate different target architectures.

31. References • Instruction-Level Parallel Processing: History, Overview and Perspective. B. Ramakrishna Rau, Joseph A. Fisher. Journal of Supercomputing, Vol. 7, No. 1, Jan. 1993, pages 9-50. • Limits of Control Flow on Parallelism. Monica S. Lam, Robert P. Wilson. 19th ISCA, May 1992, pages 19-21. • Global Code Generation for Instruction-Level Parallelism: Trace Scheduling-2. Joseph A. Fisher. Technical Report, HPLabs HPL-93-43, Jun. 1993. • VLIW at IBM Research http://www.research.ibm.com/vliw • Intel and HP hope to speed CPUs with VLIW technology that's riskier than RISC, Dick Pountain http://www.byte.com/art/9604/sec8/art3.htm • Hardware and Software Trace Scheduling http://charlotte.ucsd.edu/users/yhu/paperlist/summary.html • ILP open problems http://www.ececs.uc.edu/~ddel/projects/dss/hls_paper/node9.html • Computer Architecture A Quantitative Approach, Hennessy & Patterson, 3rd edition, M Kaufmann