Limits of Instruction-Level Parallelism

1. Presentation by: Robert Duckles CSE 520 Paper being presented: Limits of Instruction-Level Parallelism David W. Wall WRL Research Report, November 1993 Limits of Instruction-Level Parallelism

2. Instructions that do not have dependencies on each other; can be executed in any order. r1 := 0[r9] r1 := 0[r9] r2 := 17 r2 := r1 + 17 4[r3] := r6 4[r2] := r6 (has ILP) (no ILP)‏ Super-scalar machine – a machine that can issue multiple independent instructions in the same clock cycle. What is ILP?

3. Parallelism = (Number of Instructions) / (Number of Cycles it takes to execute)‏ r1 := 0[r9] r1 := 0[r9] r2 := 17 r2 := r1 + 17 4[r3] := r6 4[r2] := r6 Parallelism = 3 Parallelism = 1 Definition of Parallelism

4. That depends how hard you want to look for it... Ways to increase ILP: Register renaming Branch prediction Alias analysis Indirect-jump prediction How much parallelism is there?

5. Programs are made up of “basic blocks”—uninterrupted sequences of instructions with no branches. On average, in typical applications, basic blocks are ~10 instructions long. Each basic block has parallelism of around 3. Low estimate for ILP

6. If you look beyond a basic block, at the entire scope of a program, studies have shown that an “omniscient” scheduler can achieve parallelism of > 1000 in some numerical applications. “Omniscient” scheduling can be implemented by saving a trace of a program execution, and using an oracle to schedule it. The oracle knows what will happen, and thus can create a perfect execution schedule. Practical, achievable ILP should be between 3 and 1000. High estimate for ILP

7. Types of dependencies: * True dependency - given the computations involved, the dependency must exist * False dependency - dependency happens to exist as an artifact of the code generation engine. E.g., two independent values are allocated to the same register by the compiler. r1 := 20[r4] r2 := r1 + r4 ... ... r2 := r1 + 1 r1 := r17 - 1 (a) true data dependency (b) anti-dependency r1 := r2 * r3 if r17 = 0 goto L ... ... r1 := r2 + r3 ... r1 := 0[r7] L: (c) output dependency (d) control dependency Types of dependencies

8. The compiler's register allocation algorithm can insert false dependencies by assigning unrelated values to the same register. We can undo this damage by assigning each value to a unique register so that only true dependencies remain. However, machines have a finite number of registers, so we can never guarantee perfect parallelism. Register renaming

9. Register renaming

10. We often have registers that point to a memory location or contain a memory offset. Can two memory pointers point to the same place in memory? If so, there might be a dependency. We're not sure yet. We can try to inspect pointer values at runtime to see if they point to overlapping memory. Alias analysis

11. Alias analysis

12. Limitations of branch prediction: We can correctly predict around ~0.9 by counting which branches have been recently taken, and taking the most common one.

13. If we jump to an address that is not known at compile time--for example, if a destination address is calculated into a register at runtime. This is often the case for "return" constructs, where the the calling function's address is stored on the stack. In this case, we can do indirect-jump prediction. Indirect-jump prediction

14. Latency Multi-cycle instructions can greatly decrease parallelism

15. The window size is the maximum number of instructions that can appear in the pending cycle list. Window size

16. Overall results

17. Even with “perfect” techniques, most real applications hit an ILP limit of around 20 With reasonable, practical methods, it's even worse—it's very difficult to get an ILP above 10. Conclusions: the ILP Wall

18. Our term project is about optimization techniques for AMD64 Opteron/Athlon processors. Maximizing ILP is essential to getting the most performance out of any processor. Branch prediction, register renaming, etc., are all particularly relevant optimizations Relationship to Term Project